A deep learning-based device predictive maintenance method

Abstract

The application relates to the field of equipment diagnosis, and specifically discloses a device predictive maintenance method based on deep learning, which comprises the following steps: after time-domain amplitude normalization and frequency-domain weighted normalization are performed on training data, two-channel feature fusion is performed to obtain a fusion feature vector; a plurality of adaptive wavelet kernels are used to extract features from the fusion feature vector to obtain an adaptive time-frequency feature vector; a weight population is constructed through statistical characteristics to screen the optimal initial weight of an initial diagnosis model; the initial diagnosis model is iteratively updated through multi-scale deep feature calculation on the adaptive time-frequency feature vector, time-frequency domain attention feature fusion, fault prototype comparison learning and a pre-constructed total loss function to obtain a diagnosis model; and the diagnosis model is used to diagnose the equipment. The multi-scale feature fusion and double-path attention mechanism can cooperatively capture short-time impact and long-period patterns, and overcome the limitations of traditional methods in diversified fault scenarios.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of equipment diagnosis, and in particular to a device predictive maintenance method based on deep learning. BACKGROUND

[0002] In recent years, with the large-scale, complex and continuous operation of modern industrial equipment, higher requirements are put forward for the reliability and safety of equipment operation. Traditional equipment maintenance mode depends on periodic maintenance or after-service, which not only increases the maintenance cost, but also may cause shutdown due to sudden failure, resulting in serious economic loss. In order to realize early fault warning and health state evaluation of equipment, predictive maintenance has gradually become a research hotspot in the industry.

[0003] The prior art has the following disadvantages:

[0004] 1. Global Min-Max or single-domain normalization is commonly used, which can easily weaken the fault-sensitive frequency band features, so that the features are overwhelmed by noise, and the diagnosis accuracy is low.

[0005] 2. The traditional wavelet transform relies on manual preset parameters, cannot flexibly adapt to diversified fault modes, and leads to insufficient extraction of key time-frequency features.

[0006] 3. Randomly initialized weights can easily cause unstable convergence, fixed convolution kernel scales cannot take into account multi-time scale features, and traditional LSTM treats irrelevant noise equally, which reduces the diagnosis effect.

[0007] 4. The conventional classifier has low discrimination for similar fault categories and poor adaptability to new categories; at the same time, fixed regularization strength cannot take into account the needs of early and late training, and is easy to cause overfitting or underfitting. SUMMARY

[0008] In order to solve the above problems, the present application provides a device predictive maintenance method based on deep learning, wherein the method comprises:

[0009] After the training data is subjected to time-domain amplitude normalization and frequency-domain weighted normalization, dual-channel feature fusion is performed to obtain a fusion feature vector; the Morlet wavelet base function parameters are optimized, and a plurality of adaptive wavelet kernels are used to extract features from the fusion feature vector to obtain an adaptive time-frequency feature vector; a weight population is constructed through the statistical characteristics of the adaptive time-frequency feature vector, and the optimal initial weight of the initial diagnosis model is screened in combination with the differential evolution algorithm; the initial diagnosis model is iteratively updated through multi-scale deep feature calculation, time-frequency domain attention feature fusion, fault prototype comparison learning and pre-constructed total loss function on the adaptive time-frequency feature vector to obtain a diagnosis model; and the diagnosis model is used to diagnose the fault of the equipment to be tested.

[0010] In one example, before the time-domain amplitude normalization and the frequency-domain weighted normalization are performed on the training data, the method further comprises: collecting original vibration signals of the device in a running state at a preset sampling frequency within a preset time length through a preset vibration sensor preset on the device, and synchronously recording device working condition parameters as initial sample data; the device working condition parameters include at least one of a rotating speed, a load, and a temperature; and labeling the initial sample data in combination with a device running log, a disassembly and repair report, and an oil analysis result, wherein the labeling categories include a fault mode and a health mode, and the fault mode includes at least one of a rotor imbalance, an axis misalignment, a bearing inner ring damage, a bearing outer ring damage, a gear tooth breakage, and a gear wear.

[0011] In one example, after the time-domain amplitude normalization and the frequency-domain weighted normalization are performed on the training data, dual-channel feature fusion is performed, specifically including: performing amplitude normalization processing on the original vibration signals to obtain a time-domain normalized feature vector; performing fast Fourier transform on the original vibration signals to obtain an amplitude spectrum; performing weighting on the amplitude spectrum through a frequency weighting matrix to suppress high-frequency noise and enhance a fault-sensitive frequency band; performing logarithmic operation on the weighted result to compress the amplitude range and improve the feature separability to obtain a frequency-domain normalized feature; and performing a splicing operation on the time-domain normalized feature vector and the frequency-domain normalized feature vector to generate a fusion feature vector.

[0012] In one example, the Morlet wavelet basis function parameters are optimized, specifically including: based on a standard Morlet wavelet basis function, combining a scale parameter to control wavelet stretching, a translation parameter to control wavelet time shift, and a window function weight parameter to adjust the steepness of a Sigmoid window to construct a parameterized adaptive wavelet basis function.

