CN105467975A - Equipment fault diagnosis method - Google Patents

Abstract

An equipment fault diagnosis method successively comprises fault diagnosis training steps and fault diagnosis operation steps. The method focuses on a data form characteristic essence and combines a multi-dimensional segmentation fitting algorithm and an optimized dynamic time warping algorithm based on equipment fault data to carry out pattern expression of modeling and distance threshold extraction, and extract form characteristics from the found equipment abnormity data to carry out mode matching, so that an equipment data fault type can be identified and a reason diagnosis function is realized. In this way, the difficulties in current fault diagnosis technologies that similarity degrees among the fault data can hardly be described efficiently and accurately can be solved.

Description

Equipment fault diagnosis method

Technical Field

The invention relates to the technical field of fault diagnosis, in particular to a method for diagnosing equipment faults.

Background

In recent years, with the continuous innovation and development of scientific technology and the rapid development of social industrialization, more and more industrial industrialization systems are loaded in large-scale industrial production occasions, and huge irreplaceable productivity is created in the application environment. Maintenance work for these devices is therefore particularly important and laborious. However, many implicit influencing factors may cause the system equipment to malfunction or even fail, so that countries and enterprises pay more and more attention to the status detection and fault diagnosis of the important system equipment.

Generally, these devices are increasingly sophisticated and coupled by interaction between different parts. The method of disassembling and disassembling the equipment is time-consuming and labor-consuming, the diagnosis effect is often not ideal, and the equipment performance is easy to be unstable. The method for reading the state of the important part of the equipment in real time by installing the monitoring sensor device to observe the running trend of the relevant parameters of the equipment and contrast the index threshold value is simple, intuitive and strong in reliability, and is widely popularized and used in various production occasions such as power plants, automobiles, satellites and the like at present. The efficiency of this approach is still limited by the ability and effort of plant maintenance personnel to work with complex unknown conditions that are prominent in the plant and are less experienced.

The above method is difficult to form a standard fault diagnosis system due to the dilemma of empirical knowledge expression, so a fault diagnosis method based on mathematical mining technology is gradually introduced into the fault diagnosis work of these large and complex devices. The mathematics mining technology integrates subject knowledge of modern cybernetics, computer science, artificial intelligence, signal processing, mode recognition, statistical mathematics and the like, and excavates information which is hidden in data and is beneficial to fault condition analysis by researching historical data and current real-time data of analysis equipment. Expert systems, association rules, neuron networks, Bayesian networks and the like are mathematical mining methods which are widely applied, and a set of research systems which are divided into fault analysis, fault modeling, fault detection, fault inference, fault decision and the like are gradually formed by the methods.

In a paper "vehicle fault diagnosis expert system with fusion example and rule inference" (journal of mechanical engineering, 2002, 7 th stage: 91-95), a brand-new mixed inference method with fusion examples and rules is provided in the paper, and a whole set of expert system for solving the problem that different diagnosis units are difficult to communicate information is established. It can be seen that the expert system can solve the fault diagnosis problem in the related field with the capability similar to the human expert level in the specific field, and the main characteristic is to form the own specific rule to analyze and solve the problem by relying on the existing human expert experience knowledge. The method is a characteristic of an expert system, but is limited by the lack of experience of human professional knowledge in unknown fields, the expert system often has a series of problems of low self-learning ability, low diagnosis success rate, difficulty in obtaining system knowledge and the like, and the problems still need to be further deeply researched and optimized by the method of the expert system.

In a paper "grid fault diagnosis based on association rule data mining technology" (protection and control of power system, 2009, 9 th period: 8-14), an article applies an association rule algorithm to grid fault diagnosis. And corresponding decision attributes and condition attributes are designated according to the fault characteristics, the establishment of an original decision table is completed, and meanwhile, an association rule algorithm is applied to carry out frequent item set mining and strong rule screening on decision table data, so that reasoning diagnosis on fault information of various complex conditions is finally realized. However, the technology has some problems, for example, the common problems of the potential implicit failure mode such as poor discrimination, poor large-scale data processing capability, low efficiency of storing and updating the association rule, and the like, are still solved by the following related work.

In a paper "fault diagnosis technology research based on a BP neural network" (computer and modernization, 2009, volume 7), the article selects a general BP neural network to be applied to a fault diagnosis scene, and successfully applies the neural network to fault diagnosis according to the device level distribution characteristics under the condition of reasonably constructing a BP neural network model, so that the diagnosis efficiency and reliability are improved. The neural network has the greatest characteristic that the neural network can infinitely and nonlinearly approximate to an original data model under the condition of large sample quantity, but the inherent characteristics of overfitting, numerical randomness, unstable training and the like limit the application range of the neural network in the field of fault diagnosis, and other optimization algorithms need to be matched for diagnosis application.

In a thesis of Bayesian network fault diagnosis method for multi-feature information fusion (China mechanical engineering, 2010, 8 th: 940-945), an article takes a pump vibration signal as a research object, information fusion is carried out on multiple fault features such as a frequency domain and a time domain extracted by the signal by using a Bayesian parameter estimation method, then a complete fault classifier is established by constructing a Bayesian network, and fault types are identified by calculating a maximum posterior probability estimation value. The Bayesian network fault diagnosis method needs a priori statistical knowledge of a large amount of sample data, and meanwhile, the Bayesian network directed acyclic expression mode has a risk of error accumulation, and the method needs to be paid attention to the Bayesian network fault diagnosis method.

In view of the problems and risks of the common fault diagnosis technologies, the invention aims at fully mining the potential visual expression form of the data sample and measuring the difference effectiveness of the fault mode, and tries to better solve the problems of low efficiency and low recognition accuracy of fault diagnosis by establishing a new mode matching technology, thereby being beneficial to the continuous maintenance of the good state of the equipment.

Disclosure of Invention

The invention aims to overcome the defects of the prior art, provides an equipment fault diagnosis method, and solves the problem that the similarity degree between fault data is difficult to describe efficiently and accurately in the current fault diagnosis technology.

The invention provides an equipment fault diagnosis method, which sequentially comprises the following steps:

step 1: fault diagnosis training: acquiring fault sample information from a database, respectively obtaining a feature matrix and a mode distance threshold value of a fault sample with dimensions eliminated after calculation processing, associating the feature matrix and the mode distance threshold value with each other, and storing to generate a fault mode knowledge base;

step 2: fault diagnosis operation step: acquiring abnormal sample information from a real-time database, calculating to obtain abnormal characteristic information, sequentially calculating mode distances of sample characteristics in a fault mode knowledge base, converting the mode distances into mode similarity, and outputting a final fault diagnosis result.

Further, the specific steps of step 1 are:

(1.1) acquiring fault sample information from a database;

(1.2) sequentially carrying out piecewise linear fitting on each fault sample;

(1.3) extracting the characteristics of each section of data of the fault sample to obtain a characteristic matrix of the fault sample;

(1.4) carrying out fault characteristic conversion, eliminating characteristic dimension and obtaining a characteristic matrix of a fault sample with the dimension eliminated;

(1.5) calculating a mode distance threshold;

and (1.6) associating the fault feature matrix and the pattern distance threshold value with each other, and storing to generate a fault pattern knowledge base.

Further, the step (1.1) comprises the following specific steps: selecting a piece of research equipment with the number P of fault types being more than or equal to 2 and the occurrence frequency T of each fault being more than or equal to 2 to meet the requirements, selecting a sufficient number of observation points N, wherein N is more than or equal to 10, carrying out fault record searching on historical running state data of the equipment for a long enough time, picking out fault related measuring point information, starting and stopping time of a fault process and useful information of fault maintenance measure records from fault records by utilizing a set screening rule, and reading fault sample data from a power plant real-time database PI according to the useful information, wherein:

the fault sample data with n measuring points and m time points, all the measuring point data at the moment j are regarded as a n-dimensional column vector and are expressed as follows:

u(t_j)＝[u_j1，u_j2，u_j3，…，u_jn]

the sample data file is stored in a matrix form of m × n, and the specific form is as follows:

$F=\left[\begin{array}{c}{u}_{11},u_{12},u_{13},...,u_{1n}\\ u_{21},u_{22},u_{23},...,u_{2n}\\ ..........................\\ ..........................\\ ..........................\\ u_{m1},u_{m2},u_{m3},...,u_{\mathrm{mn}}\end{array}\right]$

the row represents m fault time, the column represents n equipment observation points, the m and n values of the row and the column between each fault sample are different, each fault sample is endowed with a fault type identifier ID, and the fault type identifier ID determination method is that if all the samples contain X types of faults, the numerical range of the fault type identifier ID is as follows: 1-X.

Further, the step (1.2) comprises the following specific steps:

(1.2.1) mean filtering operation: filtering and eliminating noise pollution doped in sample data;

(1.2.2) fault sample segmentation initialization: carrying out sectional initialization on the fault sample after filtering processing;

(1.2.3) combining every two initialized data segments to calculate a fitting error;

and (1.2.4) determining a segmentation cut point of the fault sample, and performing adaptive state segmentation on the fault sample.

Further, the step (1.3) comprises the following specific steps:

dividing the fault sample intoThe x divided data segments are selected fromData segment start feature extractionTaking, comprising the following specific operations:

due to the fact thatThe data segment is $\left[\begin{array}{c}{u}_{11},u_{12},...,u_{1n}\\ u_{21},u_{22},...,u_{2n}\\ ...................\\ ...................\\ u_{h1},u_{h2},...,u_{\mathrm{hn}}\end{array}\right]$ In the form of matrix, the row is time point number and is listed as measuring point number;

and (3) extracting features according to the dimension (column) of the measuring points, wherein the features comprise: slope k, duration l, mean m and variance v;

wherein, $\stackrel{\‾}{V_i}=\left[\begin{array}{c}{u}_{1i}\\ u_{2i}\\ ...\\ ...\\ u_{\mathrm{hi}}\end{array}\right];$

the vector is linearly fitted according to the principle of least square method, and the fitting result is a linear equation p (x) a_0i+a_1ix, so slope characteristic k ═ a_1iWherein a is_0i、a_1iIs a fitting constant;

duration characteristic l isVector length h, i.e., l ═ h;

mean characteristic m isAverage of all values of the vector, i.e.

The variance characteristic v isSum of fluctuation amplitudes of all values of the vector from the mean, i.e.N is the total number of vectors;

segmenting data segments $f_1^{*}=\left[\begin{array}{c}{u}_{11},u_{12},...,u_{1n}\\ u_{21},u_{22},...,u_{2n}\\ ...................\\ ...................\\ u_{h1},u_{h2},...,u_{\mathrm{hn}}\end{array}\right]$ Is converted into a feature matrix $\left[\begin{array}{c}{k}_1,k_2,...,k_n\\ l,l,...,l\\ m_1,m_2,...,m_n\\ v_1,v_2,...,v_n\end{array}\right];$

The rest ofExtracting characteristics according to the operation method, converting the original time domain data matrix into a mathematical characteristic matrix, and finally dividing the data segment setConversion to:

$\left[\begin{array}{c}(k_{11};l_1;m_{11};v_{11}),(k_{12};l_1;m_{12};v_{12}),...,(k_{1n};l_1;m_{1n};v_{1n})\\ (k_{21};l_2;m_{21};v_{21}),(k_{22};l_2;m_{22};v_{22}),...,(k_{2n};l_2;m_{2n};v_{2n})\\ ......\\ ......\\ (k_{x1};l_x;m_x;v_{x1}),(k_{x2};l_x;m_{x2};v_{x2}),...,(k_{\mathrm{xn}};l_x;m_{\mathrm{xn}};v_{\mathrm{xn}})\end{array}\right].$

further, the step (1.4) comprises the following specific steps: converting the slope k into an inclination angle alpha, converting the duration l into a time proportion, carrying out normalization processing on the mean value characteristic m, and carrying out characteristic matrix treatment on the fault sample

$\left[\begin{array}{c}(k_{11};l_1;m_{11};v_{11}),(k_{12};l_1;m_{12};v_{12}),...,(k_{1n};l_1;m_{1n};v_{1n})\\ (k_{21};l_2;m_{21};v_{21}),(k_{22};l_2;m_{22};v_{22}),...,(k_{2n};l_2;m_{2n};v_{2n})\\ ......\\ ......\\ (k_{x1};l_x;m_x;v_{x1}),(k_{x2};l_x;m_{x2};v_{x2}),...,(k_{\mathrm{xn}};l_x;m_{\mathrm{xn}};v_{\mathrm{xn}})\end{array}\right]$

Converted into elimination dimensions $\left[\begin{array}{c}({\mathrm{\α}}_{11};p_1;T_{11};v_{11}),({\mathrm{\α}}_{12};p_1;T_{12};v_{12}),...,({\mathrm{\α}}_{1n};p_1;T_{1n};v_{1n})\\ ({\mathrm{\α}}_{21};p_2;T_{21};v_{21}),({\mathrm{\α}}_{22};p_2;T_{22};v_{22}),...,({\mathrm{\α}}_{2n};p_2;T_{2n};v_{2n})\\ ......\\ ......\\ ({\mathrm{\α}}_{x1};p_x;T_x;v_{x1}),({\mathrm{\α}}_{x2};p_x;T_{x2};v_{x2}),...,({\mathrm{\α}}_{\mathrm{xn}};p_x;T_{\mathrm{xn}};v_{\mathrm{xn}})\end{array}\right].$

Further, the specific steps of step 2 are:

(2.1) acquiring abnormal sample information from a real-time database;

(2.2) performing piecewise linear fitting on the current abnormal sample;

(2.3) carrying out abnormal feature extraction and transformation on the data subjected to segmentation fitting;

(2.4) sequentially calculating the pattern distance of the sample features in the fault pattern knowledge base by using the abnormal feature information;

(2.5) converting the mode distance into a mode similarity using a falling ridge pacing method;

and (2.6) outputting a final fault diagnosis result.

