CNN-Adaboost multi-sensor asynchronous motor fault diagnosis method based on attention mechanism

Abstract

The invention provides a CNN-Adaboost multi-sensor asynchronous motor fault diagnosis method based on an attention mechanism. The CNN-Adaboost multi-sensor asynchronous motor fault diagnosis method comprises the steps that vibration and electromagnetic time domain signals of a motor in different operation states are collected to serve as fault diagnosis data; converting time domain data in the fault diagnosis data into a symmetrical point image form to obtain image features of vibration and electromagnetic signals; respectively inputting image data of vibration and electromagnetic signals into secondary classifiers for training to respectively obtain secondary classifiers under vibration and electromagnetic measurement modes, obtaining weights of the secondary classifiers under the vibration and electromagnetic measurement modes in combination with an attention mechanism, and carrying out weighted fusion according to the weights of the two classifiers to obtain the vibration and electromagnetic measurement modes. And a final strong classifier is obtained. According to the invention, data feature information can be effectively extracted; and randomly extracting a sample set and dividing the sample set into a training set and a test set, thereby realizing CNN-Adaboost multi-sensor asynchronous motor fault diagnosis based on the attention mechanism.

Description

Technical Field

[0001] The present invention relates to the technical field of asynchronous motor fault diagnosis, and in particular to a CNN-Adaboost multi-sensor asynchronous motor fault diagnosis method based on an attention mechanism. Background Art

[0002] Asynchronous motors have the advantages of simple structure, easy manufacturing, low cost, reliable operation, high operating efficiency and strong practicality. They are widely used in industrial fields such as mining equipment, metallurgy, fans, power plants and machine tools. The stability and reliability of asynchronous motors are not only related to the stability and long-term operation of a single mechanical equipment, but also affect the stability of normal production of the entire industrial system. Once an asynchronous motor fails, the equipment operation will be interrupted, the operating efficiency and safety stability will be reduced, and the entire industrial production system will collapse in serious cases, resulting in major economic losses and even casualties.

[0003] At present, asynchronous motor fault diagnosis can be roughly divided into model-driven and data-driven. Model-driven is to establish accurate physical models and data models based on the structure of asynchronous motors, and accurately represent the specific operating mode of faults through mathematical models. Its core is the analysis of the mathematical model of the faulty motor system. In specific research, due to the various types of motors and complex structures, it is very complicated to build corresponding mathematical models for different motors. When a motor fault occurs, due to the different types of faults, the locations of the faults are also different. At the same time, there may be multiple components or parts that fail at the same time, and nonlinear coupling relationships between different faults occur. In this case, it is very difficult to build an accurate mathematical model, which also causes certain limitations of model-driven. The asynchronous motor fault diagnosis technology based on data-driven can be further divided into time-frequency analysis method, torque analysis method, and artificial intelligence analysis method. The permanent magnet synchronous motor fault diagnosis method based on time-frequency analysis processes the fault signal and locates the fault characteristics from the sampled signal of the asynchronous motor through methods such as Hilbert transform, wavelet transform, and fast Fourier transform to complete the fault diagnosis. Permanent magnet synchronous motor fault diagnosis methods based on artificial intelligence mainly include machine learning, fuzzy logic, heuristic algorithms, etc. With the explosive and rapid development of big data and artificial intelligence technologies, related fault diagnosis research is also constantly innovating.

[0004] The motor status signal not only contains fault features, but also contains redundant signals and other interference signals. Therefore, the extraction of fault features is particularly important in motor fault diagnosis. Existing theories and methods focus on the extraction of detail features in the time domain, frequency domain or wavelet domain. The main feature extraction methods include short-time Fourier transform and continuous wavelet transform. These methods only extract detail features in one dimension, but cannot obtain two-dimensional features in the status signal. Extending the status signal to a two-dimensional image will be more conducive to fault diagnosis. The local features of the image and its descriptor are compact vector representations of the local domain. Commonly used local features include corner features and spot features. The feature description methods used for classification mainly include heat maps, shapes, etc.

[0005] The electromechanical faults of motor systems are of various types, complex structures, and difficult to monitor. Conventional signal detection methods and fault diagnosis methods have great limitations. Furthermore, with the rapid development of communication technology, signal detection and other technologies, various types of data information obtained through various sensors and other information collection channels for faults in different components and locations have shown explosive growth, and a large amount of heterogeneous information has been pouring in at the same time, which is full of overlaps, contradictions, multiple sources, and a wide variety of structures. There are great differences in accuracy, detail, structure, and extraction methods. Moreover, the performance changes caused by faults in different components and locations are different, and the external fault symptoms are also different. There is a complex nonlinear mapping relationship between them.

[0006] Most existing diagnostic methods only perform fault diagnosis on one type of sensor data. This method requires selecting sensor signals with high correlation with the fault as diagnostic signals based on the analysis of the fault mechanism in the early stage, which places high demands on the acquisition of diagnostic signals. In addition, the external fault signs are not only reflected in one diagnostic signal. Different faults may affect the data of the same sensor. Discarding sensor data with low fault correlation and using only one sensor data for fault diagnosis will also result in a waste of resources. Therefore, it is urgent to study an asynchronous motor fault diagnosis method based on multi-sensor data.

