Three-phase squirrel-cage asynchronous motor rotor broken bar small sample fault diagnosis method

Abstract

The invention discloses a three-phase squirrel-cage asynchronous motor rotor broken bar small sample fault diagnosis method. Firstly, Hilbert transform is carried out on an acquired original current signal to eliminate the influence of frequency spectrum leakage of a fundamental frequency component, and a fault characteristic frequency is highlighted; secondly, a double-branch residual shrinkage network structure is provided, noise components in the signals are eliminated, and multi-scale features in the current and vibration signals are extracted; meanwhile, aiming at the time sequence characteristics of the current signal and the vibration signal, a bidirectional circulation gating unit is introduced, and the advanced characteristics of the fused characteristics are extracted. According to the method, interference of noise and harmonic waves can be reduced, and rotor broken bar fault diagnosis and severity estimation under the small sample condition are realized by adopting the current signal and the vibration signal aiming at two aspects of regularization technology and model extraction.

Description

Technical Field

[0001] The invention belongs to the technical field of asynchronous motor rotor fault diagnosis, and in particular relates to a small sample fault diagnosis of a broken rotor bar of a three-phase squirrel cage asynchronous motor. Background Art

[0002] Three-phase asynchronous motors are widely used in different industrial fields due to their high reliability, low cost, and easy control. During the actual operation of squirrel-cage asynchronous motors, the rotor bars are subjected to alternating stresses such as rotating electromagnetic force, radial electromagnetic force, centrifugal force, and high thermal stress, which can easily lead to rotor bar failure. Broken rotor bar failure is one of the common faults of asynchronous motors, accounting for 10% to 20% of the accident rate. During the operation of the motor, the occurrence of broken rotor bars may cause local temperature rise, multiple broken rotor bars, rotor misalignment and other faults, and in severe cases may cause the motor to shut down. In some industrial application scenarios, such as nuclear power plants, the shutdown and maintenance caused by broken rotor bars in large asynchronous motors often requires a lot of manpower, material resources and economic losses. Therefore, it is of great significance to timely and effectively diagnose the fault of broken rotor bars of motors.

[0003] At present, the methods for diagnosing broken rotor bar faults in squirrel cage asynchronous motors can be mainly divided into three categories: signal analysis-based methods, mathematical model-based methods, and artificial intelligence-based methods. The signal analysis-based method is currently the most widely used method in the field of motor fault diagnosis. It directly extracts fault features from signals that can characterize the working state of the motor. Time-frequency conversion methods such as fast Fourier transform, Hilbert transform, and extended Park transform are often used. However, the signal analysis-based method relies on a lot of prior knowledge and is easily affected by the operating environment. The mathematical model-based method analyzes the characteristic mechanism when the fault occurs by establishing a theoretical model of the motor. However, the accuracy of the motor model depends on the motor parameters and will be affected by factors such as the operating environment and operating load. This method has poor robustness. In recent years, with the development of artificial intelligence technology, fault diagnosis methods based on artificial intelligence have become a new research hotspot. The artificial intelligence-based method does not require a lot of prior knowledge, and can automatically extract effective fault features from the signal, and finally realize fault diagnosis, with good generalization performance.

[0004] In recent years, domestic and foreign scholars have conducted some research on the method of motor rotor broken bar fault diagnosis based on artificial intelligence, and achieved good results. However, most methods do not realize fault diagnosis under small sample conditions. Due to the small number of fault samples in industrial sites, it is of great significance to study the rotor broken bar fault diagnosis under small sample conditions. Summary of the invention

[0005] In view of the current situation that there is little research on small sample fault diagnosis of broken rotor bar of three-phase asynchronous motor, in order to make full use of training samples to realize fault diagnosis, the present invention proposes a small sample fault diagnosis method of broken rotor bar of three-phase asynchronous motor based on double-branch residual contraction network.

[0006] In order to achieve the technical purpose, the present invention adopts the following technical solution:

[0007] A method for diagnosing a broken rotor bar fault of a three-phase squirrel cage asynchronous motor using a double-branch residual contraction network comprises the following steps:

[0008] Step S1, building a three-phase squirrel cage asynchronous motor experimental platform.

[0009] Step S2, based on the motor experimental platform, collects the motor stator current and motor drive end vibration signal under various working conditions through the data acquisition system, including the motor normal operation data and different fault severity data.

[0010] Step S3, integrating the data collected in step 2, and dividing the data set into a training set and a test set according to the ratio.

