CN113469217A - Unmanned automobile navigation sensor abnormity detection method based on deep learning - Google Patents
Unmanned automobile navigation sensor abnormity detection method based on deep learning Download PDFInfo
- Publication number
- CN113469217A CN113469217A CN202110609739.2A CN202110609739A CN113469217A CN 113469217 A CN113469217 A CN 113469217A CN 202110609739 A CN202110609739 A CN 202110609739A CN 113469217 A CN113469217 A CN 113469217A
- Authority
- CN
- China
- Prior art keywords
- data
- training
- model
- layer
- deep learning
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 71
- 238000013135 deep learning Methods 0.000 title claims abstract description 63
- 238000012549 training Methods 0.000 claims abstract description 126
- 238000000034 method Methods 0.000 claims abstract description 55
- 230000009467 reduction Effects 0.000 claims abstract description 42
- 230000002159 abnormal effect Effects 0.000 claims abstract description 36
- 238000012360 testing method Methods 0.000 claims abstract description 26
- 238000012706 support-vector machine Methods 0.000 claims abstract description 22
- 230000005856 abnormality Effects 0.000 claims abstract description 21
- 238000000605 extraction Methods 0.000 claims abstract description 20
- 230000008569 process Effects 0.000 claims abstract description 17
- 230000004927 fusion Effects 0.000 claims abstract description 15
- 238000012795 verification Methods 0.000 claims description 51
- 238000003745 diagnosis Methods 0.000 claims description 25
- 238000012545 processing Methods 0.000 claims description 23
- 238000011176 pooling Methods 0.000 claims description 21
- 238000005070 sampling Methods 0.000 claims description 16
- 239000011159 matrix material Substances 0.000 claims description 14
- 238000013527 convolutional neural network Methods 0.000 claims description 13
- 238000005259 measurement Methods 0.000 claims description 9
- 238000010200 validation analysis Methods 0.000 claims description 7
- 238000013528 artificial neural network Methods 0.000 claims description 5
- 238000002790 cross-validation Methods 0.000 claims description 4
- 230000004913 activation Effects 0.000 claims description 3
- 238000013100 final test Methods 0.000 claims description 3
- 210000002569 neuron Anatomy 0.000 claims description 3
- 238000007781 pre-processing Methods 0.000 abstract description 2
- 230000009466 transformation Effects 0.000 abstract description 2
- 238000005065 mining Methods 0.000 abstract 1
- 238000010606 normalization Methods 0.000 description 5
- 238000010586 diagram Methods 0.000 description 4
- 238000011160 research Methods 0.000 description 4
- 230000000694 effects Effects 0.000 description 3
- NVPDSZPWJFLMIC-PAMZHZACSA-N 2-amino-9-[(4s,5r)-4-hydroxy-5-(hydroxymethyl)oxolan-2-yl]-8-(pyren-2-ylamino)-3h-purin-6-one Chemical compound C1=2NC(N)=NC(=O)C=2N=C(NC=2C=C3C=CC4=CC=CC5=CC=C(C3=C54)C=2)N1C1C[C@H](O)[C@@H](CO)O1 NVPDSZPWJFLMIC-PAMZHZACSA-N 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 2
- 230000006835 compression Effects 0.000 description 2
- 238000007906 compression Methods 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 238000013178 mathematical model Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 206010039203 Road traffic accident Diseases 0.000 description 1
- 241000700605 Viruses Species 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 230000004888 barrier function Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 125000004122 cyclic group Chemical group 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000012774 diagnostic algorithm Methods 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 238000013024 troubleshooting Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Testing And Monitoring For Control Systems (AREA)
Abstract
The invention discloses a method for detecting the abnormity of a navigation sensor of an unmanned automobile based on deep learning, the method comprises collecting original multi-channel data measured by the navigation sensor of the unmanned vehicle in normal state and N abnormal states, directly inputting the original multi-channel data into the deep learning abnormal detection model, the input data layer automatically completes data preprocessing such as multi-channel data fusion, sample generation and the like, the feature extraction layer automatically completes feature mining and feature extraction, a dimensionality reduction and parameter reduction layer automatically completes a series of operations such as dimensionality transformation and dimensionality reduction and parameter reduction, a Softmax classifier is used for completing error back propagation iterative computation in a training phase of a model, a support vector machine is used in a testing phase of the model, and the whole detection process does not need any characteristic operation depending on prior knowledge of engineers, so that people can detect network attack or physical damage and other abnormalities of the unmanned automobile navigation sensor more conveniently and quickly.
Description
Technical Field
The invention relates to an intelligent driving technology, in particular to an unmanned automobile navigation sensor abnormity detection method based on deep learning.
Background
In the last decade, an unmanned automobile as a new vehicle has become an important component of an intelligent transportation system, and the unmanned automobile can automatically complete a series of manual operations such as autonomous driving, autonomous barrier, autonomous path planning and automatic parking, so that the operation time of a driver is effectively released, and traffic accidents such as fatigue driving, drunk driving, traffic jam and violation of regulations are effectively avoided and reduced; the unmanned vehicle is used as a new intelligent vehicle, the unmanned vehicle completely depends on accurate position and path data provided by a navigation sensor during running by reading and receiving road condition data from a roadbed unit, a traffic signal, a radar base station and other vehicles in real time by relying on a large number of sensors, however, the navigation sensor of the unmanned vehicle causes navigation position data abnormity or falsification due to malicious network attack or physical faults of the navigation sensor, and great threat and challenge are brought to safe running of the unmanned vehicle, so that research on abnormal detection of the navigation sensor of the unmanned vehicle is very necessary to guarantee normal running of the unmanned vehicle.
At present, researchers in the industry and scientific research institutes mainly focus on the realization of functions and performance improvement research of unmanned vehicles, and researches on network attack detection of unmanned vehicle navigation sensors are not known, some hacker attacks aiming at computer operating systems in the prior art often adopt computer virus intrusion or abnormal detection based on Kalman filter models, but the methods need to establish complex mathematical models, the detection effect depends on the model precision to a great extent, however, the establishment of high-precision mathematical models in practice is very difficult, the abnormal detection effect is very limited, and the application objects of the methods are not unmanned vehicles, so that the methods are not convenient for the real-time diagnosis and fault detection of the navigation sensor abnormity of the unmanned vehicles.
