WGS-CNN-based rotating machine fault diagnosis method, device and system, and storage medium

Abstract

The invention discloses a rotating machine fault diagnosis method, device and system based on WGS-CNN, and a storage medium, and belongs to the technical field of rotating machine fault diagnosis, and the method comprises the steps: obtaining a fault vibration signal of a rotating machine; inputting the fault vibration signal into the trained WGS-CNN model to obtain a fault diagnosis classification result of the rotating machine; the WGS-CNN model is obtained by improving an ELCNN model, and the improvement comprises the steps of initializing all convolution kernels through a wavelet primary function to obtain a wavelet initialization convolution kernel, introducing a Gaussian window into the wavelet initialization convolution kernel, and then replacing an original activation function with a square activation function. Through wavelet initialization, a Gaussian window and a square activation function, the model has better diagnosis performance and generalization ability under the conditions of strong noise, strong time variation and multi-working-condition migration, and the diagnosis accuracy is improved.

Description

Technical Field

[0001] The present invention relates to a method, device, storage medium and system for rotating machinery fault diagnosis based on WGS-CNN, belonging to the technical field of rotating machinery fault diagnosis. Background Technique

[0002] The one-dimensional convolutional neural network (1D CNN) for rotating machinery vibration signal fault diagnosis has powerful local abstraction ability and one-dimensional vibration signal adaptive learning ability, and performs excellently in extracting local features and classification performance. Existing 1D CNN fault diagnosis models usually use normal distribution to randomly initialize the convolution kernels, but there are problems such as limited time-frequency resolution, insufficient localization characteristics and weak anti-noise performance, which cannot meet the fault diagnosis requirements under complex working conditions, and the fault diagnosis accuracy is low. Summary of the Invention

[0003] The purpose of the present invention is to provide a method, device, storage medium and system for rotating machinery fault diagnosis based on WGS-CNN, so as to solve the problem of low fault diagnosis accuracy in the prior art under strong noise, strong time-variation and complex working conditions.

[0004] To achieve the above purpose, the present invention is implemented by the following technical solutions:

[0005] In the first aspect, the present invention provides a method for rotating machinery fault diagnosis based on WGS-CNN, including:

[0006] Obtain the fault vibration signal of the rotating machinery;

[0007] Input the fault vibration signal into the trained WGS-CNN model to obtain the fault diagnosis classification result of the rotating machinery;

[0008] The WGS-CNN model is obtained by improving the ELCNN model. The improvement includes: initializing all convolution kernels in the ELCNN model with the basis function of wavelet to obtain wavelet initialization convolution kernels, introducing Gaussian windows into the wavelet initialization convolution kernels, and then replacing the original activation function with a square activation function.

[0009] Furthermore, the initialization of all convolution kernels in the ELCNN model with the basis function of wavelet to obtain wavelet initialization convolution kernels is carried out through the following formula:

[0010] (1);

[0011] (2);

[0012] (3);

[0013] Among them, is the Morlet mother wavelet with the independent variable t, where t is the time variable, is the first scale parameter, is the sine carrier wave, is a Gaussian function, e is the natural constant, and a is the second scale parameter, is the Morlet wavelet basis function with the independent variable t, is the wavelet initialization convolution kernel, where j is the imaginary unit, is the angular frequency, is The Morlet wavelet is obtained by transforming through the second scale parameter a.

[0014] Further, a Gaussian window is introduced into the wavelet initialization convolution kernel through the following formula:

[0015] (4);

[0016] (5);

[0017] Among them, is the Gaussian window, x is the independent variable of the Gaussian window, and e is the natural constant, is the third scale parameter, is the windowed wavelet initialization convolution kernel obtained after introducing the Gaussian window into the wavelet initialization convolution kernel, is the wavelet initialization convolution kernel.

[0018] Further, the third scale parameter is obtained through the following method:

[0019] (6);

[0020] (7);

[0021] (8);

[0022] (9);

[0023] Among them, is the gradient of the th convolution kernel, L is the cross-entropy loss function of the convolutional neural network, is the wavelet basis function used to initialize the th convolution kernel, is the th gradient magnitude of the convolution kernel, is the The m-th value of the gradient of a convolution kernel, where k is the maximum value of m, is the average value of the gradient magnitudes of the convolution kernels, and n is the number of convolution kernels, is the updated third-scale parameter of the Gaussian window in the th wavelet-initialized convolution kernel introduced, is the third-scale parameter of the Gaussian window in the th wavelet-initialized convolution kernel before update, and α is the set adjustment coefficient.

[0024] Furthermore, the expression of the square activation function is:

[0025] (10);

[0026] where, is the square activation function, is the windowed wavelet-initialized convolution kernel, * is the convolution operation, × is the multiplication operation, and q is the input signal of the square activation function.

[0027] Furthermore, the WGS-CNN model is trained by the following method:

[0028] Obtain a time-invariant dataset and a time-varying dataset containing fault vibration signals;

[0029] Based on the time-invariant dataset and the time-varying dataset containing fault vibration signals, train the WGS-CNN model through gradient descent and backpropagation until the F1 score of the WGS-CNN model does not increase for several consecutive times.

[0030] In a second aspect, the present invention provides a rotating machinery fault diagnosis device based on WGS-CNN, including:

[0031] A fault vibration signal acquisition module, configured to: acquire the fault vibration signal of the rotating machinery;

[0032] A fault diagnosis module, configured to: input the fault vibration signal into the trained WGS-CNN model to obtain the fault diagnosis classification result of the rotating machinery;

[0033] wherein, the WGS-CNN model is obtained by improving the ELCNN model, and the improvement includes: initializing all convolution kernels in the ELCNN model through the basis function of the wavelet to obtain wavelet-initialized convolution kernels, introducing a Gaussian window in the wavelet-initialized convolution kernels, and then replacing the original activation function with a square activation function.

