Physically interpretable wavelet fusion method for motor bearing fault diagnosis

Abstract

This invention relates to a physically interpretable wavelet fusion method for motor bearing fault diagnosis, belonging to the field of motor fault diagnosis. This technical solution applies a unified wavelet basis function to each sensor channel through shared continuous wavelet convolution for time-frequency mapping, enabling heterogeneous signals to obtain a consistent physical representation scale at the feature level. This avoids the semantic misalignment problem of traditional convolution kernels between different physical quantities, providing an interpretable common basis for subsequent cross-sensor feature association. For scenarios with missing sensors, this solution employs a collaborative representation mechanism to constrain the deep semantic features of effective channels to the same joint representation space, ensuring that the reconstruction process of missing channels is limited by multi-source physical coupling laws rather than being freely generated. In the feature fusion stage, this solution uses an adaptive gating mechanism to model the spatial complementarity between sensors and a multi-head attention mechanism to capture long-term temporal evolution dependencies, representing the two perspectives independently before fusion.

Description

Technical Field

[0001] This invention relates to the field of motor fault diagnosis technology, and in particular to a physically interpretable wavelet fusion method for diagnosing motor bearing faults. Background Technology

[0002] As a core power component of industrial drive systems, the health of asynchronous motor bearings directly impacts the continuous and stable operation of production lines. Currently, multi-sensor information fusion technology is widely used for condition monitoring and fault diagnosis of asynchronous motor bearings. This technology simultaneously collects heterogeneous signals such as vibration, current, and temperature, characterizing the equipment's operating conditions from multiple physical dimensions, including mechanical dynamics, electromagnetic properties, and thermodynamics. Compared to single-signal analysis, multi-source data fusion fully utilizes the spatiotemporal correlation and information complementarity between different sensor signals, effectively overcoming the limitations of a single perspective and significantly improving the ability to identify complex faults and early, subtle fault characteristics. It has become the mainstream technical solution in the field of intelligent operation and maintenance.

[0003] However, in the complex electromagnetic environment and long-term service conditions of actual industrial sites, multi-source monitoring data often exhibits random missing or incomplete states due to factors such as sensor aging and failure, abnormal acquisition terminals, communication link interruptions, or data packet loss. Directly inputting such incomplete data into traditional fusion models will disrupt the inherent coupling and correlation structure between multiple sensors, leading to a shift in the feature space distribution and a sharp deterioration in diagnostic performance. Existing data completion methods, such as generative adversarial networks, autoencoders, or simple interpolation, mostly aim to restore the original time-domain waveform. These methods not only have difficulty in training and convergence and have high computational costs, but also easily introduce spurious noise or retain redundant information unrelated to the fault during reconstruction. They are difficult to stably extract and retain key fault features with class discriminative power under unknown and missing modes, and cannot meet the comprehensive requirements of diagnostic accuracy, robustness, and physical interpretability in industrial sites. Summary of the Invention

[0004] The purpose of this invention is to provide a physically interpretable wavelet fusion method for diagnosing motor bearing faults, so as to solve the problems existing in the prior art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A physically interpretable wavelet fusion method for diagnosing motor bearing faults includes: Multi-source sensor signals are collected during the operation of the asynchronous motor. The sensor signals are then time-synchronized, segmented, and normalized to construct a complete multi-source data sample. Incomplete multi-source data inputs are established based on sensor failure, abnormal acquisition, or communication packet loss, and the validity of the input channels is detected. The complete multi-source data sample or the incomplete multi-source data is input into the shared continuous wavelet convolution feature extraction module. Time-frequency feature mapping is performed on different sensor channels using wavelet convolution kernels with shared parameters to obtain multi-source wavelet features with unified time-frequency resolution and physical meaning. Missing patterns are identified using a missing feature reconstruction module based on collaborative representation. All valid sensor wavelet features are encoded into a deep semantic space and aggregated to obtain a multi-source joint representation. The decoder then reconstructs the complete multi-source wavelet features from the multi-source joint representation. The reconstruction process of incomplete multi-source data paths is supervised by feature knowledge generated from complete multi-source data paths. The complete multi-source wavelet features or the complete multi-source wavelet features restored by the missing feature reconstruction module are input into the dual-view feature fusion network. The spatial correlation between sensors is modeled by the multi-sensor fusion branch based on the adaptive gating mechanism, and the long-term temporal dependency is modeled by the multi-time series fusion branch and the multi-head attention mechanism. The features from the two perspectives are fused to obtain the comprehensive fault characterization. The fault comprehensive characterization input parameter shared convolutional neural network classifier extracts high-level fault semantic features through multi-level convolution, pooling and nonlinear activation, and then outputs the motor bearing health status or fault category through a fully connected layer. A multi-objective joint optimization strategy is adopted to train the fault diagnosis model. The classification loss, the distillation loss between complete and incomplete multi-source data paths, the missing feature reconstruction loss, and the cooperative representation coordination loss are jointly constructed into the overall objective function. The network parameters are updated through backpropagation to obtain a motor bearing fault diagnosis model that can handle unknown sensor missing modes.

[0006] In some implementations, the acquisition of multi-source sensor signals during the operation of the asynchronous motor, and the time synchronization, segmentation, and normalization processing of each sensor signal to construct a complete multi-source data sample, includes: The multi-source sensor signals include vibration signals, three-phase current signals, and temperature signals; the signals are input synchronously through nine channels, consisting of four vibration channels, three current channels, and two temperature channels. The original multi-source time series is sliced ​​using a sliding window of length 1024, and normalization is performed on each channel. During the training phase, complete multi-source data is used to learn the coupling laws of multi-physics fields. During the testing or deployment phase, one or more sensor channels are shielded based on actual sensor failures or random simulations to form the incomplete multi-source data input.