[0007] The invention patent application with publication number CN115114848A discloses a three-phase asynchronous motor fault diagnosis method and system based on hybrid CNN-LSTM, the diagnostic method comprising: collecting motor state information and fault type data group, processing the value of the fault type data group corresponding to the acquisition time; dividing the fault type data group into training samples and test samples, setting the label of the training sample; traversing the data sequence in the training sample, mining deep features and outputting the result; constructing a CNN-LSTM deep learning model according to the output result to train the time series data in the training sample, and confirming the model parameters of the CNN-LSTM deep learning model. The diagnostic system comprises an acquisition module, a sample grouping module, a calculation module, a modeling module and a prediction module. The disadvantage of this method is that it includes both CNN and LSTM parts, which increases its computational complexity, requires larger computing resources and memory, and also leads to a longer training time. Due to its dependence on data, this method also requires a large amount of diverse data for training, which has certain limitations in application when the amount of fault data is small and difficult to obtain. Summary of the invention

[0008] In order to solve the above technical problems, the present invention proposes a CNN-Adaboost multi-sensor asynchronous motor fault diagnosis method based on an attention mechanism, which uses a symmetrical dot pattern method with optimized parameters to expand the one-dimensional time domain signals of vibration and magnetic field in multi-sensor signals into a two-dimensional symmetrical dot image, which can effectively extract data feature information; randomly extract a sample set into a training set and a test set, and use the CNN-AdaBoost algorithm to train a secondary classifier; and use the attention mechanism to weightedly fuse the secondary classifiers to obtain the final decision classifier, thereby realizing the CNN-Adaboost multi-sensor asynchronous motor fault diagnosis based on the attention mechanism.

[0009] The purpose of the present invention is to provide a CNN-Adaboost multi-sensor asynchronous motor fault diagnosis method based on an attention mechanism, comprising collecting vibration and electromagnetic time domain signals of the motor under different operating states as fault diagnosis data, and also comprising the following steps:

[0010] Step 1: Convert the time domain data in the fault diagnosis data into a symmetrical point image to obtain image features of vibration and electromagnetic signals;

[0011] Step 2: Input the image data of vibration and electromagnetic signals into the secondary classifier for training respectively, and obtain the secondary classifiers under the two measurement methods of vibration and electromagnetic respectively. Combine the attention mechanism to obtain the weights of the secondary classifiers under the two measurement methods of vibration and electromagnetic. Perform weighted fusion according to the weights of the two classifiers to obtain the final strong classifier.

[0012] Preferably, the different operating states include a normal state, a bearing inner ring fault state, a bearing outer ring fault state and a bearing rolling element fault state.

[0013] In any of the above schemes, preferably, the collecting of vibration and electromagnetic time domain signals of the motor in different operating states as fault diagnosis data includes using an AC flux sensor and a three-axis piezoelectric acceleration sensor to respectively collect the vibration and electromagnetic signals of the motor in different operating states as fault diagnosis data.

[0014] In any of the above schemes, preferably, step 1 includes converting the collected one-dimensional vibration and electromagnetic signals into a two-dimensional image using a symmetrical dot pattern method to extract signal features and generate a sample space.

[0015] In any of the above solutions, it is preferred that the signal point z in the time domain is n When converted to polar coordinate space S(r(n),α(n),β(n)), z n The radius mapped to polar coordinates is expressed as

[0016]

[0017] Among them, z min and z max are the minimum and maximum amplitude values ​​of signal z, respectively. The counterclockwise rotation angle α(n) and clockwise rotation angle β(n) of the mirror symmetry plane are given by z n The amplitude z after the time difference α n+α Mapping, n is the nth signal.

[0018] In any of the above schemes, preferably, the formula for the counterclockwise rotation angle α(n) is expressed as

[0019]

[0020] The formula for the clockwise rotation angle β(n) is expressed as

[0021]

[0022] Among them, θ w is the angle of the wth symmetry plane, θ w =360m / k, m=0,1,2…k-1, k is the number of mirror symmetry planes, and λ is the gain angle.

[0023] In any of the above schemes, preferably, step 2 comprises the following sub-steps:

[0024] Step 21: For any sensor t among the T sensors, the input feature X t , respectively generate digital labels Y tAnd the one-hot label Z t , get T input feature databases;

[0025] Step 22: For any motor operating state x, randomly select I sets of training data sets from T databases And J group test data set Any i-th test data set is Any j-th test data set is in, is the randomly selected training set data, is the numerical label corresponding to the selected training set data, is the one-hot label corresponding to the selected training set data;

[0026] Step 23: For any t-th CNN-Adaboost classifier among T CNN-Adaboost classifiers: determine the P base classifiers in any t-th CNN-Adaboost classifier according to the sample input and output dimensions The neural network structure and parameters, and initialize the weight of any i-th group of test data as Where I is the number of groups in the training data set;

[0027] Step 24: For any t-th CNN-Adaboost classifier among the T CNN-Adaboost classifiers: perform P rounds of training using the training data set, where P base classifiers Get the prediction error According to the prediction error Calculate P base classifiers separately The classifier weight

[0028] Step 25: Combine the digital label output predicted by the base classifier under T CNN-Adaboost classifiers Get the similarity matrix of each base classifier Attention allocation weights obtained through the attention mechanism Where p is the number of the base classifier;

[0029] Step 26: For the CNN-Adaboost linear combination output H obtained by the sub-classifier in any t-th CNN-Adaboost classifier among the T CNN-Adaboost classifiers after P rounds of training, t And output one-hot label matrix The final strong classifier based on the attention mechanism allocates weights according to attention A classifier H that combines T CNN-Adaboost classifierst Get the output H of the final strong classifier based on the attention mechanism and its output fault label result

[0030] Step 27: Set J test data sets As the input of the improved CNN-Adaboost classifier based on the attention mechanism, the predicted one-hot label is converted into a digital prediction result and compared with the digital label output, and the accuracy of the CNN-Adaboost strong classifier H based on the attention mechanism is statistically analyzed.

[0031] In any of the above solutions, preferably, in step 24, P rounds of training are performed using the test data set, wherein P base classifiers When , we get the pth base classifier The prediction error The function expression is:

[0032]

[0033] in, is the weight of the i-th test data set in the t-th CNN-Adaboost classifier in the p-th round, Base classifier For input features Predicted one-hot labels, For input features The corresponding one-hot label.