[0034] In a third aspect, the present invention provides a computer-readable storage medium, on which a computer program / instructions are stored. When the computer program / instructions are executed by a processor, the steps of the method for diagnosing faults in rotating machinery based on WGS-CNN described in any one of the first aspects are implemented.

[0035] In a fourth aspect, the present invention provides a computer system, comprising:

[0036] a memory for storing computer programs / instructions;

[0037] a processor for executing the computer programs / instructions to implement the steps of the method for diagnosing faults in rotating machinery based on WGS-CNN described in any one of the first aspects.

[0038] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0039] A method, device, storage medium and system for diagnosing faults in rotating machinery based on WGS-CNN provided by the present invention improves the ELCNN convolution kernel by introducing wavelet convolution kernels, dynamically modulated Gaussian windows and square functions. Through the above improvements, the provided WGS-CNN model can meet the fault diagnosis requirements of strong noise, strong time-variation and multi-condition transfer learning. The strategy of initializing the convolution kernel with wavelets effectively solves the problem of rapid localization of time-varying faults by the synergistic effect of time-frequency localization prior injection and multi-scale feature extraction, and can obtain fault features changing with time under strong time-varying conditions; the Gaussian window is a dynamic adjustment of the convolution kernel, and the dynamic adjustment mechanism can adaptively balance stability and accuracy at different stages of training, such as the rapid learning period and the late convergence period, so that the model can avoid violent fluctuations and can be finely adjusted when approaching the optimal solution, enhancing the fault prediction ability for unknown conditions and learning fault features of unknown conditions under multi-condition transfer conditions; the features extracted by the trained convolution layer tend to be frequency-domain features, and the square function can strengthen the main peak frequency of the signal while weakening small-amplitude noise, thereby improving the robustness of the model; in summary, the above improvements greatly enhance the fault diagnosis ability of rotating machinery under strong noise, strong time-variation and complex conditions. Description of the Drawings

[0040] Figure 1 is a network structure diagram of the ELCNN model with a traditional randomly initialized convolution kernel provided by the present invention;

[0041] Figure 2 is a network structure diagram of the WGS-CNN model provided by the present invention;

[0042] Figure 3 is a waveform schematic diagram of a randomly normally distributed initialized convolution kernel and a wavelet-initialized convolution kernel provided by the present invention;

[0043] Figure 4 Schematic diagram of the Gaussian window function for adaptive learning provided by the present invention, where (a), (b), (c), and (d) are Gaussian window functions with variances of 2.14, 4.81, 4.56, and 6.34 respectively;

[0044] Figure 5 Schematic diagram of the data preprocessing process provided by the present invention;

[0045] Figure 6 Schematic diagram for comparing the fault diagnosis performance (F1) of different models provided by the present invention on the CWRU dataset (A0→A0), where (b) is the enlarged local part of (a);

[0046] Figure 7 Schematic diagram for comparing the fault diagnosis performance (F1) of different models provided by the present invention on the CWRU dataset (A1→A1), where (b) is the enlarged local part of (a);

[0047] Figure 8 Schematic diagram for comparing the fault diagnosis performance (F1) of different models provided by the present invention on the CWRU dataset (A2→A2), where (b) is the enlarged local part of (a);

[0048] Figure 9 Schematic diagram of the frequency domain characteristics of A0-IR learned by different models provided by the present invention on the CWRU dataset (A0→A0), where (a) to (h) respectively represent the frequency domain characteristics corresponding to different models;

[0049] Figure 10 Schematic diagram for comparing the fault diagnosis performance (F1) of different models provided by the present invention under the Ottawa dataset with SNR=-15;

[0050] Figure 11 Schematic diagram of the time-frequency domain characteristics of Ottawa-B0-IR learned by different models provided by the present invention on the Ottawa dataset (B0→B0);

[0051] Figure 12 Schematic diagram for comparing the fault diagnosis performance (F1) of different models provided by the present invention in the working condition migration task and on the Ottawa dataset with SNR=-15;

[0052] Figure 13 Schematic diagram of the time-frequency domain characteristics of B0-OR learned by different models provided by the present invention on the Ottawa dataset (B0→B1, B2, B3);

[0053] Figure 14It is a schematic diagram of the time-frequency domain features of B1-OR learned by different models provided by the present invention on the Ottawa dataset (B0→B1, B2, B3). Detailed implementation manners

[0054] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0055] Embodiment 1

[0056] The present invention provides a method for diagnosing faults of rotating machinery based on WGS-CNN, including:

[0057] Obtaining the fault vibration signal of the rotating machinery;

[0058] Inputting the fault vibration signal into the trained WGS-CNN model to obtain the fault diagnosis classification result of the rotating machinery;

[0059] Among them, the WGS-CNN model is obtained by improving the ELCNN model. The improvement includes: initializing all the convolutional kernels in the ELCNN model through the wavelet basis function to obtain wavelet-initialized convolutional kernels, introducing a Gaussian window into the wavelet-initialized convolutional kernels, and then replacing the original activation function with a square activation function.

[0060] The present invention initializes the convolutional kernels through the wavelet basis function, endows the model with strong feature extraction ability, enhances the multi-resolution characteristics, enables the convolutional kernels to focus on the fault center frequency according to the scientific system gradient by introducing a Gaussian window, and improves its fault prediction ability. The square activation function strengthens the main frequency features beneficial to classification, and effectively suppresses the low-amplitude noise interference. Through the above improvements, the provided WGS-CNN model can suppress noise and extract effective fault frequencies under complex working conditions. The strategy of initializing the convolutional kernels with the wavelet basis function effectively solves the problem of rapid localization of time-varying faults by the synergistic effect of time-frequency localization prior injection and multi-scale feature extraction; under strong time-varying conditions, it can obtain fault features that change with time. The Gaussian window is a dynamic adjustment of the convolutional kernels. The dynamic adjustment mechanism can adaptively balance stability and accuracy at different stages of training, such as the rapid learning period and the late convergence period, enabling the model to avoid violent fluctuations and make refined adjustments when approaching the optimal solution, enhancing the fault prediction ability for unknown working conditions, and learning the fault features of unknown working conditions under multi-working-condition migration conditions; the features extracted by the convolutional layer after training tend to be frequency-domain features, and the square function can strengthen the main peak frequency of the signal and weaken the small-amplitude noise, thereby improving the robustness of the model. In summary, the above improvements greatly enhance the fault diagnosis ability of rotating machinery under strong noise, strong time-variation and complex working conditions.

