Fault diagnosis method and diagnosis system based on ADPCW-ELCNN

Abstract

The invention discloses a fault diagnosis method based on ADPCW-ELCNN. The method comprises the following steps: obtaining a fault vibration signal of a rotating machine; inputting a fault vibration signal into the trained ADPCW-ELCNN model to obtain a fault diagnosis classification result of the rotating machine; wherein the ADPCW-ELCNN model adopts a single-layer convolution architecture, and comprises a wavelet convolution layer used for performing multi-scale feature extraction based on a wavelet convolution kernel associated with scale and length according to a fault vibration signal to obtain a first feature map; the square nonlinear mapping layer is used for performing square nonlinear mapping operation according to the first feature map to obtain a second feature map; and a global average pooling layer, a first full connection layer and a second full connection layer. According to the method, the wavelet kernel scale and length are associated, the square mapping enhancement feature is introduced, and the single-layer lightweight architecture is adopted, so that the fault diagnosis capability of the rotating machinery under strong noise and small samples is greatly improved.

Description

Technical Field

[0001] This invention relates to a fault diagnosis method and system based on ADPCW-ELCNN, belonging to the field of fault diagnosis technology. Background Technology

[0002] In the field of fault diagnosis of vibration signals in rotating machinery, one-dimensional convolutional neural networks (1DCNNs) have become a key technology choice. On the one hand, they possess powerful local abstraction capabilities, accurately capturing key local information in vibration signals; on the other hand, they have adaptive learning capabilities for one-dimensional vibration signals, thus exhibiting significant advantages in local feature extraction efficiency and fault classification accuracy. However, current mainstream 1DCNN fault diagnosis models share common design flaws: convolutional kernels are generally initialized randomly using a normal distribution. This approach easily exposes three major technical shortcomings under harsh industrial conditions of strong noise and small samples: firstly, weak feature extraction is difficult; secondly, there is a disconnect between parameters and feature adaptation; and thirdly, it is difficult to balance model generalization ability and operational efficiency. These shortcomings combined directly result in existing models failing to meet the actual fault diagnosis needs of industrial edge equipment under small sample conditions, ultimately leading to low fault diagnosis accuracy and difficulty in providing stable support for the reliable operation of rotating machinery. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a fault diagnosis method and system based on ADPCW-ELCNN. By associating wavelet kernel scale and length, introducing squared mapping to enhance features, and adopting a single-layer lightweight architecture, the fault diagnosis capability of rotating machinery under strong noise and small sample conditions is greatly improved.

[0004] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0005] On the one hand, this invention discloses a fault diagnosis method based on ADPCW-ELCNN, including:

[0006] Acquire fault vibration signals of rotating machinery;

[0007] The fault vibration signal is input into the trained ADPCW-ELCNN model to obtain the fault diagnosis classification result of the rotating machinery;

[0008] The ADPCW-ELCNN model employs a single-layer convolutional architecture and includes:

[0009] The wavelet convolutional layer is used to perform multi-scale feature extraction based on the fault vibration signal and the wavelet convolutional kernel with scale and length correlation to obtain the first feature map.

[0010] A squared nonlinear mapping layer is used to perform a squared nonlinear mapping operation on the first feature map to obtain a second feature map;

[0011] A global average pooling layer is used to perform a global average pooling operation on the second feature map to obtain a first feature vector;

[0012] The first fully connected layer is used to perform feature transformation on the first feature vector to obtain the second feature vector;

[0013] The second fully connected layer is used to perform feature classification on the second feature vector to obtain the fault diagnosis classification result.

[0014] Furthermore, the multi-scale feature extraction includes:

[0015] Calculate the scale and length of each wavelet convolution kernel;

[0016] The corresponding wavelet convolution kernel is determined based on the scale and length of each wavelet convolution kernel;

[0017] Based on multiple wavelet convolution kernels, one-dimensional convolution operations are performed on the fault vibration signal to obtain multiple first feature maps.

[0018] Furthermore, the expression for the scale of the wavelet convolution kernel is as follows:

[0019]

[0020] In the formula, This represents the scale of the i-th wavelet convolution kernel; Indicates the number of wavelet convolution kernels; Indicates the sequential index of the wavelet convolution kernel; The basic scale is expressed as follows:

[0021]

[0022] In the formula, This represents the minimum scale of the wavelet convolution kernel; This represents the maximum scale of the wavelet convolution kernel; This indicates a learnable global scale factor.

[0023] Furthermore, the expression for the scale of the wavelet convolution kernel is as follows:

[0024]

[0025] In the formula, This represents the length of the i-th wavelet convolution kernel; This represents the length adjustment factor of the i-th wavelet convolution kernel; This represents the maximum length of the wavelet convolution kernel; This represents the minimum length of the wavelet convolution kernel; This represents the floor function; This represents the base length of the wavelet convolution kernel. , This represents the learnable energy range coefficient. Indicates the basic scale. This represents the function for rounding up.

[0026] Furthermore, the expression for the squared nonlinear mapping operation is as follows:

[0027]

[0028] In the formula, This represents the z-th eigenvalue in the second feature map output by the squared nonlinear mapping layer; Represents the eigenvalues The result of performing a squared nonlinear mapping operation; This represents the z-th feature value in the first feature map output by the wavelet convolutional layer.

[0029] Furthermore, the expression for the fault diagnosis classification result is as follows:

[0030]

[0031] In the formula, This indicates the fault diagnosis classification results; Indicates a second fully connected layer; Indicates the first fully connected layer; Indicates global average pooling; Represents a squared nonlinear mapping; Represents convolution; This represents the wavelet convolution kernel.