[0034] In any of the above solutions, preferably, in step 24, according to the prediction error Calculate P base classifiers separately The classifier weight When , we get the pth base classifier The classifier weight The function expression is:

[0035]

[0036] in, is the prediction error of the p-th base classifier in the t-th CNN-Adaboost classifier.

[0037] In any of the above solutions, preferably, in step 24, the weights of the p+1th round are updated in each round of training. The function expression is:

[0038]

[0039] in, is the weight of the i-th test data set in the t-th CNN-Adaboost classifier in the p-th round, is the normalization factor of the t-th CNN-Adaboost classifier, is the pth base classifier in the tth CNN-Adaboost classifier The classifier weights are is the base classifier in the t-th CNN-Adaboost classifier For the input features in the i-th test dataset The predicted numeric label, is the input feature of the i-th test dataset in the t-th CNN-Adaboost classifier The corresponding numeric label.

[0040] In any of the above solutions, preferably, the similarity matrix of each base classifier is The expression is

[0041]

[0042] in, and are the similarity matrices corresponding to the 1st and Kth states y in the tth CNN-Adaboost classifier, and the similarity matrix corresponding to any kth state y The calculation function expression is:

[0043]

[0044] Among them, N t (k) is the sample statistics of the k-th state y in the t-th CNN-Adaboost classifier, is the cosine similarity, and the cosine similarity The calculation function expression is:

[0045]

[0046] Where I is the number of groups in the test data set, is the pth base classifier in the tth CNN-Adaboost classifier For the input features in the i-th test dataset The predicted numeric label, is the pth base classifier in the tth CNN-Adaboost classifier Input features in the i-th test dataset The corresponding numeric label.

[0047] In any of the above schemes, preferably, the attention allocation weight corresponding to any k-th state y is The expression is:

[0048]

[0049] in, represents the attention allocation weight corresponding to any k-th state y in the t-th CNN-Adaboost classifier, is the similarity matrix corresponding to any k-th state y in the t-th CNN-Adaboost classifier, and T is the number of CNN-Adaboost classifiers.

[0050] In any of the above solutions, preferably, the CNN-Adaboost linear combination output H t The expression is

[0051]

[0052] The output one-hot label matrix The expression is

[0053]

[0054] in, is the pth base classifier in the tth CNN-Adaboost classifier The classifier weights are is the pth base classifier For input features The prediction results, is the output one-hot label matrix corresponding to the i-th test data set, and P is the specified number of base classifiers included in any t-th CNN-Adaboost classifier.

[0055] In any of the above schemes, it is preferred that the expression of the output H of the final strong classifier based on the attention mechanism is:

[0056]

[0057] Among them, α t is the attention weight matrix for the pth base classifier in the T CNN-Adaboost classifiers, H t (X t ) is the input feature X of the subclassifier in the t-th CNN-Adaboost classifier t The CNN-Adaboost linear combination output, is the pth base classifier in the tth CNN-Adaboost classifier The corresponding attention allocation weight is is the pth base classifier in the tth CNN-Adaboost classifier The classifier weights are Input features for the p-th base classifier The prediction results are as follows: T is the number of CNN-Adaboost classifiers, and P is the specified number of base classifiers included in any t-th CNN-Adaboost classifier.

[0058] In any of the above solutions, preferably, step 27 includes the following sub-steps:

[0059] Step 271: Define the number of misclassifications n for any t CNN-Adaboost classifiers t , initialize the classifier number t to 1, initialize the number of misclassifications n t The value of is 0, and the data number q=1;

[0060] Step 272: Get the one-hot label predicted by the tth CNN-Adaboost classifier And the predicted one-hot label Convert to numeric labels And digital label The predicted value and the input numeric label For comparison, if the digital label The j-th predicted value of Not equal to the input numeric label The jth digital label in Then the number of misclassifications n t Add 1, otherwise the number of misclassifications n will not be modified t If the data number q is less than the number of test data sets J, the data number q is increased by 1 and step 272 is repeated;

[0061] Step 273: According to k t =n t / J Count the error rate k of the tth CNN-Adaboost classifier t , where n t is the number of misclassifications of the t-th CNN-Adaboost classifier, J is the number of groups in the test data set, if the classifier number t is less than the number T of CNN-Adaboost classifiers, add 1 to the classifier number t and execute step 272;

[0062] Step 274: Initialize the number of incorrect predictions N of the final strong classifier based on the attention mechanism to 0 and the data number q to 1;

[0063] Step 275: Obtain the one-hot label predicted by the final strong classifier based on the improved attention mechanism Predicted one-hot labels Convert to numeric labels And digital label Compare with the input value. If the digital label The j-th predicted value of Not equal to the input numeric label Y test The jth digital label Y in test(j) , then the number of wrong predictions N is increased by 1, otherwise the number of wrong predictions N remains unchanged, if the data number q is less than the number of groups J of the test data set, then the data number q is increased by 1, and step 275 is repeated;

[0064] Step 276: The error rate k of the final strong classifier based on the attention mechanism is calculated according to k=N / J, where N is the number of incorrect predictions and J is the number of groups in the test data set.

[0065] The present invention proposes a CNN-Adaboost multi-sensor asynchronous motor fault diagnosis method based on the attention mechanism, which achieves high-accuracy identification of rotor imbalance faults and solves the difficult problem of asynchronous motor rotor imbalance fault diagnosis based on multi-sensor data.

[0066] CNN-Adaboost is an improved Adaboost algorithm based on CNN as the base classifier. This algorithm linearly combines multiple CNN classifiers through weight allocation based on the attention mechanism to obtain a strong classifier that comprehensively considers the classification results from multiple CNN base classifiers.