Disclosure of Invention
The invention aims to solve the technical problem of providing an unmanned automobile navigation sensor abnormity detection method based on deep learning aiming at the defects in the prior art.
The technical scheme adopted by the invention for solving the technical problems is as follows: an unmanned vehicle navigation sensor abnormity detection method based on deep learning comprises the following steps:
1) collecting position measurement data of an unmanned automobile navigation sensor, wherein the position measurement data comprises normal state data and abnormal state data, and acquiring original abnormal state data;
2) performing multi-channel data fusion on the original abnormal state data;
3) performing data processing on the abnormal state data after the multi-channel data fusion to complete the abnormal state data expansion, and obtaining an expansion data set;
4) dividing the extended data set into a training data set, a verification data set and a test data set;
5) constructing a deep learning anomaly detection model;
the deep learning anomaly detection model includes: the system comprises an input data layer, a feature extraction layer, a dimension reduction and parameter reduction layer and a logic judgment output layer; the input data layer, the feature extraction layer, the dimension reduction and parameter reduction layer and the logic judgment output layer are sequentially connected;
the input data layer is used for reading the sample data of the training data set, the verification data set and the test data set and inputting the data into the feature extraction layer;
the feature extraction layer is a one-dimensional convolutional neural network model (1 DCNN); comprises a plurality of one-dimensional convolution layers, an activation layer and a one-dimensional pooling layer which are stacked in sequence;
the dimensionality reduction parameter reduction layer consists of 1X 1 one-dimensional convolution layer and 1 one-dimensional global mean pooling layer (1 DGAP);
the logic discrimination output layer consists of a Softmax classifier and a Support Vector Machine (SVM) classifier;
6) training the deep learning anomaly detection model to obtain weight parameters of a feature extraction layer, a dimension reduction and parameter reduction layer and a logic judgment output layer of the deep learning anomaly detection model;
training a deep learning abnormity detection model by adopting the training data set, performing cross validation on the deep learning abnormity detection model by adopting the validation data set, finishing model training until the accuracy of the validation data set reaches a set target value or training iteration times, and simultaneously storing weight parameter values of each layer of the model with the highest accuracy of the validation data set during model training;
7) and finishing final test by using the trained model to obtain a final abnormal detection result.
According to the scheme, the multi-channel data fusion in the step 2) is to convert multi-channel time sequence data from a plurality of navigation sensors into a long-time sequence two-dimensional characteristic map { T } of an original data setOriginal source；
According to the scheme, in the step 2), multi-channel time sequence data from a plurality of navigation sensors are converted into a long-time sequence original data set two-dimensional feature map; the method comprises the following specific steps:
the unmanned vehicle is provided with k navigation sensors, each sensor is provided with m data channels, each channel collects P data points, and a [ n, [ k, m ] is constructed under n state types],P]Is generated from the two-dimensional characteristic map { T } of the original dataset of the multidimensional tensor matrixOriginal source(ii) a Two-dimensional characteristic map { T } of the original data setOriginal sourceThe data ordering of (a) is set as: the channel data of the 1 st navigation sensor, the channel data of the 2 nd navigation sensor, …, and the channel data of the k-th navigation sensor are sequentially arranged.
According to the scheme, the data set is expanded in the step 3) to be a two-dimensional characteristic map { T } of the original data set of the long-time sequence multidimensional tensor matrixOriginal sourceGenerating TnumTraining samples { T } of two-dimensional feature map of extended total data set of short time sequencesGeneral assembly。
According to the scheme, the data processing in the step 3) comprises data standardization processing, data truncation processing and overlapped sampling data expansion processing which are sequentially carried out.
According to the scheme, in the step 3),
data ofNormalizing, namely compressing the magnitude of P data points of each channel to a value between 0 and 1, wherein a specific magnitude compression method adopts a maximum and minimum normalization method, and the mathematical expression of the method is as follows: { X } - (X)i-xmin)/(xmax-xmin) Wherein x isiIs the ith original data point, xminIs the data point of the minimum of the P original data points, xmaxIs the maximum data point of the P original data points, and { X } is P data points after standardization;
data truncation processing is carried out on an original data set two-dimensional characteristic map { T } of a long-time sequence multidimensional tensor matrix with the length of POriginal sourceIntercepting a plurality of short-time sequence two-dimensional feature maps { t } with the length p;
the expansion processing of the overlapped sampling data adopts a fixed sliding window W for adopting a sample data expansion method based on the overlapped sampling of a sliding window methodnumTwo-dimensional characteristic map { T } of original data set of multi-dimensional tensor matrix in long time sequenceOriginal sourceAt a fixed step SnumSliding along the direction of a time axis, generating and storing data points in a sliding window as a training sample { T } of a short-time-sequence two-dimensional feature map every time a step is moved, and accordingly generating a two-dimensional feature map { T } of an original data set of a long-time-sequence multi-dimensional tensor matrixOriginal sourceGenerating TnumTraining samples { T } of two-dimensional feature map of extended total data set of short time sequencesGeneral assembly。
According to the scheme, the feature extraction layer can be further arranged to be a cyclic neural network, a deep self-coding neural network or a deep belief neural network.
According to the scheme, the deep learning anomaly detection model is set to be in two combination forms including a training stage model combination and a testing stage model combination;
in the training stage of the deep learning anomaly detection model, the combination form of the models in the training stage is as follows:
after an input data layer, a feature extraction layer and a dimensionality reduction and parameter reduction layer are connected with a Softmax classifier in a logic discrimination output layer, a 1DCNN-1DGAP-Softmax model which is composed of a one-dimensional convolution neural network layer, a 1 multiplied by 1 one-dimensional convolution layer, a one-dimensional global mean pooling layer and the Softmax classifier is formed, and the combination is used for realizing parameter training and weight learning of neurons in each layer of the model through an error back propagation algorithm;
in the testing stage of the deep learning anomaly detection model, the combination form of the testing stage model is as follows:
and an SVM classifier is connected after the dimensionality reduction and parameter reduction layer to form a 1DCNN-1DGAP-SVM model which is formed by a one-dimensional convolutional neural network layer, a 1 multiplied by 1 one-dimensional convolutional layer, a one-dimensional global mean pooling layer and the SVM classifier together, error back propagation training is not carried out by the model, input samples are directly extracted by the one-dimensional convolutional neural network layer, the 1 multiplied by 1 one-dimensional convolutional layer and the one-dimensional global mean pooling layer and then input to the SVM classifier, and the SVM classifier finishes further promotion and output of final diagnosis results.
