CN112149498A - An online intelligent identification system and method for abnormal noise of complex automobile parts - Google Patents

Abstract

The invention discloses an on-line intelligent identification system and method for abnormal noise of complex automobile components, and relates to the technical field of signal processing and pattern identification. According to the feedback of some abnormal noise problems in the process of using the car, the invention collects the abnormal noise signal in the experimental environment of the prototype car in a targeted manner, and manually marks it through classification, arranges it into the warehouse, and stores it in the cloud server; The mobile smart device downloads the corresponding software, collects the abnormal noise signal to be identified, and after pre-processing and analysis, uploads it to the cloud server through the wireless network or mobile data network, and then passes it to the trained deep learning model to realize the online intelligent identification of abnormal noise. The invention solves the problems of cumbersome bench test of abnormal sound of complex parts of automobiles, inconsistent artificial listening, inaccurate identification results, etc., and realizes simple and efficient identification of abnormal sounds of complex parts of automobiles.

Description

An online intelligent identification system and method for abnormal noise of complex automobile parts

technical field

The invention relates to the technical field of signal processing and pattern recognition, in particular to an online intelligent recognition system and method for abnormal noise of complex parts of automobiles.

Background technique

Sound quality is one of the key performances of traditional cars and new energy vehicles, and abnormal noise is a decisive factor affecting the sound quality of the entire vehicle. Abnormal noise control level reflects the comprehensive capabilities of vehicle manufacturers in vehicle design, processing, and assembly, and has gradually become one of the key factors affecting consumers' purchasing decisions and user satisfaction.

The problem of abnormal noise of self-owned brand and joint venture brand cars is prominent, and the number of complaints remains high. Abnormal noise problems seriously affect the user's driving experience. At the same time, some abnormal noise problems are the prelude to car failures. If they are not detected in time, they will bring security risks.

Abnormal noise has the characteristics of wide source of problems, random occurrence, complex formation mechanism, and inconspicuous characteristics, and it is very difficult to identify. For the abnormal noise of the interior, body, powertrain, brake, steering and other systems, the identification and diagnosis mainly rely on subjective feelings and bench testing methods, which are less efficient and require higher experience for testers.

With the rapid development of digital signal processing and machine learning, it is of great significance in reality to use the related methods of sound recognition technology to realize the automatic recognition of abnormal car noise. In the current sound recognition methods, the extraction of feature parameters and acoustic modeling are still mostly traditional shallow models. Commonly used shallow models include Hidden Markov Model (HMM), Gaussian Mixed Model (GMM), Support Vector Machine (SVM), etc., which have the ability to model and represent complex data. Insufficient, the model training amount is large and the calculation time is long, the recognition effect and accuracy are poor, and the online recognition of abnormal noise cannot be quickly and accurately realized.

SUMMARY OF THE INVENTION

In view of the above existing problems, the present invention provides an online intelligent identification system and method for complex automobile parts, which solves the problems of cumbersome bench test of abnormal noise of automobile complex parts, inconsistent manual listening, and inaccurate identification results, so as to achieve simple, Efficiently identify the abnormal noise of complex parts of the car.

In order to achieve the above purpose, the specific technical scheme of the present invention is as follows: an online intelligent identification method for abnormal noise of complex components of automobiles, comprising the following steps:

1) Construct the WT-FBank map of abnormal noise signal samples of complex automobile parts and upload them to the cloud server.

2) Construct a convolutional neural network model CNN based on the classical model LeNet-5 architecture.

3) Use the training model to build a module, combine with the classifier support vector machine SVM, train the WT-FBank map of the abnormal noise signal sample, and obtain the trained CNN model and the trained SVM recognition model.

4) Use the abnormal sound acquisition module to collect the activated abnormal sound signal of the vehicle to be detected in real time.

5) Using the preprocessing and analysis module to preprocess the abnormal sound signal to be detected, obtain the WT-FBank characteristics of the abnormal sound signal to be detected, and construct a WT-FBank map of the abnormal sound signal to be detected.

6) Using the intelligent recognition module, input the WT-FBank map of the detected abnormal noise signal into the trained CNN model to extract the characteristics of the abnormal noise signal to be detected, and convert it into a sparse low-dimensional feature vector, and input it into the trained SVM recognition model. Intelligent matching, output recognition category results and return to the mobile smart device, and use the automatic update module to improve the database.