[0067] One-hot label is also called unique hot encoding. The rest of the positions are 0, and only one position is 1. The length of the one-hot label is equal to the number of categories of the input data. It converts the classification features into a format that is more suitable for the classification algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 The present invention is a flowchart of a preferred embodiment of the CNN-Adaboost multi-sensor asynchronous motor fault diagnosis method based on the attention mechanism.

[0069] Figure 2 It is a schematic diagram of a preferred embodiment of the CNN-Adaboost multi-sensor asynchronous motor fault diagnosis method based on the attention mechanism according to the present invention.

[0070] Figure 3 It is a schematic diagram of symmetry points of a motor in different states according to an embodiment of the CNN-Adaboost multi-sensor asynchronous motor fault diagnosis method based on the attention mechanism of the present invention.

[0071] Figure 4 The present invention is a flowchart of an embodiment of the preprocessing of training set and test set data of the CNN-Adaboost multi-sensor asynchronous motor fault diagnosis method based on the attention mechanism of the present invention.

[0072] Figure 5 It is a structural schematic diagram of an embodiment of CNN-Adaboost of the CNN-Adaboost multi-sensor asynchronous motor fault diagnosis method based on the attention mechanism of the present invention.

[0073] Figure 6 The present invention is a training flowchart of an embodiment of a CNN-Adaboost classifier of the CNN-Adaboost multi-sensor asynchronous motor fault diagnosis method based on the attention mechanism of the present invention. DETAILED DESCRIPTION

[0074] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0075] Embodiment 1

[0076] like Figure 1 As shown, step 1000 is executed to collect the vibration and electromagnetic time domain signals of the motor in different operating states as fault diagnosis data, and an AC magnetic flux sensor and a triaxial piezoelectric acceleration sensor are used to respectively collect the vibration and electromagnetic signals of the motor in different operating states as fault diagnosis data,

[0077] The different operating states include a normal state, a bearing inner ring fault state, a bearing outer ring fault state and a bearing rolling element fault state.

[0078] Execute step 1100, convert the time domain data in the fault diagnosis data into a symmetrical point image, obtain the image features of the vibration and electromagnetic signals, and convert the collected one-dimensional vibration and electromagnetic signals into a two-dimensional image using a symmetrical dot pattern method to extract signal features and generate a sample space.

[0079] The signal point z in the time domain n When converted to polar coordinate space S(r(n),α(n),β(n)), z n The radius mapped to polar coordinates is expressed as

[0080]

[0081] Among them, zmin and z max are the minimum and maximum amplitude values ​​of signal z, respectively. The counterclockwise rotation angle α(n) and clockwise rotation angle β(n) of the mirror symmetry plane are given by z n The amplitude z after the time difference α n+α Mapping, n is the nth signal.

[0082] The formula for the counterclockwise rotation angle α(n) is expressed as

[0083]

[0084] The formula for the clockwise rotation angle β(n) is expressed as

[0085]

[0086] Among them, θ w is the angle of the wth symmetry plane, θ w =360m / k, m=0,1,2…k-1, k is the number of mirror symmetry planes, and λ is the gain angle.

[0087] Execute step 1200, input the image data of vibration and electromagnetic signals into the secondary classifiers for training respectively, obtain the secondary classifiers under the vibration and electromagnetic measurement methods respectively, combine the attention mechanism to obtain the weights of the secondary classifiers under the vibration and electromagnetic measurement methods, perform weighted fusion according to the weights of the two classifiers, and obtain the final strong classifier, including the following sub-steps:

[0088] Execute step 1210, for the input feature X corresponding to any sensor t among the T types of sensors t , respectively generate digital labels Y t And the one-hot label Z t , and obtain T input feature databases.

[0089] Execute step 1220, for any motor running state x, randomly select I sets of training data sets from T databases And J group test data set Any i-th test data set is Any j-th test data set is in, is the randomly selected training set data, is the numerical label corresponding to the selected training set data, is the one-hot label corresponding to the selected training set data.

[0090] Execute step 1230, for any t-th CNN-Adaboost classifier among the T CNN-Adaboost classifiers: determine the P base classifiers in any t-th CNN-Adaboost classifier according to the sample input and output dimensions The neural network structure and parameters, and initialize the weight of any i-th group of test data as Where I is the number of groups in the training data set.

[0091] Execute step 1240, for any t-th CNN-Adaboost classifier among the T CNN-Adaboost classifiers: perform P rounds of training using the training data set, where P base classifiers Get the prediction error According to the prediction error Calculate P base classifiers separately The classifier weight

[0092] P rounds of training with the test dataset, where P base classifiers When , we get the pth base classifier The prediction error The function expression is:

[0093]

[0094] in, is the weight of the i-th test data set in the t-th CNN-Adaboost classifier in the p-th round, Base classifier For input features Predicted one-hot labels, For input features The corresponding one-hot label;

[0095] According to the prediction error Calculate P base classifiers separately The classifier weight When , we get the pth base classifier The classifier weight The function expression is:

[0096]

[0097] in, is the prediction error of the p-th base classifier in the t-th CNN-Adaboost classifier;

[0098] Update the weights of the p+1th round in each round of training The function expression is:

[0099]

[0100] in, is the weight of the i-th test data set in the t-th CNN-Adaboost classifier in the p-th round, is the normalization factor of the t-th CNN-Adaboost classifier, is the pth base classifier in the tth CNN-Adaboost classifier The classifier weights are is the base classifier in the t-th CNN-Adaboost classifier For the input features in the i-th test dataset The predicted numeric label, is the input feature of the i-th test dataset in the t-th CNN-Adaboost classifier The corresponding numeric label.