[0061] Example 2

[0062] This embodiment provides a rotating machinery fault diagnosis method based on WGS-CNN, including:

[0063] Obtain the fault vibration signal of the rotating machinery;

[0064] Input the fault vibration signal of the rotating machinery into the trained WGS-CNN model to obtain the fault diagnosis classification result of the rotating machinery;

[0065] Among them, the WGS-CNN model is obtained by improving the ELCNN (Explainable Light weight 1D CNN, an improved lightweight and explainable convolutional neural network for vibration signals) model.

[0066] Figure 1 is the existing ELCNN network structure diagram, including an input layer, 1 convolutional layer, 1 square-global average pooling layer (Global Average Pooling, GAP), and 1 fully connected layer.

[0067] Figure 2 is the structural schematic diagram of the WGS-CNN model improved in this embodiment, and its network parameters are shown in Table 1:

[0068] Table 1 - Network parameter configuration table of the WGS-CNN model using different datasets (Case Western Reserve or Ottawa)

[0069]

[0070] The convolution kernel size and filter number of the WGS-CNN model are preferably obtained according to the time characteristics of the signal. When using the non-time-varying CWRU dataset (Case Western Reserve University dataset), the full English name of CWRU is Case Western Reserve University, the convolution kernel size of the WGS-CNN model is 32, and the filter number is also 32; when using the time-varying Ottawa dataset (Ottawa dataset), the convolution kernel size of the WGS-CNN model is 128, and the filter number is 8. The sample length of the WGS-CNN model is preferably obtained according to the frequency resolution of the signal. The sample length of the CWRU dataset is 2048, while the sample length of the Ottawa dataset is 32768.

[0071] The process of rotating machinery fault diagnosis based on the WGS-CNN model is the same as that based on the ELCNN model. The fault vibration signal is first preprocessed, then local features are extracted by the convolutional layer, dimensionality reduction is performed by the pooling layer, and finally the classification task is completed by the fully connected layer. The difference lies in the improvement strategy of the convolutional layer in the WGS-CNN model, which mainly includes using the wavelet basis function to initialize the convolutional kernel, loading an adaptive learning Gaussian window for the convolutional kernel, and finally using the square function to replace the traditional activation function.

[0072] Among them, using the wavelet basis function to initialize the convolutional kernel is a static assignment to the convolutional kernel, loading an adaptive learning Gaussian window for the convolutional kernel is a dynamic adjustment to the convolutional kernel, and the square function performs nonlinear processing on the features extracted by the convolutional kernel. The detailed analysis of the above three improvement strategies is as follows:

[0073] First, the convolutional kernel is initialized using the multi-scale Morlet wavelet, and the convolutional kernel in the ELCNN model is improved to a wavelet-initialized convolutional kernel, endowing the model with strong feature extraction ability and enhancing the multi-resolution characteristics. Figure 3 For the waveforms of the convolutional kernel initialized with random normal distribution and the convolutional kernel initialized with Morlet wavelet. The WGS-CNN model uses Figure 2 the Morlet wavelet shown in (b) in Figure 2 to replace the random normal distribution convolutional kernel shown in (a) in

[0074] (1);

[0075] (2);

[0076] (3);

[0077] Among them, is the Morlet mother wavelet with independent variable t, t is the time variable, is the first scale parameter, which is used to control the width of the Gaussian envelope of the Morlet mother wavelet, thereby affecting the localization characteristics of the wavelet, is the sine carrier wave, is a Gaussian function, e is the natural constant, a is the second scale parameter, which is used to control the width of the Morlet wavelet to change the expansion and compression degree of the Morlet wavelet, is the Morlet wavelet basis function with independent variable t, is the wavelet-initialized convolutional kernel in the WGS-CNN model, j is the imaginary unit, which is used to construct the oscillatory part of the wavelet so that the wavelet can represent a complex sine wave, is the angular frequency, which determines the center frequency of the wavelet in the frequency domain. is the Morlet wavelet after being transformed by the second scale parameter a.

[0078] The Morlet mother wavelet is calculated using formula (1), and the Morlet mother wavelet obtained from formula (1) is fed into formula (2) to obtain Figure 2 a series of Morlet wavelet basis functions of different scales shown in (b) in, and the Morlet wavelet basis functions calculated from formula (2) are fed into formula (3) to initialize all the convolutional kernels of the WGS-CNN model, obtaining the improved wavelet initialization convolutional kernels.

[0079] Subsequently, an adaptive Gaussian window is introduced so that the convolutional kernel can focus on the fault center frequency according to the learnable gradient, effectively suppressing noise interference. Figure 4 is the Gaussian window of different scales for adaptive learning, where sigma is the third scale parameter. Figure 4 The Gaussian window shown in (a) in has a narrow width and can highly focus on local details, which is suitable for extracting fine local features; while the Gaussian window shown in (d) in Figure 4 has a wide width and pays more attention to the overall smoothing effect, which is beneficial to removing noise interference.