According to the scheme, in the training stage of the deep learning anomaly detection model in the step 6), while the model is trained, samples of a verification data set of the input data layer are used for carrying out real-time verification on the diagnosis accuracy of the 1DCNN-1DGAP-Softmax model in the training process, verifying the accuracy of the model on the verification set and checking whether overfitting occurs or not;
according to the scheme, the judgment of whether the 1DCNN-1DGAP-Softmax model is over-fitted in the training process in the step 6) is set as follows: if the number of iteration rounds of model training is increased, when the accuracy on the verification set and the accuracy on the training set continuously increase, the model training is normal, the model training is continued until the accuracy on the training set and the accuracy on the verification set reach the set number of iteration rounds, the model finishes the training, and the weight parameter values of each layer of the model with the highest accuracy of the verification data set in the training process are stored; if the number of iteration rounds of model training is increased, when the accuracy on a training set continuously increases and the accuracy on a verification set does not increase, and the difference between the accuracy of the training set and the accuracy of the verification set reaches a preset value, the model is diagnosed to be over-fitted, model training is stopped, the execution of skipping is carried out, the data of the training set is input into the deep learning abnormal detection model again, the super-parameters of the deep learning abnormal detection model are revised again, if the over-fitting does not occur during the re-verification, the model parameters are reasonable, the model is trained continuously until the accuracy on the training set and the accuracy on the verification set reach the set number of iteration rounds, the model finishes training, the weight parameter values of each layer of the model with the highest accuracy of the verification data set in the training process are stored, if the over-fitting occurs during the re-verification, the training is stopped, the parameters of the deep learning abnormal detection model are revised again, and the verification is continued, and repeating the steps until the verification is successful.
The invention has the following beneficial effects: according to the method, a traditional one-dimensional convolutional neural network model structure is improved, a 1 multiplied by 1 one-dimensional convolutional layer and a designed one-dimensional global mean value pooling layer (1D-GAP) are combined to replace a 2-3 layer full-connection layer structure of the traditional 1DCNN, the training parameters of the 1DCNN are effectively reduced, the diagnosis speed of the model is improved, a support vector machine is used to replace a Softmax classifier in a test stage, the diagnosis accuracy is further improved, meanwhile, data fusion can be carried out on multi-channel data collected under a plurality of navigation sensors of an automatic unmanned vehicle, and more accurate abnormity diagnosis is carried out on network attacks or physical abnormity of the navigation sensors of the unmanned vehicle by effectively utilizing the data of the channels.
Drawings
The invention will be further described with reference to the accompanying drawings and examples, in which:
FIG. 1 is a flow chart of a method of an embodiment of the present invention;
FIG. 2 is a schematic diagram of multi-sensor multi-channel data fusion according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of sample data enhancement by sliding window overlap sampling method according to an embodiment of the present invention;
FIG. 4 is an algorithm diagram of a deep learning anomaly detection model according to an embodiment of the present invention;
FIG. 5 is a diagram illustrating a deep learning unmanned vehicle navigation sensor anomaly detection scheme according to an embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail with reference to the following embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
As shown in fig. 1, the present invention provides a method for detecting an abnormality of a navigation sensor of an unmanned vehicle based on deep learning, comprising:
s101, collecting normal state data and N-type abnormal state multi-channel position measurement data of the unmanned automobile navigation sensor;
the multi-channel position measurement data of the unmanned automobile navigation sensor comprises 1-type normal state data and N-type abnormal state data, so that N-type (N is N +1) original abnormal state data is obtained.
And S102, performing multi-channel data fusion on the original abnormal state data.
Referring to FIG. 2, multi-channel data fusion is configured to transform multi-channel time series data from multiple navigation sensors into a long time series of raw data set two-dimensional feature maps { T }Original source。
In the present embodiment, assuming that the unmanned vehicle has k navigation sensors, each sensor has m data channels, each channel collects P data points, and one [ N, [ k, m ] can be constructed under N (N ═ N +1) state types],P]Is generated from the two-dimensional characteristic map { T } of the original dataset of the multidimensional tensor matrixOriginal source。
In this embodiment, two-dimensional feature map { T } of the original data setOriginal sourceThe data ordering of (a) is set as: the 1 st channel of the 1 st navigation sensor, the 2 nd channel of the 1 st navigation sensor, …, the mth channel of the 1 st navigation sensor, the 1 st channel of the 2 nd navigation sensor, the 2 nd channel of the 2 nd navigation sensor, …, the mth channel of the 2 nd navigation sensor, …, the 1 st channel of the kth navigation sensor, the 2 nd channel of the kth navigation sensor, …, the mth channel of the kth navigation sensor.
And S103, carrying out data sample expansion through data processing.
The data processing comprises data standardization processing, data truncation processing and overlapped sampling data expansion processing.
In this embodiment, in the data normalization process, in order to compress the magnitude of P data points of each channel to a value between 0 and 1, a specific magnitude compression method adopts a maximum and minimum normalization method, and a mathematical expression of the maximum and minimum normalization method is as follows: { X } - (X)i-xmin)/(xmax-xmin) Wherein x isiIs the ith original data point, xminIs the data point of the minimum of the P original data points, xmaxIs the maximum of the P raw data points, { X } is the P data points after normalization.
In the present embodiment, the data truncation process is set to two-dimensional feature map { T } of the original data set of the long-time-series multidimensional tensor matrix of length POriginal sourceAnd (4) cutting the two-dimensional feature map into a plurality of short-time sequence two-dimensional feature maps { t } with the length p.