Further, the step (2.1) comprises the following specific steps:

according to the equipment state early warning system of the power plant, discovering that certain unknown abnormal state occurs in the equipment, and carrying out the following related operations:

(2.1.1) determining the device alarm generation time t from the early warning system₁And a cut-off time t₂；

(2.1.2) determining from the early warning system the relevant observation point wp ═ x for the device alarm₁，x₂，…，x_n′]；

(2.1.3) according to the generation time t₁And a cut-off time t₂And a database sampling frequency fs, and obtaining a time point number m-fs × (t)₂-t₁) The number of device measurement points n equals length ([ x ])₁，x₂，…，x_n′]) Where length () is a calculated length function);

(2.1.4) acquiring abnormal sample data, namely the sample data with n 'of measuring points and m of time points, wherein all the measuring point data at the moment j are regarded as a column vector with n' dimension and are expressed as:

v(t_j)＝[v_j1，v_j2，v_j3，…，v_jn′]

storing the sample data file in a matrix form of m multiplied by n', wherein the specific form is as follows:

$\stackrel{\‾}{F}=\left[\begin{array}{c}{v}_{11},v_{12},v_{13},...,v_{1n\′}\\ v_{21},v_{22},v_{23},...,v_{2n\′}\\ ..........................\\ ..........................\\ ..........................\\ v_{m1},v_{m2},v_{m3},...,v_{\mathrm{mn}\′}\end{array}\right],$

where the rows represent m failure times and the columns represent n' device observation points.

The step (2.2) comprises the following specific steps:

(2.2.1) mean filtering operation: filtering and eliminating noise pollution of each measuring point doped in abnormal sample data according to a mean filtering principle;

(2.2.2) abnormal sample segmentation initialization: performing segmentation initialization on the filtered abnormal sample;

(2.2.3) combining every two initialized data segments to calculate a fitting error;

and (2.2.4) determining an abnormal sample segmentation cut point based on a recursive combination mode.

Further, the step (2.5) comprises the following specific steps:

converting the pattern distance into pattern similarity according to the principle of the falling-ridge stepwise method

$\mathrm{\ρ}\left({\stackrel{\‾}{\mathrm{dist}}}_{\mathrm{ij}}\right)=\left\{\begin{array}{c}1,{\stackrel{\‾}{\mathrm{dist}}}_{\mathrm{ij}}\≤0\\ \frac{1}{2}-\frac{1}{2}\sin (\frac{\mathrm{\π}}{{\mathrm{Th}}_i}*({\stackrel{\‾}{\mathrm{dist}}}_{\mathrm{ij}}-\frac{{\mathrm{Th}}_i}{2})),0\≤{\stackrel{\‾}{\mathrm{dist}}}_{\mathrm{ij}}\≤{\mathrm{Th}}_i\\ 0,{\stackrel{\‾}{\mathrm{dist}}}_{\mathrm{ij}}\≥{\mathrm{Th}}_i\end{array}\right.$

Wherein,representing the abnormal sample and the jth fault sample under the fault type iIdentifying the calculated pattern distance, Th_iRepresenting a mode distance threshold under fault type i;

according to the dimension of the fault type, each distance vector is divided intoConverted into similarity vectors

Finally, a similarity set of the abnormal sample FT and all samples of the fault knowledge base is obtained, and the similarity set is as follows:

$\mathrm{\ρSets}=\left\{\begin{array}{c}[\mathrm{\ρ}\left({\stackrel{\‾}{\mathrm{dist}}}_{11}\right),\mathrm{\ρ}\left({\stackrel{\‾}{\mathrm{dist}}}_{12}\right),...,\mathrm{\ρ}{\stackrel{\‾}{(\mathrm{dist}}}_{1,c1})]\\ [\mathrm{\ρ}\left({\stackrel{\‾}{\mathrm{dist}}}_{21}\right),\mathrm{\ρ}\left({\stackrel{\‾}{\mathrm{dist}}}_{22}\right),...,\mathrm{\ρ}{\stackrel{\‾}{(\mathrm{dist}}}_{2,c2})]\\ ......\\ ......\\ [\mathrm{\ρ}\left({\stackrel{\‾}{\mathrm{dist}}}_{x1}\right),\mathrm{\ρ}\left({\stackrel{\‾}{\mathrm{dist}}}_{x2}\right),...,\mathrm{\ρ}{\stackrel{\‾}{(\mathrm{dist}}}_{x,\mathrm{cx}})]\end{array}\right\}=\left\{\begin{array}{c}\mathrm{\ρ}{\mathrm{Vector}}_1\\ \mathrm{\ρ}{\mathrm{Vector}}_2\\ ...\\ ...\\ \mathrm{\ρ}{\mathrm{Vector}}_x\end{array}\right\}.$

further, the step (2.6) comprises the following specific steps:

and outputting a fault diagnosis result according to the similarity set result rho Sets and an agreed rule, wherein the rule is satisfied:

rule 1: if a plurality of fault samples with the similarity exceeding 90% with the abnormal samples exist, outputting the serial numbers of all the samples, simultaneously outputting the fault type to which the fault sample belongs, the confidence coefficient of the fault type to which the fault sample belongs, and the best matching position of the sample with the maximum similarity in the fault development stage, and if the confidence coefficient exceeds 50%, outputting suggested maintenance measures;

rule 2: if no fault sample with the similarity exceeding 90% with the abnormal sample exists and the maximum similarity value is greater than or equal to 60%, outputting the serial number of the maximum similarity sample, simultaneously outputting the fault type to which the fault sample belongs, the confidence coefficient of the fault type to which the fault sample belongs and the best matching position of the maximum similarity sample at the fault development stage, and if the confidence coefficient exceeds 50%, outputting the recommended maintenance measures;

rule 3: if no fault sample with the similarity exceeding 90% with the abnormal sample exists and the maximum similarity value is less than 60%, outputting the abnormal sample as an unusual working condition or an unknown fault type, and continuing to pay attention to the development of the abnormality.

The equipment fault diagnosis method of the invention can realize that:

1. the invention is a brand-new fault diagnosis technology independent of faults. The present invention starts from analyzing the mathematical morphological characteristics of historical fault sample data, and judges the similarity degree and the most possible development trend of the existing fault in time for the current abnormal sample.

2. The invention can efficiently and accurately carry out fault diagnosis work, realizes the data information compression target and eliminates the influence of redundant data noise by converting the sample data from the time domain space to the mathematical morphology characteristic space, thereby greatly improving the fault diagnosis efficiency.

3. In order to ensure the effectiveness of pattern expression, the invention establishes a multidimensional segmentation fitting method based on the data development state. In consideration of the characteristic that the coupling among all variables in the sample data is unclear, the optimal segmentation threshold of the current sample data is mined by adopting a self-adaptive algorithm, and the segmentation positions of all the variables are consistent, so that the development state of the segmented data in each stage is more hierarchical.

4. The invention has low requirement on sample data for fault diagnosis and does not need to carry out complex data preprocessing work. The prior related technology generally cannot realize fault diagnosis of sample data with inconsistent sample time and different variable numbers in a knowledge base, and the method can solve the problem by methods of variable trimming, dynamic time warping and the like.

5. The operation speed of the fault analysis and fault identification part is within the second level, so that not only can the fault diagnosis of a sample of a complete fault mode be realized, but also the fault location of a part of fault modes can be realized. Therefore, the invention can be completely transplanted to an equipment on-line state detection system for application.

Drawings

FIG. 1 is a comprehensive flow chart of the training phase and the operation phase of the fault diagnosis model

FIG. 2 dynamic time warping distance effect plot

FIG. 3 is a trend graph of key measuring points of three faults of a front pump of a steam pump

FIG. 4 is a diagram illustrating the effect of online fault diagnosis of abnormal samples

Detailed Description

Reference will now be made in detail to the embodiments of the present invention, the following examples of which are intended to be illustrative only and are not to be construed as limiting the scope of the invention.

Most of the prior related technologies are that each variable dimension of fault data is separately cut apart for calculation, but the invention comprehensively considers the problems of related information between different variables of sample data, realizes a segmentation fitting technology based on bottom-up, extracts various overall morphological characteristics of the fault data and simultaneously considers the numerical characteristics of the fault data. The method can realize effective compression of large-scale data, reduce the fault diagnosis time and fundamentally improve the fault diagnosis precision, so the method has the capability of quick fault diagnosis. The invention carries out the calculation of the characteristic difference of the fault samples by a dynamic time warping method, and can solve the problem that the similarity of the fault samples with different lengths is difficult to measure. The fault diagnosis method of the invention does not take the mode distance as the final measurement standard, but compares the mode distance with the similarity percentage, unifies the measurement standard of fault identification, and further improves the fault diagnosis precision again.

The invention provides an equipment fault diagnosis method based on multi-dimensional segment fitting, which is a brand-new fault diagnosis technical method for setting out the essence of data morphological features. The method mainly comprises two processes of establishing a model and operating the model.

Fig. 1 (left) is a flow chart of modeling of the present invention, and the whole modeling process mainly includes the following steps:

step 1: and acquiring fault sample information from a system database.

The obtained fault sample information comes from a device state detection system of a power plant, and the rough operation process is as follows: firstly, selecting a piece of research-available equipment with the number P (P > ═ 2) of fault types and the occurrence frequency T (T > ═ 2) of each fault meeting the requirement, selecting an observation point N (N > ═ 10) of the number of feet, and searching the fault record of the historical operation state data of the equipment for a long enough time; then, useful information such as fault related measuring point information, starting and stopping time of a fault process, fault maintenance measure records and the like are picked out from the fault records by using a set screening rule; and finally, reading fault sample data from the power plant real-time database PI according to the information.

For example, for a fault sample data with n measurement points and m time points, all measurement point data at time j can be regarded as an n-dimensional column vector, which is expressed as:

u(t_j)＝[u_j1，u_j2，u_j3，…，u_jn]

the sample data file is stored in a matrix form of m × n, and the specific form is as follows:

$F=\left[\begin{array}{c}{u}_{11},u_{12},u_{13},...,u_{1n}\\ u_{21},u_{22},u_{23},...,u_{2n}\\ ..........................\\ ..........................\\ ..........................\\ u_{m1},u_{m2},u_{m3},...,u_{\mathrm{mn}}\end{array}\right]$

according to the method, the storage form of all fault samples is as follows: the dimension of row represents m failure times, the dimension of column represents n device observation points, and the two values of the dimension (m, n) of the row and the column between each failure sample may not be the same. While each fault sample is given its fault type identification ID. The method for determining the fault type identifier ID comprises the following steps: if all samples contain X faults, the numerical range of the fault type identifier ID is as follows: 1 to X.