[0080] The method and formula for introducing the Gaussian window in the convolutional layer are as follows:

[0081] (4);

[0082] (5);

[0083] Among them, is the Gaussian window, x is the independent variable of the Gaussian window, and e is the natural constant. is the third scale parameter, which is used to control the width of the Gaussian window. is the windowed wavelet initialization convolutional kernel, which is obtained by introducing the Gaussian window on the basis of the wavelet initialization convolutional kernel. is the wavelet initialization convolutional kernel in the WGS-CNN model.

[0084] Gaussian window functions of different scales are generated through formula (4), and the Gaussian window functions generated by formula (4) are fed into formula (5) for dot product operation to obtain the windowed wavelet convolutional kernel .

[0085] In this embodiment, the third scale parameter is adaptively learned through the backpropagation of the WGS-CNN model, and its learning process is as follows:

[0086] (6);

[0087] (7);

[0088] (8);

[0089] (9);

[0090] Wherein, is the gradient of the th convolution kernel, L is the cross-entropy loss function of the convolutional neural network, is the wavelet basis function for initializing the th convolution kernel, is the gradient magnitude of the th convolution kernel, is the m-th digit value of the gradient of the th convolution kernel, k is the total number of digits of the gradient of the th convolution kernel, that is, the maximum value of m, is the average value of the gradient magnitudes of the convolution kernels, n is the number of convolution kernels, is the third scale parameter of the Gaussian window in the th wavelet-initialized convolution kernel after update, is the third scale parameter of the Gaussian window in the th wavelet-initialized convolution kernel before update, α is the set adjustment coefficient, which is 0.01 and is used to control the adjustment amplitude.

[0091] Calculate the gradient of the th convolution kernel using formula (6), Based on , calculate the gradient magnitude of the i-th convolution kernel using formula (7), Based on , calculate the average value of the gradient magnitudes of the convolution kernels using formula (8), Based on , update the third scale parameter according to formula (9). Formula (9) updates the third scale parameter in a segmented manner according to the magnitude of the gradient magnitude of the convolution kernel. When , that is, when the gradient magnitude of the current convolution kernel is greater than the average gradient magnitude, increase to obtain ; when

[0092] Finally, the square activation function is adopted to suppress low-amplitude frequency components in the full frequency band, strengthening the main frequency characteristics beneficial to classification. The expression of the square activation function is as follows:

[0093] (10);

[0094] where, is the square activation function, is the windowed wavelet initialization convolution kernel, * is the convolution operation, × is the multiplication operation, and q is the input signal of the square activation function.

[0095] The training steps of the WGS-CNN model are as follows:

[0096] This embodiment is carried out using the CWRU non-time-varying bearing dataset (referred to as the CWRU dataset or the Westinghouse dataset) and the Ottawa time-varying dataset (referred to as the Ottawa dataset or the Ottawa dataset).

[0097] Step 1, as shown, this embodiment performs preprocessing such as dataset partitioning, sequential sampling, normalization, and adding noise on the original signals in the CWRU dataset.

[0098] Step 1-1, the original signals are divided into different subsets according to the working conditions, and one of the subsets is selected as the training set, and the other subsets are used as the test sets; as shown in Table 2, the CWRU dataset mainly collects vibration signals of rolling bearings under different working conditions at 3 rotational speeds and 3 loads. According to different working conditions, the dataset is divided into 3 subsets, namely A0, A1, and A2. Each subset contains 4 bearing states, namely the normal state (Normal, referred to as N in this embodiment), the inner-race fault (Inner-race fault, referred to as IR in this embodiment), the outer-race fault (Outer-race fault, referred to as OR in this embodiment), and the balls fault (Balls fault, referred to as B in this embodiment). The sampling quantity of each signal is 200, and the sampling frequency is 12KHz.

[0099] Table 2 - Westinghouse Dataset Sample Table

[0100]

[0101] As shown in Table 3, the Ottawa dataset mainly collects vibration signals of rolling bearings in different states under 4 time-varying working conditions, namely rising, falling, rising first and then falling, and falling first and then rising. According to different working conditions, the dataset is divided into 4 subsets, namely B0, B1, B2, and B3. Each subset covers 3 bearing health states, namely N, IR, and OR. The sampling quantity of each signal is 400, and the sampling frequency is 200KHz.

[0102] Table 3 - Ottawa Dataset Sample Table

[0103]

[0104] Step 1-2, obtain a clean sample sequence by using an equally-spaced sequential sampling method.

[0105] By using the equally-spaced sequential sampling method, clean sample sequences that meet the frequency resolution are obtained from the CWRU dataset and the Ottawa dataset respectively. Meeting the frequency resolution means the time record length required to analyze the lowest fault frequency in the signal. The lowest fault frequency of the CWRU dataset is generally the cage fault frequency of 11.93Hz, and the signal length occupied by the cage fault is 1006 (12000÷11.93≈1006, 1006×2 = 2012, where 12kHz is the sampling frequency of the CWRU dataset). The lowest rotation frequency of the Ottawa dataset is 12.5Hz, and the sampling frequency is 200KHz. From this, its signal length is deduced to be 32000, and the deduction process is as follows: 200000÷12.5≈16000, 16000×2 = 32000. To ensure the integrity of the data period and avoid edge effects, in this embodiment, the sampling length of the CWRU dataset is 2048, and the sampling length of the Ottawa dataset is 1024×32 = 32768 to ensure that the signal samples have more than 2 lowest rotation frequency periods.

[0106] Step 1-3, perform normalization processing on the obtained sample signals; use the normalization operation to normalize the amplitude of the samples obtained in Step 1-2 to the interval (-1, 1) to obtain the normalized samples for training. Normalization processing facilitates the comparison of embodiments of different datasets and different models.

[0107] Step 1-4, set the sample class labels, and add additive Gaussian white noise proportional to the sample signals to the data obtained in Step 1-2 to obtain the noisy training set and test set; among them, the signal-to-noise ratios of the noisy samples in the CWRU dataset are -10dB, -5dB, 0dB, 5dB, and 10dB respectively, and the signal-to-noise ratios of the noisy samples in the Ottawa dataset are -15dB, -5dB, 0dB, 5dB, and 15dB respectively.