In this embodiment, the overlap sampling data expansion process, as shown in FIG. 3, is configured to adopt a sample data expansion method based on overlap sampling of the sliding window method, which is configured to adopt a fixed sliding window WnumTwo-dimensional characteristic map { T } of the original data set of the long-time sequence multi-dimensional tensor matrixOriginal sourceAt a fixed step SnumSliding along the direction of a time axis, generating and storing data points in a sliding window as a training sample { T } of a short-time-sequence two-dimensional feature map every time a step is moved, and accordingly generating a two-dimensional feature map { T } of an original data set of a long-time-sequence multi-dimensional tensor matrixOriginal sourceGenerating TnumTraining samples { T } of two-dimensional feature map of extended total data set of short time sequencesGeneral assembly。
In this embodiment, the method for calculating the number of new training samples generated by the sample data enhancement method of the sliding window overlap sampling method includes: t isnum＝(Inum-Wnum+1)/Snum，InumNumber of data points representing input sample, WnumRepresents the length of the sliding window, SnumRepresenting the step size of the sliding window, TnumRepresenting the number of new fault samples, the calculation result decimal point is set to round up. In FIG. 3, it is assumed that each lane contains 2000 data points, i.e., InumBy overlap sampling data truncation, 6 training samples (sliding window length W) can be obtained 2000numStep size S is 400num300, i.e., an overlap ratio of 25%), and each fault sample has a size of m,400]. As can be seen from fig. 3, the rightmost sample is incomplete, and the incomplete sample cannot constitute a valid sample and is discarded.
And S104, dividing the data set.
Specifically, training samples { T }of two-dimensional feature map of the generated extended total data setGeneral assemblyThe method comprises the following steps of dividing a training data set, a verification data set and a test data set into: setting as training samples { T }of two-dimensional feature map of the extended total data set after preprocessingGeneral assemblyRandomly selecting 30% of samples in all samples of each state category as a test data set, randomly selecting 80% of the rest 70% of samples as a training data set, and randomly selecting 20% of samples as a verification data set;
and S105, constructing a deep learning abnormity detection model.
The deep learning anomaly detection model is provided with an input data layer, a feature extraction layer, a dimension reduction and parameter reduction layer and a logic discrimination output layer.
And the input data layer is used for reading the sample data of the training data set, the verification data set and the test data set and inputting the data into the feature extraction layer.
The feature extraction layer is a one-dimensional convolutional neural network model (1DCNN), and as shown in fig. 4 and 5, includes a plurality of one-dimensional convolutional layers and one-dimensional pooling layers stacked in sequence; the dimensionality reduction parameter reduction layer is set to be composed of 1 multiplied by 1 one-dimensional convolution layer and 1 one-dimensional global mean pooling layer (1 DGAP); the logic discrimination output layer consists of a Softmax classifier and a Support Vector Machine (SVM) classifier.
In this embodiment, as shown in fig. 4 and 5, the deep learning anomaly detection model is configured to include two modules, namely a training phase model combination and a testing phase model combination, in the training phase of the deep learning anomaly detection model, a Softmax classifier is connected after the dimensionality reduction and parameter reduction layer, so as to form a 1DCNN-1DGAP-Softmax model which is composed of a one-dimensional convolutional neural network layer, a 1 × 1 one-dimensional convolutional layer, a one-dimensional global mean pooling layer, and the Softmax classifier, and the combination is used for implementing parameter training and weight learning of neurons in each layer of the model through an error back propagation algorithm; in the testing stage of the deep learning anomaly detection model, an SVM classifier is connected after the dimensionality reduction and parameter reduction layer to form a 1DCNN-1DGAP-SVM model which is formed by a one-dimensional convolutional neural network layer, a 1 × 1 one-dimensional convolutional layer, a one-dimensional global mean pooling layer and the SVM classifier, error back propagation training is not carried out by the model, input samples are directly input to the SVM after characteristics are extracted by the one-dimensional convolutional neural network layer, the 1 × 1 one-dimensional convolutional layer and the one-dimensional global mean pooling layer, and the SVM finishes further promotion and output of a final diagnosis result.
And S106, training and verifying the deep learning abnormity detection model.
Training the deep learning abnormity detection model to obtain weight parameters of a feature extraction layer, a dimension reduction and parameter reduction layer and a logic discrimination output layer of the deep learning abnormity detection model, training the deep learning abnormity detection model by adopting a training data set, and performing cross validation on the deep learning abnormity detection model by adopting a validation data set. And during model training, using samples of a verification data set of the input data layer for carrying out real-time verification on the diagnosis accuracy of the 1DCNN-1DGAP-Softmax model in the training process, verifying the accuracy of the model on the verification set and checking whether overfitting occurs.
In the training process, the judgment of whether the 1DCNN-1DGAP-Softmax model is over-fitted is set as follows: if the number of iteration rounds of model training is increased, when the accuracy on the verification set and the accuracy on the training set continuously increase, the model training is normal, the model training is continued until the accuracy on the training set and the accuracy on the verification set reach the set number of iteration rounds, the model finishes the training, and the weight parameter values of each layer of the model with the highest accuracy of the verification data set in the training process are stored; if the number of iteration rounds of model training is increased, when the accuracy on a training set continuously increases and the accuracy on a verification set does not increase, and the difference between the accuracy of the training set and the accuracy of the verification set reaches a preset value, the model is diagnosed to be over-fitted, model training is stopped, the execution of skipping is carried out, the data of the training set is input into the deep learning abnormal detection model again, the super-parameters of the deep learning abnormal detection model are revised again, if the over-fitting does not occur during the re-verification, the model parameters are reasonable, the model is trained continuously until the accuracy on the training set and the accuracy on the verification set reach the set number of iteration rounds, the model finishes training, the weight parameter values of each layer of the model with the highest accuracy of the verification data set in the training process are stored, if the over-fitting occurs during the re-verification, the training is stopped, the parameters of the deep learning abnormal detection model are revised again, and the verification is continued, and repeating the steps until the verification is successful.
And S107, completing model training and storing the optimal model parameter weight.
In the training process, when the accuracy of the verification data set reaches a set target value or training iteration times, the model training is finished, and meanwhile, the weight parameter values of each layer of the model with the highest accuracy of the verification data set during the model training are stored.
And S108, finishing the final test to obtain a final diagnosis result.
And finally, inputting the sample of the test data set or the newly acquired real-time sample of the unmanned automobile navigation sensor into a deep learning anomaly detection model for anomaly detection, and outputting a final anomaly detection result.