[0007] In some implementations, the step of inputting the complete multi-source data samples or the incomplete multi-source data into a shared continuous wavelet convolution feature extraction module, and performing time-frequency feature mapping on different sensor channels using wavelet convolution kernels with shared parameters to obtain multi-source wavelet features with unified time-frequency resolution and physical meaning, includes: For the Sensor input signal Feature extraction is performed using shared continuous wavelet convolution to obtain the corresponding wavelet features. , represented as: ; in, This represents a wavelet function with a predefined time-domain representation. and These represent the position parameter and scale parameter of the wavelet function, respectively. This represents the convolution operation. The wavelet function represents the total number of sensors included in the multi-source input; The Laplace wavelet is used to construct a shared continuous wavelet convolution kernel to extract the impact features and time-frequency variation information from the signals of each sensor. Wavelet features obtained from all sensors The output of the shared continuous wavelet convolution feature extraction module, cascaded along the channel dimension, is represented as follows: ; in, This represents the multi-source wavelet feature formed by cascading all sensor wavelet features.

[0008] In some implementations, the missing pattern identification module based on collaborative representation is used to encode all valid sensor wavelet features into a deep semantic space, aggregate them to obtain a multi-source joint representation, and the decoder reconstructs the complete multi-source wavelet features from the multi-source joint representation, including: Determining multi-source wavelet features of input using a missing feature detector Whether there are missing features; if there are missing features in the input data, continue to perform the missing pattern recognition and feature reconstruction process; if there are no missing features in the input data, skip the missing feature reconstruction process and directly enter the dual-view feature fusion network; For multi-source inputs where missing features are detected, the missing pattern recognition module analyzes the multi-source wavelet features of the input. Analysis was conducted to obtain a valid set of non-missing sensors. and the corresponding sensor missing mode , represented as: ; in, This refers to the missing pattern recognition module. Represents the set of valid, non-missing sensors. This indicates the sensor missing mode corresponding to the current input; For the set The first in One effective sensor, its wavelet features Input the corresponding convolution encoder To obtain deep semantic features , represented as: ; in, Indicates the first Convolutional encoders corresponding to each effective sensor, This represents the deep semantic features obtained after encoding; After encoding all valid sensor features, the deep semantic features from different valid data sources are concatenated, and a global average pooling operation is used to obtain a multi-source joint representation. : ; The multi-source joint representation Input the corresponding convolutional decoder to reconstruct the wavelet features of each sensor, the th The reconstructed features corresponding to each sensor are represented as follows: ; in, Indicates the first Each sensor corresponds to a convolutional decoder. The first one, obtained by reconstructing from the multi-source joint representation, represents... Wavelet features of individual sensors; All reconstructed features are concatenated along the channel dimension to form a complete reconstructed multi-source wavelet feature. .

[0009] In some implementations, the process of supervising the reconstruction of incomplete multi-source data paths with feature knowledge generated from complete multi-source data paths includes: The consistency between the deep semantic features of different effective sensors and the multi-source joint representation is constrained by the collaborative representation coordination loss, which is expressed as: ; in, Indicates coordination or loss. Indicates the first Deep semantic features of an effective sensor This represents the multi-source joint representation; The feature recovery process of incomplete multi-source data paths is supervised using multi-source wavelet features generated from complete multi-source data paths. The reconstruction loss is expressed as: ; in, This represents the loss for reconstructing missing features. This represents the complete multi-source wavelet features obtained from the reconstruction. This represents the target wavelet feature generated by the complete multi-source data path.

[0010] In some implementations, the step of modeling spatial correlations between sensors through multi-sensor fusion branches based on an adaptive gating mechanism includes: For the For each sensor, the wavelet features of other sensors besides the primary sensor are cascaded to form corresponding auxiliary features. And based on the auxiliary features Calculate the first Gating weights corresponding to each sensor : ; in, This represents the Sigmoid activation function. and They represent the first The weight matrix and bias parameters of the gating network correspond to each sensor. This indicates the generation based on information from other sensors. Individual sensor gating weights; The corresponding sensor wavelet features are adaptively weighted using the gated weights, and all weighted features are cascaded and pooled to obtain multi-sensor view fusion features. : ; in, This represents element-wise multiplication. This indicates a pooling operation.

[0011] In some implementations, the modeling of long-term temporal dependencies through multi-time-series fusion branches and multi-head attention mechanisms includes: All sensor wavelet features are concatenated along the channel direction, and coupling relationships between different data sources are established through one-dimensional convolution operations to obtain a coupled sequence representation. : ; The coupling sequence is represented The data is grouped, and a multi-head attention mechanism is used to establish global dependencies between different time positions. The attention outputs are then reorganized and fused within each group to obtain multi-time series perspective features. : ; in, This indicates a time series grouping operation. This represents multi-head attention computation. This indicates the recombination and fusion operation of features within a group.

[0012] In some implementations, the fusion of features from the two perspectives to obtain a comprehensive fault characterization includes: The multi-time series perspective features Features fused with the multi-sensor perspective The original multi-source wavelet features are then fused and incorporated. The residual connections are processed through batch normalization to obtain the final comprehensive fault characterization. : ; in, This indicates a dual-view feature fusion operation. This indicates a batch of standardized operations. This represents the final comprehensive fault characterization obtained after dual-view fusion.

[0013] In some implementations, the step of using a shared convolutional neural network classifier to extract high-level fault semantic features through multi-level convolution, pooling, and nonlinear activation, and then outputting the motor bearing health status or fault category through a fully connected layer, includes: The parameter-sharing convolutional neural network classifier consists of five cascaded convolutional modules, with the number of feature channels in each convolutional module being 32, 48, 64, 80 and 96 respectively, and the kernel size being set to 3. Each convolutional module extracts local fault semantic features through convolution operations, reduces feature dimensions and expands the effective receptive range layer by layer through max pooling, and achieves non-linear feature mapping through the ReLU activation function; After processing by a five-level convolution module, the obtained high-dimensional fault features are input into a fully connected classification layer, which outputs a category probability distribution corresponding to different motor bearing health states, and determines the final fault type based on the output results. The fault types include normal state, rotor imbalance, shaft misalignment, bearing inner ring fault, and bearing outer ring fault.

