A large motor bearing fault diagnosis method based on vibration signal analysis

Abstract

The application belongs to the field of rotating machinery fault diagnosis and signal processing technology, and relates to a large motor bearing fault diagnosis method based on vibration signal analysis. The application aims at the problem of difficult feature extraction under strong noise and small sample, and through sliding window slicing, standardization and second-order difference enhancement on the original vibration signal, a double-channel input signal is spliced. A candidate kernel library containing four kinds of wavelets is constructed, and through population initialization, sub-sampling fast evaluation and disturbance iteration, the optimal wavelet kernel parameter set is adaptively obtained and instantiated as a physical convolution kernel with frozen weights. After convolution on the input signal, five statistical features are extracted to form a fixed-length vector, which is input into a multilayer perceptron classifier to output the fault category. The application replaces the random kernel with the physically interpretable wavelet kernel, freezes the front-end feature extraction layer and only trains the rear-end light classifier, thereby reducing the trainable parameters while maintaining strong noise resistance and robustness.

Description

Technical Field

[0001] This invention belongs to the field of rotating machinery fault diagnosis and signal processing technology, and relates to a method for fault diagnosis of large motor bearings based on vibration signal analysis. Background Technology

[0002] Rolling bearings, as core components of large motors, operate under harsh conditions such as high temperature and heavy load for extended periods, making them prone to localized damage and failure. Vibration signal analysis is the most commonly used technique for bearing fault diagnosis. However, vibration signals in actual industrial settings are often affected by strong background noise, and the cost of acquiring labeled fault samples is high, resulting in weak fault characteristics and a severe shortage of effective training data.

[0003] Currently, bearing fault diagnosis methods are mainly divided into two categories: the first category is the traditional signal processing and feature engineering method, which relies on Fourier transform, wavelet transform, empirical mode decomposition and other means to extract manual statistical features. However, the feature extraction of this type of method depends on manual experience, has insufficient generalization ability, and is difficult to fully characterize weak fault information under strong noise and small sample conditions.

[0004] The second category comprises intelligent diagnostic methods based on deep learning, represented by convolutional neural networks, which can achieve end-to-end feature extraction and classification. However, existing deep learning solutions still have the following shortcomings: First, most existing deep learning models use randomly initialized convolutional kernels, lacking prior signal processing knowledge. According to matched filtering theory, the more similar the convolutional kernel is to the fault feature waveform, the higher the output signal-to-noise ratio. However, bearing faults are mostly one-sided decaying transient pulses or steady-state resonant modulations, and random convolutional kernels cannot directly match these specific waveforms, resulting in insufficient feature extraction capability under strong noise.

[0005] Secondly, obtaining fault samples in industrial settings is difficult, and signals are often accompanied by strong background noise. Existing models have a large number of parameters and lack physical constraints, making them prone to overfitting or having features overwhelmed by noise under small sample and high-noise conditions, leading to a decrease in diagnostic accuracy.

[0006] Third, in order to improve diagnostic performance, existing technologies often expand the model capacity by stacking network layers or blindly increasing the number of convolutional kernels, which leads to a surge in computational costs and introduces a large number of redundant features, severely reducing the inference efficiency of the model.

[0007] Fourth, the feature map length generated by traditional time-frequency analysis methods such as continuous wavelet transform varies with the scale parameter, making it difficult to directly connect to downstream classifiers with fixed dimensions. In addition, conventional signal preprocessing methods are unable to effectively highlight weak transient impact components in the context of strong noise, further limiting the upper limit of model performance. Summary of the Invention

[0008] In view of this, in order to solve the problems mentioned in the background technology, a fault diagnosis method for large motor bearings based on vibration signal analysis is proposed.

[0009] The objective of this invention can be achieved through the following technical solution: A fault diagnosis method for large motor bearings based on vibration signal analysis, comprising: S1, acquiring the original vibration signal of the large motor bearing, and performing sliding window slicing, standardization and second-order differential enhancement preprocessing to obtain a dual-channel input signal composed of a reference signal sequence and its corresponding differential signal sequence pairing.

[0010] S2. Construct a candidate kernel library containing multiple wavelet basis functions and optimize it using a tournament search strategy. This includes: randomly generating an initial kernel parameter population based on the candidate kernel library; quickly evaluating individuals in the population through data subsampling and a ridge classifier; and iteratively screening by parameter perturbation to obtain the optimal wavelet kernel parameter set containing the wavelet kernel type ratio, the total number of kernels, and the range of each kernel window length.

[0011] S3. The optimal wavelet kernel parameter set is sampled at equal intervals according to the wavelet basis function to generate a one-dimensional vector as the initial weight of the physical convolution kernel and frozen. The dual-channel input signal is convolved with the physical convolution kernel. Five statistical features are extracted from each convolution output sequence: positive value proportion, maximum positive span, global mean, positive value mean and standard deviation. These features are then concatenated to form a fixed-length feature vector. The physical convolution kernel is a convolution kernel with a one-dimensional vector as the initial weight.

[0012] S4. Input the fixed-length feature vector into the multilayer perceptron classifier. With the physical convolution kernel weights frozen, only update the network weights of the multilayer perceptron classifier to obtain the hybrid diagnostic model.

[0013] S5. Input the vibration signal of the large motor bearing to be diagnosed into the hybrid diagnostic model and output the fault category of the large motor bearing.