In the present embodiment, the navigation sensor types include conventional navigation sensors such as Global Positioning System (GPS) sensor, laser detection and measurement sensor (LIDAR).
In this embodiment, in order to further illustrate the feasibility and effectiveness of the method provided by the present invention in diagnosing the abnormality of the navigation sensor of the unmanned vehicle, the present embodiment takes GPS and LIDAR navigation sensors as examples to verify the method provided by the present invention.
In this embodiment, position data of two navigation sensors, namely, a GPS and a LIDAR, are selected as experimental data, and both the position data and the LIDAR data include data of three channels, namely, x, y and an attitude angle θ, wherein 5 attacks are added to the GPS data, and no attack is added to the LIDAR as a reference signal. The sampling frequency of the two sensors is 10Hz, and the testing time is 1000 seconds, so that each channel of each type of abnormality can obtain a one-dimensional time sequence data segment containing 10000 points. In the experiment, different abnormal severity degrees and uncertainty of network attack are fully considered, so that various abnormal grades are set, and the detailed description is as follows:
1) normal state and fixed point attack: no grade;
2) x-direction offset attack: five grades of +/-1 m, +/-3 m, +/-5 m, +/-7 m and +/-10 m are respectively set;
3) y-direction offset attack: five grades of +/-5 m, +/-10 m, +/-20 m, +/-30 m and +/-50 m are respectively set;
4) and (3) attack of slow change drifting in the X direction: five slope grades of +/-0.2, +/-0.4 m, +/-0.6 m, +/-0.8 m and +/-1.0 are respectively set;
5) y-direction slowly-varying drifting attack: attack levels of five slopes of + -0.2, + -0.4 m, + -0.6 m, + -0.8 m and + -1.0 are set, respectively. The above sampling time for each anomaly level is 100 seconds.
First, the raw data of each abnormal state is subjected to data fusion in the input layer, in this embodiment, 6 channel data are obtained by 2 sensors, so as to construct a [10000,6] two-dimensional feature map, finally, a raw data set of [6,10000,6] can be obtained under 5 abnormal states and 1 normal state, and the first 6 represents 6 state types. Secondly, sample cutting is performed after data standardization processing is performed on all original data, sampling frequencies of a GPS and a LIDAR of the unmanned vehicle in the experiment are both 10Hz, therefore, the length of each training sample is set to be 10 in data truncation, the original data are equally divided, and 1000 training samples can be obtained for each state type, as shown in table 2. And finally, randomly selecting 70% of samples of each type of abnormality as a training set, using 30% of samples of each type of abnormality as a test set, randomly selecting 20% of samples of each type of abnormality in the training set for cross validation, using the training set for model training, using the test set for detecting the diagnosis accuracy of the algorithm, and finally establishing an abnormal data set as shown in table 1.
TABLE 1 unmanned vehicle navigation sensor anomaly data set
According to the deep learning anomaly detection model framework shown in fig. 4 and 5, a diagnostic model is established in this embodiment, as shown in table 2, in the model, the one-dimensional convolutional neural network layer includes 3 one-dimensional convolutional layers, 3 activation layers, and 1 one-dimensional pooling layer, the dimensionality reduction parameter reduction layer includes a 1 × 1 one-dimensional convolutional layer and 1 one-dimensional global mean pooling layer, and the logical discrimination output layer includes a Softmax classification layer and 1 SVM classification layer. All the zero filling modes of the convolution kernel and the pooling kernel are set as Padding being equal to 'Same', and 200 rounds of training are carried out on 64 samples in each batch by adopting an Adam adaptive learning rate optimizer and a mini-batch training method.
TABLE 2 Hyperparameter List of diagnostic models
In the conventional 1DCNN algorithm, a 2-3 layer full-connection structure is adopted at the end of the model, in order to further verify the improvement effect of the algorithm provided by the invention, the embodiment tests a 1DCNN model with 3 full-connection layers, the number of hidden nodes of the 3 full-connection layers is respectively 256-128-6, the embodiment trains the conventional 1DCNN algorithm by using the same data set, the final diagnosis results obtained by the two models are shown in table 3, and the number of the training parameters of the two diagnosis models is shown in table 4.
TABLE 3 comparison of diagnostic results for two algorithms
TABLE 4 comparison of parameter values for two diagnostic algorithms
As can be seen by comparing the results of table 3 and table 4:
1) in the aspect of diagnosis accuracy, the accuracy of the traditional 1DCNN algorithm is 98.94%, while the diagnosis accuracy of the algorithm provided by the invention is improved to 99.78%, and the identification degree of the algorithm on a normal state, a fixed point attack, a Y-direction offset and a Y-direction gradual attack reaches 100%, so that the method has higher diagnosis accuracy compared with the traditional 1DCNN method.
2) In terms of model parameters, the total parameters of a traditional 1DCNN algorithm comprising 3 layers of fully-connected layers are 108134, and the total parameters of a 3-layer fully-connected network are 74886; the model parameters improved by adopting the one-dimensional global mean pooling provided by the invention are only 33488, compared with the traditional 1DCNN algorithm, the parameter quantity is reduced by 74304, which accounts for about 70% of the total parameter quantity, and the model parameters are effectively reduced.
3) In the aspect of diagnosis time, the training time and the testing time of the traditional method are 204.81 seconds and 0.372 seconds respectively, the training time and the testing time of the improved algorithm are 163.57 seconds and 0.223 seconds respectively, and particularly in the aspect of testing time, the algorithm provided by the invention effectively reduces the diagnosis time by about 43%, provides more alarm troubleshooting time for unmanned vehicle network attack detection, and has important practical significance for real-time diagnosis of network attacks.
The invention has the following beneficial effects: according to the method, a traditional one-dimensional convolutional neural network model structure is improved, firstly, a 1 multiplied by 1 one-dimensional convolutional layer and a designed one-dimensional global mean pooling layer are combined to replace a 2-3-layer full-connection layer structure of a traditional 1DCNN, the training parameters of the 1DCNN are effectively reduced, the diagnosis speed of the model is improved, then a support vector machine is adopted to replace a Softmax classifier in a test stage, the diagnosis accuracy is further improved, meanwhile, the method can automatically perform data fusion on multi-channel data collected under a plurality of navigation sensors of the unmanned vehicle, and more accurate abnormity diagnosis is performed on network attacks or physical abnormity of the navigation sensors of the unmanned vehicle by effectively utilizing the data of the plurality of channels.