[0013] In one example, the fusion feature vector is subjected to feature extraction using multiple sets of adaptive wavelet kernels to obtain an adaptive time-frequency feature vector, specifically including: calculating the energy of each adaptive wavelet basis function, and converting the wavelet kernel energy into a weighted coefficient through normalization processing; performing convolution operation on the fusion feature vector using multiple sets of adaptive wavelet kernels, weighting and summing the convolution results according to the energy weight coefficients after taking the modulus, and outputting an initial adaptive time-frequency feature vector; taking the variance of improving the feature distribution of different fault categories as an optimization target, and dynamically optimizing the adaptive wavelet kernel parameter set by using an evolutionary algorithm to output the adaptive time-frequency feature vector; the adaptive wavelet kernel parameter set includes at least one of a wavelet kernel scale parameter, a translation parameter, and a window weight.

[0014] In one example, the constructing a weight population through statistical characteristics of the adaptive time-frequency feature vector and screening an optimal initial weight of the initial diagnosis model in combination with a differential evolution algorithm specifically includes: performing mean vector calculation on the adaptive time-frequency feature vector to obtain a feature mean vector; performing singular value decomposition on a covariance matrix of the adaptive time-frequency feature vector to obtain a left singular matrix and a singular value diagonal matrix; generating a diversity initial weight population based on the feature mean vector, the left singular matrix and the singular value diagonal matrix in combination with a random orthogonal matrix; performing differential mutation operation on a target individual whose individual loss value is greater than a benchmark individual loss value through a mutation factor and a plurality of individual weights selected randomly to obtain an evolved weight population; evaluating loss values of all individuals in the evolved weight population and selecting a weight corresponding to an individual with the minimum loss value as the optimal initial weight of the initial diagnosis model.

[0015] In one example, the performing multi-scale deep feature calculation on the adaptive time-frequency feature vector specifically includes: performing one-dimensional convolution operation and multi-dimensional convolution operation on the adaptive time-frequency feature vector using a plurality of sizes of convolution kernels respectively to extract a short-scale feature vector and a long-scale feature vector; splicing the short-scale feature vector and the long-scale feature vector and sequentially performing batch normalization processing and Gaussian error linear unit activation function operation to output a multi-scale fusion feature.

[0016] In one example, the performing multi-scale deep feature calculation on the adaptive time-frequency feature vector and time-frequency domain attention feature fusion specifically includes: converting the multi-scale fusion feature into a query matrix through a learnable query transformation matrix, processing the query matrix through a normalized exponential function and then weighting the time domain feature to output a time domain attention feature vector; performing fast Fourier transform on the multi-scale fusion feature to obtain a frequency domain representation, calculating a frequency domain attention weight through amplitude weighting and a normalized exponential function, and outputting a frequency domain attention feature vector through inverse fast Fourier transform; adding the time domain attention feature vector and the frequency domain attention feature vector and outputting a fusion time-frequency domain attention feature vector through layer normalization operation.

[0017] In one example, the performing multi-scale deep feature calculation on the adaptive time-frequency feature vector, time-frequency domain attention feature fusion and fault prototype contrast learning specifically includes: inputting the fusion time-frequency domain attention feature vector into a multi-layer perceptron projection network to convert it into a low-dimensional projection feature vector; weighting historical fault prototype vectors based on a momentum coefficient, combining the mean value of the low-dimensional projection feature vectors of the same type of samples in the current batch, and updating the prototype vectors of each type of fault; calculating the cosine distance between the sample projection feature and the real class prototype vector, combining the distance between the sample projection feature and the prototype vectors of other classes, and obtaining a prototype contrast loss through exponential function and natural logarithm operation.

[0018] In one example, the initial diagnosis model is iteratively updated by multi-scale deep feature calculation, time-frequency domain attention feature fusion, fault prototype contrast learning and pre-constructed total loss function on the adaptive time-frequency feature vector to obtain a diagnosis model, specifically including: on the basis of the basic loss and the prototype contrast loss, a total loss function is constructed by combining an L2 regularization term decaying over time, wherein the regularization strength exponentially decays with the increase of the training round; new parameter values are generated using a multivariate normal distribution, and the decay rate parameter of the dynamic regularization coefficient is optimized by a covariance matrix adaptive evolution strategy; parent individuals are selected according to the fitness ranking, and the mean vector of the evolution distribution is updated by weighted average; based on the total loss function, an adaptive matrix estimator optimizer is used to perform iterative update of the model parameters.

[0019] The method proposed in the application can bring the following beneficial effects:

[0020] 1. The time domain amplitude feature and the frequency domain weighted feature are spliced and fused, which not only retains the original vibration amplitude feature, but also enhances the fault sensitive frequency band feature, significantly improving the feature discrimination.

[0021] 2. The Morlet wavelet base function parameters are optimized based on the evolution algorithm, a plurality of wavelet kernels are dynamically constructed and weighted fused, and adaptive time-frequency feature capture for different fault modes is realized.

[0022] 3. The weight initialization strategy optimized by differential evolution, the multi-scale convolution fusion structure and the double-path attention mechanism are proposed, so that the model can stably converge and take into account short-time impact features and long-period patterns.