[0014] In some implementations, the step of jointly constructing the overall objective function by combining the classification loss, the distillation loss between complete and incomplete multi-source data paths, the missing feature reconstruction loss, and the collaborative representation coordination loss includes: Introducing KL divergence to construct knowledge distillation loss: ; in, and Let represent the category probability distributions of the outputs from complete and incomplete multi-source data paths, respectively. Denotes KL divergence, This represents the loss from knowledge distillation. The overall objective function is expressed as: ; in, This represents the classification loss, used to measure the difference between the model's predicted fault category and the actual fault category; This represents the knowledge distillation loss between complete and incomplete multi-source data paths. This represents the loss for reconstructing missing features, used to constrain the difference between the reconstructed features and the target features of the complete path; The collaborative representation coordination loss is used to maintain consistency between the deep semantic features of each effective sensor and the joint representation; the weight of each loss term is set to 1. The shared continuous wavelet convolutional feature extraction module, the missing feature reconstruction module, the dual-view feature fusion network, and the parameter-sharing convolutional neural network classifier are jointly optimized end-to-end using a multi-objective joint loss.

[0015] The beneficial effects of the technical solution provided by this invention include at least the following: This technical solution applies a unified wavelet basis function to each sensor channel through shared continuous wavelet convolution for time-frequency mapping, enabling heterogeneous signals to achieve a consistent physical representation scale at the feature level. This avoids the semantic misalignment problem of traditional convolution kernels across different physical quantities, providing an interpretable common basis for subsequent cross-sensor feature association. For sensor missing scenarios, this solution employs a collaborative representation mechanism to constrain the deep semantic features of effective channels to the same joint representation space. This ensures that the reconstruction process of missing channels is limited by multi-source physical coupling laws rather than being freely generated. Therefore, even under unknown missing modes, it can stably recover discriminative features strongly correlated with fault categories, effectively suppressing the introduction of reconstruction noise and redundant information. In the feature fusion stage, this solution uses an adaptive gating mechanism to model the spatial complementarity between sensors and a multi-head attention mechanism to capture long-term temporal evolution dependencies. The two perspectives are represented independently before fusion, avoiding the sensitivity of a single fusion path to missing modes. This ensures the model maintains stable diagnostic output under different missing combinations. The above method achieves end-to-end training through multi-objective joint optimization, which can generalize to sensor missing patterns not seen in the training phase without the need for separate modeling for each missing combination, thus significantly reducing the cost of model deployment and maintenance. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0017] Figure 1 The diagram shows a flowchart of a physically interpretable wavelet fusion motor bearing fault diagnosis method provided by an exemplary embodiment of the present invention.

[0018] Figure 2 The diagram illustrates a logic block diagram of a physically interpretable wavelet fusion motor bearing fault diagnosis method provided by an exemplary embodiment of the present invention.

[0019] Figure 3 The figure shows a comparison of the accuracy of various fault diagnosis models under different sensor missing tasks provided by an exemplary embodiment of the present invention. Detailed Implementation

[0020] 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.

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] Figure 1 The diagram illustrates a flowchart of a physically interpretable wavelet fusion method for diagnosing motor bearing faults, provided by an exemplary embodiment of the present invention. Figure 2 The diagram illustrates a logic block diagram of a physically interpretable wavelet fusion method for diagnosing motor bearing faults, provided by an exemplary embodiment of the present invention. This physically interpretable wavelet fusion method for diagnosing motor bearing faults includes the following steps: Step 101: Collect multi-source sensor signals during the operation of the asynchronous motor, perform time synchronization, segmentation and normalization on each sensor signal, and construct a complete multi-source data sample; establish incomplete multi-source data input based on sensor failure, acquisition abnormality or communication packet loss, and detect the validity of the input channel.

[0023] In some embodiments, the above-mentioned acquisition of multi-source sensor signals during the operation of the asynchronous motor, and the time synchronization, segmentation, and normalization processing of each sensor signal to construct a complete multi-source data sample, including: The multi-source sensor signals include vibration signals, three-phase current signals, and temperature signals; the synchronous input consists of 9 channels, comprising 4 vibration channels, 3 current channels, and 2 temperature channels. The original multi-source time series is sliced ​​using a sliding window of length 1024, and normalization is performed on each channel. During the training phase, complete multi-source data is used to learn the coupling laws of multi-physics fields. During the testing or deployment phase, one or more sensor channels are shielded based on actual sensor failures or random simulations to form incomplete multi-source data input.

[0024] In this embodiment, three heterogeneous physical quantities—vibration, current, and temperature—are aligned on the time axis and sliced ​​into fixed windows, eliminating inconsistencies in representation caused by differences in sampling rates and dimensions of different sensors. Normalization compresses the dynamic range of each channel, preventing gradient imbalance caused by significant differences in numerical scales during training. During the training phase, a multiphysics coupling benchmark is established using complete data. During the testing phase, specific channels are masked to simulate real sensor failure scenarios, enabling the model to generalize to any combination of missing data without requiring separate modeling for each missing mode.

[0025] Step 102: Input the complete or incomplete multi-source data samples into the shared continuous wavelet convolution feature extraction module, and perform time-frequency feature mapping on different sensor channels using wavelet convolution kernels with shared parameters to obtain multi-source wavelet features with unified time-frequency resolution and physical meaning.

[0026] In this embodiment, the preprocessed multi-source sensor signals are input into a shared continuous wavelet convolution module for physically interpretable feature extraction. For heterogeneous signals obtained from different sensors, a parameter-shared wavelet convolution method is used for processing, enabling the extraction of physically meaningful fault features from each sensor signal at a unified time-frequency analysis scale.