Step 2: sequentially carrying out multi-dimensional piecewise linear fitting on each fault sample

The step is a key technical part of the fault diagnosis invention, and is mainly used for carrying out self-adaptive state segmentation according to the overall development trend of a fault sample and carrying out data preprocessing preparation for subsequent fault feature extraction.

Step 2.1 mean filtering operation:

in order to effectively segment fault sample data, noise pollution doped in the sample data is filtered and eliminated. To quickly and efficiently implement the filtering of noisy data, a typical linear filtering algorithm is chosen: and (5) average filtering.

According to the principle of mean filtering, the specific filtering operation process is as follows: firstly, extracting the first observation point column vector data u of the fault sample data F_-1The form is as follows:

$\left[\begin{array}{c}{u}_{11}\\ u_{21}\\ u_{31}\\ .\\ .\\ .\\ u_{m1}\end{array}\right]$

will column vector u_-1Filter template size viewed as a sequence of numbers(round is a rounding function) if L is an even number, then 1 more is needed. After the size of the filtering template is determined, the substantive filtering operation is started:

1. the center position of the filtering template is placed in u₁₁Where the sum of all data within the template is divided by L, the value is u₁₁The output value of the filtering operation of (2) is:

$U_{11}=\underset{i=1}{\overset{(L-1)/2}{\mathrm{\Σ}}}\mathrm{data}(1:(L-1)/2)/L;$

2. the center position of the filtering template is moved forward continuously, and filtering output is carried out according to the same operation until all data are completely filtered, for example, the filtering operation is carried out at u_1iThe output of the position, mean filtering is:

$U_{1i}=\underset{i=1}{\overset{L}{\mathrm{\Σ}}}\mathrm{data}(i-(L-1)/2:i+(L-1)/2)/L;$

3. finally obtained column vector data u_-1The filtered data of (2) is also a column of vectors, and the specific form is as follows:

$\left[\begin{array}{c}{U}_{11}\\ U_{21}\\ U_{31}\\ .\\ .\\ .\\ U_{m1}\end{array}\right]$

all measuring points of fault sample data are sequentially subjected to filtering operation according to the 3 steps, so that the fault sample data F is converted into F^*In the form of:

$F^{*}=\left[\begin{array}{c}{U}_{11},U_{12},U_{13},...,U_{1n}\\ U_{21},U_{22},U_{23},...,U_{2n}\\ ..........................\\ ..........................\\ ..........................\\ U_{m1},U_{m2},U_{m3},...,U_{\mathrm{mn}}\end{array}\right]$

step 2.2 Fault sample segmentation initialization

Carrying out segmentation initialization on the filtered fault sample F, dividing each 2 pieces of data into 1 data segment, and dividing the fault sample into 1 data segment because the length of the fault sample is mThe most detailed data segment; if m is odd, the last segment is composed of 3 pieces of data, and the final segment is divided intoThe most detailed data segment. The initialization effect of the segmentation is as follows:

$F^{*}=\left[\begin{array}{c}{f}_1^{*}\\ f_2^{*}\\ .\\ .\\ .\\ f_{m/2}^{*}\end{array}\right]$

wherein, $f_i^{*}=\left[\begin{array}{c}{u}_{(2*i-1)1},u_{(2*i-1)2},u_{(2*i-1)3},...,u_{(2*i-1)n}\\ u_{(2*i)1},u_{(2*i)2},u_{(2*i)3},...,u_{(2*i)n}\end{array}\right]$

step 2.3 initialization data segment pairwise combination calculation fitting error

The main purpose of this step is to generate the above step after initialization segmentation(or) And combining each data segment with the data segment adjacent to the right of the data segment, and simultaneously calculating the generated fitting error.

The detailed merge operation is as follows: such as mergingAndcomprises the following steps: $F_{\mathrm{ii}+1}^{*}=\left[\begin{array}{c}{u}_{11},u_{12},u_{13},...,u_{1n}\\ u_{21},u_{22},u_{23},...,u_{2n}\\ u_{31},u_{32},u_{33},...,u_{3n}\\ u_{41},u_{42},u_{43},...,u_{4n}\end{array}\right],$ in the form of a matrix, e.g. 4 × n, with columns being the number of test points $\left[\begin{array}{c}{u}_{1i}\\ u_{2i}\\ u_{3i}\\ u_{4i}\end{array}\right]$ And (5) fitting a straight line by a least square method. The least square method comprises the following specific steps:

the fitting objective was to fit the data $(x_j,y_j)\∈\left[\begin{array}{c}1,u_{1i}\\ 2,u_{2i}\\ 3,u_{3i}\\ 4,u_{4i}\end{array}\right]$ Using a first order polynomial p (x) ═ a₀+a₁x is expressed such that the cumulative sum of fitting errors I is (a)₀+a₁x-y_i)²And minimum.

This is apparently a solution of I ═ I (a)₀，a₁) And (4) quadratic programming of extreme values. According to the Lagrange principle to obtain

$\frac{\∂I}{\∂a_j}=2\underset{i=0}{\overset{m}{\mathrm{\Σ}}}(\underset{k=0}{\overset{n}{\mathrm{\Σ}}}{a}_k{x}_i^k-y_i)x_i^j=0,j=\mathrm{0,1}$

Namely: $\underset{k=0}{\overset{n}{\mathrm{\Σ}}}\left(\underset{i=0}{\overset{m}{\mathrm{\Σ}}}{x}_i^{j+k}\right)a_k=\underset{i=0}{\overset{m}{\mathrm{\Σ}}}{x}_i^j{y}_i,j=\mathrm{0,1}$

wherein m is 4n is 1

The above formula is a₀，a₁The linear equation system is expressed by matrix as

$\left[\begin{array}{c}m+1,\underset{i=0}{\overset{m}{\mathrm{\Σ}}}{x}_i\\ \underset{i=0}{\overset{m}{\mathrm{\Σ}}}{x}_i,\underset{i=0}{\overset{m}{\mathrm{\Σ}}}{x}_i^2\end{array}\right]\left[\begin{array}{c}{a}_0\\ a_1\end{array}\right]=\left[\begin{array}{c}\underset{i=0}{\overset{m}{\mathrm{\Σ}}}{y}_i\\ \underset{i=0}{\overset{m}{\mathrm{\Σ}}}{x}_i{y}_i\end{array}\right]$

The above formula becomes a normal equation system, and the coefficient matrix is a symmetric positive definite matrix, so that a unique solution exists. Solving for a from the above formula_k(k is 0, 1), thus obtaining a polynomial:

p(x)＝a₀+a₁x

the mean error due to the fit is called the fit error, which is recorded as:

${\left|\right|r_i\left|\right|}_2^2=\underset{i=0}{\overset{m}{\mathrm{\Σ}}}{[p\left(x_i\right)-y_i]}^2$

after all measuring points are fitted, the fitting errors generated are summed(N is the number of stations).

When all are(or) After the data segments are combined pairwise, three results are generated, namely, a data segment set is divided $F^{*}=[f_1^{*},f_2^{*},...,f_{\frac{m}{4}}^{*}];$ Second, fitting error vector $\stackrel{\‾}{R}=[R_1,R_2,...,R_{\frac{m}{4}-1}];$ Third, the sample integral segmentation position vector $\stackrel{\‾}{T}=[T_1,T_2,...,T_{\frac{m}{4}-1}].$

Step 2.4 determining the segmentation cut points of the fault samples

The main purpose of the step is to realize the determination of the fault sample segmentation position in a recursive combination mode.

1. According to the invention, a numerical value MAX _ ERROR is set according to experience as an index for judging the fitting ERROR. If the error vector is fittedIf all values in the set are greater than the maximum fitting ERROR threshold MAX _ ERROR, then operation 1 is performed: confirmationIs the final set of sample segmentation locations; otherwise, operation 2 is performed: combining fitting error vectorsThe middle minimum value position J and the segmentation data segment of the right adjacent position J +1 thereof, and updating the segmentation data segment set F^*Fitting error vectorAnd sample integral segmentation position vector

The specific updating operation is as follows: f^*The length minus 1, the preceding J-1 being kept constant, the value following J being shifted forward by 1, F 'in the position of J'_J＝[F_J；F_J+1](ii) a According to updated F^*Determining a new segmentation position vector The length is reduced by 1, the front J-1 is kept unchanged, the value behind J is shifted forward by 1 bit, and the position of J ismerge is a fitting error function generated and designed according to a least square method;

2. judging the magnitude of the updated fitting ERROR vector value and the maximum fitting ERROR threshold MAX _ ERROR again, if all the fitting ERROR vector values are larger than the threshold MAX _ ERROR, executing operation 1, and otherwise, skipping out of the cycle operation and stopping;

3. circularly combining the segmentation data segments of the minimum fitting error position and the right adjacent position thereof until the final fitting error vectorAll above the MAX ERROR value, the recursive merge loop stops.

After the above steps, after all the segmented data segments are completed by means of recursive combination, three results are generated, one is a set of segmented data segmentsSecond, fitting error vectorThird, the sample integral segmentation position vector(where x is the number of final segmentations of the sample data). Therefore, the target of adaptively segmenting the state of the fault sample is realized, and preparation is made for next fault feature extraction.

And step 3: performing feature extraction on the fault sample segmentation data segment;

the method realizes the extraction of mathematical characteristics such as slope, time proportion, mean value, variance and the like for each segmented data segment of the current fault sample.

For example, the failure sample F^*Is divided intoThe x divided data segments. FromThe data segment begins feature extraction.

The specific operation of feature extraction is as follows:

due to the fact thatThe data segment is $\left[\begin{array}{c}{u}_{11},u_{12},...,u_{1n}\\ u_{21},u_{22},...,u_{2n}\\ ...................\\ ...................\\ u_{h1},u_{h2},...,u_{\mathrm{hn}}\end{array}\right]$ In matrix form, the row is the number of time points and the column is the number of test points.

Feature extraction is carried out according to the dimension (column) of the measuring points, and the main features are as follows: slope k, duration l, mean m, variance v.

To be provided with $\stackrel{\‾}{V_i}=\left[\begin{array}{c}{u}_{1i}\\ u_{2i}\\ ...\\ ...\\ u_{\mathrm{hi}}\end{array}\right]$ For example, how to perform the relevant feature extraction work is described:

the vector is linearly fitted according to the principle of least square method, and the fitting result is a linear equation p (x) a_0i+a_1ix. So that the slope characteristic k is a_1i；

Duration characteristic l isVector length, i.e., l ═ h;

mean characteristic m isAverage of all values of the vector, i.e.

The variance characteristic v isSum of fluctuation amplitudes of all values of the vector from the mean, i.e.

Thus dividing the data segment $f_1^{*}=\left[\begin{array}{c}{u}_{11},u_{12},...,u_{1n}\\ u_{21},u_{22},...,u_{2n}\\ ...................\\ ...................\\ u_{h1},u_{h2},...,u_{\mathrm{hn}}\end{array}\right]$ Is converted into a feature matrix $\left[\begin{array}{c}{k}_1,k_2,...,k_n\\ l,l,...,l\\ m_1,m_2,...,m_n\\ v_1,v_2,...,v_n\end{array}\right].$

The rest ofAnd (4) extracting the characteristics according to the operation method, and converting the original time domain data matrix into a mathematical characteristic matrix.