[0108] Step 2: Use the training set to train the WGS-CNN model, optimize the model parameters through gradient descent and backpropagation, and trigger the early stopping mechanism according to the convergence of the model.

[0109] Input the noisy training set obtained in Steps 1-4 into the WGS-CNN model, train the model through gradient descent and backpropagation, predict the classification results, and monitor the convergence of the WGS-CNN model. When the classification effect has not improved for 10 consecutive times, trigger the early stopping mechanism to finally obtain a WGS-CNN model with better classification effect.

[0110] Step 3: Use the noisy test set to test the WGS-CNN model.

[0111] Input the noisy test set obtained in Steps 1-4 into the WGS-CNN model to output the sample classification results and the signal features learned by the model.

[0112] To better illustrate the advantages of the WGS-CNN model of the present invention, based on the CWRU dataset and the Ottawa dataset, this embodiment conducts a performance comparison of six types of models, namely CNN (Convolutional Neural Networks, traditional single-layer convolutional neural network), W-CNN (Wavelet Convolutional Neural Networks, convolutional neural network using wavelet to initialize the convolutional kernel, called: wavelet-convolutional neural network), WKN (Wavelet Kernel Net, neural network using parameter-learnable Morlet wavelet to replace the traditional convolution, called: wavelet kernel convolutional neural network), WD-CNN (Deep Convolutional Neural Networks with Wide First-layer Kernels, multi-layer wide convolutional neural network), DCA-BiGRU (Dualpath convolution with attention mechanism and Bidirectional Gated Recurrent Unit, dual-path convolution and bidirectional gated recurrent unit), and the WGS-CNN model (the WGS-CNN model improved based on wavelet, Gaussian window, and square activation function provided in this embodiment). The architectures and parameter configurations of these six types of models are shown in Table 4.

[0113] Table 4 - Architectures and parameter configurations of the models for comparison

[0114]

[0115] To accurately evaluate the model performance, this embodiment introduces accuracy (Accuracy, Acc), precision (Precision, Pre), recall (Recall, Rec), and F1 score (F1 Score, F1) to evaluate the classification results. To comprehensively evaluate the advancement of the WGS-CNN model under strong noise, strong time-variation, and multi-condition migration, this embodiment conducts a comparative analysis of the diagnostic performance of the above six types of improved CNNs based on the CWRU dataset and the Ottawa dataset.

[0116] Evaluation of fault diagnosis performance under strong noise conditions:

[0117] Figure 6 It is a schematic diagram for comparing the diagnostic performance of six types of models under strong noise conditions. Figure 6 The left part is a schematic diagram of F1 comparison of six types of models on the CWRU dataset (A0→A0). A0→A0 means that the training set is A0, the test set is A0, and → means inputting into the model sequentially from the training set to the test set. Figure 6 The right part is an enlarged schematic diagram of the framed area in the left part.

[0118] From Figure 6 the left part, it can be seen that the F1 of the WGS-CNN model provided in this embodiment is higher than that of the other five types of models at five signal-to-noise ratios such as 10dB and 5dB, indicating that the anti-noise performance of the WGS-CNN model is better than that of the other five types of models at five signal-to-noise ratios such as 10dB and 5dB. Especially from Figure 6 the right part, it can be seen that at a signal-to-noise ratio of -10dB, the F1 score of the WGS-CNN model is 92.1%, which is much higher than other models such as W-CNN and CNN. This comparative experiment fully demonstrates that the WGS-CNN model provided in this embodiment has strong anti-noise performance and excellent fault diagnosis performance under strong noise conditions.

[0119] Consistent with the anti-noise performance experiment based on the CWRU dataset (A0→A0), the WGS-CNN model also achieves high fault diagnosis performance in the Figure 7 CWRU dataset (A1→A1) shown in Figure 8 and the

[0120] CWRU dataset (A2→A2) shown in Figure 9The frequency-domain features of A0-IR learned by different models on the CWRU dataset (A0→A0), where A0-IR represents the inner-race fault in subset A0, are analyzed as follows:

[0121] Figure 9 As shown in (a), it is the clean A0-IR signal (i.e., Figure 9 the curve corresponding to Clean Raw in, and CleanRaw is the clean A0-IR signal). Its rotational frequency is 29.1 Hz, 58.5 Hz is twice the rotational frequency, 158.2 Hz is the IR fault frequency, 99.6 Hz is the sideband of the predicted IR fault, while the sideband of the actual IR fault is 99.7 Hz, which is calculated by the following method: 158.2 Hz - 58.5 Hz = 99.7 Hz.

[0122] Figure 9 As shown in (b), it is the A0-IR signal with SNR = -10. SNR is the signal-to-noise ratio. The fault frequency of 158.2 Hz is submerged in the noise, and the twice rotational frequency of 58.5 Hz and the sideband of 99.6 Hz are also interfered by the surrounding noise. The noise is Figure 9 the curve corresponding to Noise Raw in, and Noise Raw is the noise signal.

[0123] Figure 9 As shown in (c), it is the frequency-domain features of A0-IR learned by the traditional CNN. Figure 9 As shown in (d), it is the frequency-domain features of A0-IR learned by W-CNN. From Figure 9 part (c), it can be seen that CNN learned the IR fault frequency of 158.2 Hz, but did not learn the twice rotational frequency of 58.5 Hz and the sideband of 99.6 Hz. W-CNN learned the IR fault frequency of 158.2 Hz and the sideband of 99.6 Hz, but did not accurately learn the twice rotational frequency of 58.5 Hz. CNN and W-CNN are still interfered by noise, especially in the case of frequencies greater than 400 Hz.