[0011] Step S4, perform Hilbert transform on the current data to highlight the fault characteristic frequency, and use Pytorch to design a neural network model for rotor broken bar fault diagnosis; the above-mentioned dual-branch residual shrinkage network includes an improved residual shrinkage module, a dense connection mechanism, a bidirectional recurrent gating unit and a classification layer; first, the improved residual shrinkage network is used to filter out noise components for current and vibration data, improve the reference value of a small number of available training samples, and extract more effective features; secondly, a dense connection mechanism is used to connect each residual shrinkage module to achieve feature reuse to avoid the residual shrinkage module filtering out effective information; then a bidirectional recurrent gating unit is used to learn hidden information to further extract advanced features; at the same time, regularization technology is introduced: label smoothing regularization, Meta-ACON activation function, weight attenuation, Dropout, to prevent overfitting under small sample fault diagnosis; finally, the results are output through the classification layer.

[0012] Step S5, use the training set to train the neural network model. During the training process, a label-smoothed cross entropy loss function is introduced to calculate the error between the predicted label and the true label. At the same time, the error is used for back propagation to adjust the neural network parameters and structure. The experimental process is repeated for training, and the training is stopped when the set number of iterations is reached.

[0013] Step S6, use the test set to test the trained model, input the test sample into the network model, and obtain the accuracy of the diagnosis result; if the accuracy of the test sample reaches more than 95% in the case of a small sample, it meets the actual diagnosis requirements and realizes the small sample broken rotor bar fault diagnosis of the three-phase squirrel cage asynchronous motor.

[0014] Furthermore, when collecting the motor current and vibration data in step 2, in order to simulate faults of different severity, 12 operating conditions of 1Nm, 2Nm and 3Nm loads are collected when the motor is normal, half a rotor bar is broken, one rotor bar is broken, and three rotor bars are broken.

[0015] Furthermore, when dividing the data set in step 3, in order to make full use of the data and extract more effective features, data enhancement and normalization processing are used; this method uses the data enhancement method of equally spaced sliding windows to perform overlapping sampling on the original data to obtain a sample data set; 1000 samples are set for each working condition, and α is set to be the proportion of training samples, and α is set to 0.1 to 0.5 for experiments, and the sample length is 1024 data points.

[0016] Furthermore, after the rotor bar broken fault occurs in step 4, the collected motor stator current contains noise and harmonic components, and the fault characteristics are not obvious, so the Hilbert algorithm is used to enhance the fault characteristics, suppress noise and harmonics, make full use of the extracted features, and improve the learning ability of the model.

[0017] Step S41: When a motor has a broken rotor bar fault, a fault characteristic frequency with a frequency of (1±2s)f will appear in the motor stator current spectrum; since the slip rate s and the amplitude of the fault characteristic frequency are very small when the motor is running, the fault feature is not easy to be extracted; the envelope can be obtained by Hilbert transform to filter out the base frequency component, effectively eliminate the masking effect of spectrum leakage, and highlight the fault feature; for a continuous current signal x(t), its Hilbert transform is as follows:

[0018]

[0019] Among them, f is the power supply frequency of the motor power supply; s is the slip rate of the motor; t is the time variable of the function; τ is the time constant; x(τ) is the function form with the time constant as the variable;

[0020] Take the real-valued function x(t) as the real part and transform the transformed function As the imaginary part, the analytical signal z(t) can be expressed as:

[0021]

[0022] The envelope A(t) of a real-valued function x(t) can be obtained by taking the modulus value of the analytical signal as follows:

[0023]

[0024] By performing Hilbert transform on the current signal, the power frequency is converted into a DC component, and the fault characteristic frequency is converted from the original sideband frequency (1±2s)f to 2sf.

[0025] Step S42: The feature extraction process of the current and vibration signals will be affected by the noise in the signal. The interference of the noise may make it impossible to extract effective features, resulting in a decrease in diagnostic performance. The residual shrinkage network introduces the soft thresholding function in the traditional signal denoising algorithm, learns a unique set of thresholds according to the noise content in different samples, and filters out the noise components in the signal; by introducing the dilated convolution and attention mechanism ECA-Net, the traditional residual shrinkage network is improved to enhance the model's receptive field and feature extraction capabilities. The expression is as follows:

[0026]

[0027] Among them, average is the average pooling layer, i, j, c are the width, height, and number of channels of feature map x, respectively, x is the input feature, y is the output feature, z is the output of the two-layer fully connected network of the residual shrinkage module, α is the corresponding scaling parameter, and the scaling parameter is multiplied by the absolute value of x to obtain the threshold τ of the soft threshold function, which can ensure that the threshold size meets the requirements; that is, the noise component between -τ and τ is filtered out; in addition, the derivative of the soft threshold function is 1 or 0, which can effectively prevent the occurrence of gradient vanishing and gradient explosion;