[0014] In some implementations, the sliding window slicing, normalization, and second-order difference enhancement preprocessing specifically include: performing sliding window slicing, Z-Score normalization, and second-order difference enhancement preprocessing on the original vibration signal in sequence; the sliding window slicing is: sliding along the time axis of the original vibration signal with a fixed window length and a fixed step size to obtain a sample sequence of fixed length, and the overlap rate between adjacent samples is controlled by the step size;

[0015] In some implementations, the second-order difference enhancement preprocessing is as follows: calculating the difference signal sequence of the sample sequence, wherein the difference is a second-order difference; using the standardized sample sequence as the reference signal sequence, and splicing the reference signal sequence and the difference signal sequence along the channel dimension to form a dual-channel input signal.

[0016] In some implementations, the candidate kernel library includes four wavelet basis functions: Laplace wavelet, Morlet wavelet, Hermitian wavelet, and Harmonic wavelet.

[0017] In some implementations, the step of randomly generating the initial kernel parameter population includes: constructing individuals representing high-dimensional parameter vectors, wherein the components of the high-dimensional parameter vectors include the proportion of wavelet kernel types, the total number of kernels, and the length range of each kernel window; sampling the proportion of wavelet kernel types using a Dirichlet distribution and sampling the total number of kernels using a uniform distribution, and filling the sampled values ​​into the corresponding components of the high-dimensional parameter vectors; combining multiple high-dimensional parameter vectors filled with sampled values ​​to form an initial candidate pool, and performing deduplication processing on the initial candidate pool to obtain the initial kernel parameter population.

[0018] In some implementations, the step of rapidly evaluating individuals in the population through data subsampling and a ridge classifier includes: multiplying a scaling factor by the total number of samples in the training set to obtain the sample size; randomly selecting data from the training set that constitutes the sample size to form a sample subset; generating corresponding physical convolution kernels based on the initial kernel parameters and individual parameters in the population; performing convolution operations on the sample subset using the physical convolution kernels; performing pooling operations on the results to generate a feature matrix; inputting the feature matrix into a ridge classifier; the ridge classifier constructs a residual sum of squares objective function with a regularization term; solving for the weight matrix by minimizing the objective function; multiplying the feature matrix by the weight matrix to output the predicted label; comparing the predicted label with the true label to calculate the initial accuracy score for the current individual; obtaining the ratio of the number of kernels for the current individual to a preset upper limit for the number of kernels; multiplying the ratio by a penalty coefficient to obtain a kernel number penalty term; subtracting the kernel number penalty term from the initial accuracy score to calculate the difference; and using this difference as the fitness value for the individual.

[0019] In some implementations, the step of iterative screening with parameter perturbation includes: using the initial kernel parameter population as the current population; sorting the fitness values ​​of each individual in the current population in descending order, and selecting a set number of individuals with the highest ranking as the coarse screening result; performing convolution, pooling, and ridge classifier operations on each individual in the coarse screening result using the full training set data, and calculating the fitness value corresponding to each individual in the fine evaluation stage; randomly selecting a set number of individuals from the current population to form a selection group, comparing the fitness values ​​of each individual in the selection group, and determining the individual with the highest fitness value as the winner of this round; applying Gaussian perturbation to the high-dimensional parameter vector of the winner to generate offspring individuals, adding the offspring individuals to the current population, and removing the individual with the lowest fitness value in the current population to complete the population update; repeating the selection of winners, generation of offspring individuals, and population update until the set number of iterations is reached.

[0020] In some implementations, the step of obtaining the optimal wavelet kernel parameter set includes: obtaining the individual with the largest fitness value in the current population at the time of iteration termination as the optimal recipe; extracting the ratio of wavelet kernel types, the total number of kernels, and the range of window lengths of each kernel contained in the optimal recipe, and combining them to form the optimal wavelet kernel parameter set.

[0021] In some implementations, the five statistical characteristics of each convolutional output sequence—positive value proportion, maximum positive span, global mean, positive value mean, and standard deviation—specifically include: the positive value proportion is the proportion of positive points in the convolutional output sequence to the total number of points; the maximum positive span is the maximum length of a continuous positive value segment in the convolutional output sequence; the global mean is the arithmetic mean of all elements in the convolutional output sequence; the positive value mean is the arithmetic mean of the positive elements in the convolutional output sequence; and the standard deviation is the standard deviation of all elements in the convolutional output sequence.

[0022] In some implementations, the multilayer perceptron classifier sequentially comprises an input layer, a first hidden layer, a second hidden layer, and an output layer: the dimension of the input layer is equal to the dimension of the fixed-length feature vector; the first hidden layer and the second hidden layer each employ the ReLU6 activation function; the number of neurons in the output layer is equal to the number of fault categories, and the output layer employs the Softmax function to output the probability distribution of each category.

[0023] In some implementations, when iteratively updating the network weights of the multilayer perceptron classifier: the cross-entropy function is used as the loss function, and the Adam algorithm is used for optimization; during backpropagation, the gradient masking technique is used to set the weight gradients of each convolution kernel in the front-end physical convolutional layer to zero, and only the gradients of the network weights of each layer of the multilayer perceptron classifier are calculated and updated.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The present invention replaces the traditional randomly initialized convolution kernel with a wavelet basis function with a clear physical matching filter prior, so that the waveform of the convolution kernel is highly matched with the transient impact and steady-state resonance modulation characteristics of bearing faults. Validated by the SQ dataset, the average diagnostic accuracy reached 99.54% under strong noise conditions with a signal-to-noise ratio of -5dB, and the expected calibration error ECE was as low as 0.0002, confirming the significant advantage of this physical prior design in weak feature extraction.