[0101] Execute step 1250, combine the digital label output predicted by the base classifier under T CNN-Adaboost classifiers Get the similarity matrix of each base classifier Attention allocation weights obtained through the attention mechanism Where p is the number of the base classifier;

[0102] The similarity matrix of each base classifier The expression is

[0103]

[0104] in, and are the similarity matrices corresponding to the 1st and Kth states y in the tth CNN-Adaboost classifier, and the similarity matrix corresponding to any kth state y The calculation function expression is:

[0105]

[0106] Among them, N t (k) is the sample statistics of the k-th state y in the t-th CNN-Adaboost classifier, is the cosine similarity, and the cosine similarity The calculation function expression is:

[0107]

[0108] Where I is the number of groups in the test data set, is the pth base classifier in the tth CNN-Adaboost classifier For the input features in the i-th test dataset The predicted numeric label, is the pth base classifier in the tth CNN-Adaboost classifier Input features in the i-th test dataset The corresponding numeric label.

[0109] Execute step 1260, for the CNN-Adaboost linear combination output H obtained by the sub-classifier in any t-th CNN-Adaboost classifier among the T CNN-Adaboost classifiers after P rounds of training, t And output one-hot label matrix The final strong classifier based on the attention mechanism allocates weights according to attention A classifier H that combines T CNN-Adaboost classifiers t Get the output H of the final strong classifier based on the attention mechanism and its output fault label result

[0110] The attention allocation weight corresponding to any k-th state y The expression is:

[0111]

[0112] in, represents the attention allocation weight corresponding to any k-th state y in the t-th CNN-Adaboost classifier, is the similarity matrix corresponding to any k-th state y in the t-th CNN-Adaboost classifier, and T is the number of CNN-Adaboost classifiers;

[0113] The CNN-Adaboost linear combination outputs H t The expression is

[0114]

[0115] The output one-hot label matrix The expression is

[0116]

[0117] in, is the pth base classifier in the tth CNN-Adaboost classifier The classifier weights are is the pth base classifier For input features The prediction results, is the output one-hot label matrix corresponding to the i-th test data set, and P is the specified number of base classifiers included in any t-th CNN-Adaboost classifier;

[0118] The output H of the final strong classifier based on the attention mechanism is expressed as

[0119]

[0120] Among them, α t is the attention weight matrix for the pth base classifier in the T CNN-Adaboost classifiers, H t (X t ) is the input feature X of the subclassifier in the t-th CNN-Adaboost classifier t The CNN-Adaboost linear combination output, is the pth base classifier in the tth CNN-Adaboost classifier The corresponding attention allocation weight is is the pth base classifier in the tth CNN-Adaboost classifier The classifier weights are Input features for the p-th base classifier The prediction results are as follows: T is the number of CNN-Adaboost classifiers, and P is the specified number of base classifiers included in any t-th CNN-Adaboost classifier.

[0121] Execute step 1270 to set the J test data set As the input of the CNN-Adaboost classifier improved based on the attention mechanism, the predicted one-hot label is converted into a digital prediction result and compared with the digital label output, and the accuracy of the CNN-Adaboost strong classifier H based on the attention mechanism is counted, including the following sub-steps:

[0122] Execute step 1271 to define the number of misclassifications n of any t CNN-Adaboost classifiers t , initialize the classifier number t to 1, initialize the number of misclassifications n t The value of is 0, and the data number q=1;

[0123] Execute step 1272 to obtain the one-hot label predicted by the tth CNN-Adaboost classifier And the predicted one-hot label Convert to numeric labels And digital label The predicted value and the input numeric label For comparison, if the digital label The j-th predicted value of

[0124] Not equal to the input numeric label The jth digital label in Then the number of misclassifications n t Add 1, otherwise the number of misclassifications n will not be modified t If the data number q is less than the number of test data sets J, the data number q is increased by 1 and step 1272 is repeated;

[0125] Execute step 1273, according to k t =n t / J Count the error rate k of the tth CNN-Adaboost classifier t , where n t is the number of misclassifications of the t-th CNN-Adaboost classifier, J is the number of groups in the test data set, if the classifier number t is less than the number T of CNN-Adaboost classifiers, add 1 to the classifier number t and execute step 1273;

[0126] Execute step 1274 to initialize the number of incorrect predictions N of the final strong classifier based on the attention mechanism to 0 and the data number q to 1;

[0127] Execute step 1275 to obtain the one-hot label predicted by the final strong classifier improved based on the attention mechanism Predicted one-hot labels Convert to numeric labels And digital label Compare with the input value. If the digital label The j-th predicted value of Not equal to the input numeric label Y test The jth digital label Y in test(j) , then the number of wrong predictions N is increased by 1, otherwise the number of wrong predictions N remains unchanged, if the data number q is less than the number of groups J of the test data set, then the data number q is increased by 1, and step 1275 is repeated;

[0128] Execute step 1276 and calculate the error rate k of the final strong classifier based on the attention mechanism according to k=N / J, where N is the number of incorrect predictions and J is the number of groups in the test data set.

[0129] Embodiment 2

[0130] like Figure 2 As shown in the figure, a CNN-Adaboost multi-sensor asynchronous motor fault diagnosis method based on attention mechanism includes three parts: diagnostic data acquisition, diagnostic data feature extraction and classifier training:

[0131] 1) Collect the vibration and electromagnetic time domain signals of the motor under different operating conditions as fault diagnosis data.

[0132] Step 1) collecting vibration and electromagnetic time domain signals of the motor under different operating conditions as fault diagnosis data, including:

[0133] 1.1) Different operating states of the motor include normal state, bearing inner ring fault state, bearing outer ring fault state and bearing rolling element fault state.