[0028] Step S43: Since the current and vibration signals are time series signals, a bidirectional gated recurrent unit is used to extract high-level features in the signal; the bidirectional gated recurrent unit is composed of two gated recurrent units in opposite directions, one propagating information in the forward direction and the other propagating information in the reverse direction. When the past data information affects the future data information, the future information is also associated with the past information, which has good prediction and recognition capabilities and can also capture dependencies of various time scales. The expression is as follows:

[0029]

[0030] r t =σ(W r ·[h t-1 ,x t ])

[0031]

[0032] Among them, x t is the input signal, z t is the update gate, r t is the reset gate, w t and v t are the forward and backward state weight matrices, bt is the bias; in a unidirectional gated recurrent unit, the reset layer sets the hidden state h of the previous neuron t-1 With the input x of the current neuron t Combine to get candidate hidden states The update gate controls the hidden state h of the previous neuron t-1 and the candidate hidden state of the current neuron The ratio of the hidden state h t ; The output h of the bidirectional gated recurrent unit t From the front and backward Work together.

[0033] Step S44: Send the output from the bidirectional gated recurrent unit to the fully connected layer and then to the Softmax output layer to obtain the final diagnosis result.

[0034] The present invention has the following beneficial effects:

[0035] 1. The present invention proposes a deep neural network with a reasonable structure and appropriate parameters, called a double-branch residual shrinkage network, which is used for the rotor broken bar fault diagnosis of three-phase squirrel cage asynchronous motors under small samples. This method provides new possibilities for the motor rotor broken bar fault diagnosis under small sample conditions.

[0036] 2. The present invention adopts Hilbert transform, dense connection mechanism, residual shrinkage structure, dual-branch structure feature extraction, and bidirectional gated recurrent unit to filter out noise and harmonic information, thereby improving the robustness of the diagnostic system; solves the problems of gradient disappearance, gradient explosion, overfitting, etc. that may occur in the diagnostic system; and considers the timing characteristics of current signals and vibration signals, and uses bidirectional gated recurrent units to extract fused high-level features.

[0037] 3. The present invention proposes an end-to-end asynchronous motor broken rotor bar fault diagnosis system under small sample conditions. It does not require a large amount of prior knowledge and only requires single-phase current signals and vibration signals to achieve early broken rotor bar fault diagnosis under small sample conditions, effectively achieving the efficiency and accuracy of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 : Block diagram of the broken rotor bar fault diagnosis algorithm for three-phase asynchronous motor.

[0039] Figure 2 : Improved residual shrinkage module.

[0040] Figure 3 : Bidirectional recurrent gating unit, (a) is a unidirectional gated recurrent unit; (b) is a bidirectional recurrent gating unit.

[0041] Figure 4 : Broken rotor bar fault diagnosis result diagram.

[0042] Figure 5 : Confusion matrix for broken rotor bar fault diagnosis.

[0043] Figure 6 : Comparison of the accuracy of different neural network algorithms. DETAILED DESCRIPTION Specific embodiment 1

[0045] Specific embodiment 1, a three-phase asynchronous motor rotor broken bar fault diagnosis method using a double-branch residual contraction network, the algorithm block diagram is as follows Figure 1 Next, the specific implementation of this embodiment will be described in detail with reference to the accompanying drawings.

[0046] Step S1, build a three-phase asynchronous synchronous motor experimental platform. The rotor bars are cut manually, and a groove with a depth of 16 mm is cut perpendicular to the axial direction to simulate the motor rotor bar breakage fault. A total of three different rotor bar breakage fault levels are set.

[0047] Step S2, based on the motor experimental platform, the motor stator current and motor drive end vibration signals under various working conditions are collected through the data acquisition system, including the normal operation of the motor and different fault severity data. When the power supply frequency is constant at 30Hz, the data are collected under three loads of 1Nm, 2Nm, and 3Nm, including healthy, 0.5 broken rotor bars, 1 broken rotor bar, and 3 broken rotor bars. The settings of the training set and the test set are shown in Table 1, with a total of 12 categories, with labels of 0-11.