[0124] Figure 9 As shown in (d), (e), and (f) respectively, they are the frequency-domain features of A0-IR learned by the WKN, WD-CNN, and DCA-BiGRU models. From Figure 9 (d), (e), and (f), it can be seen that they are interfered by strong noise and did not learn useful fault frequencies.

[0125] Figure 9As shown in (h), the frequency-domain features of A0-IR learned by the WGS-CNN model provided in this embodiment are presented. The WGS-CNN model has a good learning effect on the fault frequencies of 158.2 Hz, twice the rotation frequency of 58.5 Hz, and the sideband frequency of 99.6 Hz, and is less affected by noise, especially in the case of frequencies greater than 400 Hz.

[0126] In summary, under strong noise conditions, 5 types of models such as CNN and W-CNN cannot accurately learn the fault features of A0-IR; while the WGS-CNN model provided in this embodiment can not only learn the fault frequencies beneficial to classification, but also has strong anti-interference ability and strong interpretability.

[0127] Fault diagnosis performance evaluation under strong time-varying conditions:

[0128] Different from the time-invariant CWRU dataset, the Ottawa dataset is a time-varying speed bearing dataset. The fault frequencies of the Ottawa dataset change with time, making feature extraction more difficult. The Ottawa dataset contains 4 subsets: B0, B1, B2, and B3. Figure 10 For the comparison of the fault diagnosis performance of different models under the Ottawa dataset with SNR = -15, the specific analysis is as follows:

[0129] The WGS-CNN model has better diagnostic performance. When both the training set and the test set of the model use B0, the F1 score of the WGS-CNN model is 97.79%, which is much higher than that of other models such as CNN (62.02%) and W-CNN (64.71%). Similar to B0, the same conclusion can be drawn from subsets B1, B2, and B3, that is, the F1 of the WGS-CNN model is better than the other 5 types of improved CNN models.

[0130] The evaluation index concentration of the WGS-CNN model is better. Whether in subset B0, B1, B2, or B3, the dispersion of the F1 score of the WGS-CNN model is small, while the F1 scores of other models are more dispersed. This shows that the WGS-CNN model has strong anti-noise performance, while other models are easily affected by noise, increasing the uncertainty of fault diagnosis.

[0131] To analyze the interpretability of the features extracted by various models under strong noise and strong time-varying conditions, this embodiment compares the Hilbert envelope spectra of the signals learned by 6 types of models.

[0132] Figure 11The time-frequency domain features of B0-IR learned by different models on the Ottawa dataset (B0→B0). B0-IR represents the inner race fault in subset B0. B0-IR is the condition of increasing speed, and the rotational frequency linearly increases from 12.5 Hz to 27.8 Hz. The inner race fault samples selected in this experiment are close to the starting rotational frequency of 12.5 Hz. Therefore, the inner race fault frequency is 12.5 Hz × 5.432 = 67.9 Hz, where 5.432 is the multiple of the fault frequency to the rotational frequency.

[0133] Figure 11 As shown in (a), the clean B0-IR signal has a relatively fuzzy fault frequency of 67.9 Hz, while the second harmonic of 135.8 Hz and the third harmonic of 203.7 Hz are relatively clear. Figure 11 In the noisy B0-IR signal shown in (b), the fault frequency has been submerged by noise.

[0134] Figure 11 Part (c) shows that CNN hardly learns the fault frequency of B0-IR. Figure 11 Part (f) shows that WD-CNN hardly learns the fault frequency of B0-IR. Figure 11 Although the W-CNN corresponding to part (d), the WKN corresponding to part (e), and the DCA-BiGRU corresponding to part (g) can learn some fault frequencies. For example, W-CNN learns the second harmonic of 135.8 Hz, and WKN and DCA-BiGRU learn the fault frequency of 67.9 Hz, the features they learn are still interfered by strong noise.

[0135] Different from other models, the WGS-CNN model can learn clearer fault frequencies of 67.9 Hz and the second harmonic of 135.8 Hz, and has excellent anti-interference ability, with almost no interference from high-frequency non-fault features.

[0136] Generally speaking, compared with other models, the WGS-CNN model can learn clear B0-IR fault features and is not affected by noise and time-varying factors. This fully shows that the WGS-CNN model has strong noise and strong time-varying characteristics.

[0137] Performance evaluation of fault diagnosis under multi-condition migration:

[0138] The goal of condition migration is to enable the model to maintain high diagnostic accuracy under other conditions after training under one condition. For example, the model is trained under the condition corresponding to subset B0 in the Ottawa dataset and can also be effectively applied under other conditions such as B1, B2, and B3. To further study the migration performance of the WGS-CNN model under complex conditions, based on the Ottawa dataset with a signal-to-noise ratio of -15 dB, the multi-condition migration performance of 6 models is compared and analyzed.

[0139] Figure 12 Performance comparison of different models in the working condition migration task for the Ottawa dataset with SNR = -15 dB, including 4 multi-working condition migration tasks, is analyzed as follows:

[0140] In the 4 multi-working condition migration tasks, the F1 of the WGS-CNN model is the best. Among them, in the transfer training of the WGS-CNN model from B0 to B1, B2, and B3, the F1 score is 97.14%, much higher than other models.

[0141] Compared with other models, the performance variance of the F1 score of the WGS-CNN model is the smallest. This shows that the WGS-CNN model can accurately extract the signal feature frequency and still has a stable working condition migration ability under strong noise and strong time-varying conditions.

[0142] To compare the model interpretability under multi-working condition migration, the time-frequency domain features extracted by 6 models are analyzed for the known fault B0-OR and the unknown fault B1-OR.