[0025] (2) This invention freezes the weights of the front-end physical convolutional kernels and trains only the shallow back-end MLP network, which significantly reduces the number of trainable parameters compared to traditional CNNs, effectively avoiding the risk of overfitting under small sample conditions. Under the stringent conditions of a signal-to-noise ratio of -10dB and only 30 training samples per class in the CWRU dataset, the average diagnostic accuracy still reaches 90.92%, with a standard deviation of only 0.95%, verifying the robustness and stability of the model under scarce samples and strong noise interference.

[0026] (3) This invention abandons the traditional deep learning approach of improving performance by stacking network layers or blindly increasing convolutional kernels. Instead, it adopts a tournament search strategy to adaptively determine the optimal number of kernels and scale range, controlling feature redundancy from the source. Experiments show that a single training session takes about 6 seconds and inference takes about 3 seconds. While ensuring diagnostic accuracy, it significantly reduces computational overhead and model complexity, meeting the engineering requirements for diagnostic efficiency in industrial settings.

[0027] (4) This invention uses second-order difference enhancement preprocessing to highlight weak transient impact components, and uses a multidimensional statistical feature aggregation operator to transform the variable-length convolution output into a fixed-length feature vector, solving the technical problem that time-frequency analysis results are difficult to connect with fixed-dimensional classifiers. On the DDS complex system dataset containing bearing and gear composite faults, the average diagnostic accuracy reached 90.41% under a signal-to-noise ratio of 0dB, verifying the generalization ability and engineering practical value of this method in complex coupled mechanical systems. Attached Figure Description

[0028] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a flowchart of a large motor bearing fault diagnosis method based on vibration signal analysis in this invention;

[0030] Figure 2 This is a radar chart showing the performance of the method of this invention and various comparison models on the SQ dataset;

[0031] Figure 3 This is a visualization of the t-SNE features of the method of this invention and various comparative models on the SQ dataset;

[0032] Figure 4 This is a visualization of the t-SNE features of the method of this invention and various comparative models on the CWRU dataset;

[0033] Figure 5 This is a comparison chart of the parallel coordinate performance of the method of this invention and various comparative models on the DDS dataset. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0036] The following description, in conjunction with the accompanying drawings, details the specific scheme of the large motor bearing fault diagnosis method based on vibration signal analysis provided by the present invention.

[0037] Please see Figure 1 As shown, the present invention includes the following steps S1 to S5.

[0038] To address the technical problems in existing bearing fault diagnosis, such as the reliance on manual experience to set convolution kernel parameters and the lack of physical interpretability, as well as the difficulty of feature extraction and high computational cost of deep learning models under strong noise and small sample conditions, a hybrid diagnostic model construction scheme that integrates physical prior wavelet basis and evolutionary search strategy is proposed.

[0039] This invention replaces traditional randomly initialized convolutional kernels with wavelet basis functions having clear physical meanings to construct a candidate kernel library. A tournament search strategy is introduced, using data subsampling and a ridge classifier to rapidly evaluate the fitness of individuals in the population. Combined with iterative screening using parameter perturbation, the optimal set of wavelet kernel parameters is adaptively optimized for the kernel type ratio, total number of kernels, and the window length range of each kernel. The optimal parameters are instantiated as physical convolutional layers with their weights frozen, serving as a fixed feature extractor. Convolution operations are performed on the input dual-channel vibration signal, extracting five statistical features from the convolution output sequence and concatenating them into a fixed-length feature vector. This feature vector is then input into a multilayer perceptron classifier. During training, only the weights of the MLP network are updated, while the weights of the front-end convolutional layers remain frozen.

[0040] This scheme abandons the traditional mode of random initialization of convolutional kernels and end-to-end full update, and combines the advantages of wavelet physical matching filtering with lightweight network. It retains the physical interpretability and strong noise resistance of feature extraction, while greatly reducing the number of trainable parameters, and achieves high-precision, high-efficiency and robust fault diagnosis with small sample size.

[0041] S1. Obtain the original vibration signal of the large motor bearing, and perform sliding window slicing, normalization and second-order differential enhancement preprocessing to obtain a dual-channel input signal composed of a reference signal sequence and its corresponding differential signal sequence.

[0042] In this embodiment, an accelerometer is used to collect the original vibration signal of a large motor bearing. Since the vibration signal collected in industrial sites is usually a long sequence and contains a lot of redundancy and noise, the original vibration signal is first processed by sliding window slicing, Z-Score normalization, and differential enhancement in sequence: 1) Sliding window slicing: In order to increase the sample size and capture transient signals, the sliding window technique is used to split the long original vibration signal sequence into samples of fixed length. Let the length of the original vibration signal sequence be... Sample length The default setting is 2048 data points, with a step size of [missing value]. Then the i-th sample can be represented as: .

[0043] in, Indicates the first The starting index of each sample in the original sequence; This indicates the end index (excluding this position), meaning the original sequence is truncated from the index at this point. arrive The continuous data points constitute the sample .

[0044] The overlap rate between adjacent samples is defined as: By setting the step size This allows for the increase in the number of samples while maintaining the continuity of signal features.

[0045] 2) Z-Score Normalization: Due to significant differences in vibration signal amplitude under different operating conditions, it is necessary to eliminate the influence of dimensions and give the signal a uniform scale. For the sliced ​​samples... The standardized result for: .