[0048] Table 1 Experimental dataset

[0049]

[0050]

[0051] Step S3, collating the sample data set of step 2, and dividing the sample data set into two parts: a training set and a test set. In order to make full use of the data and extract more effective features, the data enhancement and normalization processing methods are adopted, and the data enhancement technology of the equally spaced sliding window is used to overlap the original data to obtain the sample data set. 1000 samples are set for each working condition, and the length of each sample is 1024 data points. Integrate the data collected in step 2, and divide the data set into a training set and a test set according to the proportion.

[0052] Step S4, perform Hilbert transform on the current data to highlight the fault characteristic frequency, and use Pytorch to design a neural network model for rotor broken bar fault diagnosis; the above-mentioned dual-branch residual shrinkage network includes an improved residual shrinkage module, a dense connection mechanism, a bidirectional recurrent gating unit and a classification layer; first, the improved residual shrinkage network is used to filter out noise components for current and vibration data, improve the reference value of a small number of available training samples, and extract more effective features; secondly, a dense connection mechanism is used to connect each residual shrinkage module to achieve feature reuse to avoid the residual shrinkage module filtering out effective information; then a bidirectional recurrent gating unit is used to learn hidden information to further extract advanced features; at the same time, regularization technology is introduced: label smoothing regularization, Meta-ACON activation function, weight attenuation, Dropout, to prevent overfitting under small sample fault diagnosis; finally, the results are output through the classification layer.

[0053] Step S41: When a motor has a broken rotor bar fault, a fault characteristic frequency with a frequency of (1±2s)f will appear in the motor stator current spectrum; since the slip rate s and the amplitude of the fault characteristic frequency are very small when the motor is running, the fault feature is not easy to be extracted; the envelope can be obtained by Hilbert transform to filter out the base frequency component, effectively eliminate the masking effect of spectrum leakage, and highlight the fault feature; for a continuous current signal x(t), its Hilbert transform is as follows:

[0054]

[0055] Among them, f is the power supply frequency of the motor power supply; s is the slip rate of the motor; t is the time variable of the function; τ is the time constant; x(τ) is the function form with the time constant as the variable;

[0056] Take the real-valued function x(t) as the real part and transform the transformed function As the imaginary part, the analytical signal z(t) can be expressed as:

[0057]

[0058] The envelope A(t) of a real-valued function x(t) can be obtained by taking the modulus value of the analytical signal as follows:

[0059]

[0060] By performing Hilbert transform on the current signal, the power frequency is converted into a DC component, and the fault characteristic frequency is converted from the original sideband frequency (1±2s)f to 2sf;

[0061] Step S42: The feature extraction process of the current and vibration signals will be affected by the noise in the signal. The interference of the noise may make it impossible to extract effective features, resulting in a decrease in diagnostic performance. The residual shrinkage network introduces the soft thresholding function in the traditional signal denoising algorithm, learns a unique set of thresholds according to the noise content in different samples, and filters out the noise components in the signal; by introducing the dilated convolution and attention mechanism ECA-Net, the traditional residual shrinkage network is improved to enhance the model's receptive field and feature extraction capabilities. The expression is as follows:

[0062]

[0063] Among them, x is the input feature, y is the output feature, z is the output of the two-layer fully connected network of the residual shrinkage module, α is the corresponding scaling parameter, and the scaling parameter is multiplied by the absolute value of x to obtain the threshold τ of the soft threshold function, which can ensure that the threshold size meets the requirements; that is, the noise component between -τ and τ is filtered out; in addition, the derivative of the soft threshold function is 1 or 0, which can effectively prevent the occurrence of gradient vanishing and gradient explosion;

[0064] Step S43: Since the current and vibration signals are time series signals, a bidirectional gated recurrent unit is used to extract high-level features in the signal; the bidirectional gated recurrent unit is composed of two gated recurrent units in opposite directions, one propagating information in the forward direction and the other propagating information in the reverse direction. When the past data information affects the future data information, the future information is also associated with the past information, which has good prediction and recognition capabilities and can also capture dependencies of various time scales. The expression is as follows:

[0065]

[0066] r t =σ(W r ·[h t-1 ,x t ])

[0067]

[0068] Among them, x t is the input signal, z t is the update gate, r t is the reset gate, w t and v t are the forward and backward state weight matrices, b t is the bias; in a unidirectional gated recurrent unit, the reset layer sets the hidden state h of the previous neuron t-1 With the input x of the current neuron t Combine to get candidate hidden states The update gate controls the hidden state h of the previous neuront-1 and the candidate hidden state of the current neuron The ratio of the hidden state h t ; The output h of the bidirectional gated recurrent unit t From the front and backward Work together.