The invention provides a deep learning-based abnormality detection method for an unmanned automobile navigation sensor, which automatically performs multi-channel data fusion on multi-channel position measurement data of the unmanned automobile navigation sensor, directly inputs original multi-channel measurement data of the unmanned automobile navigation sensor into a constructed deep learning abnormality detection model, automatically completes a series of operations of multi-channel data fusion, data standardization processing, data truncation, sample data enhancement, sample generation, feature extraction, dimension transformation, dimension reduction, parameter reduction and the like in an input data layer, does not need any feature operation depending on the priori knowledge of engineers in the whole detection process, has excellent self-adaptability and flexibility universality in an end-to-end algorithm structure, has higher detection speed and higher diagnosis accuracy, is more suitable for network attack or abnormal real-time diagnosis and rapid detection of physical damage of the unmanned automobile navigation sensor, the abnormity detection of the pilotless automobile navigation sensor by people is more convenient and quicker.
It will be understood that modifications and variations can be made by persons skilled in the art in light of the above teachings and all such modifications and variations are intended to be included within the scope of the invention as defined in the appended claims.
Claims (10)
1. A method for detecting abnormality of a navigation sensor of an unmanned vehicle based on deep learning is characterized by comprising the following steps:
1) collecting position measurement data of an unmanned automobile navigation sensor, wherein the position measurement data comprises normal state data and abnormal state data, and acquiring original abnormal state data;
2) performing multi-channel data fusion on the original abnormal state data;
3) performing data processing on the abnormal state data after the multi-channel data fusion to complete the abnormal state data expansion, and obtaining an expansion data set;
4) dividing the extended data set into a training data set, a verification data set and a test data set;
5) constructing a deep learning anomaly detection model;
the deep learning anomaly detection model includes: the system comprises an input data layer, a feature extraction layer, a dimension reduction and parameter reduction layer and a logic judgment output layer; the input data layer, the feature extraction layer, the dimension reduction and parameter reduction layer and the logic judgment output layer are sequentially connected;
the input data layer is used for reading the sample data of the training data set, the verification data set and the test data set and inputting the data into the feature extraction layer;
the dimension reduction parameter reduction layer comprises a 1 x 1 one-dimensional convolution layer and a one-dimensional global mean value pooling layer;
the logic discrimination output layer consists of a Softmax classifier and a Support Vector Machine (SVM) classifier;
6) training the deep learning anomaly detection model to obtain weight parameters of a feature extraction layer, a dimension reduction and parameter reduction layer and a logic judgment output layer of the deep learning anomaly detection model;
training a deep learning abnormity detection model by adopting the training data set, performing cross validation on the deep learning abnormity detection model by adopting the validation data set, finishing model training until the accuracy of the validation data set reaches a set target value or training iteration times, and simultaneously storing weight parameter values of each layer of the model with the highest accuracy of the validation data set during model training;
7) and finishing final test by using the trained model to obtain a final abnormal detection result.
2. The method for detecting the abnormality of the navigation sensor of the unmanned vehicle based on the deep learning as claimed in claim 1, wherein the multi-channel data fusion in the step 2) is to convert multi-channel time series data from a plurality of navigation sensors into a long time series two-dimensional feature map { T } of a raw data setOriginal source。
3. The method for detecting the abnormality of the navigation sensor of the unmanned vehicle based on the deep learning of claim 1, wherein the step 2) converts the multi-channel time series data from a plurality of navigation sensors into a long time series two-dimensional feature map of the original data set; the method comprises the following specific steps:
the unmanned vehicle is provided with k navigation sensors, each sensor is provided with m data channels, each channel collects P data points, and a [ n, [ k, m ] is constructed under n state types],P]Is generated from the two-dimensional characteristic map { T } of the original dataset of the multidimensional tensor matrixOriginal source(ii) a Two-dimensional characteristic map { T } of the original data setOriginal sourceThe data ordering of (a) is set as: the channel data of the 1 st navigation sensor, the channel data of the 2 nd navigation sensor, …, and the channel data of the k-th navigation sensor are sequentially arranged.
4. The method for detecting the abnormality of the sensor for navigating the unmanned vehicle based on the deep learning as claimed in claim 1, wherein the step 3) is to expand the data set into a two-dimensional feature map { T } of an original data set of a long time series multidimensional tensor matrixOriginal sourceGenerating TnumTraining samples { T } of two-dimensional feature map of extended total data set of short time sequencesGeneral assembly。
5. The method for detecting the abnormality of the sensor for unmanned vehicle navigation based on deep learning of claim 1, wherein the data processing in step 3) comprises sequentially performing data standardization processing, data truncation processing and overlapped sampling data expansion processing.
6. The method for detecting the abnormality of the navigation sensor of the unmanned vehicle based on the deep learning as claimed in claim 1, wherein in the step 3), the data processing is specifically as follows:
data standardization processing is carried out, namely the magnitude of P data points of each channel is compressed to a value between 0 and 1;
data truncation processing is carried out on an original data set two-dimensional characteristic map { T } of a long-time sequence multidimensional tensor matrix with the length of POriginal sourceIntercepting a plurality of short-time sequence two-dimensional feature maps { t } with the length p;
the expansion processing of the overlapped sampling data adopts a fixed sampling data expansion method based on the overlapped sampling of a sliding window methodFixed sliding window WnumTwo-dimensional characteristic map { T } of original data set of multi-dimensional tensor matrix in long time sequenceOriginal sourceAt a fixed step SnumSliding along the direction of a time axis, generating and storing data points in a sliding window as a training sample { T } of a short-time-sequence two-dimensional feature map every time a step is moved, and accordingly generating a two-dimensional feature map { T } of an original data set of a long-time-sequence multi-dimensional tensor matrixOriginal sourceGenerating TnumTraining samples { T } of two-dimensional feature map of extended total data set of short time sequencesGeneral assembly。
7. The unmanned aerial vehicle navigation sensor abnormality detection method based on deep learning of claim 1, wherein the feature extraction layer is a one-dimensional convolutional neural network model; comprises a plurality of one-dimensional convolution layers, an activation layer and a one-dimensional pooling layer which are stacked in sequence.