[0046] in, It is the arithmetic mean of all data points in the current sample. The standard deviation of the sample is 1. For a minimum value (e.g., 1×10) -12 This is used to prevent division by zero errors. The standardized signal... The distribution satisfies a mean of 0 and a standard deviation of 1.

[0047] 3) Differential Enhancement Processing: After completing the Z-Score normalization process, this invention further constructs a dual-channel input signal, assuming the normalized reference signal sequence is... Differential enhancement processing is performed. As a basic implementation method, first-order forward difference can be used. .

[0048] Furthermore, in a preferred embodiment of the present invention, in order to effectively remove the low-frequency trend term and highlight the secondary transient impact characteristics caused by the fault under strong background noise, a second-order forward differential operation is used to obtain a differential signal sequence. : .

[0049] Zero-padding is used at the boundaries to maintain the sequence length. Second-order differencing further suppresses slowly varying components in the signal, while sharp spikes caused by impulses are relatively amplified. Compared to zero-order or first-order differencing, second-order differencing results in the lowest model diagnostic variance and the best robustness, and is therefore the default configuration in this invention.

[0050] This invention abandons the traditional single-channel input or simple feature-level fusion method, and instead adopts a data-level channel splicing strategy. The reference signal sequence and the differential signal sequence are spliced ​​along the channel dimension to form a dual-channel input signal.

[0051] This dual-channel input design preserves the complete amplitude envelope and low-frequency steady-state components of the vibration signal through the reference channel, capturing the smoothness characteristics of bearing rotation. The differential channel, equivalent to a high-pass filter in the frequency domain, effectively suppresses low-frequency trend terms and relatively enhances high-frequency transient abrupt components caused by fault impacts. Parallel input of the two signals allows the subsequent feature extraction module to simultaneously acquire complementary information on global energy distribution and local abrupt changes, thereby improving the model's ability to detect weak fault pulses.

[0052] S2. Construct a candidate kernel library containing multiple wavelet basis functions and optimize it using a tournament search strategy. This includes: randomly generating an initial kernel parameter population based on the candidate kernel library; quickly evaluating individuals in the population through data subsampling and a ridge classifier; and iteratively screening by parameter perturbation to obtain the optimal wavelet kernel parameter set containing the wavelet kernel type ratio, the total number of kernels, and the range of each kernel window length.

[0053] This step simulates natural selection to efficiently optimize the recipe space, minimizing kernel redundancy while maintaining classification accuracy. Therefore, a tournament search strategy is used to optimize the wavelet kernel parameter space and obtain the optimal set of wavelet kernel parameters.

[0054] First, a candidate kernel library containing various wavelet basis functions is constructed. This library includes four wavelet basis functions: Laplace wavelet, Morlet wavelet, Hermitian wavelet, and Harmonic wavelet. The Laplace wavelet is essentially a unilaterally exponentially decaying oscillating function, and its waveform is highly consistent with the unilateral transient impact waveform caused by bearing failure. The Morlet wavelet is obtained by multiplying a complex exponential signal with a Gaussian window function, and it excels at capturing the resonant frequency band modulation components excited by the fault. The Hermitian wavelet, relying on the second-order Gaussian derivative structure, is more sensitive to signal peaks and singularities, complementing the Laplace wavelet. The Harmonic wavelet, constructed from sine and cosine waves, can accurately extract low-frequency steady-state components such as the fundamental frequency and its harmonics of the rotating shaft.

[0055] Subsequently, an initial kernel parameter population is generated based on the candidate kernel library. Let the population size be M, then M (e.g., M=100) distinct convolutional kernel combinations are randomly generated. This serves as the initial candidate pool. Each individual R... i The population is represented as a high-dimensional parameter vector, with components encompassing the wavelet kernel type ratio (mix), the total number of kernels (N), and the window length range (L) of each kernel. The population initialization process is defined as follows: .

[0056] Each parameter is randomly sampled from its probability distribution. Specifically, since the sum of the proportions must be 1, a Dirichlet distribution is used to sample the proportions of the wavelet kernel types, and a uniform distribution is used to sample the total number of kernels. The sampled values ​​are then filled into the corresponding components of the high-dimensional parameter vector. Multiple high-dimensional parameter vectors filled with sampled values ​​are combined to form an initial candidate pool, and deduplication is performed on the initial candidate pool to ensure that each recipe in the candidate pool is unique, thus obtaining the initial kernel parameter population.

[0057] Furthermore, since training each recipe on the full dataset would result in excessive computational overhead, a rapid evaluation of the population is performed using data subsampling and a ridge classifier. The specific steps are as follows: First, the sampling sample size is obtained by multiplying the scaling factor η (η=0.2 in this example) by the total number of samples in the training set. Data of this sampling sample size is then randomly selected from the training set to form a sample subset. Individuals in the population are extracted as the initial kernel parameters and used as the convolution kernel. The convolution kernel is then applied to the sample subset, and the result is pooled to generate a feature matrix. The feature matrix is ​​input into the ridge classifier, which constructs an objective function of residual sum of squares with L2 regularization: .

[0058] in, For the true labels corresponding to the sample subset, This is the weight matrix. The regularization coefficient is used. The weight matrix is ​​solved by minimizing the objective function. The predicted label is output by multiplying the feature matrix and the weight matrix. The predicted label is then compared with the true label to calculate the initial accuracy score for the current individual. .