[0143] Figure 13 Time-frequency features of B0-OR learned by different models on the Ottawa dataset (B0→B1, B2, B3). The working condition corresponding to B0-OR is the speed increase condition, and the rotation frequency linearly increases from 14.8 Hz to 27.1 Hz. The inner race fault samples selected in this experiment are close to the starting rotation frequency of 14.8 Hz, so the inner race fault frequency is 52.9 Hz, and the calculation method is: 3.571×14.8 Hz = 52.9 Hz, where 3.571 is the multiple of the fault frequency to the rotation frequency. From Figure 13 The following conclusions can be drawn:

[0144] The clean B0-OR contains rich frequency features, such as the fault frequency of 52.9 Hz, sidebands of 82.5 Hz (52.85 Hz + 14.8 Hz×2 = 82.5 Hz), 188.2 Hz (52.85 Hz×3 + 14.8 Hz×2 = 188.2 Hz), etc. The fault frequency of the noisy B0-OR has been submerged by the noise.

[0145] The CNN learned the sidebands at 82.5 Hz, 490.5 Hz (52.85 Hz × 9 + 14.8 Hz = 490.5 Hz), and 619.4 Hz (52.85 Hz × 12 - 14.8 Hz = 619.4 Hz). The W-CNN learned many features of B0-OR, such as the sidebands at 82.5 Hz, 241.0 Hz (52.85 Hz × 4 + 14.8 Hz × 2 = 241.0 Hz), and the multiple frequency at 581.4 Hz (52.85 Hz × 11 = 581.4 Hz). The WKN also learned many B0-OR sideband features, such as 188.2 Hz, 279.1 Hz (52.85 Hz × 5 + 14.8 Hz = 279.1 Hz), etc. The WD-CNN learned features such as the multiple frequency at 105.7 Hz (52.85 Hz × 2 = 105.7 Hz) and the sideband at 302.3 Hz (52.85 Hz × 6 - 14.8 Hz = 302.3 Hz). The DCA-BiGRU learned 82.5 Hz, 135.3 Hz (52.85 Hz × 2 + 14.8 Hz × 2 = 135.3 Hz), and the multiple frequency at 634.2 Hz (52.85 Hz × 12 = 634.2 Hz). However, the above models are still affected by strong noise, strong time-variation, and working condition migration, and the learned feature resolution is not high, and the noise interference is serious.

[0146] The WGS-CNN model can not only learn the fault frequency of 52.9 Hz, but also learn the clear sidebands at 82.5 Hz and the multiple frequency at 158.6 Hz (52.85 Hz × 3 = 158.6 Hz), and the learned feature resolution ratio is high and the interference ability is strong.

[0147] Figure 14 The time-frequency domain features of B1-OR learned by different models on the Ottawa dataset (B0→B1, B2, B3). The working condition corresponding to B1-OR is the speed reduction condition, and the rotation frequency linearly decreases from 24.9 Hz to 9.8 Hz. The selected inner race fault samples in this experiment are close to the starting rotation frequency of 24.9 Hz, so the inner race fault frequency is 52.9 Hz, 3.571 × 24.9 Hz = 88.9 Hz, where 3.571 is the multiple of the fault frequency to the rotation frequency. From Figure 14 The following conclusions can be drawn:

[0148] CNN, W-CNN, and WKN can learn relatively clear fault frequencies and sidebands from the noisy B1-OR, such as 64.1 Hz, 202.7 Hz, and 88.9 Hz. However, the resolution of these features is still low, and they are greatly affected by strong noise, strong time-variation, and working condition migration. In addition to the above models, WD-CNN and DCA-BiGRU are more severely affected by complex working conditions, especially in the frequency band above 500 Hz.

[0149] The WGS-CNN model can clearly learn the fault frequency characteristics, such as the sideband frequency of 88.9 Hz, the double frequency of 177.8 Hz (88.9179 Hz × 2 = 177.8 Hz), and 266.8 Hz (88.9179 Hz × 3 = 266.8 Hz).

[0150] By comparing the B0-OR and B1-OR characteristics learned by different models, it can be seen that: compared with other models, the WGS-CNN model can learn clear fault characteristics and sidebands, has strong multi-condition migration performance, and has good anti-noise performance and strong time-varying fault diagnosis ability.

[0151] Generally speaking, compared with other models, the WGS-CNN model proposed in the present invention has better diagnostic performance, generalization performance and interpretability under strong noise, strong time-varying and multi-condition transfer learning, providing an effective solution for fault diagnosis under complex working conditions.

[0152] Embodiment 3

[0153] This embodiment provides a rotating machinery fault diagnosis device based on WGS-CNN, including:

[0154] A fault vibration signal acquisition module, configured to: acquire the fault vibration signal of the rotating machinery;

[0155] A fault diagnosis module, configured to: input the fault vibration signal into the trained WGS-CNN model to obtain the fault diagnosis classification result of the rotating machinery;

[0156] Wherein, the WGS-CNN model is obtained by improving the ELCNN model, and the improvement includes: initializing all convolution kernels in the ELCNN model through the wavelet basis function to obtain wavelet-initialized convolution kernels, introducing Gaussian windows into the wavelet-initialized convolution kernels, and then replacing the original activation function with a square activation function.

[0157] Embodiment 4

[0158] This embodiment provides a computer-readable storage medium, on which computer programs / instructions are stored. When the computer programs / instructions are executed by a processor, the steps of the rotating machinery fault diagnosis method based on WGS-CNN provided in Embodiment 1 are implemented:

[0159] Acquire the fault vibration signal of the rotating machinery;

[0160] Input the fault vibration signal into the trained WGS-CNN model to obtain the fault diagnosis classification result of the rotating machinery;

[0161] Among them, the WGS-CNN model is obtained by improving the ELCNN model. The improvement includes: initializing all the convolutional kernels in the ELCNN model with the wavelet basis function to obtain wavelet-initialized convolutional kernels, introducing a Gaussian window into the wavelet-initialized convolutional kernels, and then replacing the original activation function with a square activation function.