7) Determine whether the abnormal sound signal detection is over, if not, go to step 4), continue to detect and identify new abnormal sound signals, otherwise end the abnormal sound signal detection.

Further, constructing the WT-FBank map of the abnormal noise signal sample of complex automobile parts as described in the above step 1) and uploading it to the cloud server includes the following steps:

1.1) Use the sound acquisition device to collect the abnormal noise signal samples of the complex parts of the automobile under the starting state.

1.2) Manually label the abnormal noise signal samples, and establish a sample library of abnormal noise signals for complex automobile parts.

1.3) Use the preprocessing and analysis module to preprocess the abnormal noise signal samples, that is, extract the WT-FBank features of the abnormal noise signal samples, construct the WT-FBank map of the abnormal noise signal samples, and upload them to the cloud server for storage.

Further, in the above-mentioned step 3), the trained CNN model and the trained SVM recognition model are obtained, including the following steps:

3.1) Input the WT-FBank map of the abnormal noise signal sample into the CNN model to obtain the features after global average pooling.

3.2) Input the feature after global average pooling into the Softmax function, and calculate the error between the output value and the expected value.

3.3) According to the error, the back propagation algorithm is used to reversely update the weight and bias parameter parameters of each layer layer by layer until the training error accuracy is reached, and the trained CNN model is obtained.

3.4) Input the WT-FBank map of abnormal noise signal samples of complex automobile parts into the trained CNN model for feature extraction, convert it into a low-dimensional feature vector, and input the corresponding category label.

3.5) Input the extracted low-dimensional feature vector and the corresponding label to the SVM classifier to complete the training, and obtain a trained SVM recognition model.

Further, in above-mentioned step 1.3) and described step 5), the preprocessing and analysis module preprocessing process includes the following steps:

4.1) Use wavelet threshold noise reduction and empirical mode decomposition noise reduction methods to perform noise reduction processing on the abnormal noise signal collected by the acquisition module, and filter out unimportant information and background noise.

4.2) Pre-emphasize the noise-reduced signal through a first-order high-pass filter to compensate for the excessive attenuation of the high-frequency component of the signal during the transmission process.

4.3) Add a Hamming window to the pre-emphasized signal to form a short-term signal x(i), and the frame length is N.

4.4) Perform wavelet transform on the short-term signal x(i), convert it into a frequency domain signal, and calculate its short-term energy spectrum W _i (ω), the calculation formula is as follows:

W _i (ω)=|W _i (f)| ² =|X _i (e ^jω )| ² , 0≤i≤N-1

In the formula, _i refers to the _i -th frame, Wi ( ^f ) is the frequency domain signal after wavelet transformation, and Wi ₍ ω) is the short-term energy spectrum.

4.5) Use the Mel filter bank to filter the obtained energy spectrum Wi (ω), and calculate the logarithmic energy _s (m) output by the filter, that is, the WT-FBank feature. The calculation formula is as follows:

where s(m) is the logarithmic energy, H _m (ω) is the frequency response of the mel filter, where m refers to the mth mel filter.

4.6) Splicing the WT-FBank features of each frame together to obtain the WT-FBank map.

Further, the present invention also provides an online intelligent identification system for abnormal noise of complex automobile components, including an abnormal noise acquisition module, a preprocessing and analysis module, a database construction module, a training model construction module, an intelligent identification module, and an automatic update module; Among them, the abnormal sound acquisition module is used to collect the abnormal sound to be identified in the problem vehicle that is started; the preprocessing and analysis module is used to preprocess and analyze the abnormal sound signal of the component collected by the abnormal sound acquisition module, and obtain a signal that can characterize the abnormal sound of the component. The feature map of the sound; the database building module is used to use the abnormal noise data to build the abnormal noise database of complex parts of the car and store it in the cloud server; the training model building module is used to build the training model; the intelligent recognition module is used to identify the feature map of the abnormal noise signal of the part , and output the classification results; the automatic update module regularly adds the newly identified various abnormal noise data into the cloud database, and retrains the model for database expansion.

Compared with the prior art, the present invention overcomes the shortcomings of the traditional method that the bench experiment is cumbersome and relies too much on the prior knowledge of experts. The invention provides a system construction idea with the help of mobile smart devices and cloud servers, which can not only provide abnormal noise identification and diagnosis services for various OEMs and 4S stores, but also provide users with an application scheme for detecting abnormal noises, further improving Sound quality performance of traditional and new energy vehicles.