[0032] Furthermore, the training steps of the ADPCW-ELCNN model are as follows:

[0033] A sample set of fault vibration signals of rotating machinery is obtained, and the sample set is preprocessed to construct a training set;

[0034] Based on the training set, the pre-built ADPCW-ELCNN model is iteratively trained using gradient descent and backpropagation algorithms to optimize the model parameters of the ADPCW-ELCNN model.

[0035] The training continues until the preset training termination condition is met, at which point the trained ADPCW-ELCNN model is output.

[0036] The gradient descent and backpropagation algorithm includes:

[0037] The training samples of the training set are input into the ADPCW-ELCNN model to obtain the fault diagnosis classification training results;

[0038] The loss function is calculated based on the fault diagnosis classification training results and the true labels of the training samples.

[0039] Calculate the gradient of the loss function with respect to each parameter of the model; wherein, for the wavelet convolutional layer, calculate the gradient of the loss function with respect to the scale and length of each wavelet convolutional kernel respectively;

[0040] Based on the gradient, the gradient descent algorithm, which uses a global-local parameter adaptation mechanism, updates the parameters of the model.

[0041] Furthermore, the expression for the gradient of the loss function with respect to the scale parameter of the wavelet convolution kernel is as follows:

[0042]

[0043] In the formula, This represents the gradient of the loss function with respect to the scale of the i-th wavelet convolution kernel;

[0044] Represents the partial differential operator; Represents the loss function; This represents the scale of the i-th wavelet convolution kernel; This represents the second feature map output by the squared nonlinear mapping layer; This represents the first feature map output by the wavelet convolutional layer.

[0045] Furthermore, the expression for the gradient of the loss function with respect to the length of the wavelet convolution kernel is as follows:

[0046]

[0047] In the formula, This represents the gradient of the loss function with respect to the length of the i-th wavelet convolution kernel;

[0048] Represents the partial differential operator; Represents the loss function; This represents the length of the i-th wavelet convolution kernel; This represents the second feature map output by the squared nonlinear mapping layer; This represents the first feature map output by the wavelet convolutional layer.

[0049] On the other hand, this invention discloses a fault diagnosis system based on ADPCW-ELCNN, applicable to the aforementioned fault diagnosis method based on ADPCW-ELCNN, comprising:

[0050] The data acquisition module is used to acquire fault vibration signals of rotating machinery;

[0051] The fault diagnosis module is used to input the fault vibration signal into the trained ADPCW-ELCNN model to obtain the fault diagnosis classification result of the rotating machinery.

[0052] The ADPCW-ELCNN model employs a single-layer convolutional architecture and includes:

[0053] The wavelet convolutional layer is used to perform multi-scale feature extraction based on the fault vibration signal and the wavelet convolutional kernel with scale and length correlation to obtain the first feature map.

[0054] A squared nonlinear mapping layer is used to perform a squared nonlinear mapping operation on the first feature map to obtain a second feature map;

[0055] A global average pooling layer is used to perform a global average pooling operation on the second feature map to obtain a first feature vector;

[0056] The first fully connected layer is used to perform feature transformation on the first feature vector to obtain the second feature vector;

[0057] The second fully connected layer is used to perform feature classification on the second feature vector to obtain the fault diagnosis classification result.

[0058] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0059] The fault diagnosis method and system based on ADPCW-ELCNN of the present invention firstly extracts multi-scale features based on wavelet convolution kernels with scale-length correlation, drawing on statistical methods. The principle of interval quantization ensures that longer convolutional kernels correspond to large-scale analysis of low-frequency signals, while shorter kernels correspond to small-scale analysis of high-frequency signals, thus fundamentally solving the problem of the disconnect between parameters and features. Secondly, the introduction of a squared nonlinear mapping operation enhances energy-concentrated fault features and suppresses noise interference with dispersed energy. Finally, a single-layer convolutional architecture is employed, which, while ensuring accurate preservation and adaptive optimization of fault features in the time-frequency domain, significantly reduces the number of model parameters and computational complexity, achieving an effective balance between high diagnostic accuracy and edge deployment efficiency. Attached Figure Description

[0060] Figure 1This is a flowchart of the fault diagnosis method based on ADPCW-ELCNN provided in Embodiment 1 of the present invention;

[0061] Figure 2 This is a flowchart of the training process of the ADPCW-ELCNN model provided in Embodiment 1 of the present invention;

[0062] Figure 3 This is a diagram illustrating the correlation between wavelet kernel scale and length, as well as the training process of energy range coefficients provided by this invention.

[0063] Figure 4 These are comparison diagrams of wavelet kernel waveforms of different scales and lengths provided by this invention;

[0064] Figure 5 This is a schematic diagram of the data preprocessing process provided by the present invention;

[0065] Figure 6 This is a comparison chart of the diagnostic performance of different improved models provided by this invention;

[0066] Figure 7 This is a schematic diagram showing the comparison of the noise resistance performance of different improved models based on the Ottawa dataset provided by this invention;

[0067] Figure 8 This is a schematic diagram showing the comparison of the noise resistance performance of different improved models based on the MCC5-THU dataset provided by this invention.

[0068] Figure 9 This is a schematic diagram showing the comparison of small-sample generalization performance of different improved models based on the Ottawa dataset under a signal-to-noise ratio of -10dB, as provided by this invention.