[0023] 4. The prototype representation and contrast loss function are updated by momentum to improve the small sample diagnosis capability, and a regularization strategy exponentially decaying with the training process is introduced, taking into account early learning and late overfitting prevention. BRIEF DESCRIPTION OF DRAWINGS

[0024] The drawings described herein are used to provide further understanding of the application, and form a part of the application. The illustrative embodiments of the application and their descriptions serve to explain the application, and do not constitute an improper limitation on the application. In the drawings:

[0025] Figure 1 a flowchart of a device predictive maintenance method based on deep learning in an embodiment of the application;

[0026] Figure 2 a schematic diagram of several original vibration signals in an embodiment of the application;

[0027] Figure 3 a schematic diagram of several conventional frequency spectrum analysis in an embodiment of the application;

[0028] Figure 4 Fig. 6 is a diagram illustrating the influence of different normalization methods on the accuracy of fault diagnosis in embodiments of the present application;

[0029] Figure 5 Fig. 7 is a diagram illustrating the diagnostic performance of different methods on various faults in embodiments of the present application. DETAILED DESCRIPTION

[0030] To make the objectives, technical solutions, and advantages of the present application clearer, the technical solutions of the present application will be described below in conjunction with specific embodiments of the present application and corresponding drawings. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0031] The technical solutions provided by the embodiments of the present application will be described in detail below with reference to the drawings.

[0032] Figure 1 Fig. 1 is a flowchart of a deep learning-based equipment predictive maintenance method provided by one or more embodiments of the present specification. The method can be applied to different types of industrial equipment, such as bearings, gearboxes, or motors, etc. The flowchart can be executed by a computing device (such as a control module or a cloud server, etc.) connected to the industrial equipment. Some input parameters or intermediate results in the flowchart allow manual intervention to adjust to help improve accuracy.

[0033] The implementation of the analysis method related to the embodiments of the present application can be a terminal device or a server, and the present application does not make special limitations thereon. For the convenience of understanding and description, the following embodiments are described in detail with the server as an example. It should be noted that the server can be a separate device or a system composed of multiple devices, i.e., a distributed server, and the present application does not make specific limitations thereon.

[0034] As shown in Figure 1 Fig. 1, the present application provides a deep learning-based equipment predictive maintenance method, which includes:

[0035] S101: After time-domain amplitude normalization and frequency-domain weighted normalization of the training data, double-channel feature fusion is performed to obtain a fusion feature vector.

[0036] First, the diagnostic model needs to be trained according to different types of equipment to be tested. When training the diagnostic model, first, the training data is obtained, and after time-domain amplitude normalization and frequency-domain weighted normalization of the training data, double-channel feature fusion is performed to obtain a fusion feature vector.

[0037] In one embodiment, when acquiring training data, a triaxial vibration sensor needs to be deployed in key parts of the equipment (such as bearings, gearboxes, or motor shaft ends) to continuously acquire the original vibration signal of the equipment under operating conditions at a sampling frequency of not less than 20kHz; the acquisition time of each initial sample data is 0.2 seconds, generating an original vibration signal vector with a dimension of 4096, and synchronously recording the equipment operating parameters (such as speed, load, and temperature).

[0038] When labeling the initial sample data, domain experts combined equipment operation logs, disassembly and maintenance reports, and oil analysis results to complete the labeling. The labeling categories cover six typical failure modes (including rotor imbalance, shaft misalignment, bearing inner ring damage, bearing outer ring damage, gear tooth breakage, and gear wear) and health status, forming a labeled sample library.

[0039] All annotation processes must ensure that the physical mechanisms of the fault modes and vibration characteristics strictly correspond. For example, the annotation of bearing inner ring damage must correspond to the characteristic frequency sideband modulation phenomenon. Finally, a standardized training dataset containing time-domain waveforms, operating parameters and multi-category fault labels is constructed.

[0040] In one embodiment, the original signals and their timely spectra of health status, bearing inner ring damage, and gear tooth breakage are as follows: Figure 2 As shown, conventional spectrum analysis is as follows: Figure 3 As shown, it is difficult to distinguish the characteristic frequencies of similar faults.

[0041] In one embodiment, vibration sensor monitoring data contains a mixture of high-frequency noise and low-frequency fault features. Conventional processing methods, such as Min-Max normalization, use global normalization to eliminate the influence of dimensions, but this weakens the characteristic representation of fault-related frequency bands, causing fault features to be submerged by noise and affecting the accuracy of subsequent diagnosis.

[0042] This invention calculates both time-domain and frequency-domain normalized feature vectors, and then merges them into a highly discriminative feature vector through a concatenation operation. The specific steps are as follows:

[0043] First, the original vibration signal is normalized in amplitude to obtain a time-domain normalized feature vector. Specifically, this is done by subtracting the minimum value of the vector and dividing by the difference between the maximum and minimum values, resulting in a time-domain normalized feature vector with values ​​ranging from 0 to 1. This process eliminates the influence of dimensions while preserving the amplitude characteristics of the original waveform, as shown below:

[0044]

[0045] in, This represents the original vibration signal vector, with a dimension of 4096; Representing vectors The minimum value; denotes the maximum value of the vector . is a time-domain normalized feature vector, and the value range is .

[0046] Then the original vibration signal is subjected to frequency domain weighted normalization to output a frequency domain normalized feature vector. Specifically, a fast Fourier transform is performed on the original vibration signal vector to obtain an amplitude spectrum, the amplitude spectrum is weighted by a frequency weighting matrix to suppress high-frequency noise and enhance fault-sensitive frequency bands, and a logarithmic operation is performed on the weighted result to compress the amplitude range and improve feature separability, and finally a frequency domain normalized feature vector is output, denoted as:

[0047]

[0048] In the formula, denotes a fast Fourier transform operation; denotes the amplitude spectrum of the fast Fourier transform result; is a frequency weighting matrix, defined as a diagonal matrix ; denotes an operator for constructing a diagonal matrix; denotes a-0.5 power operation on the first frequency component after Fourier transform; denotes a-0.5 power operation on the second frequency component after Fourier transform; denotes a-0.5 power operation on the 2048th frequency component after Fourier transform; is a logarithmic function with base 10, used to compress the amplitude range and improve feature separability; is a frequency domain normalized feature vector, with a dimension of 2048.