Description of drawings

FIG. 1 is a schematic diagram of the composition of the online identification system for abnormal noise of complex automobile components according to the present invention.

FIG. 2 is a flow chart of the method for online intelligent identification of abnormal noise in the present invention.

Fig. 3 is a flow chart of extracting the characteristic map of abnormal noise signal WT-FBank in the present invention.

Detailed ways

The present invention will be described in detail below with reference to the various embodiments shown in the accompanying drawings, but these embodiments do not limit the present invention, and those of ordinary skill in the art can make structural, method, or functional transformations according to these embodiments. All are included in the protection scope of the present invention.

Fig. 1 is a schematic diagram showing the composition of an online intelligent identification system for abnormal noise of complex automobile parts according to the present invention, and the online identification system for abnormal noise of complex automobile parts comprises an abnormal noise acquisition module, a preprocessing and analysis module, and a database construction module , training model building module, intelligent recognition module, automatic update module. in:

Abnormal sound acquisition module, which uses a mobile smart device (smartphone or tablet) to connect a product-level microphone to collect the abnormal sound to be identified in the problem vehicle that is started. Compared with the traditional abnormal sound acquisition equipment based on laboratory professional data acquisition equipment and high-precision acoustic sensors, the hand-held equipment is more convenient and simple to operate. It can meet the acquisition accuracy and cost less, and it is also easy to determine the performance of abnormal sound sources. Obvious location area. The abnormal sound problem of a vehicle covers various complex components, including the main and auxiliary instrument panels, seat rails, door and window openings, body accessories, chassis and power transmission systems.

The preprocessing and analysis module is used for preprocessing and analyzing the abnormal sound signal of the component collected by the abnormal sound acquisition module to obtain a characteristic map that can characterize the abnormal sound of the component. That is, pre-processing and analyzing the collected abnormal noise signal of the component to be identified through the mobile smart device APP, obtaining a feature map that can characterize the abnormal sound of the component, and displaying the map on the screen of the mobile smart device.

The database building module is used to use the abnormal noise data to construct the abnormal noise database of complex automobile parts and store it in the cloud server. That is, according to the feedback of the user's problems in the process of using the car, the abnormal noise signals of common complex parts are collected in the prototype vehicle experimental environment in a targeted manner, and classified and manually marked, and the abnormal noise database of complex parts of the car is constructed and stored in the cloud server.

The training model building module is used to build the training model. That is, the data in the established abnormal noise database is trained based on the deep learning method, and the trained model is obtained for the input of the abnormal noise signal to be measured, so as to realize the online identification function.

The intelligent identification module is used to identify the feature map of the abnormal noise signal of the component and output the classification result. The spectrum obtained after pre-processing and analysis that can characterize the abnormal noise of a certain auto part is uploaded to the cloud server through the wireless network or mobile data network, input to the trained recognition model for online intelligent matching, and the output classification result is returned to Mobile smart device APP.

The module is automatically updated, and the newly identified abnormal noise data is regularly added to the cloud database, and the model is retrained for database expansion.

The functions of the above abnormal noise acquisition module and preprocessing and analysis module are mainly realized by mobile smart devices, and the functions of database building module, training model building module, intelligent identification module and automatic update module are all realized by cloud server.

As shown in Figure 2, the present invention is an online intelligent identification method for abnormal noise of complex automobile components, comprising the following steps:

1) construct the WT-FBank map of the abnormal noise signal sample of complex automobile parts, and upload it to the cloud server; as a preferred embodiment of the present invention, it includes the following steps:

1.1) Use the sound acquisition device to collect the abnormal noise signal samples of the complex parts of the automobile under the starting state;

1.2) Manually mark abnormal noise signal samples, and establish a sample library of abnormal noise signals for complex automobile parts;

In the specific embodiment of the present invention, labels are manually defined for the collected abnormal noise signals of various complex parts of the automobile, and the collected samples specifically include abnormal noise of seat vibration, abnormal noise of side door trim, abnormal noise of window sealing strip, abnormal noise of air conditioner outlet, For the abnormal noise of the main and auxiliary instrument panels and the abnormal noise of the body accessories, the above samples are respectively defined as digital labels 1-6, and different digital labels represent different types of abnormal noises.