[0069] Figure 10 This is a schematic diagram showing the comparison of small-sample generalization performance of different improved models based on the MCC5-THU dataset under a signal-to-noise ratio of -5dB, provided by the present invention.

[0070] Figure 11 This is a schematic diagram comparing the waveforms of convolution kernels learned by different improved models provided by this invention.

[0071] Figure 12 This is a comparative diagram showing the time-frequency domain features learned by different improved models provided by this invention for gear pitting faults.

[0072] Figure 13 This is a comparative diagram showing the time-frequency domain features learned by different improved models provided by this invention for tooth loss faults.

[0073] Figure 14 This is a comparative diagram showing the time-frequency domain features learned by different improved models provided by this invention for gear breakage faults. Detailed Implementation

[0074] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0075] Example 1

[0076] This embodiment 1 provides a fault diagnosis method based on ADPCW-ELCNN, including:

[0077] Acquire fault vibration signals of rotating machinery;

[0078] The fault vibration signal is input into the trained ADPCW-ELCNN model to obtain the fault diagnosis and classification results of the rotating machinery.

[0079] The ADPCW-ELCNN model employs a single-layer convolutional architecture and includes:

[0080] The wavelet convolutional layer is used to extract multi-scale features based on the fault vibration signal and the wavelet convolutional kernel with scale and length correlation to obtain the first feature map.

[0081] The squared nonlinear mapping layer is used to perform a squared nonlinear mapping operation on the first feature map to obtain the second feature map;

[0082] A global average pooling layer is used to perform a global average pooling operation on the second feature map to obtain the first feature vector;

[0083] The first fully connected layer is used to perform feature transformation on the first feature vector to obtain the second feature vector;

[0084] The second fully connected layer is used to perform feature classification on the second feature vector to obtain the fault diagnosis classification result.

[0085] The technical concept of this invention is as follows: First, multi-scale feature extraction is performed based on wavelet convolution kernels that are correlated with scale and length, drawing on statistical methods. The principle of interval quantization ensures that longer convolutional kernels correspond to large-scale analysis of low-frequency signals, while shorter kernels correspond to small-scale analysis of high-frequency signals, thus fundamentally solving the problem of the disconnect between parameters and features. Secondly, the introduction of a squared nonlinear mapping operation enhances energy-concentrated fault features and suppresses noise interference with dispersed energy. Finally, a single-layer convolutional architecture is employed, which, while ensuring accurate preservation and adaptive optimization of fault features in the time-frequency domain, significantly reduces the number of model parameters and computational complexity, achieving an effective balance between high diagnostic accuracy and edge deployment efficiency.

[0086] like Figure 1 As shown, the specific steps are as follows:

[0087] Step 1: Obtain the fault vibration signal of the rotating machinery;

[0088] Step 2: Input the fault vibration signal into the trained ADPCW-ELCNN model to obtain the fault diagnosis classification results of the rotating machinery.

[0089] The ADPCW-ELCNN (Adaptive Dual-Parameter Collaborative WaveletExplainable Lightweight Convolutional Neural Network) model in this embodiment adopts a single-layer convolutional architecture.

[0090] 2.1 Structure of the ADPCW-ELCNN model:

[0091] The ADPCW-ELCNN model includes:

[0092] 2.1.1 Input layer, used to input fault vibration signals.

[0093] 2.1.2 Wavelet convolutional layer, used to extract multi-scale features based on the fault vibration signal and the wavelet convolution kernel with scale and length correlation to obtain the first feature map.

[0094] Multi-scale feature extraction is performed, including:

[0095] Calculate the scale and length of each wavelet convolution kernel;

[0096] The corresponding wavelet convolution kernel is determined based on the scale and length of each wavelet convolution kernel;

[0097] Based on multiple wavelet convolution kernels, one-dimensional convolution operations are performed on the fault vibration signal to obtain multiple first feature maps.

[0098] The expression for the scale of the wavelet convolution kernel is as follows:

[0099]

[0100] In the formula, This represents the scale of the i-th wavelet convolution kernel; Indicates the number of wavelet convolution kernels; Indicates the sequential index of the wavelet convolution kernel; The basic scale is expressed as follows:

[0101]

[0102] In the formula, This represents the minimum scale of the wavelet convolution kernel; This represents the maximum scale of the wavelet convolution kernel; This indicates a learnable global scale factor.

[0103] The expression for the length of the wavelet convolution kernel is as follows:

[0104]

[0105] In the formula, This represents the length of the i-th wavelet convolution kernel; This represents the length adjustment factor of the i-th wavelet convolution kernel; This represents the maximum length of the wavelet convolution kernel; This represents the minimum length of the wavelet convolution kernel; This represents the floor function; This represents the base length of the wavelet convolution kernel. , This represents the learnable energy range coefficient. This represents the function for rounding up.

[0106] This invention establishes a quantitative correlation between wavelet kernel scale and length, defining the effective coverage range of wavelet energy, thus laying a theoretical foundation for the collaborative optimization of scale and length parameters. The training process for this correlation and the energy range coefficients is as follows: Figure 3 As shown, where Figure 3 (a) shows the correlation between scale and length distribution. Figure 3 (b) is the training curve for the energy range coefficient.

[0107] In the above correlation, scale and length follow a clear physical matching rule: large scales are suitable for analyzing low-frequency signals and require longer convolution kernels; small scales are suitable for analyzing high-frequency signals and can correspond to shorter convolution kernels. This rule ensures the effective extraction of signal features from different frequency bands.