[0049] Then the time domain normalized feature vector and the frequency domain normalized feature vector need to be spliced to generate a fusion feature vector, denoted as:

[0050]

[0051] In the formula, denotes a vector splicing operation; is a fusion feature vector.

[0052] Different normalization methods are analyzed and compared to evaluate the performance difference between the frequency-time domain hybrid normalization method proposed in the application and traditional normalization techniques. For example Figure 4The comparison results of four normalization methods on fault diagnosis accuracy are shown: the global normalization as the conventional technology performs the weakest; the time domain normalization (only retains the original waveform amplitude feature) and the frequency domain normalization (only suppresses high frequency noise) each has an improvement but the effect is limited; and the mixed normalization method of the application significantly improves the diagnosis accuracy by fusing the time domain amplitude feature and the frequency domain weighted feature. It can be seen in the figure that the accuracy bar chart of the mixed normalization method is significantly higher than that of other methods, and the error bar length is the shortest, indicating that its diagnosis result stability is the best. The unit percentage of the vertical coordinate "diagnosis accuracy (%)" directly reflects the technical advantages of the method in eliminating the influence of dimension and enhancing the fault feature performance, solving the core problem that the fault feature is overwhelmed by noise in the conventional method.

[0053] S102: Optimize the Morlet wavelet base function parameters, use multiple adaptive wavelet kernels to extract features from the fusion feature vector to obtain an adaptive time-frequency feature vector.

[0054] Conventional wavelet transform needs to preset base function parameters such as scale and translation, which is difficult to adapt to the time-frequency feature expression requirement under variable fault modes, resulting in insufficient fault feature extraction and inability to effectively capture fault sensitive information.

[0055] The application dynamically optimizes the wavelet base function parameters through an evolutionary algorithm to construct an adaptive wavelet kernel group, and the specific steps are as follows: first, the Morlet wavelet base function parameters need to be optimized, specifically, based on the standard Morlet wavelet base function, the scale parameter is used to control the wavelet stretching, the translation parameter is used to control the wavelet time shift, and the window function weight parameter is used to adjust the steepness of the Sigmoid window, to construct a parameterized adaptive wavelet base function, which is represented as:

[0056]

[0057] In the formula, Adaptive wavelet base function is represented as Standard Morlet wavelet base function is represented as Scale parameter is represented as , control wavelet stretching Translation parameter is represented as Window function weight parameter is represented as Sigmoid activation function is represented as Time variable is represented as

[0058] After the wavelet basis function parameters are dynamically optimized by the evolutionary algorithm, a plurality of adaptive wavelet kernels are used to extract features from the fusion feature vector to obtain an adaptive time-frequency feature vector. At this time, the wavelet kernel energy weight needs to be calculated first, that is, the energy of each adaptive wavelet basis function is calculated, and the wavelet kernel energy is converted into a weighting coefficient through normalization processing, which is used for subsequent feature fusion and is expressed as:

[0059]

[0060] In the formula, denotes the normalized energy weight of the th wavelet kernel; denotes the energy of the th wavelet kernel, and the calculation method is expressed as ; denotes the energy of the th wavelet kernel; is the total number of wavelet kernels, such as ; denotes the th adaptive wavelet basis function; is the th adaptive wavelet basis function.

[0061] Then the time-frequency feature needs to be extracted, that is, a plurality of adaptive wavelet kernels are used to perform convolution operation on the fusion feature vector, and the adaptive time-frequency feature vector is output after the convolution result is weighted and summed according to the energy weight coefficient, which is expressed as:

[0062]

[0063] In the formula, is the adaptive time-frequency feature vector; denotes the convolution operation.

[0064] Finally, the wavelet basis parameters can be optimized. At this time, the evolutionary algorithm can be used to dynamically optimize the parameter set , so that the variance between fault classes is maximized. The optimization goal is to maximize the variance of the feature distribution of different fault classes, and the optimization parameters include the scale of the th wavelet kernel, the translation of the th wavelet kernel, and the window weight of the th wavelet kernel. Specifically, the standard evolutionary algorithm is used to update the parameter population through mutation, crossover and selection operations. The termination condition is to reach the preset evolution number or the fitness converges, and the adaptive time-frequency feature vector corresponding to the optimal parameter set is output .

[0065] S103: Construct a weight population through the statistical characteristics of the adaptive time-frequency feature vector, and combine the differential evolution algorithm to screen the optimal initial weight of the initial diagnosis model.

[0066] If the deep model adopts completely random weight initialization in the initial training stage, the training process will often be unstable and even unstable convergence due to the mismatch between the weight distribution and the input features, especially in the case of high-dimensional vibration features, random initialization is difficult to reflect the statistical law of the features, thereby prolonging the model convergence time and reducing the final diagnosis performance.