[0059] To achieve a balance between accuracy and feature dimensionality, the number of kernels for the current individual is obtained. With the preset maximum number of cores The ratio of the two factors is compared with the penalty coefficient. Multiplying yields a kernel count penalty term, which is used to adjust the initial accuracy score. The difference is calculated by subtracting the core quantity penalty term, and this difference is used as the fitness value of the individual. The specific formula is as follows: .

[0060] Since the search space is huge and it is necessary to avoid getting trapped in local optima, iterative screening is carried out in conjunction with parameter perturbation. A two-stage screening and tournament selection mechanism is adopted, specifically: the initial kernel parameter population is used as the current population.

[0061] Coarse screening stage: Sort the fitness values ​​of each individual in the current population in descending order of size, and select the top k individuals as the coarse screening results (k=10 in this example).

[0062] Fine screening stage: For each individual in the coarse screening results, convolution operations, pooling operations, and ridge classifier classification are performed on the full training set data to calculate the fitness value of each individual in the fine evaluation stage. The best individual in the fine screening results is directly added to the current population, and the individual with the lowest fitness value in the current population is removed to maintain the population size.

[0063] Tournament Selection: q individuals (q=5 in this example) are randomly selected from the current population to form a selection group. The fitness values ​​of each individual in the selection group after the fine-tuning update are compared, and the individual with the highest fitness value is determined as the winner of this round. To generate offspring, apply a Gaussian perturbation to the high-dimensional parameter vector of the winner: Apply a Gaussian perturbation to the winner's parameters with a relatively high probability (e.g., 0.8) to generate offspring. .in, The disturbance amplitude (in this embodiment) ).

[0064] Since the sum of wavelet ratio components must always be 1, the total number of kernels must be a positive integer, and the window length must be within a fixed range, a validity correction is performed on the perturbed offspring individuals: the wavelet ratio components are normalized so that their sum is 1. The normalization formula is as follows: ,in The total number of wavelet types, For the first The wavelet-like proportion components; rounding the total number of kernels to the nearest integer; truncating the length of each window to its value range.

[0065] Add offspring individuals to the current population and remove the individual with the lowest fitness value in the current population to complete the population update. Repeat the process of selecting winners, generating offspring individuals, and updating the population until the set number of iterations T is reached (T=50 in this embodiment).

[0066] The individual with the highest fitness value in the current population at the time of iteration termination is taken as the optimal recipe. The ratio of wavelet kernel types, the total number of kernels, and the range of window lengths of each kernel contained in the optimal recipe are extracted and combined to form the optimal wavelet kernel parameter set.

[0067] It should be noted that the tournament search process described above is only an independent parameter optimization stage, aiming to determine the optimal kernel type ratio, total number of kernels, and window length range for each kernel from the parameter space. Changes in the total number of kernels during the search process are only used for fitness evaluation and do not affect the subsequent classifier structure construction. When the tournament search terminates, the total number of kernels N contained in the individual with the highest fitness value in the current population is fixed as the final optimal value, and subsequent feature dimensions and classifier structures are instantiated and trained based on this fixed N value.

[0068] S3. The optimal wavelet kernel parameter set is sampled at equal intervals according to the wavelet basis function to generate a one-dimensional vector as the initial weight of the physical convolution kernel and frozen. The dual-channel input signal is convolved with the physical convolution kernel. Five statistical features are extracted from each convolution output sequence: positive value proportion, maximum positive span, global mean, positive value mean and standard deviation. These features are then concatenated to form a fixed-length feature vector. The physical convolution kernel is a convolution kernel with a one-dimensional vector as the initial weight.

[0069] Specifically, the obtained optimal wavelet kernel parameter set is instantiated as the initial weight values ​​for the convolutional layers in the neural network. This involves sampling at equal intervals along the time axis based on the wavelet kernel function, with the number of sampling points equal to the window length of the convolutional kernel. The function values ​​at each time point are arranged in chronological order to form a one-dimensional vector, which represents the initial weight values ​​for the physical convolutional kernel. For complex wavelets (such as Morlet wavelets), the real part is taken as the convolutional kernel weights.

[0070] In this invention, the physical convolutional layer instantiated by the optimal wavelet kernel parameter set is defined as the Evolved Wavelet Kernel Layer (hereinafter referred to as the E-wake Layer). During subsequent model training, the weights of the E-wake Layer remain fixed, i.e., frozen, and do not participate in gradient updates.

[0071] Let the input vibration signal be X, and let the E-wake layer contain the j-th physical convolution kernel obtained by optimization in step S2. The output of this layer is the feature map. It is obtained by convolution operation between the signal and the corresponding physical convolution kernel: ;in, This represents the convolution operation. This refers to all the parameters of the convolution kernel (including wavelet type, scale range, and window length, etc.).

[0072] Because the weights of this layer are frozen during training, i.e., satisfying... (L is the total loss function), this E-wake layer functions solely as a fixed physical feature extractor. This design significantly reduces the number of trainable parameters while preserving the physical feature representation capability of the wavelet kernel.

[0073] Because the window lengths of each physical convolution kernel are different, the lengths of the output sequences after the convolution operation vary, making direct concatenation into a fixed-length vector impossible. Therefore, this invention designs a set of multidimensional statistical feature aggregation operators. For the sequence output by this convolution operation... Further extract the following 5 features to form the feature vector. (1) Positive proportion (PPV, The proportion of positive values ​​in the convolution output sequence reflects the frequency of waveform matching between the signal and the wavelet basis function.