[0108] Figure 4 The invention demonstrates a comparison of four wavelet kernel waveforms with different combinations of scale and length parameters, exhibiting a matching relationship of "short kernels for small scales and long kernels for large scales." Figure 4 (a) Wavelet kernel with a scale of 3.54 and a length of 27; Figure 4 (b) Wavelet kernel with a scale of 9.63 and a length of 71; Figure 4 (c) Wavelet kernel with a scale of 16.96 and a length of 109; Figure 4 (d) Wavelet kernel with a scale of 22.51 and a length of 183. By comparison, the differences in waveform morphology of wavelet kernels under different parameter configurations can be observed intuitively.

[0109] 2.1.3. The squared nonlinear mapping layer is used to perform a squared nonlinear mapping operation on the first feature map to obtain the second feature map.

[0110] The expression for the squared nonlinear mapping operation is as follows:

[0111]

[0112] In the formula, This represents the z-th eigenvalue in the second feature map output by the squared nonlinear mapping layer; Represents the eigenvalues The result of performing a squared nonlinear mapping operation; This represents the z-th feature value in the first feature map output by the wavelet convolutional layer.

[0113] This square transformation essentially achieves coordinated optimization of the convolution kernel scale parameter and length parameter by reshaping the feature energy distribution and gradient propagation path; where the scale parameter determines the perceptual range of the feature, and the length parameter corresponds to the physical dimension of the feature.

[0114] 2.1.4 Global average pooling layer, used to perform global average pooling operation on the second feature map to obtain the first feature vector.

[0115] 2.1.5 The first fully connected layer is used to perform feature transformation on the first feature vector to obtain the second feature vector.

[0116] 2.1.6 The second fully connected layer is used to perform feature classification on the second feature vector to obtain the fault diagnosis classification result.

[0117] The expression for the fault diagnosis classification result is as follows:

[0118]

[0119] In the formula, This indicates the fault diagnosis classification results; Indicates a second fully connected layer; Indicates the first fully connected layer; Indicates global average pooling; Represents a squared nonlinear mapping; Represents convolution; Represents the wavelet convolution kernel;

[0120] 2.1.7 Output layer, used to output fault diagnosis classification results.

[0121] The parameters of the ADPCW-ELCNN model in this embodiment are shown in Table 1:

[0122] Table 1 - ADPCW-ELCNN Model Parameter Configuration Table

[0123]

[0124] The kernel length of the ADPCW-ELCNN model is dynamically learned, and the number of filters is optimized based on the temporal characteristics of the signal. When using the University of Ottawa bearing vibration dataset (Ottawa dataset) or the MCC5 Group and Tsinghua University gearbox vibration dataset (MCC5-THU dataset), the ADPCW-ELCNN model has 8 filters in both cases. The sample length of the ADPCW-ELCNN model is optimized based on the frequency resolution of the signal. The Ottawa dataset has a sample length of 32,768, while the MCC5-THU dataset has a sample length of 256,000.

[0125] 2.2 Training of the ADPCW-ELCNN model.

[0126] like Figure 2 As shown, the training steps for the ADPCW-ELCNN model are as follows:

[0127] 2.2.1 Obtain a sample set of fault vibration signals of rotating machinery, and preprocess the sample set to construct a training set.

[0128] In this embodiment, the University of Ottawa Bearing Vibration Dataset (hereinafter referred to as the Ottawa Dataset) and the China Metallurgical Group-Tsinghua University Gearbox Vibration Dataset (hereinafter referred to as the MCC5-THU Dataset), which cover complex working conditions of "variable speed" and "variable load", are selected as the sample set of fault vibration signals of rotating machinery.

[0129] like Figure 5 As shown, this embodiment performs preprocessing on the original signals in the sample set, including dataset partitioning, sequential sampling, normalization, and adding noise, to obtain the training set.

[0130] The specific steps are as follows:

[0131] 2.2.1.1 In this embodiment, four dynamic speed curves and three types of vibration signals in healthy states are selected from the Ottawa dataset.

[0132] As shown in Table 2, the dataset contains four dynamic speed curves: increasing speed condition A0, decreasing speed condition A1, increasing-then-decreasing speed condition A2, and decreasing-then-increasing speed condition A3, as well as three health states: normal (H in this embodiment), inner-race fault (IF in this embodiment), and outer-race fault (OF in this embodiment). The data sampling frequency is 200kHz, and the number of samples for each signal is 400. The training / validation / test sets are divided in a 7:2:1 ratio.

[0133] Table 2 - Sample Table of Ottawa Dataset

[0134]

[0135] As shown in Table 3, this embodiment selects four time-varying operating conditions from the MCC5-THU dataset: condition B0, condition B1, condition B2, and condition B3. Condition B0 corresponds to a speed of 0→1000→500 rpm and a load of 10 Nm; condition B1 corresponds to a speed of 0→1000→500 rpm and a load of 20 Nm; condition B2 corresponds to a speed of 1000 rpm and a load of 0→10 Nm; and condition B3 corresponds to a speed of 1000 rpm and a load of 0→20 Nm. The dataset contains vibration signals for three types of gear fault states: gear pitting (GP in this embodiment), missing teeth (MT in this embodiment), and gear breakage (TB in this embodiment). The signal sampling frequency is 12.8 kHz, and the number of samples and their division ratio are consistent with the aforementioned Ottawa dataset.