[0067] The application constructs a weight population by using the statistical characteristics of the adaptive time-frequency feature, and combines the differential evolution algorithm to screen the optimal initial weight, the specific steps are as follows:

[0068] First, the mean vector of the adaptive time-frequency feature vector is calculated, and the left singular matrix and the singular value diagonal matrix are obtained by performing singular value decomposition on the covariance matrix, which is represented as:

[0069]

[0070]

[0071] In the formula, indicates the mean value of the feature vector; is the feature mean vector; indicates the calculation of the feature covariance matrix; indicates the singular value decomposition operation; is the left singular matrix; is the singular value diagonal matrix.

[0072] Then, based on the feature mean vector, the left singular matrix and the singular value diagonal matrix, a diversity initial weight population is generated by combining a random orthogonal matrix, which is represented as:

[0073]

[0074] In the formula, indicates the initial weight matrix of the th individual, . is the th random orthogonal matrix; traverse all individual indexes, . is the population size, such as, .

[0075] Then, for the individual whose current individual loss value is greater than the benchmark individual loss value, the evolutionary weight is generated by differential mutation operation through the mutation factor and the three individual weights selected randomly, otherwise the original weight is retained, which is represented as:​​

[0076]

[0077] In the formula, Indicates the first The evolutionary weight of each individual; The index of the first body selected randomly; The index of the second body is randomly selected; The index of the third body is randomly selected; For example, as a variable factor, ; The loss function is calculated based on the cross-entropy loss of a 5% subset of data.

[0078] Finally, the loss values ​​of all individuals in the evolutionary weight population can be evaluated, and the weight corresponding to the individual with the smallest loss value can be selected as the optimal initial weight, expressed as:

[0079]

[0080] In the formula, The optimal initial weights; This indicates the search for parameters that minimize the objective function; For the first The loss value of each evolved individual.

[0081] S104: The initial diagnostic model is iteratively updated by performing multi-scale deep feature calculation, time-frequency domain attention feature fusion, fault prototype comparison learning, and a pre-constructed total loss function on the adaptive time-frequency feature vector to obtain the diagnostic model.

[0082] Traditional convolutional neural networks use fixed-size convolutional kernels, which can only extract features at a single scale. For vibration signals, there are both short-term impact features and long-period patterns. Single-scale convolutional kernels often cannot capture features at different time scales at the same time, resulting in insufficient ability of the model to represent features at multiple scales.

[0083] This application provides a dual-branch convolutional structure that extracts short-scale local features and long-scale global features separately, and further generates multi-scale deep features that combine local and global information. The specific steps are as follows:

[0084] When performing short-scale feature extraction, a small-sized convolution kernel can be used to perform one-dimensional convolution operations on the adaptive time-frequency feature vector to extract the short-scale feature vector, as shown below:

[0085]

[0086] In the formula, Represents short-scale eigenvectors; represents a one-dimensional convolution operation; is a convolution kernel size, such as, ; is a convolution step size, such as, .

[0087] When performing long-scale feature extraction, a large-size convolution kernel is used for a dilated convolution operation on the adaptive time-frequency feature vector, and the receptive field expansion degree is controlled by the dilated rate, and a long-scale feature vector is output, denoted as:

[0088]

[0089] In the formula, represents the output long-scale feature vector; represents a dilated convolution operation; is a dilated rate, such as, , controls the receptive field expansion degree.

[0090] Then, the short-scale feature vector and the long-scale feature vector are spliced, and batch normalization processing and Gaussian error linear unit activation function operation are sequentially performed, and a multi-scale fusion feature is output, denoted as:

[0091]

[0092] In the formula, is the output multi-scale fusion feature; represents batch normalization processing; is a Gaussian error linear unit activation function.

[0093] The traditional long short-term memory network treats all time or frequency components equally when processing feature sequences, but in the fault diagnosis scenario, key features are often concentrated in certain time segments or frequency components, and equal treatment will cause irrelevant noise to occupy a large proportion, thereby reducing the diagnosis accuracy.

[0094] The application introduces a double-channel attention mechanism on the basis of the fusion feature, and models the time domain and the frequency domain respectively, and the specific steps are as follows:

[0095] First, the time domain attention is calculated, specifically: the multi-scale fusion feature is converted into a query matrix through a learnable query transformation matrix, the self-attention weight is calculated combined with the key matrix and the value matrix, and the time domain attention feature is output after being weighted with the time domain feature after normalization exponential function processing, denoted as:

[0096]

[0097] In the formula, represents the time domain attention output feature; is a learnable query transformation matrix; is a query matrix, which is linearly transformed from input features, and the calculation mode is represented as is a key matrix; is a value matrix; is a scaling factor, such as , to prevent the dot product from being too large; represents a normalized exponential function; represents transpose of

[0098] Then frequency domain attention calculation is needed, specifically: the multi-scale fusion features are subjected to fast Fourier transform to obtain frequency domain representation, the frequency domain attention weight is calculated through amplitude weighting and normalized exponential function, and the frequency domain attention feature is output through inverse fast Fourier transform, represented as:

[0099]

[0100] In the formula, represents the frequency domain attention output feature; represents element-wise multiplication; represents inverse fast Fourier transform.

[0101] Then attention feature fusion is performed, specifically: the time domain attention feature and the frequency domain attention feature are added, and the fused time-frequency domain attention feature vector is output through layer normalization operation, represented as:

[0102]

[0103] In the formula, is a fused time-frequency domain attention feature vector; represents layer normalization operation.

[0104] In the scene of few-shot fault diagnosis, traditional classifiers are difficult to effectively distinguish similar fault modes, and have poor adaptability to new fault categories, which will lead to a significant decrease in recognition accuracy under small sample conditions.