Final segmentation of data segment setsConversion to the form:

$\left[\begin{array}{c}(k_{11};l_1;m_{11};v_{11}),(k_{12};l_1;m_{12};v_{12}),...,(k_{1n};l_1;m_{1n};v_{1n})\\ (k_{21};l_2;m_{21};v_{21}),(k_{22};l_2;m_{22};v_{22}),...,(k_{2n};l_2;m_{2n};v_{2n})\\ ......\\ ......\\ (k_{x1};l_x;m_x;v_{x1}),(k_{x2};l_x;m_{x2};v_{x2}),...,(k_{\mathrm{xn}};l_x;m_{\mathrm{xn}};v_{\mathrm{xn}})\end{array}\right]$

and 4, step 4: carrying out fault characteristic conversion and eliminating characteristic dimension;

1 the slope k is converted into an inclination angle α, since k has a value range of [ - ∞, + ∞]In order to prevent the similar two fault samples from generating excessive difference of slope values due to the existence of fitting errors and finally causing misjudgment, the slope values are converted into angle values through an arc tangent function to eliminate the characteristic dimension, and the slope characteristic conversion formula is α_ij＝arctan(k_ij) Wherein i ∈ [1, x]；j∈[1，n；

Duration l translates into time ratio p: under the condition that the lengths of the fault samples are different, the segmentation duration cannot represent the real and effective segmentation stage length, and misjudgment is easy to occur, so that the fault samples need to be judgedWhen divided by the whole length of the original fault sample, the dimension of the time length characteristic is eliminated. The time length characteristic conversion formula is as follows:wherein i ∈ [1, x]；

3 mean feature m normalization: under the condition that the segmentation numbers of the fault samples are different, the average value cannot represent the real and effective numerical characteristics of the segmentation stage, and misjudgment is easy to generate, so the total numerical quantity of the original fault samples should be divided to eliminate the dimension of the average value characteristics. The time length characteristic conversion formula is as follows:wherein i ∈ [1, x]；j∈[1，n]；

Feature matrix of final fault samples $\left[\begin{array}{c}(k_{11};l_1;m_{11};v_{11}),(k_{12};l_1;m_{12};v_{12}),...,(k_{1n};l_1;m_{1n};v_{1n})\\ (k_{21};l_2;m_{21};v_{21}),(k_{22};l_2;m_{22};v_{22}),...,(k_{2n};l_2;m_{2n};v_{2n})\\ ......\\ ......\\ (k_{x1};l_x;m_x;v_{x1}),(k_{x2};l_x;m_{x2};v_{x2}),...,(k_{\mathrm{xn}};l_x;m_{\mathrm{xn}};v_{\mathrm{xn}})\end{array}\right]$ Converted into elimination dimensions $\left[\begin{array}{c}({\mathrm{\α}}_{11};p_1;T_{11};v_{11}),({\mathrm{\α}}_{12};p_1;T_{12};v_{12}),...,({\mathrm{\α}}_{1n};p_1;T_{1n};v_{1n})\\ ({\mathrm{\α}}_{21};p_2;T_{21};v_{21}),({\mathrm{\α}}_{22};p_2;T_{22};v_{22}),...,({\mathrm{\α}}_{2n};p_2;T_{2n};v_{2n})\\ ......\\ ......\\ ({\mathrm{\α}}_{x1};p_x;T_x;v_{x1}),({\mathrm{\α}}_{x2};p_x;T_{x2};v_{x2}),...,({\mathrm{\α}}_{\mathrm{xn}};p_x;T_{\mathrm{xn}};v_{\mathrm{xn}})\end{array}\right]---\left(1\right)$ Such a matrix form;

and 5: calculating a mode distance threshold;

the purpose of this step is to find the mode distance between every two of all the samples obtained in step 1 and calculate the distance threshold for each fault category.

Through the four steps, all fault sample data { F¹，F²，…，F^sAll the data are converted into a feature matrix form set { fm } similar to the formula (1)¹，fm²，…，fm^s}. And then calculating the mode distance of two different samples in the same type of fault samples by using a Dynamic Time Warping (DTW) method, and determining the mode distance threshold of each type of fault according to the calculation result.

The specific operation of calculating the mode distance by the Dynamic Time Warping (DTW) method is described below:

1 there are two sample feature matrices I under the assumption of failure type I_AAnd I_BThe specific form is as follows:

$I_A=\left[\begin{array}{c}({\mathrm{\α}}_{11};p_1;T_{11};v_{11}),({\mathrm{\α}}_{12};p_1;T_{12};v_{12}),...,({\mathrm{\α}}_{1n};p_1;T_{1n};v_{1n})\\ ({\mathrm{\α}}_{21};p_2;T_{21};v_{21}),({\mathrm{\α}}_{22};p_2;T_{22};v_{22}),...,({\mathrm{\α}}_{2n};p_2;T_{2n};v_{2n})\\ ......\\ ......\\ ({\mathrm{\α}}_{x1};p_x;T_x;v_{x1}),({\mathrm{\α}}_{x2};p_x;T_{x2};v_{x2}),...,({\mathrm{\α}}_{\mathrm{xn}};p_x;T_{\mathrm{xn}};v_{\mathrm{xn}})\end{array}\right]=\left[\begin{array}{c}{a}_1\\ a_2\\ ...\\ ...\\ a_x\end{array}\right]$

$I_B=\left[\begin{array}{c}({\stackrel{\‾}{\mathrm{\α}}}_{11};{\stackrel{\‾}{p}}_1;{\stackrel{\‾}{T}}_{11};{\stackrel{\‾}{v}}_{11}),({\stackrel{\‾}{\mathrm{\α}}}_{12};{\stackrel{\‾}{p}}_1;{\stackrel{\‾}{T}}_{12};{\stackrel{\‾}{v}}_{12}),...,({\stackrel{\‾}{\mathrm{\α}}}_{1n};{\stackrel{\‾}{p}}_1;{\stackrel{\‾}{T}}_{1n};{\stackrel{\‾}{v}}_{1n})\\ ({\stackrel{\‾}{\mathrm{\α}}}_{21};{\stackrel{\‾}{p}}_2;{\stackrel{\‾}{T}}_{21};{\stackrel{\‾}{v}}_{21}),({\stackrel{\‾}{\mathrm{\α}}}_{22};{\stackrel{\‾}{p}}_2;{\stackrel{\‾}{T}}_{22};{\stackrel{\‾}{v}}_{22}),...,({\stackrel{\‾}{\mathrm{\α}}}_{2n};{\stackrel{\‾}{p}}_2;{\stackrel{\‾}{T}}_{2n};{\stackrel{\‾}{v}}_{2n})\\ ......\\ ......\\ ({\stackrel{\‾}{\mathrm{\α}}}_{x^{\′}1};{\stackrel{\‾}{p}}_{x^{\′}};{\stackrel{\‾}{T}}_{x^{\′}1};{\stackrel{\‾}{v}}_{x^{\′}1}),({\stackrel{\‾}{\mathrm{\α}}}_{x^{\′}2};{\stackrel{\‾}{p}}_{x^{\′}};{\stackrel{\‾}{T}}_{x^{\′}2};{\stackrel{\‾}{v}}_{x^{\′}2}),...,({\stackrel{\‾}{\mathrm{\α}}}_{x^{\′}n};{\stackrel{\‾}{p}}_{x^{\′}};{\stackrel{\‾}{T}}_{x^{\′}n};{\stackrel{\‾}{v}}_{x^{\′}n})\end{array}\right]=\left[\begin{array}{c}{b}_1\\ b_2\\ ...\\ ...\\ b_{x^{\′}}\end{array}\right]$

wherein, a_i(i＝1，2，…，x)，b_j(j ═ 1, 2, …, x') for two sample feature matrices I, respectively_AAnd I_BThe row vector of (a) represents the mathematical characterization of each segmented data segment of the respective fault sample in all the measurement point dimensions.

2 two samples S under the same fault type_AAnd S_BCharacteristic matrix I of_A、I_BAccording to the dynamic time warping principle, and the optimal path is determined by combining the mode distance I with the mode distance I shown in the figure 2_AAnd I_BThe row vectors are copied and stretched to meet the condition that the length of the characteristic matrix columns of the two samples are aligned equally I_AAnd I_BThe mode distance calculation result of (2) is minimum. S_AAnd S_BThe formula for calculating the mode distance is as follows:

${\mathrm{MD}}_{\mathrm{DTW}}(S_A,S_B)={\mathrm{MD}}_{\mathrm{base}}(a_1,b_1)+\min \left\{\begin{array}{c}{\mathrm{MD}}_{\mathrm{DTW}}(I_A,I_B[2:-\left]\right),\\ {\mathrm{MD}}_{\mathrm{DTW}}\left(I_A\right[2:-],I_B),\\ {\mathrm{MD}}_{\mathrm{DTW}}\left(I_A\right[2:-],I_B[2:-\left]\right).\end{array}\right.$

wherein, T [ i: represents a multivariate sequence consisting of the ith row vector to the last row vector of the feature matrix T;

MD_base(a_i，b_j) The Euclidean distance from the segmented data segment is represented, and the specific calculation method is as follows:

${\mathrm{MD}}_{\mathrm{base}}(a_i,b_j)=\underset{t=1}{\overset{n}{\mathrm{\Σ}}}\mathrm{\ϵ}*\mathrm{eud}({\mathrm{\α}}_i,{\stackrel{\‾}{\mathrm{\α}}}_j)+\mathrm{\λ}*\mathrm{eud}(p_i,{\stackrel{\‾}{p}}_j)+\mathrm{\ρ}*\mathrm{eud}(T_i,{\stackrel{\‾}{T}}_j)+\mathrm{\ω}*\mathrm{eud}(v_i,{\stackrel{\‾}{v}}_j)$

wherein eud (X, Y) represents the Euclidean distance between two vectors X, Y, i.e. the Euclidean distanceThe four variables of lambda, rho and omega represent the weight values of the four characteristics of the inclination angle α, the time proportion p, the mean value proportion T and the variance v in the calculation mode distance, and the sum of the four variables of lambda, rho and omega is 1, and all the variables are set to be 0.25

According to the DTW calculation method described above, two samples S can be obtained_AAnd S_BHas a mode distance of

If there are d fault samples in the fault type I, the d fault samples are obtained according to the permutation and combination mode and the calculation mode $\stackrel{\‾}{n}=[{\mathrm{Md}}_{[s_1,s_2]},{\mathrm{Md}}_{[s_1,s_3]},...,{\mathrm{Md}}_{[s_{d-1},s_d]}]$ In totalIndividual mode distance results. And a pattern distance threshold Th of fault type I₁Should be a vectorMaximum value of

According to the above method, the mode distance threshold of all fault types can be calculated, and a mode distance threshold vector is formed:

Thsets＝[Th₁，Th₂，…，Th_x]

step 6: integrating useful information to generate a failure mode knowledge base;

the useful information in the fault knowledge base should include the following parts:

1 feature matrix set FM ═ { FM ] extracted from all fault samples¹，fm²，…，fm^s} (s is the number of failed samples);

2 measuring point name set PT ═ { PT associated with each fault type₁，pt₂，…，pt_x} (x is the number of categories of the type of failure);

3 pattern distance threshold vector Thsets corresponding to each fault type [ Th ]₁，Th₂，…，Th_x](x is the number of categories of fault types).

In order to correlate the two pieces of information (fault signature matrix and pattern distance threshold) with each other, the invention stores the fault information in a fault knowledge base in the following manner:

$\mathrm{faultMessageSets}=\left\{\begin{array}{c}{\mathrm{fm}}^{11},{\mathrm{fm}}^{12},...,{\mathrm{fm}}^{1,c1},{\mathrm{pt}}_1,{\mathrm{Th}}_1\\ {\mathrm{fm}}^{21},{\mathrm{fm}}^{22},...,{\mathrm{fm}}^{2,c2},{\mathrm{pt}}_2,{\mathrm{Th}}_2\\ ......\\ ......\\ {\mathrm{fm}}^{x1},{\mathrm{fm}}^{x2},...,{\mathrm{fm}}^{x,c1},{\mathrm{pt}}_x,{\mathrm{Th}}_x\end{array}\right\}$

fig. 1 (right) is a flow chart of the model operation of the present invention, and the whole operation process mainly includes the following 6 steps:

step 1: acquiring abnormal sample information from a real-time database;

according to the condition early warning system of the power plant, the condition of the equipment is found to be in an unknown abnormal condition, and the following relevant operations are carried out: 1 determining the alarm generation time t of the equipment from the early warning system₁And a cut-off time t₂；

2, determining a relevant observation point wp ═ x of the equipment alarm from the early warning system₁，x₂，…，x_n′]；

According to the generation time t₁And a cut-off time t₂And database sampling frequency fs, the number of available time points m ═ fs × (t)₂-t₁) The number of device measurement points n equals length ([ x ])₁，x₂，…，x_n′]) (length calculation length function).