[0027] In some embodiments, the above-mentioned continuous wavelet convolution feature extraction module, which shares complete or incomplete multi-source data samples, performs time-frequency feature mapping on different sensor channels using wavelet convolution kernels with shared parameters, to obtain multi-source wavelet features with unified time-frequency resolution and physical meaning, including: For the Sensor input signal Feature extraction is performed using shared continuous wavelet convolution to obtain the corresponding wavelet features. , represented as: ; in, This represents a wavelet function with a predefined time-domain representation. and These represent the position parameter and scale parameter of the wavelet function, respectively. This represents the convolution operation. Indicates the total number of sensors included in a multi-source input; wavelet function The Laplace wavelet is used to construct a shared continuous wavelet convolution kernel to extract the impact features and time-frequency variation information from the signals of each sensor. Wavelet features obtained from all sensors The output of the cascaded, shared continuous wavelet convolution feature extraction module along the channel dimension is represented as follows: ; in, This represents the multi-source wavelet feature formed by cascading all sensor wavelet features.

[0028] In this embodiment, by applying a shared wavelet convolution rule to signals from different physical sources, the differences in feature scales of heterogeneous data can be reduced, providing a unified physical feature basis for subsequent missing pattern detection and missing feature reconstruction.

[0029] In this embodiment, the shared continuous wavelet convolution feature extraction module uses the Laplacian wavelet as the shared convolution kernel and applies the same time-frequency analysis basis function to all sensor channels. This allows heterogeneous signals such as vibration, current, and temperature to be mapped to wavelet coefficients with clear physical meaning at the same scale. The parameter sharing mechanism forces the features of each channel to be in a unified resolution space, eliminating semantic misalignment caused by different sensor types and providing an alignable feature basis for subsequent cross-channel collaborative reconstruction. The cascaded output retains the physical correspondence of the original channels, facilitating the fusion network to select effective information as needed and avoiding false associations introduced by missing channels.

[0030] Step 103: Identify missing patterns using the missing feature reconstruction module based on collaborative representation, encode all valid sensor wavelet features into the deep semantic space, aggregate them to obtain a multi-source joint representation, and reconstruct the complete multi-source wavelet features from the multi-source joint representation using the decoder; supervise the reconstruction process of the incomplete multi-source data path (i.e., the forward computation process of incomplete multi-source data processed by each module) with the feature knowledge generated by the complete multi-source data path (i.e., the forward computation process of incomplete multi-source data processed by each module).

[0031] In this embodiment, the integrity of multi-source inputs is judged by a missing feature reconstruction module based on cooperative representation, and the missing features are reconstructed based on the fault information contained in the valid sensors.

[0032] In some embodiments, the missing feature reconstruction module based on collaborative representation identifies missing patterns, encodes all valid sensor wavelet features into a deep semantic space, aggregates them to obtain a multi-source joint representation, and the decoder reconstructs the complete multi-source wavelet features from the multi-source joint representation, including: Determining multi-source wavelet features of input using a missing feature detector Whether there are missing features; if there are missing features in the input data, continue to perform the missing feature recognition and feature reconstruction process; if there are no missing features in the input data, skip the missing feature reconstruction process and directly enter the dual-view feature fusion network; For multi-source inputs where missing features are detected, the missing pattern recognition module analyzes the multi-source wavelet features of the input. Analysis was conducted to obtain a valid set of non-missing sensors. and the corresponding sensor missing mode , represented as: ; in, This indicates the missing pattern recognition module. Represents the set of valid, non-missing sensors. This indicates the sensor missing mode corresponding to the current input; For sets The first in One effective sensor, its wavelet features Input the corresponding convolution encoder To obtain deep semantic features , represented as: ; in, Indicates the first Convolutional encoders corresponding to each effective sensor, This represents the deep semantic features obtained after encoding; After encoding all valid sensor features, the deep semantic features from different valid data sources are concatenated, and a global average pooling operation is used to obtain a multi-source joint representation. : ; Multi-source joint representation Input the corresponding convolutional decoder to reconstruct the wavelet features of each sensor, the th The reconstructed features corresponding to each sensor are represented as follows: ; in, Indicates the first Each sensor corresponds to a convolutional decoder. The first term is represented by the reconstructed form obtained from the multi-source joint representation. Wavelet features of individual sensors; All reconstructed features are concatenated along the channel dimension to form a complete reconstructed multi-source wavelet feature. .

[0033] In some embodiments, the feature knowledge generated from the complete multi-source data path supervises the reconstruction process of the incomplete multi-source data path, including: By using collaborative representation to coordinate the consistency between deep semantic features from different effective sensors and multi-source joint representations, the collaborative representation coordination loss is expressed as follows: ; in, Indicates coordination or loss. Indicates the first Deep semantic features of an effective sensor Indicates multi-source joint representation; The feature recovery process of incomplete multi-source data paths is supervised using multi-source wavelet features generated from complete multi-source data paths. The reconstruction loss is expressed as: ; in, This represents the loss for reconstructing missing features. This represents the complete multi-source wavelet features obtained from the reconstruction. This represents the target wavelet feature generated by the complete multi-source data path.

[0034] In this embodiment, by combining the coordination loss and the missing feature reconstruction loss through collaborative representation, the incomplete multi-source data path can learn the fault knowledge in the complete multi-source data, and the adaptive reconstruction of missing fault features can be achieved by utilizing the nonlinear correlation and physical complementarity between effective sensors.