[0105] To this end, the momentum update mechanism is used to construct the fault prototype representation, and the contrast loss function is used to pull the features of the same class samples and their corresponding prototypes closer, while pushing the distance of prototypes of different categories farther apart. First, feature projection transformation is needed

[0106] The fused time-frequency domain attention feature vector is input into the multi-layer perceptron projection network to be converted into a low-dimensional projection feature vector, represented as:

[0107]

[0108] In the formula, represents the first​ The fused time-frequency domain attention feature vector corresponding to each sample; This represents a multilayer perceptron projection network; For the first The low-dimensional projective features corresponding to each sample.

[0109] Then, the historical fault prototype vectors are weighted based on the momentum coefficient, and combined with the mean of the projected features of the current batch of similar samples, the prototype vector of each type of fault is updated, as follows:

[0110]

[0111] In the formula, Indicates the first Class of faults in The prototype vector updated in the next iteration; Indicates the first Class of faults in The prototype vector updated in the next iteration; For example, the momentum coefficient, Control the degree to which historical information is retained; Indicates that the current batch belongs to the first... A collection of sample indexes for a class; Indicates that the current batch belongs to the first... The number of samples in the sample index set of the class; For fault category index, ; This represents the total number of fault categories.

[0112] Finally, the contrastive loss is calculated, which involves calculating the cosine distance between the projected features of the sample and the true class prototype vector, combined with the distances to prototype vectors of other classes, and then using an exponential function and the natural logarithm to obtain the prototype contrastive loss, expressed as:

[0113]

[0114] In the formula, Indicates prototype contrast loss; Indicates the first The sample corresponds to the real number. Prototype vectors of each category; Indicates the first The sample corresponds to the first Prototype vectors of each category; Represents the cosine distance function; For example, temperature coefficient, Control the sharpness of the distribution; Represents an exponential function; It represents the natural logarithm.

[0115] Because in the model training process, if the fixed regularization strength is adopted, it is often difficult to meet the needs of different training stages. In the early stage, too strong constraint may inhibit the learning ability of the model, and in the later stage, too weak constraint may lead to overfitting. The fixed strategy is not good enough in the dynamic training process, especially in the vibration fault diagnosis.

[0116] Based on this, the total loss function is constructed based on the basic loss and prototype contrast loss combined with the L2 regularization term decaying with time. First, define the dynamic regularization loss function, specifically: based on the basic loss and prototype contrast loss, combined with the L2 regularization term decaying with time to construct the total loss function, wherein the regularization strength decays exponentially with the increase of training rounds, denoted as:

[0117]

[0118] In the formula, denotes the total loss function; denotes the L2 norm; is the model weight parameter; is the time-varying regularization coefficient, which decays with time, and the calculation method is denoted as ; denotes the current training round; denotes the total preset training round; is the initial regularization strength, such as ; is the minimum regularization strength, such as ; denotes the new parameter value of the th generation.

[0119] Then optimize the decay rate parameter, which needs to generate new parameter values using multivariate normal distribution, and optimize the decay rate parameter of dynamic regularization coefficient through covariance matrix adaptive evolution strategy, denoted as:

[0120]

[0121] In the formula, denotes the new parameter value of the th generation; denotes the multivariate normal distribution; is the distribution mean vector of the th generation; is the step parameter of the th generation; is the covariance matrix of the th generation; denotes the evolution generation index.

[0122] Finally, the evolutionary distribution parameters can be updated, i.e., the parent individuals are selected according to the fitness ranking, and the mean vector of the evolutionary distribution is updated by weighted average, denoted as:

[0123]

[0124] wherein, is the distribution mean vector of the th generation, and represents the parameter distribution center of the th generation calculated by weighted average; is the distribution mean vector of the th generation, and represents the parameter distribution center of the th generation calculated by weighted average; is the th individual in the th generation ranked by fitness, is the th individual in the th generation ranked by fitness, represents the index ranked by fitness in the population; represents the number of selected parents; represents the th weight coefficient. In the process of model iterative updating and parameter training, the adaptive matrix estimation optimizer can be used to perform iterative updating of the model parameters based on the total loss function described above. In each training round, first, the preprocessed vibration feature samples are loaded in small batches, and the multi-scale fusion features, attention enhanced features and low-dimensional projection features are calculated by forward propagation. Then, the sum of the basic classification loss, the prototype contrast loss and the dynamic regularization loss is calculated as the total loss. The gradient of the total loss with respect to all weight parameters of the model is calculated by the back propagation algorithm, and the gradient clipping technique is applied to limit the gradient norm to stabilize the training process. The optimizer dynamically adjusts the learning rate of each parameter by combining the first-order moment estimation and the second-order moment estimation of the gradient, and simultaneously incorporates a dynamic weight decay mechanism to adapt to the time-varying regularization constraint.

[0125] During the training process, the fault diagnosis accuracy and loss value on the validation set are continuously monitored. If the validation loss does not improve for consecutive multiple rounds, the early stopping mechanism is triggered. Finally, the model parameter snapshot with the optimal performance on the validation set is saved, and the end-to-end training of the deep learning model is completed.