Thus, the obtained abnormal sample data is sample data with n 'measuring points and m time points, and all the measuring point data at the moment j can be regarded as a column vector with n' dimension, which is expressed as:

v(t_j)＝[v_j1，v_j2，v_j3，…，v_jn′]

the sample data file is stored in a matrix form of m × n', and the specific form is as follows:

$\stackrel{\‾}{F}=\left[\begin{array}{c}{v}_{11},v_{12},v_{13},...,v_{{1n}^{\′}}\\ v_{21},v_{22},v_{23},...,v_{{2n}^{\′}}\\ ..........................\\ ..........................\\ ..........................\\ v_{m1},v_{m2},v_{m3},...,v_{{\mathrm{mn}}^{\′}}\end{array}\right]$

according to the method, the obtained abnormal sample storage form is as follows: the row dimension represents m failure times and the column dimension represents n' device observation points.

Step 2: performing piecewise linear fitting on the current abnormal sample;

the method mainly carries out adaptive state segmentation according to the overall development trend of the abnormal sample, and carries out data preprocessing preparation for subsequent fault feature extraction. The method is operated according to piecewise linear fitting of a model training stage:

2.1 mean value Filter operation

Setting an abnormal sampleThe length of the sample is 0.1 times as the length of the filtering template, and the noise pollution doped in each measuring point in the abnormal sample is filtered and eliminated according to the mean filtering principle. The specific operation is seen in the model training phase step 2.1.

All the measuring points of the abnormal sample data finish filtering operation according to corresponding steps in sequence, so that the abnormal sample dataIs converted intoThe form is as follows:

${\stackrel{\‾}{F}}^{*}=\left[\begin{array}{c}{V}_{11},V_{12},V_{13},...,V_{1n}\\ V_{21},V_{22},V_{23},...,V_{2n}\\ ..........................\\ ..........................\\ ..........................\\ V_{m1},V_{m2},V_{m3},...,V_{\mathrm{mn}}\end{array}\right]$

2.2 Exception sample segmentation initialization

For filtered abnormal samplesSegment initialization is carried out, every 2 pieces of data are divided into 1 data segment, and the abnormal sample data is divided into 1 data segment due to the fact that the length of the abnormal sample data is mThe most detailed data segment; if m is odd, the last segment is composed of 3 pieces of data, and the final segment is divided intoThe most detailed data segment. The initialization effect of the segmentation is as follows:

${\stackrel{\‾}{F}}^{*}=\left[\begin{array}{c}{\stackrel{\‾}{f}}_1^{*}\\ {\stackrel{\‾}{f}}_2^{*}\\ .\\ .\\ .\\ {\stackrel{\‾}{f}}_{m/2}^{*}\end{array}\right]$

wherein, ${\stackrel{\‾}{f}}_i^{*}=\left[\begin{array}{c}{V}_{(2*i-1)1},V_{(2*i-1)2},V_{(2*i-1)3},...,V_{(2*i-1)n}\\ V_{(2*i)1},V_{(2*i)2},V_{(2*i)3},...,V_{(2*i)n}\end{array}\right].$

2.3 computing fitting error by combining two initialized data segments

Generated after initializing the step 2.2 abnormal sample segment(or) And each data segment and the right adjacent data segment are fitted according to the least square principle to realize combination, and the fitting error sum generated by all the measuring points is calculated.

When all are(or) After the data segments are combined pairwise, three results are generated, 1, the data segment set is divided ${\stackrel{\‾}{F}}^{*}=[{\stackrel{\‾}{f}}_1^{*},{\stackrel{\‾}{f}}_2^{*},...,{\stackrel{\‾}{f}}_{\frac{m}{4}}^{*}];$ 2 fitting error vector $\stackrel{=}{R}=[R_1,R_2,...,R_{\frac{m}{4}-1}];$ 3-sample global segmentation position vector $\stackrel{=}{T}=[T_1,T_2,...,T_{\frac{m}{4}-1}].$

2.4 recursive merging to determine outlier sample segmentation cut points

This step specifically operates the reference model construction phase, divided aboveContinuously merging two adjacent data segments generating the minimum fitting error by the recursive merging mode of each data segment to generate 3 updated results, wherein 1 is a set of divided data segmentsSecond, fitting error vectorThird, the sample integral segmentation position vector(where x is the number of final segmentations of the sample data). The goal of adaptive state segmentation of the anomaly samples is achieved,and extracting the operation object for the next step of abnormal features.

And step 3: carrying out abnormal feature extraction and transformation on the data subjected to segmentation fitting;

the method realizes extraction of 4 mathematical characteristics of slope, time span, mean value and variance of each data segment after the current abnormal sample is segmented.

For example, abnormal samplesIs divided intoThe x divided data segments. The specific operation of feature extraction may refer to step 3 of the model construction phase.

Each of the divided data segments ${\stackrel{\‾}{f}}_i^{*}=\left[\begin{array}{c}{u}_{11},u_{12},...,u_{1n}\\ u_{21},u_{22},...,u_{2n}\\ ...................\\ ...................\\ u_{h1},u_{h2},...,u_{\mathrm{hn}}\end{array}\right]$ Extracting characteristics such as corresponding slope k, time span l, mean value m, variance v and the like according to the dimension of the measuring point, wherein the result after extraction is as follows:

$\left[\begin{array}{c}{u}_{11},u_{12},...,u_{1n}\\ u_{21},u_{22},...,u_{2n}\\ ...................\\ ...................\\ u_{h1},u_{h2},...,u_{\mathrm{hn}}\end{array}\right]\⇒\begin{array}{}\end{array}\left[\begin{array}{c}{k}_1,k_2,...,k_n\\ l,l,...,l\\ m_1,m_2,...,m_n\\ v_1,v_2,...,v_n\end{array}\right]$

when all the segmentation data segments of the abnormal sample are subjected to feature extraction according to the method, the segmentation data segment vector with the length of xCan be converted into the following characteristic matrix form:

$\left[\begin{array}{c}(k_{11};l_1;m_{11};v_{11}),(k_{12};l_1;m_{12};v_{12}),...,(k_{1n};l_1;m_{1n};v_{1n})\\ (k_{21};l_2;m_{21};v_{21}),(k_{22};l_2;m_{22};v_{22}),...,(k_{2n};l_2;m_{2n};v_{2n})\\ ......\\ ......\\ (k_{x1};l_x;m_x;v_{x1}),(k_{x2};l_x;m_{x2};v_{x2}),...,(k_{\mathrm{xn}};l_x;m_{\mathrm{xn}};v_{\mathrm{xn}})\end{array}\right]$

in order to prevent the abnormal samples and similar fault samples from causing final misjudgment when calculating the pattern distance due to the following reasons:

1, the existence of fitting errors causes overlarge difference of slope values, and the solution is to convert the slope value k into an angle value alpha through an arc tangent function so as to eliminate the characteristic dimension;

2, under the condition that the lengths of the fault samples are different, the segmentation time length l cannot represent the proportion of a real and effective segmentation stage, misjudgment is easy to generate, and the segmentation time length l is divided by the whole length m of the original fault sample to eliminate the dimension of time length characteristics;

3 under the condition that the number of the divided fault samples is different, the mean value cannot represent the real and effective numerical characteristics of the divided stages, and misjudgment is easy to occur, so the total numerical quantity of the original fault samples should be divided to eliminate the dimension of the mean value characteristics.

The specific operation method for eliminating the dimension can be seen in the characteristic transformation formula of step 4 of the model construction stage. Thus transformed, the features are transformed into the following dimensionless form:

$\stackrel{\‾}{\mathrm{FT}}=\left[\begin{array}{c}({\mathrm{\α}}_{11};p_1;T_{11};v_{11}),({\mathrm{\α}}_{12};p_1;T_{12};v_{12}),...,({\mathrm{\α}}_{1n};p_1;T_{1n};v_{1n})\\ ({\mathrm{\α}}_{21};p_2;T_{21};v_{21}),({\mathrm{\α}}_{22};p_2;T_{22};v_{22}),...,({\mathrm{\α}}_{2n};p_2;T_{2n};v_{2n})\\ ......\\ ......\\ ({\mathrm{\α}}_{x1};p_x;T_x;v_{x1}),({\mathrm{\α}}_{x2};p_x;T_{x2};v_{x2}),...,({\mathrm{\α}}_{\mathrm{xn}};p_x;T_{\mathrm{xn}};v_{\mathrm{xn}})\end{array}\right]$

abnormal sampleConversion into the above feature matrixThen, the step of calculating the distance of the sample mode of the fault knowledge base can be carried out.

And 4, step 4: sequentially calculating pattern distances for sample features in a fault pattern knowledge base by using the abnormal feature information;

firstly, extracting the stored information faultMessageSets in the fault knowledge base, wherein the specific mode is as follows:

$\mathrm{faultMessageSets}=\left\{\begin{array}{c}{\mathrm{fm}}^{11},{\mathrm{fm}}^{12},...,{\mathrm{fm}}^{1,c1},{\mathrm{pt}}_1,{\mathrm{Th}}_1\\ {\mathrm{fm}}^{21},{\mathrm{fm}}^{22},...,{\mathrm{fm}}^{2,c2},{\mathrm{pt}}_2,{\mathrm{Th}}_2\\ ......\\ ......\\ {\mathrm{fm}}^{x1},{\mathrm{fm}}^{x2},...,{\mathrm{fm}}^{x,c1},{\mathrm{pt}}_x,{\mathrm{Th}}_x\end{array}\right\}$

then, a feature matrix FT and a measuring point information vector wp of the abnormal sample and first-type fault information V₁＝{fm¹¹，fm¹²，…，fm^1，c1，pt₁，Th₁And fourthly, performing fault identification work. The specific operation is as follows:

a comparison measuring point information vector wp and a fault type fault measuring point information vector pt₁The repetition rate f is calculated by the following formula: $\mathrm{\η}=\frac{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{wp}}_i\∩{\mathrm{pt}}_{1i}}{\mathrm{length}\left({\mathrm{pt}}_1\right)}*100\%$

sequentially carrying out pattern distance numerical dist on the feature matrix FT of the two abnormal samples and each feature matrix under the fault type 1 according to the following method_1iAnd (3) calculating:

the abnormal sample may be a complete fault sample of a certain type or a partial fault sample in a development process of a certain type. For accurately calculating dist_1iAnd the step of accurately positioning the fault part by using a telescopic window sliding search technology, so that the mode recognition function of the abnormal sample is realized, and the development stage information of the abnormal sample is captured.

2.1 first, all possible values of the shrink window WT are determined. The following contents are set in the invention:

1, all fault samples in a fault knowledge base are samples of a complete process;

2, the expansion ratio of the expansion window is xi-20%;

the value range of WT may be defined as round (length (ft)) 80% to round (length (ft)) 120%, and WT ═ n is set₁，n₂，…，n_x]；

Wherein round is a rounding function; length is a function of calculating the number of sample segments, i.e., the number of rows of the feature matrix.

2.2 then from the telescopic window n₁Initially, at each sample feature matrix fm for fault type 1¹ⁱAnd the mode distance is calculated by sliding.

If fm¹ⁱIf the length of the window is m, the sliding frequency of the telescopic window is m, and the range of pattern recognition on the fault sample is as follows:

[1，n₁]～[m-n₁+1，m]。

abnormal samples from [1, n ] of faulty samples₁]The position starts to calculate the mode distance, each sliding step L is set to 1, and FT and fm are calculated¹ⁱ[1：n₁]The mode distance formula of (1):

dist₁₁＝κ*MD_DTW(FT，fm¹ⁱ[1：n₁]

wherein k is 2 if the repetition rate η is equal to 0, and k is 2 if the repetition rate η is not equal to 0MD_DTW(FT，fm¹ⁱ[1，n₁]) To find FT and fm¹ⁱ[1，n₁]DTW distance function of (d).