[0035] In this embodiment, the missing feature reconstruction module based on collaborative representation pre-checks the input integrity using a missing feature detector, triggering the reconstruction process only when a channel is missing, thus avoiding redundant computation under complete input. The convolutional encoder maps the wavelet features of each effective channel to a unified deep semantic space, and aggregates them into a multi-source joint representation through global average pooling (GAP), compressing fault information from different physical sources into a compact representation that carries the coupling relationship between sensors. The convolutional decoder then recovers the missing channel features, utilizing the physical complementarity between effective sensors rather than free generation. The collaborative representation coordination loss constrains the semantic features of each channel to converge to the joint representation, and the reconstruction loss uses the complete path features as the supervision target. This dual constraint ensures that the reconstruction result is consistent with the real multi-source features at the semantic level, effectively suppressing false discriminative information introduced by missing channels.

[0036] Step 104: Input the complete multi-source wavelet features or the complete multi-source wavelet features restored by the missing feature reconstruction module into the dual-view feature fusion network. The spatial correlation between sensors is modeled by the multi-sensor fusion branch based on the adaptive gating mechanism, and the long-term temporal dependency is modeled by the multi-time series fusion branch and the multi-head attention mechanism. The features from the two perspectives are fused to obtain the comprehensive fault characterization.

[0037] In this embodiment, the dual-view feature fusion network includes a multi-sensor fusion branch and a multi-time-series fusion branch. The multi-sensor fusion branch employs an adaptive gating mechanism to dynamically evaluate the importance of each data source based on the correlation between different sensors and selectively enhances sensor features. The multi-time-series fusion branch is used to further establish long-distance dependencies between different time locations.

[0038] In some embodiments, the above-described modeling of spatial correlations between sensors through multi-sensor fusion branches based on an adaptive gating mechanism includes: For the For each sensor, the wavelet features of other sensors besides the primary sensor are cascaded to form corresponding auxiliary features. And based on auxiliary features Calculate the first Gating weights corresponding to each sensor : ; in, This represents the Sigmoid activation function. and They represent the first The weight matrix and bias parameters of the gating network correspond to each sensor. This indicates the generation based on information from other sensors. Individual sensor gating weights; Adaptive weighting of the corresponding sensor wavelet features is performed using various gating weights, and all weighted features are concatenated and pooled to obtain multi-sensor view fusion features. : ; in, This represents element-wise multiplication. This indicates a pooling operation.

[0039] In this embodiment, a gating mechanism is used to dynamically adjust the feature weights of each sensor based on contextual information provided by other data sources, thereby highlighting high-confidence information related to the current fault state and suppressing redundant information and unreliable sensor features.

[0040] In some embodiments, the above-described modeling of long-term temporal dependencies using multi-time-series fusion branches and multi-head attention mechanisms includes: All sensor wavelet features are concatenated along the channel direction, and coupling relationships between different data sources are established through one-dimensional convolution operations to obtain a coupled sequence representation. : ; Representation of coupled sequences The data is grouped, and a multi-head attention mechanism is used to establish global dependencies between different time positions. The attention outputs are then reorganized and fused within each group to obtain multi-time series perspective features. : ; in, This indicates a time series grouping operation. This represents multi-head attention computation. This indicates the recombination and fusion operation of features within a group.

[0041] In some embodiments, the above-described fusion of features from two perspectives yields a comprehensive fault characterization, including: Features from multiple time series perspectives Features fused with multi-sensor perspectives The original multi-source wavelet features are then fused and incorporated. The residual connections are processed through batch normalization to obtain the final comprehensive fault characterization. : ; in, This indicates a dual-view feature fusion operation. This indicates a batch of standardized operations. This represents the final comprehensive fault characterization obtained after dual-view fusion.

[0042] In this embodiment, through the above-mentioned dual-perspective feature fusion process, the reconstructed multi-source wavelet features are deeply fused from two complementary perspectives: sensor spatial correlation and long-term time series dependence, thereby obtaining more comprehensive, stable, and fault-discriminative integrated features.

[0043] In this embodiment, the dual-view feature fusion network extracts fault discrimination information from two complementary dimensions: space and time. The multi-sensor fusion branch uses the features of the remaining channels as context, dynamically adjusting the weights of each sensor through Sigmoid gating to automatically strengthen high-confidence sources and suppress abnormal or redundant channels, avoiding misjudgments caused by fixed weighting. The multi-time-series fusion branch first establishes cross-channel coupling through one-dimensional convolution, then captures long-range temporal dependencies using multi-head attention, compensating for insufficient receptive fields in local convolutions. After feature fusion, the original wavelet feature residuals are superimposed and batch-normalized, preserving both low-level physical details and stabilizing high-level semantics, making the final representation robust to missing reconstruction errors and operational fluctuations.

[0044] Step 105: Input the fault comprehensive representation into the parameter-sharing convolutional neural network classifier, extract high-level fault semantic features through multi-level convolution, pooling and nonlinear activation, and then output the motor bearing health status or fault category through a fully connected layer.

[0045] In some embodiments, the above-mentioned fault comprehensive characterization input parameter shared convolutional neural network classifier extracts high-level fault semantic features through multi-level convolution, pooling, and nonlinear activation, and then outputs the motor bearing health status or fault category through a fully connected layer, including: The parameter-sharing convolutional neural network classifier consists of five cascaded convolutional modules, with the number of feature channels in each convolutional module being 32, 48, 64, 80 and 96 respectively, and the kernel size being set to 3. Each convolutional module extracts local fault semantic features through convolution operations, reduces feature dimensions and expands the effective receptive range layer by layer through max pooling, and achieves non-linear feature mapping through the ReLU activation function; After processing by a five-level convolution module, the obtained high-dimensional fault features are input into a fully connected classification layer, which outputs the category probability distribution corresponding to different motor bearing health states, and determines the final fault type based on the output results. The fault types include normal state, rotor imbalance, shaft misalignment, bearing inner ring fault, and bearing outer ring fault.