[0134] 1.2) Use AC flux sensor and triaxial piezoelectric acceleration sensor to collect vibration and electromagnetic signals of the motor under different operating conditions as fault diagnosis data.

[0135] 2) Convert the obtained time domain data into a symmetrical point image to obtain the image features of vibration and electromagnetic signals. The symmetrical point image is as follows: Figure 3 shown.

[0136] Step 2) converting the obtained time domain data into a symmetrical point image to obtain image features of vibration and electromagnetic signals, including:

[0137] The collected one-dimensional vibration and electromagnetic signals are converted into two-dimensional images using the symmetrical dot pattern method to extract signal features and generate sample space:

[0138] The signal point z in the time domain n When converted to polar coordinate space S(r(n),α(n),β(n)), z n The radius mapped to polar coordinates is expressed as

[0139]

[0140] Among them, z min and z max are the minimum and maximum amplitude values ​​of signal z respectively. The counterclockwise rotation angle α(n) and clockwise rotation angle β(n) of the mirror symmetry plane are given by z n The amplitude z after the time difference α n+α Mapping, the formula is expressed as

[0141]

[0142]

[0143] Among them, θ w is the angle of the wth symmetry plane, θ w =360m / k, m = 0, 1, 2…k-1, k is the number of mirror symmetry planes, generally 6), λ is the gain angle (λ≤θ w ), α(n) and β(n) jointly determine the range of polar coordinates. SDP images are used to reflect the hidden morphological features in one-dimensional vibration and electromagnetic signals.

[0144] 3) The image data of vibration and electromagnetic signals are respectively input into the secondary classifier for training, and the secondary classifiers under the vibration and electromagnetic measurement methods are obtained respectively. The weights of the secondary classifiers under the vibration and electromagnetic measurement methods are obtained by combining the attention mechanism. The weighted fusion is performed according to the weights of the two classifiers to obtain the final strong classifier.

[0145] Step 3) The image data of vibration and electromagnetic signals are respectively input into the secondary classifier for training, and the secondary classifiers under the vibration and electromagnetic measurement methods are obtained respectively. The weights of the secondary classifiers under the vibration and electromagnetic measurement methods are obtained by combining the attention mechanism. The weighted fusion is performed according to the weights of the two classifiers to obtain the final strong classifier, including:

[0146] 3.1) For any sensor t among the T sensors, the input feature X t , respectively generate digital labels Y t And the one-hot label Z t , get T input feature databases;

[0147] like Figure 4 As shown, step 3.1) includes:

[0148] 3.1.1) Training sample set input;

[0149] 3.1.2) Expand the training data sample set into a two-dimensional symmetrical dot matrix image feature set;

[0150] 3.1.3) Randomly select I training data set and J test data set from the two-dimensional symmetrical dot matrix image feature set;

[0151] 3.1.4) Generate digital labels Y for different image features of different sensors t And the one-hot label Z t .

[0152] 3.2) For any motor running state x, randomly select I sets of training data sets from T databases And J group test data set Any i-th test data set is Any j-th test data set is

[0153] 3.3) The structural diagram of CNN-Adaboost is as follows Figure 5 As shown, for any t-th CNN-Adaboost classifier among T CNN-Adaboost classifiers: determine the P base classifiers in any t-th CNN-Adaboost classifier according to the sample input and output dimensions The neural network structure and parameters, and initialize the weight of any i-th group of test data as Where I is the number of groups in the training data set.

[0154] 3.4) For any t-th CNN-Adaboost classifier among T CNN-Adaboost classifiers: perform P rounds of training using the training dataset, where P base classifiers Get the prediction error According to the prediction error Calculate P base classifiers separately The classifier weight

[0155] In step 3.4), P rounds of training are performed on the test data set, where P base classifiers When , we get the pth base classifier The prediction error The function expression is:

[0156]

[0157] in, is the weight of the i-th test data set in the t-th CNN-Adaboost classifier in the p-th round, Base classifier For input features Predicted one-hot labels, For input features The corresponding one-hot label;

[0158] The prediction error Calculate P base classifiers separately The classifier weight When , we get the pth base classifier The classifier weight The function expression is:

[0159]

[0160] in, is the prediction error of the p-th base classifier in the t-th CNN-Adaboost classifier;

[0161] And in each round of training, the weights of the p+1th round are updated The function expression is:

[0162]

[0163] in, is the weight of the i-th test data set in the t-th CNN-Adaboost classifier in the p-th round, is the normalization factor of the t-th CNN-Adaboost classifier, is the pth base classifier in the tth CNN-Adaboost classifier The classifier weights are is the base classifier in the t-th CNN-Adaboost classifier For the input features in the i-th test dataset The predicted numeric label, is the input feature of the i-th test dataset in the t-th CNN-Adaboost classifier The corresponding numeric label.

[0164] 3.5) Combine the digital label output predicted by the base classifier under T CNN-Adaboost classifiers Get the similarity matrix of each base classifier Attention allocation weights obtained through the attention mechanism Where p is the number of the base classifier;

[0165] The similarity matrix of each base classifier obtained in step 3.5) The expression is

[0166]

[0167] in, and are the similarity matrices corresponding to the 1st and Kth states y in the tth CNN-Adaboost classifier, and the similarity matrix corresponding to any kth state y The calculation function expression is:

[0168]

[0169] Among them, N t (k) is the sample statistics of the k-th state y in the t-th CNN-Adaboost classifier, is the cosine similarity, and the cosine similarity The calculation function expression is:

[0170]

[0171] Where I is the number of groups in the test data set, is the pth base classifier in the tth CNN-Adaboost classifier For the input features in the i-th test dataset The predicted numeric label, is the pth base classifier in the tth CNN-Adaboost classifier Input features in the i-th test dataset The corresponding numeric label.