1.3) Use the preprocessing and analysis module to preprocess the abnormal noise signal samples, that is, extract the WT-FBank features of the abnormal noise signal samples, construct the WT-FBank map of the abnormal noise signal samples, and upload them to the cloud server for storage;

In the specific embodiment of the present invention, the feature map is established on the basis of the above preprocessing, and the extracted FBank (Filter Banks) feature optimized and improved by the wavelet transform (Wavelet Transform, WT), referred to as the WT-FBank feature. FBank is a parameter proposed in the field of speech recognition to simulate the auditory characteristics of the human ear. It is a traditional acoustic feature parameter, Mel Frequency Cepstrum Coefficient (MFCC), when calculating the features obtained by excluding discrete cosine transform, reducing the amount of calculation. , so that the subsequent recognition model can make better use of the feature correlation, so as to achieve a good recognition effect. Similarly, unlike traditional feature extraction, which is based on fast Fourier transform, the time and frequency resolution in the fast Fourier transform process remains unchanged, and the best results cannot be achieved at the same time. It has better frequency resolution at low frequencies and better time resolution at high frequencies, and wavelet transform makes up for the above shortcomings. It has the ability to represent local characteristics in both time and frequency domains, and can fully adapt to the abnormal noise signal time-frequency characteristics.

As shown in Figure 3, as a preferred embodiment of the present invention, extracting the abnormal noise signal WT-FBank features, and constructing a WT-FBank map, includes the following steps:

1.3.1) Use wavelet threshold noise reduction and empirical mode decomposition noise reduction methods to perform noise reduction processing on the abnormal noise signal collected by the acquisition module, and filter out unimportant information and background noise;

1.3.2) Pre-emphasize the denoised signal through a first-order high-pass filter to compensate for the excessive attenuation of the high-frequency component of the signal during transmission;

1.3.3) Add a Hamming window to the pre-emphasized signal and frame it into a short-term signal x(i), and the frame length is N;

1.3.4) Perform wavelet transform on the short-term signal x(i), convert it into a frequency domain signal, and calculate its short-term energy spectrum W _i (ω), the calculation formula is as follows:

W _i (ω)=|W _i (f)| ² =|X _i (e ^jω )| ² , 0≤i≤N-1

In the formula, i refers to the i-th frame, W _i (f), namely X _i (e ^jω ), is the frequency domain signal after wavelet transformation, and W _i (ω) is the short-term energy spectrum;

1.3.5) Use the Mel filter bank to filter the obtained energy spectrum W _i (ω), and calculate the logarithmic energy s(m) output by the filter, that is, the WT-FBank feature. The calculation formula is as follows:

In the formula, s(m) is the logarithmic energy, H _m (ω) is the frequency response of the mel filter, where m refers to the mth mel filter (there are M in total).

In the specific embodiment of the present invention, the filter used is a triangular filter. The mel filter bank is formed by adding triangular filters to the filter bank within the mel frequency, and the center frequency of each triangular filter is distributed at equal intervals on the frequency axis.

1.3.6) Assemble the WT-FBank features of each frame together to obtain the WT-FBank map.

2) Construct a convolutional neural network model (Convolutional Neural Network, CNN) based on the classical model LeNet-5 architecture. Among them, the CNN model includes: data input layer, feature extraction layer and classification output layer.

In the specific embodiment of the present invention, the data input layer is the WT-FBank map and is cut to obtain a two-dimensional feature map input with a size of 36*36; Cyclic stacking structure, the output of the previous layer is used as the input of the latter layer, so as to build a deep feature extraction layer, set the size of the convolution kernel in the convolution layer to 3*3, the number to 10, the sliding step size of the convolution window to 1, and the pool The maximum pooling method is used in the operation, and the size of the pooling window is 2*2, and the modified linear unit (ReLU) is used as the activation function of the activation layer after the convolution; the classification output layer is formed by the combination of the fully connected layer and the Softmax classifier, and the model is set. The number of iterations is 50, and the training accuracy is required to be above 90%. Considering that the fully connected layer has too many structural parameters and a large amount of calculation, it is easy to cause the test to take too long, which is not conducive to real-time rapid identification, so a new method is adopted here: the convolution layer of the last set of feature extraction layers Then, the global average pooling layer is used to replace the traditional fully connected layer, that is, the pixels of the feature map of the previous layer are replaced by the average value, so as to obtain a low-dimensional feature vector.