[0046] In this embodiment, the parameter-sharing convolutional neural network classifier adopts a five-level cascaded convolutional structure, with the number of channels increasing progressively from 32 to 96. Combined with a fixed-size 3 convolutional kernel and max pooling, it expands the receptive field while abstracting high-level semantic representations corresponding to bearing failure modes layer by layer. The parameter-sharing convolutional neural network classifier uses uniform network weight parameters, applicable to features from both complete multi-source data paths and reconstructed incomplete multi-source data paths. This parameter-sharing mechanism avoids the redundancy of parameters caused by training independent classifiers for each sensor missing mode, ensuring that the model can output a consistent fault category probability distribution based on the same set of discrimination criteria when facing known or unknown sensor missing conditions. This significantly improves the model's deployment efficiency and generalization ability for unknown missing modes. ReLU activation introduces non-linear mapping capabilities, enhancing the discriminative power for complex fault boundaries. The fully connected layer maps high-dimensional features to probability distributions of five health states, achieving end-to-end discrimination from comprehensive representation to specific fault categories. The output results can be directly used for operation and maintenance decisions.

[0047] Step 106: Train the fault diagnosis model using a multi-objective joint optimization strategy. Combine the classification loss, the distillation loss between complete and incomplete multi-source data paths, the missing feature reconstruction loss, and the cooperative representation coordination loss to form the overall objective function. Update the network parameters through backpropagation to obtain a motor bearing fault diagnosis model that can handle unknown sensor missing modes.

[0048] In some embodiments, the above-mentioned classification loss, distillation loss between complete and incomplete multi-source data paths, missing feature reconstruction loss, and collaborative representation coordination loss are jointly constituted into the overall objective function, including: Introducing KL divergence to construct knowledge distillation loss: ; in, and Let represent the category probability distributions of the outputs from complete and incomplete multi-source data paths, respectively. Denotes KL divergence, This represents the loss from knowledge distillation. Finally, the classification loss, knowledge distillation loss, missing feature reconstruction loss, and collaborative representation coordination loss are combined, and the overall objective function is expressed as: ; in, This represents the classification loss, used to measure the difference between the model's predicted fault category and the actual fault category; This represents the knowledge distillation loss between complete and incomplete multi-source data paths. This represents the loss for reconstructing missing features, used to constrain the difference between the reconstructed features and the target features of the complete path; The collaborative representation coordination loss is used to maintain consistency between the deep semantic features of each effective sensor and the joint representation; the weight of each loss term is set to 1. End-to-end joint optimization is performed on the shared continuous wavelet convolutional feature extraction module, the missing feature reconstruction module, the dual-view feature fusion network, and the parameter-sharing convolutional neural network classifier using a multi-objective joint loss. This enables the model to simultaneously achieve physically interpretable feature extraction capability, missing feature recovery capability, multi-source information deep fusion capability, and motor bearing fault classification capability.

[0049] For example, by using the backpropagation algorithm to jointly update all network parameters, the model can not only complete the fault category identification, but also take into account the recovery of missing information, cross-sensor collaborative representation learning, and knowledge transfer from complete multi-source data paths to incomplete multi-source data paths.

[0050] In this embodiment, the training strategy combines four losses—classification, distillation, reconstruction, and collaboration—with equal weights to form an end-to-end multi-objective optimization framework. Classification loss ensures accurate fault identification; knowledge distillation loss guides incomplete paths with the probability distribution of complete paths, allowing missing scenarios to inherit the discriminative knowledge of complete data; reconstruction loss constrains the recovery of features to approximate the true multi-source representation; and collaboration loss coordinates the convergence of semantics from each channel to the joint representation. The synergistic effect of these four losses enables the model to simultaneously improve feature extraction, missing data recovery, cross-source fusion, and state recognition capabilities during unified parameter updates, ultimately achieving robust diagnostic performance for missing patterns from unknown sensors.

[0051] Figure 3 The figure shows a comparison of the accuracy of various fault diagnosis models under different sensor missing tasks provided by an exemplary embodiment of the present invention. In this embodiment, the experiment uses 9-channel asynchronous motor multi-source data collected by a laboratory-built asynchronous motor test platform under a constant load of 0Nm for verification. The experiment sets up a progressive task with complete data and randomly missing 1 to 5 sensor channels.

[0052] Experimental results show that as the degree of sensor loss increases, the method proposed in this invention can maintain fault diagnosis capability relatively stably, indicating that the shared continuous wavelet convolution, collaborative representation of missing features reconstruction, and dual-view feature fusion mechanism can effectively utilize the remaining sensor information and improve the robustness of the model under the condition of incomplete multi-source data.

[0053] Comparison of different models shows that this invention maintains good fault identification performance in both tasks with complete data and tasks with varying degrees of sensor loss, exhibiting strong stability even under conditions of severe information loss. Further ablation experiments verify that the shared continuous wavelet feature extraction module, the missing feature reconstruction module, and the dual-view feature fusion module all contribute positively to the final diagnostic performance.

[0054] As a supplementary explanation Figure 3 The models shown include: Model of this application: Proposed (i.e., the "physically interpretable wavelet fusion motor bearing fault diagnosis method" proposed in this invention); PCFNN (referring to Parallel Convolutional Fuzzy Neural Network); DPWFN (referring to Deep Parallel Wavelet Fuzzy Network); MSFNN (referring to Multi-Scale Fuzzy Neural Network); MS-WFN (referring to Multi-Scale Wavelet Fuzzy Network); GAN-WFN (referring to Generative Adversarial Network based Wavelet FuzzyNetwork); DAE-WFN (referring to Denoising Autoencoder based Wavelet Fuzzy Network).

[0055] This invention addresses the problems of random failure of multi-source sensors and incomplete data in complex industrial environments. Through physically interpretable feature extraction, collaborative missing reconstruction, multi-perspective information fusion, and joint optimization, it achieves fault diagnosis of motor bearings under incomplete multi-source data conditions, and can provide technical support for condition monitoring and intelligent operation and maintenance of industrial equipment.