[0172] The attention allocation weight obtained by the attention mechanism When the attention weight corresponding to any k-th state y is The expression is:

[0173]

[0174] in, represents the attention allocation weight corresponding to any k-th state y in the t-th CNN-Adaboost classifier, is the similarity matrix corresponding to any k-th state y in the t-th CNN-Adaboost classifier, and T is the number of CNN-Adaboost classifiers.

[0175] 3.6) For the CNN-Adaboost linear combination output H obtained by any t-th CNN-Adaboost classifier in T CNN-Adaboost classifiers after P rounds of training, t And output one-hot label matrix The final strong classifier based on the attention mechanism allocates weights according to attention A classifier H that combines T CNN-Adaboost classifiers t Get the output H of the final strong classifier based on the attention mechanism and its output fault label result

[0176] The CNN-Adaboost linear combination output H obtained in step 3.6) t And output one-hot label matrix The function expression is:

[0177]

[0178]

[0179] in, is the pth base classifier in the tth CNN-Adaboost classifier The classifier weights are is the pth base classifier For input features The prediction results, is the output one-hot label matrix corresponding to the i-th test data set, and P is the specified number of base classifiers included in any t-th CNN-Adaboost classifier;

[0180] The output H of the final strong classifier improved based on the attention mechanism and its output fault result label result The function expression is:

[0181]

[0182] Among them, α t is the attention weight matrix for the pth base classifier in the T CNN-Adaboost classifiers, H t (X t ) is the input feature X of the subclassifier in the t-th CNN-Adaboost classifier t The CNN-Adaboost linear combination output, is the pth base classifier in the tth CNN-Adaboost classifier The corresponding attention allocation weight is is the pth base classifier in the tth CNN-Adaboost classifier The classifier weights are Input features for the p-th base classifier The prediction results are as follows: T is the number of CNN-Adaboost classifiers, and P is the specified number of base classifiers included in any t-th CNN-Adaboost classifier.

[0183] 3.7) Set J test data sets As the input of the CNN-Adaboost classifier improved based on the attention mechanism, the predicted one-hot label is converted into a digital prediction result and compared with the digital label output, and the accuracy of the CNN-Adaboost strong classifier H based on the attention mechanism is counted;

[0184] Step 3.7) includes:

[0185] 3.7.1) Define the number of misclassifications n for any t CNN-Adaboost classifiers t , initialize the classifier number t to 1, initialize the number of misclassifications n t The value of is 0, and the data number q=1;

[0186] 3.7.2) Get the one-hot label predicted by the tth CNN-Adaboost classifier And the predicted one-hot label Convert to numeric labels And digital label The predicted value and the input numeric label For comparison, if the digital label The j-th predicted value of Not equal to the input numeric label The jth digital label in Then the number of misclassifications n t Add 1, otherwise the number of misclassifications n will not be modified t If the data number q is less than the number of test data sets J, then add 1 to the data number q and repeat 3.7.2), otherwise execute 3.7.3);

[0187] 3.7.3) According to k t =n t / J Count the error rate k of the tth CNN-Adaboost classifier t , where n t is the number of misclassifications of the t-th CNN-Adaboost classifier, J is the number of groups in the test data set, if the classifier number t is less than the number of CNN-Adaboost classifiers T, add 1 to the classifier number t and jump to 3.7.2), otherwise, the error rate k of T CNN-Adaboost classifiers t After the calculation is completed, jump to the next step;

[0188] 3.7.4) Initialize the number of incorrect predictions N of the final strong classifier based on the attention mechanism to 0 and the data number q to 1;

[0189] 3.7.5) Obtain the one-hot label predicted by the final strong classifier based on the improved attention mechanism Predicted one-hot labels Convert to numeric labels And digital label Compare with the input value. If the digital label The j-th predicted value of Not equal to the input numeric label Y tes The jth digital label Y in t test(j) , then increase the number of wrong predictions N by 1, otherwise keep the number of wrong predictions N unchanged. If the data number q is less than the number of groups J of the test data set, then increase the data number q by 1 and repeat 3.7.5), otherwise execute 3.7.6);

[0190] 3.7.6) The error rate k of the final strong classifier based on the attention mechanism is calculated according to k = N / J, where N is the number of incorrect predictions and J is the number of groups in the test data set.

[0191] In order to better understand the present invention, the above is described in detail in conjunction with the specific embodiments of the present invention, but it is not intended to limit the present invention. Any simple modifications made to the above embodiments based on the technical essence of the present invention still fall within the scope of the technical solution of the present invention. Each embodiment in this specification focuses on the differences from other embodiments, and the same or similar parts between the embodiments can be referenced to each other. For the system embodiment, since it basically corresponds to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

Claims

1. A CNN-Adaboost multi-sensor asynchronous motor fault diagnosis method based on attention mechanism, comprising collecting vibration and electromagnetic time domain signals of the motor under different operating states as fault diagnosis data, characterized in that: The following steps are also included: Step 1: Convert the time domain data in the fault diagnosis data into a symmetrical point image to obtain image features of vibration and electromagnetic signals; Step 2: Input the image data of vibration and electromagnetic signals into the secondary classifier for training respectively, and obtain the secondary classifiers under the two measurement methods of vibration and electromagnetic respectively. Combine the attention mechanism to obtain the weights of the secondary classifiers under the two measurement methods of vibration and electromagnetic. Perform weighted fusion according to the weights of the two classifiers to obtain the final strong classifier.

2. The CNN-Adaboost multi-sensor asynchronous motor fault diagnosis method based on the attention mechanism as claimed in claim 1, characterized in that: The step 1 includes converting the collected one-dimensional vibration and electromagnetic signals into two-dimensional images using a symmetrical dot matrix pattern method to extract signal features and generate a sample space.