[0126] In one embodiment, the performance analysis of multi-fault type diagnosis is performed, and the performance of the model is evaluated by using the confusion matrix as shown in

[0127]

[0128] In one embodiment, the performance analysis of multi-fault type diagnosis is performed, and the performance of the model is evaluated by using the confusion matrix as shown in Figure 5 ​​The illustrated grouped column chart verifies the generalization ability of the application on seven typical fault types (including healthy state). The horizontal coordinate lists all fault types, and the vertical coordinate "diagnostic accuracy (%)" quantifies the diagnostic performance. The traditional support vector machine method (left column) is obviously weak in complex faults such as "gear tooth breakage"; the conventional deep learning method (middle column) has overall improvement, but still has shortcomings in "bearing outer ring damage" and other categories; the application method (right column) maintains the highest column height in all fault types, especially in "bearing inner ring damage" (which requires identification of characteristic frequency sideband modulation) and "gear wear" (long period pattern), and the experimental results show that the multi-scale feature fusion and dual-path attention mechanism can cooperatively capture short-term impact and long-period pattern, overcoming the limitations of traditional methods in diversified fault scenarios.

[0129] S105: Diagnose the fault of the to-be-tested equipment through the diagnostic model.

[0130] The trained deep learning model is deployed to an edge computing terminal or a cloud diagnosis platform. In the online diagnosis stage, the equipment vibration signal is collected in real time, and S1 data collection, S2 frequency-time domain mixed normalization, S3 adaptive time-frequency feature extraction and S4 deep learning model forward inference are sequentially performed, outputting the fault category probability distribution and confidence score. The diagnostic result is mapped to a specific fault mode, such as "bearing outer ring damage, probability 92%", and a health state index is generated synchronously.

[0131] The maintenance decision system generates the optimal maintenance strategy according to the fault type, such as "replace the bearing", and pushes it to the equipment management system and links to the spare parts inventory, realizing closed-loop management from condition monitoring, fault diagnosis to maintenance scheduling.

[0132] Each of the embodiments in the present application is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other. Each embodiment mainly explains the difference from other embodiments. Especially, for the device and medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the related parts can be referred to the part of the method embodiment.

[0133] The device and medium provided by the embodiments of the present application are one-to-one corresponding to the method, so the device and medium also have similar beneficial technical effects as the corresponding method. Since the beneficial technical effects of the method have been described in detail above, the beneficial technical effects of the device and medium will not be described here.

[0134] Those skilled in the art will appreciate that embodiments of the present application can be readily used as software, hardware, or a combination of software and hardware. In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0135] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more functions specified in the flowchart block or blocks. Figure 1 means for functionally implementing one or more functions specified in the flowchart block or blocks.

[0136] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more functions specified in the flowchart block or blocks. Figure 1 means for functionally implementing one or more functions specified in the flowchart block or blocks.

[0137] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more functions specified in the flowchart block or blocks. Figure 1 means for functionally implementing one or more functions specified in the flowchart block or blocks.

[0138] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0139] The memory can include non-persistent memory and / or volatile memory, such as random access memory (RAM) and / or cache memory, non-volatile memory, such as read-only memory (ROM), optical storage, and / or flash memory. The memory is an example of computer readable storage media.

[0140] Computer-readable media includes permanent and non-permanent, movable and non-movable media that can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic tape disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information that can be accessed by a computing device. According to the definition herein, computer-readable media does not include transitory media such as modulated data signals and carriers.

[0141] It should also be noted that the terms "comprising", "containing", or any other variant thereof are intended to cover non-exclusive inclusions, so that a process, method, article or apparatus that includes a list of elements does not only include those elements, but also includes other elements not explicitly listed, or further includes elements inherent in such a process, method, article or apparatus. Without more limitations, the element defined by the statement "comprising a" does not exclude the presence of additional identical elements in the process, method, article or apparatus that includes the element.

[0142] The above only describes the embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application shall be included in the scope of claims of the present application.