The rest is [2, n ]₂]～[m-n₁+1，m]The ranges of (a) and (b) are all calculated according to the steps 2.1-2.2, and finally a mode distance vector is obtainedAt window size n₁In the case of (2), the local best mode distance is dist₁₁＝min(distVector₁) The local best matching position is:

L₁₁＝[n_min-n₁+1，n_min]。

2.3 last other expansion windows x-1 value cases n₂～n_xCalculating the local optimal mode distance dist under each window value_1iAnd local best matching position L_1i＝[n_x-n_i+1，n_x]。

2.4 after all window values are matched, the following two important results are generated:

1 local mode distance vector

2 local optimum position vector

Selecting mode distance minimum from local mode distance vectorAnd corresponding best matching positionFeature matrices FT and fm as abnormal samples and failure samples¹ⁱThe pattern recognition end result of (1).

After the three abnormal sample feature matrixes FT are matched with all feature matrixes of other fault samples one by one, the pattern distance vector is generatedAnd a position vector

Remaining x-1 type faults V_iSolving the pattern distance vector according to the mode and the abnormal samples FT and wp, and finally obtaining a pattern distance set $\mathrm{distSets}=\left\{\begin{array}{c}{\stackrel{\‾}{\mathrm{distVector}}}_1\\ {\stackrel{\‾}{\mathrm{distVector}}}_2\\ ...\\ ...\\ {\stackrel{\‾}{\mathrm{distVector}}}_x\end{array}\right\}$ And a set of position vectors $\mathrm{LSets}=\left\{\begin{array}{c}\stackrel{\‾}{L_1}\\ {\stackrel{\‾}{L}}_2\\ ...\\ ...\\ {\stackrel{\‾}{L}}_x\end{array}\right\}.$

And 5: converting the mode distance into mode similarity by using a falling-ridge distribution method;

the pattern distance obtained in the above step is the distance obtained by applying the DTW algorithm, which is essentially the euclidean distance, and the euclidean distance is related to the time length of the samples, and the greater the time length is, the euclidean distance is correspondingly increased, so that the similarity between the two samples cannot be visually judged, and therefore, the final fault diagnosis result can be obtained only after the pattern distance obtained in the above step is converted into the pattern similarity by adopting the ridge-descending distribution method in the present step.

The mode distance set composition component obtained in step 4 is abc mode distancePattern distance is converted to similarity according to the principle of the falling-ridge stepwise method according to the following formula

$\mathrm{\ρ}\left({\stackrel{\‾}{\mathrm{dist}}}_{\mathrm{ij}}\right)=\left\{\begin{array}{c}1,{\stackrel{\‾}{\mathrm{dist}}}_{\mathrm{ij}}\≤0\\ \frac{1}{2}-\frac{1}{2}\sin (\frac{\mathrm{\π}}{{\mathrm{Th}}_i}*({\stackrel{\‾}{\mathrm{dist}}}_{\mathrm{ij}}-\frac{{\mathrm{Th}}_i}{2})),0\≤{\stackrel{\‾}{\mathrm{dist}}}_{\mathrm{ij}}\≤{\mathrm{Th}}_i\\ 0,{\stackrel{\‾}{\mathrm{dist}}}_{\mathrm{ij}}\≥{\mathrm{Th}}_i\end{array}\right.$

Wherein,represents the pattern distance calculated by pattern recognition of the abnormal sample and the jth fault sample under the fault type i, and Th_iRepresenting the mode distance threshold for fault type i.

Thus, each distance vector is calculated according to the dimension of the fault typeConverted into similarity vectors

Finally, the similarity set of the obtained abnormal sample FT and all samples of the fault knowledge base is as follows:

$\mathrm{\ρSets}=\left\{\begin{array}{c}[\mathrm{\ρ}\left({\stackrel{\‾}{\mathrm{dist}}}_{11}\right),\mathrm{\ρ}\left({\stackrel{\‾}{\mathrm{dist}}}_{12}\right),...,\mathrm{\ρ}{\stackrel{\‾}{(\mathrm{dist}}}_{1,c1})]\\ [\mathrm{\ρ}\left({\stackrel{\‾}{\mathrm{dist}}}_{21}\right),\mathrm{\ρ}\left({\stackrel{\‾}{\mathrm{dist}}}_{22}\right),...,\mathrm{\ρ}{\stackrel{\‾}{(\mathrm{dist}}}_{2,c2})]\\ ......\\ ......\\ [\mathrm{\ρ}\left({\stackrel{\‾}{\mathrm{dist}}}_{x1}\right),\mathrm{\ρ}\left({\stackrel{\‾}{\mathrm{dist}}}_{x2}\right),...,\mathrm{\ρ}{\stackrel{\‾}{(\mathrm{dist}}}_{x,\mathrm{cx}})]\end{array}\right\}=\left\{\begin{array}{c}\mathrm{\ρ}{\mathrm{Vector}}_1\\ \mathrm{\ρ}{\mathrm{Vector}}_2\\ ...\\ ...\\ \mathrm{\ρ}{\mathrm{Vector}}_x\end{array}\right\}$

step 6: and outputting a final fault diagnosis result.

According to the similarity set result rho Sets obtained in the step 5, the most effective and reliable fault diagnosis result is output according to the agreed rule:

rule 1: if a plurality of fault samples with the similarity exceeding 90% with the abnormal sample exist, outputting the serial numbers of all the samples, simultaneously outputting the type vector of the fault sample, the confidence level of the fault type, the best matching position Location of the sample with the maximum similarity in the fault development stage, and outputting the recommended maintenance action if the confidence level exceeds 50%;

rule 2: if no fault sample with the similarity exceeding 90% with the abnormal sample exists and the maximum similarity value is greater than or equal to 60%, outputting the serial number of the maximum similarity sample, and simultaneously outputting the type of the fault to which the fault sample belongs, the confidence level of the fault type to which the fault sample belongs, and the best matching position Location of the maximum similarity sample in the fault development stage; if the confidence level exceeds 50%, recommending action of the adopted maintenance measure;

rule 3: if no fault sample with the similarity exceeding 90% with the abnormal sample exists and the maximum similarity value is less than 60%, outputting the abnormal sample as an unusual working condition or an unknown fault type, and recommending to continuously pay attention to the development of the abnormality;

wherein the typeVector value satisfies the above ruleThe value of i in (1); the confidence level value of a certain fault type is equal to the number of the fault types meeting the rule divided by the total number of the fault samples of the fault type; the maintenance measure repairAction can be obtained by inquiring the operation rule of the equipment according to the fault type.

The steam pump preposition pump of a #2 unit of a certain thermal power plant in the north is taken as a state monitoring object, the steam pump preposition pump is an important component of a steam feed water pump, the function is important, and the performance safety of boiler equipment is maintained. The equipment belongs to exposed facilities, has more observation parts and is easy to cause various faults, and the characteristic is suitable for the equipment fault diagnosis method designed by the invention. The implementation process of the present invention is further illustrated by the detailed description of the embodiment.

The implementation steps of the embodiment of the invention for the fault diagnosis of the pre-pump equipment of the steam pump of a certain power plant are as follows:

fault diagnosis knowledge base construction process of steam pump pre-pump equipment

Step 1: acquiring failure sample information of a front pump of the steam pump from a power plant PI database;

selecting 21 key parameters related to the operation of a front pump of the air pump, wherein the key parameters comprise actual power (MW), radial bearing temperature (DEG C) of a motor transmission end, thrust bearing temperature (DEG C), motor current (A) and the like, and therefore each piece of observation data of the equipment is 21-dimensional row vector:

u(t_j)＝[u_j1，u_j2，u_j3，…，u_j21]

searching fault records of the equipment from the 10 months in 2013 to the 10 months in 2015, determining that the fault type P is greater than or equal to 2 and the fault occurrence frequency T is greater than or equal to 2 according to a screening rule, and extracting the following information from the fault records of the equipment according to the screening rule: in the period, 3 faults (radial bush temperature sudden rise of a motor transmission end, thrust bush temperature sudden drop and motor current sudden change) occur totally, the number of the faults is 86 (radial bush temperature sudden rise of the motor transmission end, thrust bush temperature sudden drop of the motor transmission end, and motor current sudden change of the motor transmission end is 25, 28 and 33), fault related measuring points are different (the radial bush temperature sudden rise of the motor transmission end is from measuring point 6 to measuring point 14, the thrust bush temperature sudden drop is from measuring point 15 to measuring point 17, the motor current sudden change is from measuring point 1 to measuring point 6 and from measuring point 18 to measuring point 21), and the fault duration range is 2 to 14 hours. (ii) a And finally, reading all fault sample data from the power plant real-time database PI according to the information.

The finally obtained fault samples comprise three parts of fault type number information (the radial tile temperature sudden rise number of a motor transmission end is 1, the thrust tile temperature sudden fall number is 2, and the motor current sudden change number is 3), fault sample data and fault type related measuring point information, wherein the sequence of each type of fault samples is sequenced according to the sequence of fault occurrence time:

$\mathrm{FaultSets}=\left\{\begin{array}{c}F_1^1,F_2^1,...,F_{25}^1;\\ F_1^2,F_2^2,...,F_{28}^2;\\ F_1^3,F_2^3,...,F_{33}^3;\end{array}\right\};$

wherein,ⁱF_jthe specific form of (1) is a matrix form, the row m represents the sample length, and the column represents the measuring point. As follows:

$F_j^i=\left[\begin{array}{c}{u}_{11},u_{12},u_{13},...,u_{121}\\ u_{21},u_{22},u_{23},...,u_{221}\\ ..........................\\ ..........................\\ ..........................\\ u_{m1},u_{m2},u_{m3},...,u_{m21}\end{array}\right]$

step 2: performing piecewise linear fitting on each fault sample in the FaultSets;

from the sample¹F₁Starting until the sample³F₃₃And each fault sample adopts a piecewise linear fitting technology, so that the self-adaptive state segmentation is realized according to the overall development trend of the fault sample, and data preprocessing preparation is carried out for subsequent fault feature extraction.

First, a mean filtering operation is performed. Sequentially extractingⁱF_jPer station column vector u_-iFiltering and eliminating noise pollution doped in sample data according to the operation steps of the mean filtering principle in the specific embodiment to obtain filtered dataⁱF_j ^*。

After the filtering operation, the maximum fitting ERROR threshold MAX _ ERROR is empirically set to 0.2. The bottom-up state partitioning is implemented in a recursive merge manner. The following three results were obtained: 1 partitioning a set of data segments2 fitting error vector $\stackrel{\‾}{R}=[R_1,R_2,...,R_{\frac{m}{4}-1}];$ 3-sample global segmentation position vector $\stackrel{\‾}{T}=[T_1,T_2,...,T_{\frac{m}{4}-1}].$

Finally, each sample of the steam pump is arranged in front of the pumpⁱF_jAll are processed into

And step 3: extracting the characteristics of each section of data of each segmented fault sample;

because the sample data time length is in the hour level, in order to facilitate the rapid and effective operation of the fault diagnosis, the mathematical characteristics such as the slope k, the time length l, the mean value m, the variance v and the like are extracted from each section of data of the fault sample to carry out data compression and the transformation of the feature space.