[0069] Step S44: Send the output from the bidirectional gated recurrent unit to the fully connected layer and then to the Softmax output layer to obtain the final diagnosis result. The specific parameters of the proposed network are shown in Table 2.

[0070] Table 2 Network parameter settings

[0071]

[0072] Step S5, use the training set to train the neural network model. During the training process, the cross entropy loss function using label smoothing is introduced to calculate the error between the predicted label and the true label. At the same time, the error is used for back propagation, and the neural network parameters and structure are adjusted. The experimental process is repeated for training, and the training is stopped when the set number of iterations is reached. The model training uses the Adam optimization algorithm, the number of iterations (Epoch) is set to 50 times, the initial value of the learning rate is set to 1e-4, and the batch size (batch size) is set to 32.

[0073] Step S6, use the test set to perform a performance test on fault diagnosis, input the proportionally divided test samples into the network, and the final diagnosis result can be obtained. The test accuracy of different training set proportions (α) is shown in Table 3. The training accuracy curve and test accuracy curve when the proportion is 0.2 are shown in Table 3. Figure 4 As shown in the figure, when the proportion is the smallest, the test accuracy can still reach 97.65%. When the proportion increases, the test accuracy can reach more than 99%.

[0074] In order to make the diagnosis more intuitive, the confusion matrix is ​​used to visualize the classification of each type of fault, such as Figure 5 The figure shows the confusion matrix with a training set ratio of 0.2, which shows the correspondence between the model's predicted labels and the true labels on different categories. The horizontal axis represents the predicted labels, the vertical axis represents the actual labels, and the values ​​on the diagonal are the accuracy ratios of each class. In the confusion matrix, the fault classification accuracy under the same load conditions reaches 100%, and only the 12th type of fault has a classification error, and the overall accuracy is significantly higher.

[0075] Figure 6The figure shows the performance comparison of rotor broken bar fault diagnosis under different algorithms. Four methods including dual-branch convolutional neural network (DCNN), attention mechanism-bidirectional recurrent gating unit (DCA-BiGRU), synthetic minority oversampling technology (SMOTE) and generative adversarial network (GAN) are selected for experimental comparison. Figure 6 It can be seen that in the case of small samples, the algorithm proposed by the present invention has a higher accuracy rate. The rotor broken bar fault diagnosis method under small samples provided by the present invention can effectively filter out the noise components in the current and vibration signals, extract the effective features in the signals, and finally realize the fault diagnosis classification.

[0076] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "illustrative embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0077] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A small sample fault diagnosis method for a three-phase squirrel cage asynchronous motor rotor bar breakage, characterized in that: The steps include: Step S1, building a three-phase squirrel cage asynchronous motor experimental platform; Step S2, based on the motor experimental platform, collecting the motor stator current and motor drive end vibration signal under various working conditions through the data acquisition system, including the normal operation data of the motor and data of different fault severity; Step S3, collecting the data collected in step 2, and dividing the data set into a training set and a test set according to the ratio; Step S4, performing Hilbert transform on the current data to highlight the fault characteristic frequency, and using Pytorch to design a neural network model for rotor bar fault diagnosis; the above-mentioned double-branch residual contraction network includes an improved residual contraction module, a dense connection mechanism, a bidirectional recurrent gating unit BiGRU and a classification layer; firstly, the improved residual contraction network is used to filter out noise components from the current and vibration data, improve the reference value of a small number of available training samples, and extract more effective features; Secondly, a dense connection mechanism is used to connect each residual shrinkage module to achieve feature reuse to avoid the residual shrinkage module filtering out effective information; then a bidirectional recurrent gating unit is used to learn hidden information to further extract advanced features; at the same time, regularization techniques are introduced: label smoothing regularization, dynamic learnable activation function Meta-ACON, weight decay, random inactivation Dropout, to prevent overfitting in small sample fault diagnosis; finally, the results are output through the classification layer; Step S5, using the training set to train the neural network model, introducing a label smoothing cross entropy loss function to calculate the error between the predicted label and the true label during the training process, and using the error to perform back propagation, adjusting the neural network parameters and structure, and repeatedly performing the experimental process for training, and stopping the training when the set number of iterations is reached; Step S6, using the test set to test the trained model, inputting the test sample into the trained network model, and obtaining the accuracy of the diagnosis result; If the accuracy of the test sample reaches more than 95% in the case of small samples, the actual diagnosis requirements are met and the small sample broken rotor bar fault diagnosis of the three-phase squirrel cage asynchronous motor is realized.