8. The unmanned vehicle navigation sensor abnormality detection method based on deep learning of claim 1, wherein the deep learning abnormality detection model is set to include two combination forms of a training phase model combination and a testing phase model combination;
in the training stage of the deep learning anomaly detection model, the combination form of the models in the training stage is as follows:
after an input data layer, a feature extraction layer and a dimensionality reduction and parameter reduction layer are connected with a Softmax classifier in a logic discrimination output layer, a 1DCNN-1DGAP-Softmax model which is composed of a one-dimensional convolution neural network layer, a 1 multiplied by 1 one-dimensional convolution layer, a one-dimensional global mean pooling layer and the Softmax classifier is formed, and the combination is used for realizing parameter training and weight learning of neurons in each layer of the model through an error back propagation algorithm;
in the testing stage of the deep learning anomaly detection model, the combination form of the testing stage model is as follows:
and an SVM classifier is connected after the dimensionality reduction and parameter reduction layer to form a 1DCNN-1DGAP-SVM model which is formed by a one-dimensional convolutional neural network layer, a 1 multiplied by 1 one-dimensional convolutional layer, a one-dimensional global mean pooling layer and the SVM classifier together, error back propagation training is not carried out by the model, input samples are directly extracted by the one-dimensional convolutional neural network layer, the 1 multiplied by 1 one-dimensional convolutional layer and the one-dimensional global mean pooling layer and then input to the SVM classifier, and the SVM classifier finishes further promotion and output of final diagnosis results.
9. The deep learning based unmanned vehicle navigation sensor abnormality detection method of claim 8, wherein in the step 6), in the training phase of the deep learning abnormality detection model, during model training, samples of the verification dataset of the input data layer are used for performing real-time verification on the diagnosis accuracy of the 1DCNN-1DGAP-Softmax model during training, verifying the accuracy of the model on the verification set and checking whether overfitting occurs.
10. The method for detecting the abnormality of the sensor for the unmanned vehicle navigation based on the deep learning of claim 9, wherein the step 6) of determining whether the 1DCNN-1DGAP-Softmax model is over-fitted during the training process is performed by: if the number of iteration rounds of model training is increased, when the accuracy on the verification set and the accuracy on the training set continuously increase, the model training is normal, the model training is continued until the accuracy on the training set and the accuracy on the verification set reach the set number of iteration rounds, the model finishes the training, and the weight parameter values of each layer of the model with the highest accuracy of the verification data set in the training process are stored; if the number of iteration rounds of model training is increased, when the accuracy on a training set continuously increases and the accuracy on a verification set does not increase, and the difference between the accuracy of the training set and the accuracy of the verification set reaches a preset value, the model is diagnosed to be over-fitted, model training is stopped, the execution of skipping is carried out, the data of the training set is input into the deep learning abnormal detection model again, the super-parameters of the deep learning abnormal detection model are revised again, if the over-fitting does not occur during the re-verification, the model parameters are reasonable, the model is trained continuously until the accuracy on the training set and the accuracy on the verification set reach the set number of iteration rounds, the model finishes training, the weight parameter values of each layer of the model with the highest accuracy of the verification data set in the training process are stored, if the over-fitting occurs during the re-verification, the training is stopped, the parameters of the deep learning abnormal detection model are revised again, and the verification is continued, and repeating the steps until the verification is successful.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110609739.2A CN113469217A (en) | 2021-06-01 | 2021-06-01 | Unmanned automobile navigation sensor abnormity detection method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110609739.2A CN113469217A (en) | 2021-06-01 | 2021-06-01 | Unmanned automobile navigation sensor abnormity detection method based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN113469217A true CN113469217A (en) | 2021-10-01 |
Family
ID=77871940
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110609739.2A Pending CN113469217A (en) | 2021-06-01 | 2021-06-01 | Unmanned automobile navigation sensor abnormity detection method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113469217A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113859250A (en) * | 2021-10-14 | 2021-12-31 | 泰安北航科技园信息科技有限公司 | Intelligent automobile information security threat detection system based on driving behavior abnormity identification |
| CN113988202A (en) * | 2021-11-04 | 2022-01-28 | 季华实验室 | Mechanical arm abnormal vibration detection method based on deep learning |
| CN116828413A (en) * | 2023-08-24 | 2023-09-29 | 长河信息股份有限公司 | Heterogeneous multi-source data signal receiving and transmitting system and method |
| CN117173459A (en) * | 2023-08-29 | 2023-12-05 | 哈尔滨工业大学 | Interconnected indium column defect multi-classification method based on transfer learning |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104616033A (en) * | 2015-02-13 | 2015-05-13 | 重庆大学 | Fault diagnosis method for rolling bearing based on deep learning and SVM (Support Vector Machine) |
| CN106874957A (en) * | 2017-02-27 | 2017-06-20 | 苏州大学 | A kind of Fault Diagnosis of Roller Bearings |
| CN107066759A (en) * | 2017-05-12 | 2017-08-18 | 华北电力大学（保定） | A kind of Vibration Fault Diagnosis of Turbine Rotor method and device |
| CN107067026A (en) * | 2017-02-16 | 2017-08-18 | 国网江苏省电力公司电力科学研究院 | Electrical equipment fault detection method based on deep neural network |
| CN108520080A (en) * | 2018-05-11 | 2018-09-11 | 武汉理工大学 | Automatic system of marine diesel-generator failure predication and health status online evaluation system and method |