[0074] (2) Maximum positive span (Max Stretch, ): The maximum length of consecutive positive values ​​in the convolution output sequence, characterizing the periodic stability of fault impacts and reflecting the regularity of impact intervals.

[0075] (3) Global Mean : The overall energy level of the convolution output, which measures the overall correlation between the signal and the kernel function.

[0076] (4) Positive Mean ): The average intensity is calculated only for the local waveforms that are successfully matched.

[0077] (5) Standard deviation (Std, The degree of dispersion of the convolution response reflects the intensity of the signal energy distribution.

[0078] This 5-dimensional feature aggregation operator deconstructs the physical convolution output from three dimensions: statistical distribution, polarity bias, and energy fluctuation. Ultimately, if N optimal wavelet kernels are evolved, the feature dimension input to the subsequent classifier is 5N, which are then concatenated to form a fixed-length feature vector.

[0079] S4. Input the fixed-length feature vector into the multilayer perceptron classifier. With the physical convolution kernel weights frozen, only update the network weights of the multilayer perceptron classifier to obtain the hybrid diagnostic model.

[0080] Since the front-end E-wake layer has already extracted high-quality, linearly separable features, the back-end classifier does not require an overly deep network structure. This invention designs a shallow multilayer perceptron (MLP) containing three fully connected layers. The specific structural parameters are as follows: Input layer: Dimension equal to the dimension of the fixed-length feature vector. Since each physical convolutional kernel extracts 5 statistical features; when the number of optimal wavelet kernels obtained through the tournament search strategy is N, the 5 statistical features corresponding to each kernel are concatenated to form a fixed-length feature vector with a total dimension of 5N. Therefore, the number of neurons in this input layer is set to 5N.

[0081] First hidden layer: 256 neurons, ReLU6 activation function.

[0082] Second hidden layer: 128 neurons, ReLU6 activation function.

[0083] Output layer: The number of neurons equals the number of fault categories. Softmax is used to output the probability distribution, specifically: .

[0084] in, These are the trainable weight matrices for the first hidden layer, the second hidden layer, and the output layer, respectively. For the corresponding bias vector, It is a fixed-length eigenvector.

[0085] Subsequently, the hybrid diagnostic model was trained using the standard cross-entropy loss function: .

[0086] Where M represents the batch size. For true labels represented by one-hot encoding, The model is used to predict probabilities. The optimizer uses the Adam algorithm, which dynamically adjusts the learning rate using first and second moment estimates to accelerate convergence. The initial learning rate is set to 0.001, and the batch size is set to 32.

[0087] During backpropagation, gradient masking is used to zero out the weight gradients of each convolutional kernel in the front-end physical convolutional layer, and gradients are only calculated and updated for the network weights of each layer of the multilayer perceptron classifier. This frozen kernel + shallow MLP design significantly reduces model complexity and training difficulty, and is suitable for small sample scenarios.

[0088] S5. Input the vibration signal of the large motor bearing to be diagnosed into the hybrid diagnostic model and output the fault category of the large motor bearing.

[0089] The preprocessed signal to be diagnosed is input into the trained hybrid diagnostic model. The model propagates forward and outputs the probability distribution of the fault category. The category with the highest probability is taken as the final diagnostic result.

[0090] Finally, to verify the effectiveness of the method of this invention, experiments were conducted on three bearing datasets: SQ, CWRU, and DDS. The total number of original samples in the three datasets exceeded 15,000, 2,000, and 16,000, respectively. To simulate the actual working conditions where fault samples are scarce in industrial settings, 150 samples were randomly selected from each fault category to construct an experimental subset, which was then divided into training and test sets at a ratio of 2:8. During the training phase, the number of effective training samples for each category was controlled to be 30, and the remaining 120 samples were all used as unseen test data to verify the generalization accuracy of the model.

[0091] Meanwhile, to verify the anti-interference capability of the invention under strong background noise, Gaussian white noise with signal-to-noise ratios of 0dB, -5dB, and -10dB was added to the test set, respectively. All experiments were repeated 10 times, and the final results are presented in the form of "mean ± standard deviation" to eliminate the random bias caused by a single random division.

[0092] Comparative studies show that, on the SQ dataset with an SNR of -5dB, the present invention achieves an average accuracy of 99.54% and an ECE of 0.0002; on the CWRU dataset with an SNR of -10dB, the accuracy reaches 90.92% and the ECE is 0.0010; and on the DDS dataset with an SNR of 0dB, the accuracy reaches 90.41%. These results demonstrate that the present invention outperforms the comparative models in diagnostic accuracy, noise robustness, and confidence calibration (ECE), while also requiring a moderate training time.

[0093] Figure 2 This is a radar chart showing the performance metrics of the proposed method and various benchmark models on the SQ dataset. The Chinese names and English abbreviations of the models in the chart are as follows: Time-Frequency Network (TFN), Multi-Scale Wavelet Kernel Network (MWDFN), Multi-Scale Random Convolutional Feature Network (MultiROCKET), Lightweight Random Convolutional Feature Network (MiniROCKET), Wavelet Perceptron Network (CWT_MLP), Linear Convolutional Transformer (LiConvFormer), Multi-Channel Self-Attention Transformer (MCSAT), and the proposed hybrid diagnostic model (E-wake-MLP) is referred to as the proposed model. It should be noted that, regarding the previously defined comparison model names, subsequent... Figures 3-5 The same model names appearing in the text are all represented by their English names.