3. The CNN-Adaboost multi-sensor asynchronous motor fault diagnosis method based on the attention mechanism as claimed in claim 2, characterized in that: The signal point z in the time domain n When converted to polar coordinate space S(r(n),α(n),β(n)), z n The radius mapped to polar coordinates is expressed as The counterclockwise rotation angle α(n) and the clockwise rotation angle β(n) of the mirror symmetry plane are given by z n The amplitude z after the time difference α n+α Mapping, Among them, z min and z max are the minimum and maximum amplitudes of signal z respectively, and n is the nth signal.

4. The CNN-Adaboost multi-sensor asynchronous motor fault diagnosis method based on the attention mechanism as claimed in claim 3, characterized in that: The formula for the counterclockwise rotation angle α(n) is expressed as The formula for the clockwise rotation angle β(n) is expressed as Among them, θ w is the angle of the wth symmetry plane, θ w =360m / k, m=0,1,2…k-1, k is the number of mirror symmetry planes, and λ is the gain angle.

5. The CNN-Adaboost multi-sensor asynchronous motor fault diagnosis method based on the attention mechanism as claimed in claim 4, characterized in that: The step 2 includes the following sub-steps: Step 21: For any sensor t among the T sensors, the input feature X t , respectively generate digital labels Y t And the one-hot label Z t , get T input feature databases; Step 22: For any motor operating state x, randomly select I sets of training data sets from T databases And J group test data set Any i-th test data set is Any j-th test data set is in, is the randomly selected training set data, is the numerical label corresponding to the selected training set data, is the one-hot label corresponding to the selected training set data; Step 23: For any t-th CNN-Adaboost classifier among the T CNN-Adaboost classifiers, determine the P base classifiers in any t-th CNN-Adaboost classifier according to the sample input and output dimensions. The neural network structure and parameters, and initialize the weight of any i-th group of test data as Where, I is the number of groups in the training data set; Step 24: For any t-th CNN-Adaboost classifier among the T CNN-Adaboost classifiers: perform P rounds of training using the training data set, where P base classifiers Get the prediction error According to the prediction error Calculate P base classifiers separately The classifier weight Step 25: Combine the digital label output predicted by the base classifier under T CNN-Adaboost classifiers Get the similarity matrix of each base classifier Attention allocation weights obtained through the attention mechanism Where p is the number of the base classifier; Step 26: For the CNN-Adaboost linear combination output H obtained by the sub-classifier in any t-th CNN-Adaboost classifier among the T CNN-Adaboost classifiers after P rounds of training, t And output one-hot label matrix The final strong classifier based on the attention mechanism allocates weights according to attention A classifier H that combines T CNN-Adaboost classifiers t Get the output H of the final strong classifier based on the attention mechanism and its output fault label result Step 27: Set J test data sets As the input of the improved CNN-Adaboost classifier based on the attention mechanism, the predicted one-hot label is converted into a digital prediction result and compared with the digital label output, and the accuracy of the CNN-Adaboost strong classifier H based on the attention mechanism is statistically analyzed.

6. The CNN-Adaboost multi-sensor asynchronous motor fault diagnosis method based on the attention mechanism as claimed in claim 5, characterized in that: In step 24, P rounds of training are performed using the test data set, where P base classifiers When , we get the pth base classifier The prediction error The function expression is: in, is the weight of the i-th test data set in the t-th CNN-Adaboost classifier in the p-th round, Base classifier For input features Predicted one-hot labels, For input features The corresponding one-hot label.

7. The CNN-Adaboost multi-sensor asynchronous motor fault diagnosis method based on the attention mechanism as claimed in claim 6, characterized in that: In step 24, according to the prediction error Calculate P base classifiers separately The classifier weight When , we get the pth base classifier The classifier weight The function expression is: in, is the prediction error of the p-th base classifier in the t-th CNN-Adaboost classifier.

8. The CNN-Adaboost multi-sensor asynchronous motor fault diagnosis method based on the attention mechanism as claimed in claim 7, characterized in that: In step 24, the weights of the p+1th round are updated in each round of training. The function expression is: in, is the weight of the i-th test data set in the t-th CNN-Adaboost classifier in the p-th round, is the normalization factor of the t-th CNN-Adaboost classifier, is the pth base classifier in the tth CNN-Adaboost classifier The classifier weights are is the base classifier in the t-th CNN-Adaboost classifier For the input features in the i-th test dataset The predicted numeric label, is the input feature of the i-th test dataset in the t-th CNN-Adaboost classifier The corresponding numeric label.

9. The CNN-Adaboost multi-sensor asynchronous motor fault diagnosis method based on the attention mechanism as claimed in claim 7, characterized in that: The similarity matrix of each base classifier The expression is in, and are the similarity matrices corresponding to the 1st and Kth states y in the tth CNN-Adaboost classifier, and the similarity matrix corresponding to any kth state y The calculation function expression is: Among them, N t (k) is the sample statistics of the k-th state y in the t-th CNN-Adaboost classifier, is the cosine similarity, and the cosine similarity The calculation function expression is: Where I is the number of groups in the test data set, is the pth base classifier in the tth CNN-Adaboost classifier For the input features in the i-th test dataset The predicted numeric label, is the pth base classifier in the tth CNN-Adaboost classifier Input features in the i-th test dataset The corresponding numeric label.

10. The CNN-Adaboost multi-sensor asynchronous motor fault diagnosis method based on the attention mechanism as claimed in claim 7, characterized in that: The attention allocation weight corresponding to any k-th state y The expression is: in, represents the attention allocation weight corresponding to any k-th state y in the t-th CNN-Adaboost classifier, is the similarity matrix corresponding to any k-th state y in the t-th CNN-Adaboost classifier, and T is the number of CNN-Adaboost classifiers.

Images