[0136] Table 3 - Sample Datasheet of China Fifth Metallurgical Construction Group - Tsinghua University

[0137]

[0138] 2.2.1.2. Clean sample sequences are obtained by using an equal-interval sequential sampling method.

[0139] Clean sample sequences meeting frequency resolution were obtained from the Ottawa dataset and the MCC5-THU dataset using an equally spaced sequential sampling method. Meeting frequency resolution refers to the time length required to analyze the lowest fault frequency in the signal. In this embodiment, the Ottawa sampling length was set to 32768 to ensure a minimum rotation frequency of 12.5Hz; the MCC5-THU sampling length was set to 256000 to ensure a frequency domain resolution of 0.05Hz.

[0140] 2.2.1.3. Normalize the obtained sample signals; use a normalization operation to normalize the amplitude of the samples obtained in step 2.2.1.2 to the (-1, 1) interval, and obtain normalized samples for training. Normalization facilitates comparison of implementation examples of different datasets and different models.

[0141] 2.2.1.4 Set sample category labels. Add additive white Gaussian noise proportional to the sample signal to the data obtained in step 2.2.1.2 to obtain a noisy training set and test set; where the signal-to-noise ratio of the noisy samples is -10dB, -5dB, 0dB, 5dB, 10dB and 20dB.

[0142] 2.2.2 Based on the training set, the pre-built ADPCW-ELCNN model is iteratively trained using gradient descent and backpropagation algorithms to optimize the model parameters of the ADPCW-ELCNN model.

[0143] The gradient descent and backpropagation algorithms include:

[0144] The training samples from the training set are input into the ADPCW-ELCNN model to obtain the fault diagnosis classification training results;

[0145] The loss function is calculated based on the fault diagnosis classification training results and the true labels of the training samples.

[0146] Calculate the gradient of the loss function with respect to each parameter of the model; for wavelet convolutional layers, calculate the gradient of the loss function with respect to the scale and length of each wavelet convolutional kernel respectively.

[0147] Based on the gradient, the gradient descent algorithm, which uses a global-local parameter adaptation mechanism, updates the corresponding parameters of the model.

[0148] Specifically, the expression for the gradient of the loss function with respect to the scale parameter of the wavelet convolution kernel is as follows:

[0149]

[0150] In the formula, This represents the gradient of the loss function with respect to the scale of the i-th wavelet convolution kernel;

[0151] Represents the partial differential operator; Represents the loss function; This represents the scale of the i-th wavelet convolution kernel; This represents the second feature map output by the squared nonlinear mapping layer; This represents the first feature map output by the wavelet convolutional layer. It is a term introduced by the squared mapping, which can enhance the gradient in high-amplitude fault regions to accelerate small-scale updates, suppress low-amplitude noise gradients, and suppress extreme gradients in small samples to avoid overfitting.

[0152] The expression for the gradient of the loss function with respect to the length of the wavelet convolution kernel is as follows:

[0153]

[0154] In the formula, This represents the gradient of the loss function with respect to the length of the i-th wavelet convolution kernel;

[0155] Represents the partial differential operator; Represents the loss function; This represents the length of the i-th wavelet convolution kernel; This represents the second feature map output by the squared nonlinear mapping layer; This represents the first feature map output by the wavelet convolutional layer. The length gradient is optimized by accelerating gradient updates in high-frequency fault regions to drive length reduction, while the gradient in low-frequency background regions is flattened to ensure global feature integration.

[0156] In this embodiment, firstly, the wavelet energy coverage range is defined by quantitatively correlating the wavelet kernel scale and length, thus establishing a theoretical basis for the dual-parameter synergistic optimization of scale and length. Specifically, the wavelet kernel scale and length must follow a clear physical matching rule: large scales are suitable for analyzing low-frequency signals, and therefore require longer convolution kernels; small scales are suitable for analyzing high-frequency signals, and can correspond to shorter convolution kernels.

[0157] Based on this, a global-local parameter adaptation mechanism is constructed. This mechanism matches the frequency characteristics of the signal through scale-level learning on the one hand, and covers the time-domain energy range through length co-optimization on the other hand, ultimately achieving dynamic parameter adaptation and accurate extraction of multi-scale features.

[0158] 2.2.3 Continue until the preset training termination condition is met, then output the trained ADPCW-ELCNN model.

[0159] In this embodiment, the training process of the ADPCW-ELCNN model employs an early stopping mechanism as the termination condition. Specifically, during training, after each training cycle or after a predetermined number of iterations, the current model performance is evaluated using an independent validation set. If the model's classification performance metrics on the validation set do not improve within ten consecutive training cycles, the model is considered to have converged, and the early stopping mechanism is triggered to terminate the training process. The currently converged model parameters are then saved and used as the output of the final trained ADPCW-ELCNN model. This mechanism effectively avoids model overfitting and ensures its generalization ability.

[0160] To better illustrate the advantages of the ADPCW-ELCNN model of this invention, this embodiment compares the performance of six models based on the Ottawa dataset and the MCC5-THU dataset: traditional three-layer convolutional neural networks (CNNs), wavelet convolutional neural networks (W-CNNs) that initialize convolutional kernels with wavelets, wavelet kernel convolutional neural networks (WKNs) that replace traditional convolutional kernels with learnable Morlet wavelets, deep convolutional neural networks with wide first-layer kernels (WDCNNs), dual path convolution with attention mechanism and bidirectional gated recurrent units (DCA-BiGRUs), and the ADPCW-ELCNN model provided in this embodiment. The architecture and parameter configuration of these six models are shown in Table 4.