3) Use the training model to build a module, combine the classifier support vector machine SVM, train the WT-FBank map of the abnormal noise signal sample, and obtain the trained CNN model and the trained SVM recognition model in turn. As a preferred embodiment of the present invention, it includes the following steps:

3.1) Input the constructed abnormal noise signal sample WT-FBank map into the CNN model to obtain the features after global average pooling;

3.2) Input the feature after global average pooling into the Softmax function, and calculate the error between the output value and the expected value.

In an achievable way, the cross entropy function is selected as the error cost function, and the calculation formula is as follows:

In the formula, n is the number of training samples, t is the output value of the training model, y is the expected value, and E is the error between the expected value and the output value.

3.3) According to the error, the back propagation algorithm is used to reversely update the weight and bias parameter parameters of each layer layer by layer, so as to achieve the purpose of minimizing the calculation error, and stop training when the training error accuracy or the number of iteration steps meet the preset target. . At this point, the parameters optimized and improved by the model will be saved to obtain the trained CNN model.

In an achievable way, the gradient descent method is used to solve the minimum error, and the calculation formula is as follows:

In the formula, ω′ and b′ are the updated weight and bias item parameters, respectively, ω and b are the current network weight and bias item parameters, respectively, and ε is the current network learning rate parameter.

3.4) Input the WT-FBank map of abnormal noise signal samples of complex automobile parts into the trained CNN model for feature extraction, convert it into a low-dimensional feature vector, and input the corresponding category label.

3.5) Input the extracted low-dimensional feature vector and the corresponding label to the SVM classifier to complete the training, and obtain the final trained SVM recognition model.

4) Use the abnormal sound acquisition module to collect the activated abnormal sound signal of the vehicle to be detected in real time. In the specific embodiment of the present invention, the abnormal noise collection module is formed by an external microphone of the intelligent handheld device.

5) Using the preprocessing and analysis module to preprocess the abnormal noise signal to be detected, that is, to obtain the WT-FBank feature of the abnormal noise signal of the vehicle to be detected, and construct the WT-FBank map of the abnormal noise signal of the vehicle to be detected;

6) Using the intelligent recognition module, input the WT-FBank map of the detected abnormal noise signal into the trained CNN model to extract the characteristics of the abnormal noise signal to be detected, and convert it into a sparse low-dimensional feature vector, and input it into the trained SVM recognition model. Intelligent matching, output the recognition category results and return them to the mobile smart device, and finally use the automatic update module to continuously improve the database;

7) Determine whether the abnormal sound signal detection is over, if not, proceed to step 4), detect and identify a new abnormal sound signal, otherwise end the abnormal sound signal detection.

Different from the traditional CNN, the global average pooling layer is designed in the improved CNN algorithm model to replace the fully connected layer in the traditional CNN, which greatly reduces the amount of model training parameters and computing time. Then, after the global mean pooling layer, the Softmax layer and the SVM classifier are connected in parallel for the final recognition and classification result output. Combining the powerful deep feature extraction and data mining capabilities of CNN and the good generalization ability of SVM for limited samples, the proposed method is more suitable for online identification of abnormal car noises.

Claims (5)