2. The method according to claim 1, characterized in that When collecting the motor current and vibration data in step 2, in order to simulate faults of different severity, 12 operating conditions of 1Nm, 2Nm and 3Nm loads are collected when the motor is normal, half a rotor bar is broken, one rotor bar is broken, and three rotor bars are broken.

3. The method according to claim 1, characterized in that: When dividing the data set in step 3, in order to make full use of the data and extract more effective features, data enhancement and normalization processing are used; this method uses the data enhancement method of equally spaced sliding windows to perform overlapping sampling on the original data to obtain a sample data set; 1000 samples are set for each working condition, and α is set to be the proportion of the training set, and α is set to 0.1 to 0.5 for experiments, and the sample length is 1024 data points.

4. The method according to claim 1, characterized in that: After the rotor bar failure occurs in step 4, the motor stator current is collected and there are noise and harmonic components, and the fault characteristics are not obvious, so the Hilbert algorithm is used to enhance the fault characteristics, suppress noise and harmonics, make full use of the extracted features, improve the learning ability of the model, and facilitate diagnosis; the specific steps are as follows: Step S41: When a motor has a broken rotor bar fault, a fault characteristic frequency with a frequency of (1±2s)f will appear in the motor stator current spectrum; since the slip rate s and the amplitude of the fault characteristic frequency are very small when the motor is running, the fault feature is not easy to be extracted; the envelope can be obtained by Hilbert transform to filter out the base frequency component, effectively eliminate the masking effect of spectrum leakage, and highlight the fault feature; for a continuous current signal x(t), its Hilbert transform is as follows: Among them, f is the power supply frequency of the motor power supply; s is the slip rate of the motor; t is the time variable of the function; τ is the time constant; x(τ) is the function form with the time constant as the variable; Take the real-valued function x(t) as the real part and transform the transformed function As the imaginary part, the analytical signal z(t) can be expressed as: The envelope A(t) of a real-valued function x(t) can be obtained by taking the modulus value of the analytical signal as follows: By performing Hilbert transform on the current signal, the power frequency is converted into a DC component, and the fault characteristic frequency is converted from the original sideband frequency (1±2s)f to 2sf; Step S42: The feature extraction process of the current and vibration signals will be affected by the noise in the signal. The interference of the noise may make it impossible to extract effective features, resulting in a decrease in diagnostic performance. The residual shrinkage network introduces the soft thresholding function in the traditional signal denoising algorithm, learns a unique set of thresholds according to the noise content in different samples, and filters out the noise components in the signal; by introducing the dilated convolution and attention mechanism ECA-Net, the traditional residual shrinkage network is improved to enhance the model's receptive field and feature extraction capabilities. The expression is as follows: Among them, average is the average pooling layer, i, j, c are the width, height, and number of channels of feature map x, respectively, x is the input feature, y is the output feature, z is the output of the two-layer fully connected network of the residual shrinkage module, α is the corresponding scaling parameter, and the scaling parameter is multiplied by the absolute value of x to obtain the threshold τ of the soft threshold function, which can ensure that the threshold size meets the requirements; that is, the noise component between -τ and τ is filtered out; in addition, the derivative of the soft threshold function is 1 or 0, which can effectively prevent the occurrence of gradient vanishing and gradient explosion; Step S43: Since the current and vibration signals are time series signals, a bidirectional gated recurrent unit is used to extract high-level features in the signal; the bidirectional gated recurrent unit is composed of two gated recurrent units in opposite directions, one propagating information in the forward direction and the other propagating information in the reverse direction. When the past data information affects the future data information, the future information is also associated with the past information, which has good prediction and recognition capabilities and can also capture dependencies of various time scales. The expression is as follows: r t =σ(W r ·[h t-1 ,x t ]) Where σ is the sigmoid function, W r ,W is the weight matrix, GRU is a unidirectional gated recurrent unit, x t is the input signal, z t is the update gate, r t is the reset gate, w t and v t are the forward and backward state weight matrices, b t is the bias; in a unidirectional gated recurrent unit, the reset layer sets the hidden state h of the previous neuron t-1 With the input x of the current neuron t Combine to get candidate hidden states The update gate controls the hidden state h of the previous neuron t-1 and the candidate hidden state of the current neuron The ratio of the hidden state h t ; The output h of the bidirectional gated recurrent unit t From the front and backward Work together. Step S44: Send the output from the bidirectional gated recurrent unit to the fully connected layer and then to the Softmax output layer to obtain the final diagnosis result.

Images