| CN109115501A (en) * | 2018-07-12 | 2019-01-01 | 哈尔滨工业大学（威海） | A kind of Civil Aviation Engine Gas path fault diagnosis method based on CNN and SVM |
| CN110443488A (en) * | 2019-07-31 | 2019-11-12 | 中国石油大学(华东) | The recognition methods of drilling well spill hazard, system and equipment based on convolutional neural networks |
| US20200285900A1 (en) * | 2019-03-06 | 2020-09-10 | Wuhan University | Power electronic circuit fault diagnosis method based on optimizing deep belief network |
| CN112067916A (en) * | 2019-09-20 | 2020-12-11 | 武汉理工大学 | Time series data intelligent fault diagnosis method based on deep learning |
-
2021
- 2021-06-01 CN CN202110609739.2A patent/CN113469217A/en active Pending
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104616033A (en) * | 2015-02-13 | 2015-05-13 | 重庆大学 | Fault diagnosis method for rolling bearing based on deep learning and SVM (Support Vector Machine) |
| CN107067026A (en) * | 2017-02-16 | 2017-08-18 | 国网江苏省电力公司电力科学研究院 | Electrical equipment fault detection method based on deep neural network |
| CN106874957A (en) * | 2017-02-27 | 2017-06-20 | 苏州大学 | A kind of Fault Diagnosis of Roller Bearings |
| CN107066759A (en) * | 2017-05-12 | 2017-08-18 | 华北电力大学（保定） | A kind of Vibration Fault Diagnosis of Turbine Rotor method and device |
| CN108520080A (en) * | 2018-05-11 | 2018-09-11 | 武汉理工大学 | Automatic system of marine diesel-generator failure predication and health status online evaluation system and method |
| CN109115501A (en) * | 2018-07-12 | 2019-01-01 | 哈尔滨工业大学（威海） | A kind of Civil Aviation Engine Gas path fault diagnosis method based on CNN and SVM |
| US20200285900A1 (en) * | 2019-03-06 | 2020-09-10 | Wuhan University | Power electronic circuit fault diagnosis method based on optimizing deep belief network |
| CN110443488A (en) * | 2019-07-31 | 2019-11-12 | 中国石油大学(华东) | The recognition methods of drilling well spill hazard, system and equipment based on convolutional neural networks |
| CN112067916A (en) * | 2019-09-20 | 2020-12-11 | 武汉理工大学 | Time series data intelligent fault diagnosis method based on deep learning |
Non-Patent Citations (6)
| Title |
|---|
| GONG W等: ""A novel deep learning method for intelligent fault diagnosis of rotating machinery based on improved CNN-SVM and multichannel data fusion"", 《SENSORS》 * |
| 周剑飞: ""基于深度学习的设备异常检测方法研究"", 《中国优秀硕士学位论文全文数据库工程科技Ⅱ辑》 * |
| 宫文峰等: ""基于改进CNN-GAP-SVM的船舶电力变换器快速故障诊断方法"", 《计算机集成制造系统》 * |
| 宫文峰等: ""基于改进卷积神经网络的滚动轴承智能故障诊断研究"", 《振动工程学报》 * |
| 宫文峰等: ""基于深度学习的电机轴承微小故障智能诊断方法"", 《仪器仪表学报》 * |
| 宫文峰等: ""基于深度学习的船舶机械微小故障快速诊断方法"", 《计算机集成制造系统》 * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113859250A (en) * | 2021-10-14 | 2021-12-31 | 泰安北航科技园信息科技有限公司 | Intelligent automobile information security threat detection system based on driving behavior abnormity identification |
| CN113988202A (en) * | 2021-11-04 | 2022-01-28 | 季华实验室 | Mechanical arm abnormal vibration detection method based on deep learning |
| CN116828413A (en) * | 2023-08-24 | 2023-09-29 | 长河信息股份有限公司 | Heterogeneous multi-source data signal receiving and transmitting system and method |
| CN116828413B (en) * | 2023-08-24 | 2023-12-08 | 长河信息股份有限公司 | Heterogeneous multi-source data signal receiving and transmitting system and method |
| CN117173459A (en) * | 2023-08-29 | 2023-12-05 | 哈尔滨工业大学 | Interconnected indium column defect multi-classification method based on transfer learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113469217A (en) | Unmanned automobile navigation sensor abnormity detection method based on deep learning | |
| JP6855091B2 (en) | A method for acquiring a sample image for label acceptance among auto-labeled images used for neural network learning, and a sample image acquisition device using the sample image. | |
| CN106407931B (en) | A kind of depth convolutional neural networks moving vehicle detection method | |
| CN109214349B (en) | Object detection method based on semantic segmentation enhancement | |
| CN102463990B (en) | For the system and method for tracking object | |
| CN107491736A (en) | A kind of pavement adhesion factor identifying method based on convolutional neural networks | |
| CN111639678B (en) | INS/GPS integrated navigation fault detection and diagnosis method based on integrated neural network | |
| CN109285348A (en) | A kind of vehicle behavior recognition methods and system based on two-way length memory network in short-term | |
| CN112669594B (en) | Method, device, equipment and storage medium for predicting traffic road conditions | |
| US20240244066A1 (en) | Iov intrusion detection method and device based on improved convolutional neural network | |
| CN113188807B (en) | Automatic abs result judging algorithm | |
| CN107238500A (en) | Vehicle handling stability tests RES(rapid evaluation system) method for building up | |
| CN111967308A (en) | Online road surface unevenness identification method and system | |
| CN114648633A (en) | Method for determining semantic segmentation of a vehicle environment | |
| CN113159115A (en) | Vehicle fine-grained identification method, system and device based on neural architecture search | |
| Islam et al. | Comparative analysis of steering angle prediction for automated object using deep neural network | |
| CN106339720A (en) | Automobile engine failure detection method | |
| EP3640857B1 (en) | Method, vehicle, system, and storage medium for indicating anomalous vehicle scenario using encoder network and discriminator network intermediate layer activation | |
| CN106056150A (en) | System and method for establishing part division remote damage assessment of different vehicle types based on artificial intelligence random forest method | |
| CN117237911A (en) | Image-based dynamic obstacle rapid detection method and system | |
| CN116738318A (en) | Convolutional neural network road surface unevenness identification method based on tracked vehicle vibration characteristics | |
| CN116152754A (en) | Expressway accident judging method based on DCNN | |
| CN114818845A (en) | Noise-stable high-resolution range profile feature selection method | |
| CN115563607A (en) | Network attack abnormity diagnosis system method for autonomous driving unmanned ship navigation sensor | |
| CN118736425B (en) | Remote sensing image building change detection method and system based on morphological constraint |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20211001 |
|
| RJ01 | Rejection of invention patent application after publication |