1. an on-line intelligent identification method for abnormal noise of complex parts of automobiles, is characterized in that, comprises the steps: 1) Build the WT-FBank map of abnormal noise signal samples of complex automobile parts and upload them to the cloud server; 2) Construct a convolutional neural network model CNN based on the classic model LeNet-5 architecture; 3) Use the training model to build a module, combine the classifier support vector machine SVM, train the WT-FBank map of the abnormal noise signal sample, and obtain the trained CNN model and the trained SVM recognition model; 4) Use the abnormal sound acquisition module to collect the activated abnormal sound signal of the vehicle to be detected in real time; 5) Using the preprocessing and analysis module to preprocess the abnormal sound signal to be detected, obtain the WT-FBank characteristics of the abnormal sound signal to be detected, and construct the WT-FBank map of the abnormal sound signal to be detected; 6) Using the intelligent recognition module, input the WT-FBank map of the detected abnormal noise signal into the trained CNN model to extract the characteristics of the abnormal noise signal to be detected, and convert it into a sparse low-dimensional feature vector, and input it into the trained SVM recognition model. Intelligent matching, output recognition category results and return to the mobile smart device, and use the automatic update module to improve the database; 7) Determine whether the abnormal sound signal detection is over, if not, go to step 4), continue to detect and identify new abnormal sound signals, otherwise end the abnormal sound signal detection. 2. The online intelligent identification method for abnormal noise of complex automobile parts as claimed in claim 1, characterized in that, described in the step 1), construct a WT-FBank map of the abnormal noise signal sample of complex automobile parts, and upload it to the cloud The server includes the following steps: 1.1) Use the sound acquisition device to collect the abnormal noise signal samples of the complex parts of the automobile under the starting state; 1.2) Manually mark abnormal noise signal samples, and establish a sample library of abnormal noise signals for complex automobile parts; 1.3) Use the preprocessing and analysis module to preprocess the abnormal noise signal samples, that is, extract the WT-FBank features of the abnormal noise signal samples, construct the WT-FBank map of the abnormal noise signal samples, and upload them to the cloud server for storage. 3. the online intelligent identification method for abnormal noise of complex automobile parts as claimed in claim 1, is characterized in that, in described step 3), obtain trained CNN model and trained SVM recognition model, comprise the steps: 3.1) Input the WT-FBank map of the abnormal noise signal sample into the CNN model to obtain the features after global average pooling; 3.2) Input the feature after global average pooling into the Softmax function, and calculate the error between the output value and the expected value; 3.3) According to the error, the back propagation algorithm is used to reversely update the weight and bias parameter parameters of each layer layer by layer until the training error accuracy is reached, and a trained CNN model is obtained; 3.4) Input the WT-FBank map of the abnormal noise signal sample of complex automobile parts into the trained CNN model for feature extraction, convert it into a low-dimensional feature vector, and input the corresponding category label; 3.5) Input the extracted low-dimensional feature vector and the corresponding label to the SVM classifier to complete the training, and obtain a trained SVM recognition model. 4. the online intelligent identification method for abnormal noise of complex automobile parts as claimed in claim 1, is characterized in that, in described step 1.3) and described step 5), pre-processing and analysis module pre-processing process comprises the steps: 4.1) Use wavelet threshold noise reduction and empirical mode decomposition noise reduction methods to perform noise reduction processing on the abnormal noise signal collected by the acquisition module, and filter out unimportant information and background noise; 4.2) Pre-emphasize the denoised signal through a first-order high-pass filter to compensate for the excessive attenuation of the high-frequency component of the signal during transmission; 4.3) Add a Hamming window to the pre-emphasized signal and frame it into a short-term signal x(i), and the frame length is N; 4.4) Perform wavelet transform on the short-term signal x(i), convert it into a frequency domain signal, and calculate its short-term energy spectrum W _i (ω), the calculation formula is as follows: W _i (ω)=|W _i (f)| ² =|X _i (e ^jω )| ² , 0≤i≤N-1 In the formula, i refers to the i-th frame, W _i (f), namely X _i (e ^jω ), is the frequency domain signal after wavelet transformation, and W _i (ω) is the short-term energy spectrum; 4.5) Use the Mel filter bank to filter the obtained energy spectrum Wi (ω), and calculate the logarithmic energy _s (m) output by the filter, that is, the WT-FBank feature. The calculation formula is as follows:

where s(m) is the logarithmic energy, H _m (ω) is the frequency response of the Mel filter, where m refers to the mth Mel filter; 4.6) Splicing the WT-FBank features of each frame together to obtain the WT-FBank map. 5. An online intelligent identification system for abnormal noise of complex automobile components, characterized in that it comprises an abnormal noise acquisition module, a preprocessing and analysis module, a database construction module, a training model construction module, an intelligent identification module, and an automatic update module; The abnormal sound acquisition module is used to collect the abnormal sound to be identified in the problem vehicle that is started; the preprocessing and analysis module is used to preprocess and analyze the abnormal sound signal of the component collected by the abnormal sound acquisition module, and obtain a signal that can characterize the abnormal sound. The characteristic map of the abnormal sound of the parts; the database building module is used to construct the abnormal noise database of complex automobile parts by using the abnormal noise data and store it in the cloud server; the training model building module is used to build the training model; the intelligent identification The module is used to identify the feature map of the abnormal noise signal of the component, and output the classification result; the automatic update module regularly adds the newly identified various abnormal noise data into the cloud database, and retrains the model for database expansion.

Images