[0161] Table 4 - Architecture and parameter configuration of the models used for comparison

[0162]

[0163] To accurately evaluate model performance, this embodiment introduces accuracy (Acc), precision (Pre), recall (Rec), and F1 score (F1) to assess classification results. To comprehensively evaluate the advantages of the ADPCW-ELCNN model in small-sample fault diagnosis tasks, this embodiment uses the Ottawa dataset and the MCC5-THU dataset, and, in conjunction with Table 5, comprehensively analyzes the performance of the above six improved models in terms of lightweight design, computational efficiency, diagnostic performance, and training speed.

[0164] (1) Significantly lightweight: In Table 5, ADPCW-ELCNN has 9.19K parameters, which is only 0.02%-0.21% of the comparison model, and the model size is 55KB, which is 0.09%-83.33% of the comparison model;

[0165] (2) Outstanding computational efficiency: The average floating-point computation of ADPCW-ELCNN is 0.02×10³M, which is only 2.54‰-34.48‰ of the comparison model;

[0166] (3) Superior diagnostic performance: ADPCW-ELCNN has an average accuracy of 99.98%, which is slightly better than WKN's 99.84% and significantly surpasses other models;

[0167] (4) High training speed: The average convergence time is 71.26s, which is only 39.61%-90.67% of the comparison model.

[0168] Table 5 - Comparison of lightweight performance of different improved models

[0169]

[0170] Figure 6 The detailed comparison of multiple metrics shows that the ADPCW-ELCNN model outperforms all four categories of datasets in terms of accuracy (Accc), precision (Pre), recall (Rec), and F1 score, confirming its comprehensive performance advantage.

[0171] To further verify the state-of-the-art performance of the proposed ADPCW-ELCNN model under strong noise conditions, six types of improved models were compared... Figure 7 and Figure 8 The noise reduction performance under different working conditions was compared.

[0172] Figure 7 This is a schematic diagram showing the comparison of the noise resistance performance of different improved models based on the Ottawa dataset provided by this invention. Figure 7 (a) Comparison results under the A0 condition of the Ottawa dataset; Figure 7 (b) Comparison results under condition A1; Figure 7(c) Comparison results under the A2 working condition; Figure 7 (d) Comparison results under the A3 working condition.

[0173] Figure 8 This is a schematic diagram showing the comparison of the noise resistance performance of different improved models based on the MCC5-THU dataset provided by this invention. Figure 8 (a) Comparison results under the B0 working condition of the corresponding MCC5-THU dataset; Figure 8 (b) Comparison results under condition B1; Figure 8 (c) Comparison results under working condition B2; Figure 8 (d) Comparison results under working condition B3.

[0174] Figure 7 This shows a comparison of the F1 scores of the six models on the Ottawa dataset. Figure 8 This section presents a comparison of F1 scores on the MCC5-THU dataset. The noise resistance performance of each model shows a completely consistent trend across both datasets: when the signal-to-noise ratio drops to -10 dB, the F1 scores of the CNN, W-CNN, and WDCNN models plummet to failure levels, while the WKN and DCA-BiGRU models exhibit significant performance degradation. In contrast, the ADPCW-ELCNN model achieves F1 scores of 96.98%-99.97% across all scenarios on the Ottawa dataset, and also performs superiorly across all scenarios on the MCC5-THU dataset, fully demonstrating its diagnostic stability in noisy environments.

[0175] To further illustrate the advancement of the ADPCW-ELCNN model under small sample conditions, six improved models are compared. Figure 9 and Figure 10 A comparison of generalization performance with small samples was conducted.

[0176] Figure 9 This diagram illustrates the comparison of small-sample generalization performance of different improved models based on the Ottawa dataset under a signal-to-noise ratio of -10dB, as provided by this invention. Figure 9 (a) Comparison results under the A0 working condition; Figure 9 (b) Comparison results under condition A1; Figure 9 (c) Comparison results under the A2 working condition; Figure 9 (d) Comparison results under the A3 working condition.

[0177] Figure 10 This diagram illustrates the comparison of small-sample generalization performance of different improved models based on the MCC5-THU dataset under a signal-to-noise ratio of -5dB, as provided by this invention. Figure 10 (a) Comparison results under the B0 operating condition; Figure 10(b) Comparison results under condition B1; Figure 10 (c) Comparison results under working condition B2; Figure 10 (d) Comparison results under working condition B3.

[0178] Figure 9 The corresponding results for the Ottawa dataset at a signal-to-noise ratio (SNR) of -10 dB are as follows. Figure 10 The results correspond to the MCC5-THU dataset with a signal-to-noise ratio (SNR) of -5dB. The comparison results show that the small-sample performance trends of all the compared models on both datasets are consistent, and the ADPCW-ELCNN model has a particularly significant advantage in generalization ability in scenarios with scarce samples.

[0179] In extremely limited sample sizes, with only 20 samples per class as an example: CNN, W-CNN, WDCNN, and DCA-BiGRU models, relying on large amounts of data for training, struggle to learn complete fault features, resulting in a significant performance decline. While the WKN model possesses some generalization ability thanks to wavelet prior knowledge, its F1 score is still significantly lower than the ADPCW-ELCNN model, failing to achieve high-precision diagnosis. In contrast, even with extremely limited sample sizes, the ADPCW-ELCNN model consistently maintains high-precision diagnosis, further demonstrating its superior generalization ability in small-sample scenarios.