After feature extractionⁱF_jEach piece of data ofAll converted into four-dimensional elements (k, l, m, v). The specific feature extraction formula can be seen in step 3 of the modeling stage of the specific embodiment. Thus each failure sampleThe process that the sample is converted from the time domain space to the feature space is realized:

$\{f_1^{*},f_2^{*},...,f_{\frac{m}{4}}^{*}\}\⇒\begin{array}{}\end{array}\left[\begin{array}{c}(k_{11};l_1;m_{11};v_{11}),(k_{12};l_1;m_{12};v_{12}),...,(k_{121};l_1;m_{121};v_{121})\\ (k_{21};l_2;m_{21};v_{21}),(k_{22};l_2;m_{22};v_{22}),...,(k_{221};l_2;m_{221};v_{221})\\ ......\\ ......\\ (k_{\frac{m}{4}1};l_{\frac{m}{4}};m_{\frac{m}{4}};v_{\frac{m}{4}1}),(k_{\frac{m}{4}2};l_{\frac{m}{4}};m_{\frac{m}{4}2};v_{\frac{m}{4}2}),...,(k_{\frac{m}{4}21};l_{\frac{m}{4}};m_{\frac{m}{4}21};v_{\frac{m}{4}21})\end{array}\right]$

and 4, step 4: carrying out fault characteristic conversion and eliminating characteristic dimension;

in order to prevent misjudgment of different fault samples due to dimension reasons when the mode distance is calculated, three characteristics of the slope, the duration and the mean value need to be converted into a slope angle, a time proportion and a mean value proportion. The specific operation method for eliminating the dimension can be seen in a characteristic transformation formula in the step 4 of the modeling stage. Through transformation, the characteristics are transformed into the following dimensionless forms:

$\stackrel{\‾}{\mathrm{FT}_j^i}=\left[\begin{array}{c}({\mathrm{\α}}_{11};p_1;T_{11};v_{11}),({\mathrm{\α}}_{12};p_1;T_{12};v_{12}),...,({\mathrm{\α}}_{121};p_1;T_{121};v_{121})\\ ({\mathrm{\α}}_{21};p_2;T_{21};v_{21}),({\mathrm{\α}}_{22};p_2;T_{22};v_{22}),...,({\mathrm{\α}}_{221};p_2;T_{221};v_{221})\\ ......\\ ......\\ ({\mathrm{\α}}_{\frac{m}{4}1};p_{\frac{m}{4}};T_{\frac{m}{4}1};v_{\frac{m}{4}1}),({\mathrm{\α}}_{\frac{m}{4}2};p_{\frac{m}{4}};T_{\frac{m}{4}2};v_{\frac{m}{4}2}),...,({\mathrm{\α}}_{\frac{m}{4}21};p_{\frac{m}{4}};T_{\frac{m}{4}21};v_{\frac{m}{4}21})\end{array}\right];$

so far, the conversion from the time domain space to the feature space is realized by all samples, and the conversion of the fault sample set FaultSets into the fault sample feature set is as follows:

$\mathrm{FaultFeatureSets}=\left\{\begin{array}{c}\stackrel{\‾}{\mathrm{FT}_1^1},\stackrel{\‾}{\mathrm{FT}_2^1},...,\stackrel{\‾}{\mathrm{FT}_{25}^1};\\ \stackrel{\‾}{\mathrm{FT}_1^2},\stackrel{\‾}{\mathrm{FT}_2^2},...,\stackrel{\‾}{\mathrm{FT}_{28}^2};\\ \stackrel{\‾}{\mathrm{FT}_1^3},\stackrel{\‾}{\mathrm{FT}_2^3},...,\stackrel{\‾}{\mathrm{FT}_{33}^3};\end{array}\right\}$

and 5: calculating a mode distance threshold value under each fault type;

through the 4 steps, all fault sample data are converted into a characteristic matrix form set form. In the step, a Dynamic Time Warping (DTW) method is used for calculating the mode distance of two different samples in the same type of fault samples, and the mode distance threshold of each type of fault is determined according to the calculation result.

For example, there are 25 sample feature matrices in the type 1 fault in the fault sample, and every two feature matrices calculate the DTW distance to obtain a vector DtwVector₁＝[Md₁，Md₂，…，Md₂₅]. Pattern distance threshold Th of fault type 1₁＝min(DtwVector₁). The pattern distance threshold Th for fault types 2, 3 is obtained in the same way₂、Th₃。

Step 6: generating a steam pump gas quality pump fault mode knowledge base;

summarizing the information obtained in the steps, the steam pump pre-pump fault knowledge base should include the following 3 parts of information:

1. and extracting a feature matrix set FaultFeatureSets from all fault samples.

2. Relevant measuring point name set PT ═ { PT) of three types of faults₁，pt₂，pt₃}；

3. Pattern distance threshold values Thsets ═ Th corresponding to three types of faults₁，Th₂，…，Th_x]。

And integrating the above 3 parts of information to generate a complete steam pump preposed pump fault knowledge base in the following form:

$\mathrm{faultMessageSets}=\left\{\begin{array}{c}\stackrel{\‾}{\mathrm{FT}_1^1},\stackrel{\‾}{\mathrm{FT}_2^1},...,\stackrel{\‾}{\mathrm{FT}_{25}^1},{\mathrm{pt}}_1,{\mathrm{Th}}_1\\ \stackrel{\‾}{\mathrm{FT}_1^2},\stackrel{\‾}{\mathrm{FT}_2^2},...,\stackrel{\‾}{\mathrm{FT}_{28}^2},{\mathrm{pt}}_2,{\mathrm{Th}}_2\\ \stackrel{\‾}{\mathrm{FT}_1^3},\stackrel{\‾}{\mathrm{FT}_2^3},...,\stackrel{\‾}{\mathrm{FT}_{33}^3},{\mathrm{pt}}_3,{\mathrm{Th}}_3\end{array}\right\}$

fault on-line diagnosis process of steam pump pre-pump equipment

The plant-level monitoring system of the power plant finds that unknown abnormality occurs at a certain measuring point of the steam pump pre-pump equipment when 03 is 12 months and 15 days in 2014. In order to better help the professional personnel to carry out inspection and maintenance, the power plant personnel retrieve abnormal sample data from the real-time database according to the starting and ending time of the known abnormality and the related alarm measuring point information ptAnd the fault diagnosis method of the invention is used for carrying out advanced prejudgment.

Abnormal sample dataAfter data processing operations such as piecewise linear fitting, abnormal feature extraction and feature dimension elimination, calculating to obtain abnormal sample dataCorresponding feature matrixAnd (4) matrix. The above operations can be referred to as corresponding steps of the model operation phase of the embodiment.

Abnormal sample informationAnd pt is used for matching the sample information in the fault mode knowledge base one by one, and converting the mode distance into corresponding fault similarity.

The abnormal samples are matched with all fault samples in the knowledge base, and the DTW distance is calculated by using a sliding telescopic time window method to serve as the final mode distance. The set of available distances is in the form:

$\mathrm{PatternDistSets}=\left\{\begin{array}{c}\stackrel{\‾}{\mathrm{Md}_1^1},\stackrel{\‾}{\mathrm{Md}_2^1},...,\stackrel{\‾}{\mathrm{Md}_{25}^1};\\ \stackrel{\‾}{\mathrm{Md}_1^2},\stackrel{\‾}{\mathrm{Md}_2^2},...,\stackrel{\‾}{\mathrm{Md}_{28}^2};\\ \stackrel{\‾}{\mathrm{Md}_1^3},\stackrel{\‾}{\mathrm{Md}_2^3},...,\stackrel{\‾}{\mathrm{Md}_{33}^3};\end{array}\right\}$

and (4) mapping the distances between the [0, 1] numerical values by operating a falling-ridge distribution method, and converting the mode distances into similarity according to a unified standard. Embodiments A falling ridge distribution method of the model run phase is used herein. The available similarity sets are in the form:

$\mathrm{PatternSimSets}=\left\{\begin{array}{c}\stackrel{\‾}{\mathrm{Sd}_1^1},\stackrel{\‾}{\mathrm{Sd}_2^1},...,\stackrel{\‾}{\mathrm{Sd}_{25}^1};\\ \stackrel{\‾}{\mathrm{Sd}_1^2},\stackrel{\‾}{\mathrm{Sd}_2^2},...,\stackrel{\‾}{\mathrm{Sd}_{28}^2};\\ \stackrel{\‾}{\mathrm{Sd}_1^3},\stackrel{\‾}{\mathrm{Sd}_2^3},...,\stackrel{\‾}{\mathrm{Sd}_{33}^3};\end{array}\right\}$

outputting a judgment result of fault diagnosis according to an agreed fault diagnosis rule: the abnormity is pre-determined to be the radial tile temperature sudden-rise fault of the motor transmission end, and the determination reason is as follows: 15 fault samples with similarity of more than 90% with the abnormal sample in the front-mounted pump of the steam pump are (Etc.), the confidence coefficient of the diagnosis fault type 1 is 60% and exceeds the confidence coefficient threshold value, and the fault sample with the maximum similarity is that the number of the segmentation segments is 7The optimal matching positions are 1-4 segments, and the actual fault diagnosis effect diagram is shown in figure 4. The integral trends of the two samples are matched and consistent from the beginning to the middle position, so that the fault development stage is determined as the middle stage of fault development, and the query operation regulation suggests to take maintenance measures as a method for forcibly cooling the transmission end of the cold water impact motor

The fact shows that the diagnosis result after the professional staff arrive in time is really the radial tile temperature sudden-rise fault of the motor transmission end, and the operation of the front-mounted pump equipment of the steam pump is recovered to be normal after the front-mounted pump of the steam pump is forced to be cooled to a reasonable temperature range by adopting a cold water cooling method.

Although exemplary embodiments of the present invention have been described for illustrative purposes, those skilled in the art will appreciate that various modifications, additions, substitutions and the like can be made in form and detail without departing from the scope and spirit of the invention as disclosed in the accompanying claims, all of which are intended to fall within the scope of the claims, and that various steps in the various sections and methods of the claimed product can be combined together in any combination. Therefore, the description of the embodiments disclosed in the present invention is not intended to limit the scope of the present invention, but to describe the present invention. Accordingly, the scope of the present invention is not limited by the above embodiments, but is defined by the claims or their equivalents.

Claims (10)

1. The equipment fault diagnosis method is characterized by sequentially comprising the following steps of:

step 1: fault diagnosis training: acquiring fault sample information from a database, respectively obtaining a feature matrix and a mode distance threshold value of a fault sample with dimensions eliminated after calculation processing, associating the feature matrix and the mode distance threshold value with each other, and storing to generate a fault mode knowledge base;

step 2: fault diagnosis operation step: acquiring abnormal sample information from a real-time database, calculating to obtain abnormal characteristic information, sequentially calculating mode distances of sample characteristics in a fault mode knowledge base, converting the mode distances into mode similarity, and outputting a final fault diagnosis result.

2. The method of claim 1, wherein: the specific steps of the step 1 are as follows:

(1.1) acquiring fault sample information from a database;

(1.2) sequentially carrying out piecewise linear fitting on each fault sample;

(1.3) extracting the characteristics of each section of data of the fault sample to obtain a characteristic matrix of the fault sample;

(1.4) carrying out fault characteristic conversion, eliminating characteristic dimension and obtaining a characteristic matrix of a fault sample with the dimension eliminated;

(1.5) calculating a mode distance threshold;

and (1.6) associating the fault feature matrix and the pattern distance threshold value with each other, and storing to generate a fault pattern knowledge base.

3. The method of claim 2, wherein: the step (1.1) comprises the following specific steps: selecting a piece of research equipment with the number P of fault types being more than or equal to 2 and the occurrence frequency T of each fault being more than or equal to 2 to meet the requirements, selecting a sufficient number of observation points N, wherein N is more than or equal to 10, carrying out fault record searching on historical running state data of the equipment for a long enough time, picking out fault related measuring point information, starting and stopping time of a fault process and useful information of fault maintenance measure records from fault records by utilizing a set screening rule, and reading fault sample data from a power plant real-time database PI according to the useful information, wherein:

the fault sample data with n measuring points and m time points, all the measuring point data at the moment j are regarded as a n-dimensional column vector and are expressed as follows:

u(t_j)＝[u_j1,u_j2,u_j3,...,u_jn]

the sample data file is stored in a matrix form of m × n, and the specific form is as follows:

$F=\left[\begin{array}{c}{u}_{11},u_{12},u_{13},...,u_{1n}\\ u_{21},u_{22},u_{23},...,u_{2n}\\ \mathrm{...........................}\\ \mathrm{...........................}\\ \mathrm{...........................}\\ u_{m1},u_{m2},u_{m3},...,u_{mn}\end{array}\right]$

the row represents m fault time, the column represents n equipment observation points, the m and n values of the row and the column between each fault sample are different, each fault sample is endowed with a fault type identifier ID, and the fault type identifier ID determination method is that if all the samples contain X types of faults, the numerical range of the fault type identifier ID is as follows: 1-X.