[0094] In the graph, all metrics (average accuracy, training time, testing time, expected calibration error (ECE), F1 score, and recall) are normalized to the [0, 1] interval. The training time, testing time, and ECE metrics are inverted (i.e., smaller values ​​correspond to higher scores), ensuring that all metrics on the radar chart consistently show that larger values ​​represent better performance. Figure 2 It can be intuitively seen that the method of this invention (E-wake-MLP) forms the largest envelope area around the radar image. Its accuracy, F1 score and recall are close to the outer circle extreme value. At the same time, it also maintains a significant advantage in training time, testing time and ECE. This shows that the method has no significant shortcomings in multiple dimensions such as diagnostic accuracy, noise resistance, confidence calibration and computational efficiency. Its overall performance is better than all the comparison models.

[0095] like Figure 3 and Figure 4 As shown, t-SNE is used to visualize and analyze the results of the second-to-last output layer of the model. On the SQ and CWRU datasets, comparisons are made with benchmark models (MCSAT, LiConvFormer, CWT_MLP, MiniROCKET, MultiROCKET, MWDFN, TFN, Supervised Contrastive Learning Network (SUPCON), Convolutional-Attention Hybrid Network (CNN_Transf), Residual Wavelet Kernel Network (Reswkn), etc.). The visualization results show that sample points of each category overlap to varying degrees, without forming obvious category clusters. However, the visualization results of the method in this invention show clear separation boundaries between different fault categories, and sample points of each category form compact clusters, confirming that the evolutionary wavelet kernel-multilayer perceptron hybrid diagnostic model constructed in this invention has excellent feature discrimination ability.

[0096] Furthermore, to verify the comprehensive diagnostic performance of this invention in complex coupled mechanical systems containing combined bearing and gear faults, comparative experiments were conducted on the DDS dataset, and the results are as follows: Figure 5 As shown in the figure, this graph uses a parallel coordinate plot to comprehensively compare the normalized performance of each model across six dimensions: accuracy, training time, testing time, expected calibration error (ECE), F1 score, and recall. The training time, testing time, and ECE metrics have been inverted so that all metrics on the parallel coordinate plot show a consistent trend where larger values ​​represent better performance.

[0097] from Figure 5It can be intuitively seen that the proposed method (E-wake-MLP) maintains high values ​​across all indicator dimensions, and its curves exhibit small cross-vertical inflections on each coordinate axis, indicating a significant positive correlation among the evaluation indicators and reflecting the balanced overall performance of the model. Specifically, under a signal-to-noise ratio (SNR) of 0 dB, the proposed method achieves an average diagnostic accuracy of 90.41%, with an expected calibration error (ECE) as low as 0.0038, and both the F1 score and recall rate reach 90.39%, significantly outperforming comparative models such as TFN, MWDFN, MultiROCKET, MiniROCKET, LiConvFormer, and SUPCON. Particularly noteworthy is that when the SNR drops to -5 dB, the diagnostic accuracy of the comparative models shows a significant decrease, while the proposed method still maintains an accuracy of 79.79%, fully validating the strong robustness and engineering practical value of the proposed method under complex transmission paths and multi-component coupled noise interference.

[0098] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0099] Those skilled in the art will recognize that the algorithmic steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0100] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0101] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0102] Finally, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for fault diagnosis of bearings in large electric motors based on vibration signal analysis, characterized in that, include: S1. Obtain the original vibration signal of the large motor bearing, and perform sliding window slicing, normalization and second-order differential enhancement preprocessing to obtain a dual-channel input signal composed of a reference signal sequence and its corresponding differential signal sequence. S2. Construct a candidate kernel library containing multiple wavelet basis functions and optimize it using a tournament search strategy. This includes: randomly generating an initial kernel parameter population based on the candidate kernel library; quickly evaluating individuals in the population through data subsampling and a ridge classifier; and iteratively screening with parameter perturbation to obtain the optimal wavelet kernel parameter set containing the wavelet kernel type ratio, the total number of kernels, and the range of each kernel window length. S3. The optimal wavelet kernel parameter set is sampled at equal intervals according to the wavelet basis functions to generate a one-dimensional vector as the initial weight of the physical convolution kernel and frozen. The dual-channel input signal is convolved with the physical convolution kernel. Five statistical features are extracted from each convolution output sequence: the proportion of positive values, the maximum positive span, the global mean, the positive mean, and the standard deviation. These features are then concatenated to form a fixed-length feature vector. The physical convolution kernel is a convolution kernel with a one-dimensional vector as the initial weight. S4. Input the fixed-length feature vector into the multilayer perceptron classifier. Under the condition that the physical convolution kernel weights are kept frozen, only the network weights of the multilayer perceptron classifier are iteratively updated to obtain the hybrid diagnostic model. S5. Input the vibration signal of the large motor bearing to be diagnosed into the hybrid diagnostic model and output the fault category of the large motor bearing. The steps for rapid evaluation of individuals in the population using data subsampling and a ridge classifier include: The sampling sample size is obtained by multiplying the scaling factor by the total number of samples in the training set. Data of this sampling sample size are randomly selected from the training set to form a sample subset. Based on the individual parameters in the initial kernel parameter population, generate corresponding physical convolution kernels, perform convolution operations on a subset of samples using the physical convolution kernels, and perform pooling operations on the operation results to generate a feature matrix; The feature matrix is ​​input into the ridge classifier, which constructs a residual sum of squares objective function with a regularization term. The weight matrix is ​​solved by minimizing the objective function. The feature matrix and the weight matrix are multiplied to output the predicted label. The predicted label is compared with the true label to calculate the initial accuracy score of the current individual. Obtain the ratio of the current number of kernels to the preset upper limit of the number of kernels. Multiply the ratio by the penalty coefficient to obtain the kernel number penalty term. Subtract the kernel number penalty term from the initial accuracy score to calculate the difference. Use this difference as the fitness value of the individual.