Claims

1. A deep learning-based device predictive maintenance method, characterized by, The method comprises the following steps: After time-domain amplitude normalization and frequency-domain weighted normalization are performed on the training data, dual-channel feature fusion is performed to obtain a fusion feature vector; Optimize the Morlet wavelet basis function parameters, and use multiple sets of adaptive wavelet kernels to extract features from the fusion feature vector to obtain an adaptive time-frequency feature vector; Construct a weight population based on the statistical characteristics of the adaptive time-frequency feature vector, and combine a differential evolution algorithm to screen the optimal initial weight of the initial diagnosis model; Iteratively update the initial diagnosis model through multi-scale deep feature calculation, time-frequency domain attention feature fusion, fault prototype comparison learning, and a pre-constructed total loss function on the adaptive time-frequency feature vector to obtain a diagnosis model; Diagnose the fault of the to-be-tested equipment through the diagnosis model; The optimization of the Morlet wavelet basis function parameters specifically includes: Based on the standard Morlet wavelet basis function, control the wavelet stretching by combining the scale parameter, control the wavelet time shift by combining the shift parameter, and adjust the steepness of the Sigmoid window by combining the window function weight parameter to construct a parameterized adaptive wavelet basis function; The use of multiple sets of adaptive wavelet kernels to extract features from the fusion feature vector to obtain an adaptive time-frequency feature vector specifically includes: Calculate the energy of each adaptive wavelet basis function, and convert the energy of each wavelet kernel into a weight coefficient through normalization processing; Convolve the fusion feature vector using multiple sets of adaptive wavelet kernels, weight and sum the convolution results according to the energy weight coefficients after taking the modulus, and output the initial adaptive time-frequency feature vector; The variance of improving the feature distribution of different fault categories is taken as the optimization target, and an evolutionary algorithm is used to dynamically optimize the adaptive wavelet kernel parameter set to output the adaptive time-frequency feature vector; The adaptive wavelet kernel parameter set includes at least one of the wavelet kernel scale parameter, the shift parameter, and the window weight; The construction of the weight population based on the statistical characteristics of the adaptive time-frequency feature vector, and the combination of the differential evolution algorithm to screen the optimal initial weight of the initial diagnosis model specifically includes: Calculate the mean vector of the adaptive time-frequency feature vector to obtain a feature mean vector; Perform singular value decomposition on the covariance matrix of the adaptive time-frequency feature vector to obtain a left singular matrix and a singular value diagonal matrix; Based on the feature mean vector, the left singular matrix, and the singular value diagonal matrix, generate a diversity initial weight population combined with a random orthogonal matrix; For a target individual whose individual loss value is greater than a benchmark individual loss value, perform a differential mutation operation on the mutation factor and a plurality of individual weights selected randomly to obtain an evolutionary weight population; Evaluate the loss values of all individuals in the evolutionary weight population, and select the weight corresponding to the individual with the smallest loss value as the optimal initial weight of the initial diagnosis model; The multi-scale deep feature calculation and time-frequency domain attention feature fusion on the adaptive time-frequency feature vector specifically includes: Convert the multi-scale fusion feature into a query matrix through a learnable query transformation matrix, process the query matrix through a normalized exponential function, and output a time domain attention feature vector by weighting the time domain feature. The multi-scale fusion feature is subjected to fast Fourier transform to obtain a frequency domain representation, the frequency domain attention weight is calculated through amplitude weighting and a normalized exponential function, and the frequency domain attention feature vector is output through inverse fast Fourier transform; The time domain attention feature vector and the frequency domain attention feature vector are added, and a fusion time-frequency domain attention feature vector is output through layer normalization operation; The adaptive time-frequency feature vector is subjected to multi-scale deep feature calculation, time-frequency domain attention feature fusion, and fault prototype contrast learning, specifically including: The fusion time-frequency domain attention feature vector is input into a multi-layer perceptron projection network to be converted into a low-dimensional projection feature vector; The historical fault prototype vectors are weighted based on momentum coefficients, and the prototype vectors of each type of fault are updated in combination with the mean value of the low-dimensional projection feature vectors of the same type of sample in the current batch; The cosine distance between the sample projection feature and the real class prototype vector is calculated, and the distance between the sample projection feature and the prototype vectors of other classes is combined to obtain a prototype contrast loss through an exponential function and a natural logarithm operation; The initial diagnosis model is iteratively updated based on the adaptive time-frequency feature vector, multi-scale deep feature calculation, time-frequency domain attention feature fusion, fault prototype contrast learning, and a pre-constructed total loss function to obtain a diagnosis model, specifically including: On the basis of the basic loss and the prototype contrast loss, a total loss function is constructed in combination with an L2 regularization term that decays over time, wherein the regularization strength exponentially decays with the increase of the training round; A new parameter value is generated using a multivariate normal distribution, and the decay rate parameter of the dynamic regularization coefficient is optimized through a covariance matrix adaptive evolution strategy; Parent individuals are selected according to the fitness ranking, and the mean vector of the evolution distribution is updated through weighted averaging; Based on the total loss function, an adaptive matrix estimator optimizer is used to perform iterative updating of the model parameters.

2. The method of claim 1, wherein, Before the time domain amplitude normalization and the frequency domain weighted normalization of the training data, the method further includes: Collecting, through a preset vibration sensor on the equipment, an original vibration signal in a preset time length at a preset sampling frequency under the running state of the equipment, and synchronously recording equipment operating parameters as initial sample data; the equipment operating parameters include at least one of rotating speed, load, and temperature; Labeling the initial sample data in combination with equipment operation logs, disassembly and repair reports, and oil analysis results; the labeling categories include fault modes and health modes, and the fault modes include at least one of rotor imbalance, shaft misalignment, bearing inner ring damage, bearing outer ring damage, gear tooth breakage, and gear wear.

3. The method of claim 1, wherein, After the time domain amplitude normalization and the frequency domain weighted normalization of the training data, performing double-channel feature fusion, specifically including: Performing amplitude normalization processing on the original vibration signal to obtain a time domain normalized feature vector; Performing fast Fourier transform on the original vibration signal to obtain an amplitude spectrum; Weighting the amplitude spectrum through a frequency weighting matrix to suppress high-frequency noise and enhance fault-sensitive frequency bands; Performing logarithmic operation on the weighted result to compress the amplitude range and improve feature separability to obtain a frequency domain normalized feature; The time-domain normalized feature vector is spliced with the frequency-domain normalized feature vector to generate a fusion feature vector.

4. The method of claim 1, wherein, The multi-scale deep feature calculation on the adaptive time-frequency feature vector specifically includes: One-dimensional convolution operation and multi-dimensional convolution operation are respectively performed on the adaptive time-frequency feature vector using convolution kernels of multiple sizes to extract a short-scale feature vector and a long-scale feature vector. The short-scale feature vector and the long-scale feature vector are spliced, and batch normalization processing and Gaussian error linear unit activation function operation are sequentially performed, and a multi-scale fusion feature is output.

Images