[0162] Embodiment 5

[0163] The present invention provides a computer system, including:

[0164] A memory for storing computer programs / instructions;

[0165] A processor for executing the computer programs / instructions to implement the steps of the method for diagnosing faults of rotating machinery based on WGS-CNN provided in Embodiment 1:

[0166] Obtain the fault vibration signal of the rotating machinery;

[0167] Input the fault vibration signal into the trained WGS-CNN model to obtain the fault diagnosis classification result of the rotating machinery;

[0168] Among them, the WGS-CNN model is obtained by improving the ELCNN model. The improvement includes: initializing all the convolutional kernels in the ELCNN model with the wavelet basis function to obtain wavelet-initialized convolutional kernels, introducing a Gaussian window into the wavelet-initialized convolutional kernels, and then replacing the original activation function with a square activation function.

[0169] Those skilled in the art should understand that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0170] The present application is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, and the combination of processes and / or blocks in the flowchart and / or block diagram can be realized by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate for realizing the processes Figure 1 one process or multiple processes and / or blocks Figure 1A device for the functions specified in one or more boxes.

[0171] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory produce a manufactured article including an instruction device that implements the functions specified in one Figure 1 process or more processes and / or boxes Figure 1 A box or the functions specified in multiple boxes.

[0172] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are performed on the computer or other programmable device to generate a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one Figure 1 process or more processes and / or boxes Figure 1 A box or the functions specified in multiple boxes.

[0173] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the technical principles of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the protection scope of the present invention.

Claims

1. A rotating machinery fault diagnosis method based on WGS-CNN, characterized in that, Including: Obtain the fault vibration signal of the rotating machinery; Input the fault vibration signal into the trained WGS-CNN model to obtain the fault diagnosis classification result of the rotating machinery; The WGS-CNN model is obtained by improving the ELCNN model. The improvement includes: initializing all the convolutional kernels in the ELCNN model with the wavelet basis function to obtain wavelet-initialized convolutional kernels, introducing a Gaussian window into the wavelet-initialized convolutional kernels, and then replacing the original activation function with a square activation function.

2. The method for diagnosing faults of rotating machinery based on WGS-CNN according to claim 1, wherein The initializing of all the convolutional kernels in the ELCNN model with the wavelet basis function to obtain wavelet-initialized convolutional kernels is carried out through the following formula: （1）； （2）； （3）； Among them, is the Morlet mother wavelet with the independent variable t, where t is the time variable, is the first scale parameter, is the sine carrier wave, is a Gaussian function, e is the natural constant, and a is the second scale parameter, is the Morlet wavelet basis function with the independent variable t, is the wavelet initialization convolution kernel, and j is the imaginary unit, is the angular frequency, is the Morlet wavelet obtained by transformation with the second scale parameter a.

3. The method for diagnosing faults of rotating machinery based on WGS-CNN according to claim 1, characterized in that The introducing of a Gaussian window into the wavelet-initialized convolutional kernels is carried out through the following formula: （4）； （5）； Among them, is a Gaussian window, x is the independent variable of the Gaussian window, and e is the natural constant. is the third scale parameter. is the windowed wavelet initialization convolution kernel obtained by introducing the Gaussian window into the wavelet initialization convolution kernel. is the wavelet initialization convolution kernel.

4. The method for diagnosing faults of rotating machinery based on WGS-CNN according to claim 3, wherein, The third-scale parameter is obtained through the following method: （6）； （7）； （8）； （9）； Among them, is the gradient of the th convolutional kernel, L is the cross-entropy loss function of the convolutional neural network, is the wavelet basis function used to initialize the th convolutional kernel, is the gradient magnitude of the th convolutional kernel, is the m-th digit value of the gradient of the th convolutional kernel, k is the maximum value of m, is the average value of the gradient magnitudes of the convolutional kernels, n is the number of convolutional kernels, is the third scale parameter of the Gaussian window in the th wavelet-initialized convolutional kernel after update, is the third scale parameter of the Gaussian window in the th wavelet-initialized convolutional kernel before update, and α is the set adjustment coefficient.

5. The method for diagnosing faults of rotating machinery based on WGS-CNN according to claim 1, wherein, The expression of the square activation function is: （10）； Among them, is the square activation function, is the windowed wavelet initialization convolution kernel, * is the convolution operation, × is the multiplication operation, and q is the input signal of the square activation function.

6. The method for diagnosing faults of rotating machinery based on WGS-CNN according to claim 1, characterized in that, The WGS-CNN model is trained through the following method: Obtain a non-time-varying data set and a time-varying data set containing fault vibration signals; Based on the non-time-varying data set and the time-varying data set containing fault vibration signals, train the WGS-CNN model through gradient descent and backpropagation until the F1 score of the WGS-CNN model does not increase for several consecutive times.

7. A rotating machinery fault diagnosis device based on WGS-CNN, characterized in that, Including: A fault vibration signal acquisition module configured to: obtain the fault vibration signal of the rotating machinery; A fault diagnosis module configured to: input the fault vibration signal into the trained WGS-CNN model to obtain the fault diagnosis classification result of the rotating machinery; Wherein, the WGS-CNN model is obtained by improving the ELCNN model. The improvement includes: initializing all the convolutional kernels in the ELCNN model with the wavelet basis function to obtain wavelet-initialized convolutional kernels, introducing a Gaussian window into the wavelet-initialized convolutional kernels, and then replacing the original activation function with a square activation function.

8. A computer-readable storage medium having computer programs / instructions stored thereon, characterized in that, When the computer program / instructions are executed by a processor, the steps of the WGS-CNN-based rotating machinery fault diagnosis method according to any one of claims 1 to 6 are implemented.

9. A computer system, characterized in that, Including: A memory for storing computer programs / instructions; A processor for executing the computer programs / instructions to implement the steps of the WGS-CNN-based rotating machinery fault diagnosis method according to any one of claims 1 to 6.

Images