[0180] To further illustrate the interpretability advantages of the ADPCW-ELCNN model of this invention, this embodiment compares the convolution kernel morphology learned by the six types of models and their corresponding time-frequency feature maps.

[0181] Figure 11 The waveforms of convolutional kernels from different improved models are shown. Traditional models, including CNN, WDCNN, and DCA-BiGRU, exhibit chaotic waveforms in their convolutional kernels; some kernels in the W-CNN model undergo waveform distortion during backpropagation; and the WKN model lacks sensitivity to impact features. In contrast, the ADPCW-ELCNN model, relying on the second derivative properties of the Mexican Hat wavelet and combining two-parameter adaptive learning and length co-optimization mechanisms, enables its convolutional kernels to dynamically adapt to the input signal, thus possessing stronger feature capture targeting capabilities.

[0182] Figures 12 to 14 The paper presents the time-frequency feature maps of three different faults learned by different models under the B0 condition of the MCC5-THU dataset, with a signal-to-noise ratio (SNR) of -5 and a sample size of Num=100. The specific analysis is as follows.

[0183] All three types of fault characteristics were affected by the meshing frequency of 474.1Hz and its second harmonic of 948.2Hz and third harmonic of 1422.3Hz. Among them, the time-frequency feature maps generated by the CNN model, W-CNN model, WDCNN model, and DCA-BiGRU model were severely affected by noise, only partially capturing the meshing frequency and its harmonic components, and completely failing to clearly identify the fault characteristic frequencies. While the WKN model responded to the fault characteristic frequencies, its ability to identify high-frequency faults, especially gear breakage faults, was insufficient. In contrast, the ADPCW-ELCNN model performed exceptionally well in terms of time-frequency feature maps: for gear pitting faults, its features showed a compact and regular cluster in the time-frequency domain; for tooth loss faults, its characteristic frequency of 138.4Hz was clearly identifiable; and for gear breakage faults, it formed transient impact regions with no significant noise interference at 286.3Hz and 1896.4Hz.

[0184] The performance differences among the models stem from the fault feature modeling capabilities of their underlying architectures: traditional CNN models lack prior knowledge of signal time and frequency, resulting in a blind feature extraction process; the wavelet convolution kernels in the W-CNN model are prone to distortion during backpropagation, making it difficult to effectively distinguish between noise and real signals; the Molet wavelet used in the WKN model has a low matching degree for high-frequency impact features and limited noise suppression capabilities; the multi-layer architecture of the WDCNN model is prone to over-smoothing features under strong noise, thus losing subtle fault features; the attention mechanism in the DCA-BiGRU model is difficult to effectively focus on fault features in a strong noise background, leading to feature redundancy.

[0185] In contrast, the ADPCW-ELCNN model, with its ability to accurately locate fault feature frequencies using the Mexican Hat wavelet function and the coordinated adjustment of both scale and length adaptive parameters, achieves clear extraction of fault features under extreme operating conditions.

[0186] In summary, under real-world operating conditions characterized by both strong noise and small sample sizes, the five comparative models, including the W-CNN and WKN models, struggle to accurately capture key fault features in conditions such as B0 of MCC5-THU. However, the ADPCW-ELCNN model proposed in this embodiment not only accurately extracts feature frequencies effective for fault classification but also possesses excellent small-sample generalization ability and strong model interpretability, thus providing a practical and effective new solution for intelligent fault diagnosis of equipment under small-sample operating conditions.

[0187] Example 2

[0188] Example 2 provides a fault diagnosis system based on ADPCW-ELCNN, applicable to the fault diagnosis method based on ADPCW-ELCNN in Example 1, including:

[0189] The data acquisition module is used to acquire fault vibration signals of rotating machinery;

[0190] The fault diagnosis module is used to input the fault vibration signal into the trained ADPCW-ELCNN model to obtain the fault diagnosis classification results of the rotating machinery.

[0191] The ADPCW-ELCNN model employs a single-layer convolutional architecture and includes:

[0192] The wavelet convolutional layer is used to extract multi-scale features based on the fault vibration signal and the wavelet convolutional kernel with scale and length correlation to obtain the first feature map.

[0193] The squared nonlinear mapping layer is used to perform a squared nonlinear mapping operation on the first feature map to obtain the second feature map;

[0194] A global average pooling layer is used to perform a global average pooling operation on the second feature map to obtain the first feature vector;

[0195] The first fully connected layer is used to perform feature transformation on the first feature vector to obtain the second feature vector;

[0196] The second fully connected layer is used to perform feature classification on the second feature vector to obtain the fault diagnosis classification result.

[0197] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied 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.

[0198] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0199] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0200] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0201] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A fault diagnosis method based on ADPCW-ELCNN, characterized by: include: Acquire fault vibration signals of rotating machinery; The fault vibration signal is input into the trained ADPCW-ELCNN model to obtain the fault diagnosis classification result of the rotating machinery; The ADPCW-ELCNN model employs a single-layer convolutional architecture and includes: The wavelet convolutional layer is used to perform multi-scale feature extraction based on the fault vibration signal and the wavelet convolutional kernel with scale and length correlation to obtain the first feature map. A squared nonlinear mapping layer is used to perform a squared nonlinear mapping operation on the first feature map to obtain a second feature map; A global average pooling layer is used to perform a global average pooling operation on the second feature map to obtain a first feature vector; The first fully connected layer is used to perform feature transformation on the first feature vector to obtain the second feature vector; The second fully connected layer is used to perform feature classification on the second feature vector to obtain the fault diagnosis classification result.