[0056] It is understood that the specific examples in this document are only intended to help those skilled in the art better understand the invention, and not to limit the scope of the invention.

[0057] It is understood that in the various embodiments described in this specification, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of this invention.

[0058] It is understood that the various implementation methods described in this specification can be implemented individually or in combination, and the present invention does not limit them.

[0059] Unless otherwise stated, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of this specification. The term "and / or" as used in this specification includes any and all combinations of one or more of the associated listed items. The singular forms "a," "the," and "the" as used in this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0060] The above description is merely a specific embodiment of this specification, but the scope of protection of this invention 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 specification should be included within the scope of protection of this specification. Therefore, the scope of protection of this invention is determined by the scope of the claims.

Claims

1. A physically interpretable wavelet fusion method for diagnosing motor bearing faults, characterized in that, include: Multi-source sensor signals are collected during the operation of the asynchronous motor. The sensor signals are then time-synchronized, segmented, and normalized to construct a complete multi-source data sample. Based on sensor failure, abnormal data acquisition, or communication packet loss, an incomplete multi-source data input is established, and the effectiveness of the input channels is detected. The complete multi-source data sample or the incomplete multi-source data is input into the shared continuous wavelet convolution feature extraction module. Time-frequency feature mapping is performed on different sensor channels using wavelet convolution kernels with shared parameters to obtain multi-source wavelet features with unified time-frequency resolution and physical meaning. Missing patterns are identified by a missing feature reconstruction module based on collaborative representation. All effective sensor wavelet features are encoded into a deep semantic space and aggregated to obtain a multi-source joint representation. The decoder then reconstructs the complete multi-source wavelet features from the multi-source joint representation. The feature knowledge generated from complete multi-source data paths is used to supervise the reconstruction process of incomplete multi-source data paths; The complete multi-source wavelet features or the complete multi-source wavelet features restored by the missing feature reconstruction module are input into the dual-view feature fusion network. The spatial correlation between sensors is modeled by the multi-sensor fusion branch based on the adaptive gating mechanism, and the long-term temporal dependency is modeled by the multi-time series fusion branch and the multi-head attention mechanism. The features from the two perspectives are fused to obtain the comprehensive fault characterization. The fault comprehensive characterization input parameter shared convolutional neural network classifier extracts high-level fault semantic features through multi-level convolution, pooling and nonlinear activation, and then outputs the motor bearing health status or fault category through a fully connected layer. A multi-objective joint optimization strategy is adopted to train the fault diagnosis model. The classification loss, the distillation loss between complete and incomplete multi-source data paths, the missing feature reconstruction loss, and the cooperative representation coordination loss are jointly constructed into the overall objective function. The network parameters are updated through backpropagation to obtain a motor bearing fault diagnosis model that can handle unknown sensor missing modes.

2. The physically interpretable wavelet fusion method for diagnosing motor bearing faults according to claim 1, characterized in that, The process involves collecting multi-source sensor signals during the operation of the asynchronous motor, performing time synchronization, segmentation, and normalization on each sensor signal, and constructing a complete multi-source data sample, including: The multi-source sensor signals include vibration signals, three-phase current signals, and temperature signals; the signals are input synchronously through 9 channels, consisting of 4 vibration channels, 3 current channels, and 2 temperature channels. The original multi-source time series is sliced ​​using a sliding window of length 1024, and normalization is performed on each channel. During the training phase, complete multi-source data is used to learn the coupling laws of multi-physics fields. During the testing or deployment phase, one or more sensor channels are shielded based on actual sensor failures or random simulations to form the incomplete multi-source data input.

3. The physically interpretable wavelet fusion motor bearing fault diagnosis method according to claim 1, characterized in that, The step of inputting the complete multi-source data samples or the incomplete multi-source data into the shared continuous wavelet convolution feature extraction module, and performing time-frequency feature mapping on different sensor channels using wavelet convolution kernels with shared parameters, to obtain multi-source wavelet features with unified time-frequency resolution and physical meaning includes: For the Sensor input signal Feature extraction is performed using shared continuous wavelet convolution to obtain the corresponding wavelet features. , represented as: ； in, This represents a wavelet function with a predefined time-domain representation. and These represent the position parameter and scale parameter of the wavelet function, respectively. This represents the convolution operation. The wavelet function represents the total number of sensors included in the multi-source input; The Laplace wavelet is used to construct a shared continuous wavelet convolution kernel to extract the impact features and time-frequency variation information from the signals of each sensor. Wavelet features obtained from all sensors The output of the shared continuous wavelet convolution feature extraction module, cascaded along the channel dimension, is represented as follows: ； in, This represents the multi-source wavelet feature formed by cascading all sensor wavelet features.

4. The physically interpretable wavelet fusion motor bearing fault diagnosis method according to claim 1, characterized in that, The missing feature reconstruction module based on collaborative representation identifies missing patterns, encodes all valid sensor wavelet features into a deep semantic space, aggregates them to obtain a multi-source joint representation, and the decoder reconstructs the complete multi-source wavelet features from the multi-source joint representation, including: Determining the multi-source wavelet features of the input using a missing feature detector Whether there are missing features; if there are missing features in the input data, continue to perform the missing pattern recognition and feature reconstruction process; if there are no missing features in the input data, skip the missing feature reconstruction process and directly enter the dual-view feature fusion network; For multi-source inputs with detected missing features, the missing pattern recognition module analyzes the multi-source wavelet features of the input. Analysis was conducted to obtain a valid set of non-missing sensors. and the corresponding sensor missing mode , represented as: ； in, This refers to the missing pattern recognition module. Represents the set of valid, non-missing sensors. This indicates the sensor missing mode corresponding to the current input; For the set The first in One effective sensor, its wavelet features Input the corresponding convolution encoder To obtain deep semantic features , represented as: ； in, Indicates the first Convolutional encoders corresponding to each effective sensor, This represents the deep semantic features obtained after encoding; After encoding all valid sensor features, the deep semantic features from different valid data sources are concatenated, and a global average pooling operation is used to obtain a multi-source joint representation. : ； The multi-source joint representation Input the corresponding convolutional decoder to reconstruct the wavelet features of each sensor, the th The reconstructed features corresponding to each sensor are represented as follows: ； in, Indicates the first Each sensor corresponds to a convolutional decoder. The first one, obtained by reconstructing from the multi-source joint representation, represents... Wavelet features of individual sensors; All reconstructed features are concatenated along the channel dimension to form a complete reconstructed multi-source wavelet feature. .