2. The method for fault diagnosis of large motor bearings based on vibration signal analysis as described in claim 1, characterized in that, The sliding window slicing, normalization, and second-order difference enhancement preprocessing specifically include: The original vibration signal was preprocessed sequentially by sliding window slicing, Z-Score normalization, and second-order difference enhancement. The sliding window slicing is performed by sliding along the time axis of the original vibration signal with a fixed window length and a fixed step size to obtain a sample sequence of fixed length, and the overlap rate between adjacent samples is controlled by the step size. The second-order difference enhancement preprocessing is as follows: calculate the difference signal sequence of the sample sequence, wherein the difference is second-order difference; take the standardized sample sequence as the reference signal sequence, and splice the reference signal sequence and the difference signal sequence along the channel dimension to form a dual-channel input signal.

3. The method for diagnosing bearing faults in large electric motors based on vibration signal analysis as described in claim 1, characterized in that, The candidate kernel library contains four wavelet basis functions: Laplace wavelet, Morlet wavelet, Hermitian wavelet, and Harmonic wavelet.

4. The method for diagnosing bearing faults in large electric motors based on vibration signal analysis as described in claim 1, characterized in that, The step of randomly generating the initial kernel parameter population includes: Construct an individual representing a high-dimensional parameter vector. The components of the high-dimensional parameter vector include the ratio of wavelet kernel types, the total number of kernels, and the range of window lengths for each kernel. Dirichlet distribution is used to sample the proportion of wavelet kernel types, and uniform distribution is used to sample the total number of kernels. The sampled values ​​are then filled into the corresponding components of the high-dimensional parameter vector. Multiple high-dimensional parameter vectors filled with sampled values ​​are combined to form an initial candidate pool. The initial candidate pool is then deduplicated to obtain the initial kernel parameter population.

5. The method for diagnosing bearing faults in a large electric motor based on vibration signal analysis as described in claim 4, characterized in that, The iterative screening step involving the perturbation of the matching parameters includes: Use the initial kernel parameter population as the current population; Sort the fitness values ​​of each individual in the current population in descending order of size, and select a set number of individuals with the highest ranking as the coarse screening result; For each individual in the coarse screening results, convolution, pooling and ridge classifier operations are performed on the full training set data to calculate the fitness value of each individual in the fine evaluation stage. A set number of individuals are randomly selected from the current population to form a selection group. The fitness values ​​of each individual in the selection group are compared, and the individual with the highest fitness value is determined as the winner in the corresponding selection group. Apply a Gaussian perturbation to the high-dimensional parameter vector of the winner to generate offspring individuals. Add the offspring individuals to the current population and remove the individual with the lowest fitness value in the current population to complete the population update. Repeat the process of selecting winners, generating offspring individuals, and updating the population until the set number of iterations is reached.

6. The method for diagnosing bearing faults in a large electric motor based on vibration signal analysis as described in claim 5, characterized in that, The steps for obtaining the optimal wavelet kernel parameter set include: The individual with the highest fitness value in the current population at the time of iteration termination is taken as the optimal recipe; Extract the proportion of wavelet kernel types, the total number of kernels, and the window length range of each kernel in the optimal formula, and combine them to form the optimal wavelet kernel parameter set.

7. The method for diagnosing bearing faults in large electric motors based on vibration signal analysis as described in claim 1, characterized in that, The five statistical features of each convolutional output sequence—the proportion of positive values, the maximum positive span, the global mean, the mean of positive values, and the standard deviation—specifically include: The positive value ratio is the proportion of positive points in the convolution output sequence out of the total number of points. The maximum positive span is the maximum length of a consecutive positive segment in the convolution output sequence; The global mean is the arithmetic mean of all elements in the convolution output sequence; The positive mean is the arithmetic mean of the positive elements in the convolution output sequence; The standard deviation is the standard deviation of all elements in the convolution output sequence.

8. The method for diagnosing bearing faults in large electric motors based on vibration signal analysis as described in claim 1, characterized in that, The multilayer perceptron classifier sequentially comprises an input layer, a first hidden layer, a second hidden layer, and an output layer: The dimension of the input layer is equal to the dimension of the fixed-length feature vector; The first hidden layer and the second hidden layer both use the ReLU6 activation function; The number of neurons in the output layer is equal to the number of fault categories, and the output layer uses the Softmax function to output the probability distribution of each category.

9. The method for fault diagnosis of large motor bearings based on vibration signal analysis as described in claim 1, characterized in that, When iteratively updating the network weights of the multilayer perceptron classifier: The Adam algorithm is used to perform optimization, with the cross-entropy function as the loss function. During backpropagation, gradient masking is used to set the weight gradients of each convolutional kernel in the front-end physical convolutional layer to zero, and only the gradients of the network weights of each layer of the multilayer perceptron classifier are calculated and updated.

Images