2. The fault diagnosis method based on ADPCW-ELCNN according to claim 1, characterized in that, The multi-scale feature extraction includes: Calculate the scale and length of each wavelet convolution kernel; The corresponding wavelet convolution kernel is determined based on the scale and length of each wavelet convolution kernel; Based on multiple wavelet convolution kernels, one-dimensional convolution operations are performed on the fault vibration signal to obtain multiple first feature maps.

3. The fault diagnosis method based on ADPCW-ELCNN according to claim 2, characterized in that, The expression for the scale of the wavelet convolution kernel is as follows: ； In the formula, This represents the scale of the i-th wavelet convolution kernel; Indicates the number of wavelet convolution kernels; Indicates the sequential index of the wavelet convolution kernel; The basic scale is expressed as follows: ； In the formula, This represents the minimum scale of the wavelet convolution kernel; This represents the maximum scale of the wavelet convolution kernel; This indicates a learnable global scale factor.

4. The fault diagnosis method based on ADPCW-ELCNN according to claim 2, characterized in that, The expression for the scale of the wavelet convolution kernel is as follows: ； In the formula, This represents the length of the i-th wavelet convolution kernel; This represents the length adjustment factor of the i-th wavelet convolution kernel; This represents the maximum length of the wavelet convolution kernel; This represents the minimum length of the wavelet convolution kernel; This represents the floor function; This represents the base length of the wavelet convolution kernel. , This represents the learnable energy range coefficient. Indicates the basic scale. This represents the function for rounding up.

5. The fault diagnosis method based on ADPCW-ELCNN according to claim 1, characterized in that, The expression for the squared nonlinear mapping operation is as follows: ； In the formula, This represents the z-th eigenvalue in the second feature map output by the squared nonlinear mapping layer; Represents the eigenvalues The result of performing a squared nonlinear mapping operation; This represents the z-th feature value in the first feature map output by the wavelet convolutional layer.

6. The fault diagnosis method based on ADPCW-ELCNN according to claim 1, characterized in that, The expression for the fault diagnosis classification result is as follows: ； In the formula, This indicates the fault diagnosis classification results; Indicates a second fully connected layer; Indicates the first fully connected layer; Indicates global average pooling; Represents a squared nonlinear mapping; Represents convolution; This represents the wavelet convolution kernel.

7. The fault diagnosis method based on ADPCW-ELCNN according to claim 1, characterized in that, The training steps for the ADPCW-ELCNN model are as follows: A sample set of fault vibration signals of rotating machinery is obtained, and the sample set is preprocessed to construct a training set; Based on the training set, the pre-built ADPCW-ELCNN model is iteratively trained using gradient descent and backpropagation algorithms to optimize the model parameters of the ADPCW-ELCNN model. The training continues until the preset training termination condition is met, at which point the trained ADPCW-ELCNN model is output. The gradient descent and backpropagation algorithm includes: The training samples of the training set are input into the ADPCW-ELCNN model to obtain the fault diagnosis classification training results; The loss function is calculated based on the fault diagnosis classification training results and the true labels of the training samples. Calculate the gradient of the loss function with respect to each parameter of the model; wherein, for the wavelet convolutional layer, calculate the gradient of the loss function with respect to the scale and length of each wavelet convolutional kernel respectively; Based on the gradient, the gradient descent algorithm, which uses a global-local parameter adaptation mechanism, updates the parameters of the model.

8. The fault diagnosis method based on ADPCW-ELCNN according to claim 7, characterized in that, The expression for the gradient of the loss function with respect to the scale parameter of the wavelet convolution kernel is as follows: ； In the formula, This represents the gradient of the loss function with respect to the scale of the i-th wavelet convolution kernel; Represents the partial differential operator; Represents the loss function; This represents the scale of the i-th wavelet convolution kernel; This represents the second feature map output by the squared nonlinear mapping layer; This represents the first feature map output by the wavelet convolutional layer.

9. The fault diagnosis method based on ADPCW-ELCNN according to claim 7, characterized in that, The expression for the gradient of the loss function with respect to the length of the wavelet convolution kernel is as follows: ； In the formula, This represents the gradient of the loss function with respect to the length of the i-th wavelet convolution kernel; Represents the partial differential operator; Represents the loss function; This represents the length of the i-th wavelet convolution kernel; This represents the second feature map output by the squared nonlinear mapping layer; This represents the first feature map output by the wavelet convolutional layer.

10. A fault diagnosis system based on ADPCW-ELCNN, applicable to the fault diagnosis method based on ADPCW-ELCNN as described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire fault vibration signals of rotating machinery; The fault diagnosis module is used to input the fault vibration signal into the trained ADPCW-ELCNN model to obtain the fault diagnosis classification result of the rotating machinery. The ADPCW-ELCNN model employs a single-layer convolutional architecture and includes: The wavelet convolutional layer is used to perform multi-scale feature extraction based on the fault vibration signal and the wavelet convolutional kernel with scale and length correlation to obtain the first feature map. A squared nonlinear mapping layer is used to perform a squared nonlinear mapping operation on the first feature map to obtain a second feature map; A global average pooling layer is used to perform a global average pooling operation on the second feature map to obtain a first feature vector; The first fully connected layer is used to perform feature transformation on the first feature vector to obtain the second feature vector; The second fully connected layer is used to perform feature classification on the second feature vector to obtain the fault diagnosis classification result.

Images