4. The method of claim 3, wherein: the step (1.2) comprises the following specific steps:

(1.2.1) mean filtering operation: filtering and eliminating noise pollution doped in sample data;

(1.2.2) carrying out fault sample segmentation initialization, namely carrying out segmentation initialization on the fault sample after filtering processing;

(1.2.3) combining every two initialized data segments to calculate a fitting error;

and (1.2.4) determining a segmentation cut point of the fault sample, and performing adaptive state segmentation on the fault sample.

5. The method of claim 4, wherein: the step (1.3) comprises the following specific steps:

dividing the fault sample intoThe x divided data segments are selected fromThe data segment starts to carry out feature extraction, and the specific operations are as follows:

due to the fact thatThe data segment is $\left[\begin{array}{c}{u}_{11},u_{12},...,u_{1n}\\ u_{21},u_{22},...,u_{2n}\\ ...................\\ ...................\\ u_{h1},u_{h2},...,u_{hn}\end{array}\right]$ In the form of matrix, the row is time point number and is listed as measuring point number;

and (3) extracting features according to the dimension (column) of the measuring points, wherein the features comprise: slope k, duration l, mean m and variance v;

wherein, $\stackrel{\‾}{V_i}=\left[\begin{array}{c}{u}_{1i}\\ u_{2i}\\ ...\\ ...\\ u_{hi}\end{array}\right];$

the vector is linearly fitted according to the principle of least square method, and the fitting result is a linear equation p (x) a_0i+a_1ix, so slope characteristic k ═ a_1iWherein a is_0i、a_1iIs a fitting constant;

duration characteristic l isVector length h, i.e., l ═ h;

mean characteristic m isAverage of all values of the vector, i.e.

The variance characteristic v isSum of fluctuation amplitudes of all values of the vector from the mean, i.e.N is the total number of vectors;

segmenting data segments $f_1^{*}=\left[\begin{array}{c}{u}_{11},u_{12},...,u_{1n}\\ u_{21},u_{22},...,u_{2n}\\ \mathrm{....................}\\ \mathrm{....................}\\ \mathrm{....................}\\ u_{h1},u_{h2},...,u_{hn}\end{array}\right]$ Is converted into a feature matrix $\left[\begin{array}{c}{k}_1,k_2,\mathrm{...},k_n\\ l,l,\mathrm{...},l\\ m_1,m_2,\mathrm{...},m_n\\ v_1,v_2,\mathrm{...},v_n\end{array}\right];$

The rest ofExtracting characteristics according to the operation method, converting the original time domain data matrix into a mathematical characteristic matrix, and finally dividing the data segment setConversion to:

$\left[\begin{array}{c}(k_{11};l_1;m_{11};v_{11}),(k_{12};l_1;m_{12};v_{12}),\mathrm{...},(k_{1n};l_1;m_{1n};v_{1n})\\ (k_{21};l_2;m_{21};v_{21}),(k_{22};l_2;m_{22};v_{22}),\mathrm{...},(k_{2n};l_2;m_{2n};v_{2n})\\ \mathrm{......}\\ \mathrm{......}\\ (k_{x1};l_x;m_{x1};v_{x1}),(k_{x2};l_x;m_{x2};v_{x2}),\mathrm{...},(k_{xn};l_x;m_{xn};v_{xn})\end{array}\right].$

6. the method of claim 5, wherein: the step (1.4) comprises the following specific steps: converting the slope k into an inclination angle alpha, converting the duration l into a time proportion, carrying out normalization processing on the mean value characteristic m, and carrying out characteristic matrix treatment on the fault sample

$\left[\begin{array}{c}(k_{11};l_1;m_{11};v_{11}),(k_{12};l_1;m_{12};v_{12}),\mathrm{...},(k_{1n};l_1;m_{1n};v_{1n})\\ (k_{21};l_2;m_{21};v_{21}),(k_{22};l_2;m_{22};v_{22}),\mathrm{...},(k_{2n};l_2;m_{2n};v_{2n})\\ \mathrm{......}\\ \mathrm{......}\\ (k_{x1};l_x;m_{x1};v_{x1}),(k_{x2};l_x;m_{x2};v_{x2}),\mathrm{...},(k_{xn};l_x;m_{xn};v_{xn})\end{array}\right]$

Converted into elimination dimensions $\left[\begin{array}{c}({\mathrm{\α}}_{11};p_1;T_{11};v_{11}),({\mathrm{\α}}_{12};p_1;T_{12};v_{12}),...,({\mathrm{\α}}_{1n};p_1;T_{1n};v_{1n})\\ ({\mathrm{\α}}_{21};p_2;T_{21};v_{21}),({\mathrm{\α}}_{22};p_2;T_{22};v_{22}),...,({\mathrm{\α}}_{2n};p_2;T_{2n};v_{2n})\\ ......\\ ......\\ ({\mathrm{\α}}_{x1};p_x;T_x;v_{x1}),({\mathrm{\α}}_{x2};p_x;T_{x2};v_{x2}),...,({\mathrm{\α}}_{xn};p_x;T_{xn};v_{xn})\end{array}\right].$

7. The method of claim 1, wherein: the specific steps of the step 2 are as follows:

(2.1) acquiring abnormal sample information from a real-time database;

(2.2) performing piecewise linear fitting on the current abnormal sample;

(2.3) carrying out abnormal feature extraction and transformation on the data subjected to segmentation fitting;

(2.4) sequentially calculating the pattern distance of the sample features in the fault pattern knowledge base by using the abnormal feature information;

(2.5) converting the mode distance into a mode similarity using a falling ridge pacing method;

and (2.6) outputting a final fault diagnosis result.

8. The method of claim 7, wherein: the step (2.1) comprises the following specific steps:

according to the equipment state early warning system of the power plant, discovering that certain unknown abnormal state occurs in the equipment, and carrying out the following related operations:

(2.1.1) determining device alerts from early warning systemsGenerating time t₁And a cut-off time t₂；

(2.1.2) determining from the early warning system the relevant observation point wp ═ x for the device alarm₁,x₂,...,x_n']；

(2.1.3) according to the generation time t₁And a cut-off time t₂And a database sampling frequency fs, and obtaining a time point number m-fs × (t)₂-t₁) The number of device measurement points n equals length ([ x ])₁,x₂,...,x_n']) Where length () is a calculated length function);

(2.1.4) acquiring abnormal sample data, namely the sample data with n 'of measuring points and m of time points, wherein all the measuring point data at the moment j are regarded as a column vector with n' dimension and are expressed as:

v(t_j)＝[v_j1,v_j2,v_j3,...,v_jn']

storing the sample data file in a matrix form of m multiplied by n', wherein the specific form is as follows:

$\stackrel{\‾}{F}=\left[\begin{array}{c}{v}_{11},v_{12},v_{13},...,v_{1n^{\′}}\\ v_{21},v_{22},v_{23},...,v_{2n^{\′}}\\ ........................\mathrm{..}\\ ........................\mathrm{..}\\ ........................\mathrm{..}\\ v_{m1},v_{m2},v_{m3},...,v_{{\mathrm{mn}}^{\′}}\end{array}\right],$

where the rows represent m failure times and the columns represent n' device observation points.

The step (2.2) comprises the following specific steps:

(2.2.1) mean filtering operation: filtering and eliminating noise pollution of each measuring point doped in abnormal sample data according to a mean filtering principle;

(2.2.2) carrying out segmentation initialization on the abnormal sample after filtering treatment;

(2.2.3) combining every two initialized data segments to calculate a fitting error;

and (2.2.4) determining an abnormal sample segmentation cut point based on a recursive combination mode.

9. The method of claim 8, wherein: the step (2.5) comprises the following specific steps:

converting the pattern distance into pattern similarity according to the principle of the falling-ridge stepwise method

$\mathrm{\ρ}\left({\stackrel{\‾}{dist}}_{ij}\right)=\left\{\begin{array}{c}1,{\stackrel{\‾}{dist}}_{ij}\≤0\\ \frac{1}{2}-\frac{1}{2}\sin (\frac{\mathrm{\π}}{{\mathrm{Th}}_i}*({\stackrel{\‾}{dist}}_{ij}-\frac{{\mathrm{Th}}_i}{2}\left)\right),0\≤{\stackrel{\‾}{dist}}_{ij}\≤{\mathrm{Th}}_i\\ 0,{\stackrel{\‾}{dist}}_{ij}\≥{\mathrm{Th}}_i\end{array}\right.$

Wherein,represents the pattern distance calculated by pattern recognition of the abnormal sample and the jth fault sample under the fault type i, and Th_iRepresenting a mode distance threshold under fault type i;

according to the dimension of the fault type, each distance vector is divided into $\stackrel{\‾}{{\mathrm{distVector}}_i}=\[{\stackrel{\‾}{dist}}_{i1},{\stackrel{\‾}{dist}}_{i2},...,{\stackrel{\‾}{dist}}_{i,ch}\],i\∈\[1,x\],$ Converted into similarity vectors ${\mathrm{\ρVector}}_i=\[\mathrm{\ρ}\left({\stackrel{\‾}{dist}}_{i1}\right),\mathrm{\ρ}\left({\stackrel{\‾}{dist}}_{i2}\right),...,\mathrm{\ρ}{\stackrel{\‾}{(dist}}_{i,ch})\],i\∈\[1,x\];$

Finally, a similarity set of the abnormal sample FT and all samples of the fault knowledge base is obtained, and the similarity set is as follows:

$\mathrm{\ρ}Sets=\left\{\begin{array}{c}\[\mathrm{\ρ}\left({\stackrel{\‾}{dist}}_{11}\right),\mathrm{\ρ}\left({\stackrel{\‾}{dist}}_{12}\right),\mathrm{...},\mathrm{\ρ}\left({\stackrel{\‾}{dist}}_{1,c1}\right)\]\\ \[\mathrm{\ρ}\left({\stackrel{\‾}{dist}}_{21}\right),\mathrm{\ρ}\left({\stackrel{\‾}{dist}}_{22}\right),\mathrm{...},\mathrm{\ρ}{\stackrel{\‾}{(dist}}_{2,c2})\]\\ \mathrm{......}\\ \mathrm{......}\\ \[\mathrm{\ρ}\left({\stackrel{\‾}{dist}}_{x1}\right),\mathrm{\ρ}\left({\stackrel{\‾}{dist}}_{x2}\right),\mathrm{...},\mathrm{\ρ}\left({\stackrel{\‾}{dist}}_{x,cx}\right)\]\end{array}\right\}=\left\{\begin{array}{c}{\mathrm{\ρVector}}_1\\ {\mathrm{\ρVector}}_2\\ \mathrm{...}\\ \mathrm{...}\\ {\mathrm{\ρVector}}_x\end{array}\right\}.$

10. the method of claim 9, wherein: the step (2.6) comprises the following specific steps:

and outputting a fault diagnosis result according to the similarity set result rho Sets and an agreed rule, wherein the rule is satisfied:

rule 1: if a plurality of fault samples with the similarity exceeding 90% with the abnormal samples exist, outputting the serial numbers of all the samples, simultaneously outputting the fault type to which the fault sample belongs, the confidence coefficient of the fault type to which the fault sample belongs, and the best matching position of the sample with the maximum similarity in the fault development stage, and if the confidence coefficient exceeds 50%, outputting suggested maintenance measures;

rule 2: if no fault sample with the similarity exceeding 90% with the abnormal sample exists and the maximum similarity value is greater than or equal to 60%, outputting the serial number of the maximum similarity sample, simultaneously outputting the fault type to which the fault sample belongs, the confidence coefficient of the fault type to which the fault sample belongs and the best matching position of the maximum similarity sample at the fault development stage, and if the confidence coefficient exceeds 50%, outputting the recommended maintenance measures;

rule 3: if no fault sample with the similarity exceeding 90% with the abnormal sample exists and the maximum similarity value is less than 60%, outputting the abnormal sample as an unusual working condition or an unknown fault type, and continuing to pay attention to the development of the abnormality.