5. The physically interpretable wavelet fusion motor bearing fault diagnosis method according to claim 4, characterized in that, The process of supervising the reconstruction of incomplete multi-source data paths with feature knowledge generated from complete multi-source data paths includes: The consistency between the deep semantic features of different effective sensors and the multi-source joint representation is constrained by the collaborative representation coordination loss, which is expressed as: ； in, Indicates coordination or loss. Indicates the first Deep semantic features of an effective sensor This represents the multi-source joint representation; The feature recovery process of incomplete multi-source data paths is supervised using multi-source wavelet features generated from complete multi-source data paths. The reconstruction loss is expressed as: ； in, This represents the loss for reconstructing missing features. This represents the complete multi-source wavelet features obtained from the reconstruction. This represents the target wavelet feature generated by the complete multi-source data path.

6. The physically interpretable wavelet fusion method for diagnosing motor bearing faults according to claim 1, characterized in that, The method of modeling spatial correlations between sensors through multi-sensor fusion branches based on an adaptive gating mechanism includes: For the For each sensor, the wavelet features of other sensors besides the primary sensor are cascaded to form corresponding auxiliary features. And based on the auxiliary features Calculate the first Gating weights corresponding to each sensor : ； in, This represents the Sigmoid activation function. and They represent the first The weight matrix and bias parameters of the gating network correspond to each sensor. This indicates the generation based on information from other sensors. Individual sensor gating weights; The corresponding sensor wavelet features are adaptively weighted using the gated weights, and all weighted features are cascaded and pooled to obtain multi-sensor view fusion features. : ； in, This represents element-wise multiplication. This indicates a pooling operation.

7. The physically interpretable wavelet fusion motor bearing fault diagnosis method according to claim 1, characterized in that, The method of modeling long-term temporal dependencies through multi-time-series fusion branches and multi-head attention mechanisms includes: All sensor wavelet features are concatenated along the channel direction, and coupling relationships between different data sources are established through one-dimensional convolution operations to obtain a coupled sequence representation. : ； The coupling sequence is represented The data is grouped, and a multi-head attention mechanism is used to establish global dependencies between different time positions. The attention outputs are then reorganized and fused within each group to obtain multi-time series perspective features. : ； in, This indicates a time series grouping operation. This represents multi-head attention computation. This indicates the recombination and fusion operation of features within a group.

8. The physically interpretable wavelet fusion motor bearing fault diagnosis method according to claim 6 or 7, characterized in that, The fusion of features from the two perspectives to obtain a comprehensive fault characterization includes: The multi-time series perspective features Features fused with the multi-sensor perspective The original multi-source wavelet features are then fused and incorporated. The residual connections are processed through batch normalization to obtain the final comprehensive fault characterization. : ； in, This indicates a dual-view feature fusion operation. This indicates a batch of standardized operations. This represents the final comprehensive fault characterization obtained after dual-view fusion.

9. The physically interpretable wavelet fusion method for diagnosing motor bearing faults according to claim 1, characterized in that, The method involves using a shared convolutional neural network classifier to extract high-level fault semantic features through multi-level convolution, pooling, and non-linear activation, and then outputting the motor bearing health status or fault category through a fully connected layer. The parameter-sharing convolutional neural network classifier consists of five cascaded convolutional modules, with the number of feature channels in each convolutional module being 32, 48, 64, 80 and 96 respectively, and the kernel size being set to 3. Each convolutional module extracts local fault semantic features through convolution operations, reduces feature dimensions and expands the effective receptive range layer by layer through max pooling, and achieves non-linear feature mapping through the ReLU activation function; After processing by a five-level convolution module, the obtained high-dimensional fault features are input into a fully connected classification layer, which outputs a category probability distribution corresponding to different motor bearing health states, and determines the final fault type based on the output results. The fault types include normal state, rotor imbalance, shaft misalignment, bearing inner ring fault, and bearing outer ring fault.

10. The physically interpretable wavelet fusion method for diagnosing motor bearing faults according to claim 1, characterized in that, The overall objective function is jointly constructed by the classification loss, the distillation loss between complete and incomplete multi-source data paths, the missing feature reconstruction loss, and the collaborative representation coordination loss, including: Introducing KL divergence to construct knowledge distillation loss: ； in, and Let represent the category probability distributions of the outputs from complete and incomplete multi-source data paths, respectively. Denotes KL divergence, This represents the loss from knowledge distillation. The overall objective function is expressed as: ； in, This represents the classification loss, used to measure the difference between the model's predicted fault category and the actual fault category; This represents the knowledge distillation loss between complete and incomplete multi-source data paths. This represents the loss for reconstructing missing features, used to constrain the difference between the reconstructed features and the target features of the complete path; The collaborative representation coordination loss is used to maintain consistency between the deep semantic features of each effective sensor and the joint representation; the weight of each loss term is set to 1. The shared continuous wavelet convolutional feature extraction module, the missing feature reconstruction module, the dual-view feature fusion network, and the parameter-sharing convolutional neural network classifier are jointly optimized end-to-end using a multi-objective joint loss.

Images