Motor intelligent diagnosis method, system and medium
By transforming the multi-source heterogeneous information of the motor to the frequency domain and using physical mechanisms to map the spectral attention mask, and adding physical simulation and generative models to generate targeted enhancement data, a Transformer hybrid diagnostic network is constructed. This solves the problem of independent integration of mechanism and data methods in existing motor fault diagnosis, and achieves high-precision and interpretable motor fault diagnosis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HANGZHOU REBOTECH
- Filing Date
- 2026-04-14
- Publication Date
- 2026-07-10
AI Technical Summary
Existing motor fault diagnosis technologies lack interpretability and automation in mechanistic methods that rely on expert experience, and data-driven methods have insufficient generalization ability in small sample scenarios. Furthermore, the independent integration of mechanistic and data methods has failed to achieve deep synergy, resulting in limited diagnostic accuracy and adaptability.
By acquiring multi-source heterogeneous information of the motor and transforming it to the frequency domain, and using the learnable spectral attention mask mapped by the physical mechanism of motor faults for weighting, and combining physical simulation and generative model to generate targeted enhancement data, a hybrid diagnostic network containing Transformer is constructed to extract global temporal features and perform multi-task decoding to output diagnostic results.
It achieves high-precision diagnosis of motor faults under complex working conditions and small sample scenarios, improves the robustness and interpretability of the model, can accurately identify early weak faults and compound faults, and provides comprehensive and reliable diagnostic basis.
Smart Images
Figure CN122017561B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of motor technology, and in particular to a method, system and medium for intelligent diagnosis of motors. Background Technology
[0002] As a core power source in industrial production, the real-time monitoring and accurate diagnosis of the motor's operating status are crucial for ensuring production safety and reducing maintenance costs. Currently, motor fault diagnosis technologies mainly fall into two categories: The first is based on physical mechanism models, which rely on expert experience and analyze specific frequency components of vibration signals (such as the characteristic frequencies of bearing faults) to determine faults. This type of method has strong interpretability, but its diagnostic effectiveness is highly dependent on the accuracy of expert knowledge and it struggles to identify early, subtle faults and complex compound faults. The second is purely data-driven methods, which utilize machine learning (such as support vector machines and convolutional neural networks) to train massive amounts of historical data and automatically learn fault patterns. This type of method can effectively handle complex nonlinear relationships, but it is essentially a black-box model; the diagnostic results lack physical basis, have poor interpretability, and in real-world industrial scenarios where fault data samples are scarce, the model's generalization ability and diagnostic accuracy drop sharply.
[0003] To combine the advantages of both, existing technologies attempt to integrate mechanistic and data-driven methods. However, this typically employs a shallow fusion approach: feature extraction followed by model concatenation. This involves manually extracting mechanistic features from signals based on expert experience, then using these features as input to a data classification model for fault identification. This approach not only fails to overcome reliance on expert experience and suffers from low automation, but also maintains that mechanistic knowledge and the data model operate independently during training, failing to achieve deep synergy at the algorithmic level. Consequently, the model's diagnostic accuracy and adaptability remain limited when facing individual equipment differences, varying operating conditions, and small sample scenarios. Summary of the Invention
[0004] To enable more accurate diagnosis of motors, embodiments of this application provide a method, system, and medium for intelligent motor diagnosis.
[0005] Firstly, a method for intelligent diagnosis of a motor is provided, the method comprising:
[0006] The process involves acquiring multi-source heterogeneous information of the motor and its real-time operating parameters, transforming the multi-source heterogeneous information to the frequency domain to obtain a spectral tensor, mapping the real-time operating parameters to a learnable spectral attention mask based on the physical mechanism of the motor fault, and weighting the spectral tensor with the spectral attention mask to obtain a mechanism-weighted spectrum.
[0007] The multi-source heterogeneous information and the real-time operating parameters are processed by physical simulation and generative modeling to obtain targeted enhancement data for physical information enhancement;
[0008] A hybrid diagnostic network containing a Transformer is constructed, taking the mechanism-weighted spectrum and the target enhancement data as input, and extracting global temporal features in the Transformer using the spectral attention mask as a bias of the self-attention mechanism;
[0009] The global time-series features are input into a multi-task decoder to output the motor's fault type probability, fault quantification parameters, and fault feature heatmap used to characterize the diagnostic basis in parallel. The diagnostic result is obtained based on the fault type probability, fault quantification parameters, and fault feature heatmap.
[0010] In some embodiments, transforming the multi-source heterogeneous information to the frequency domain to obtain a spectral tensor includes:
[0011] Customized hierarchical preprocessing adapted to the characteristics of each type of signal in the multi-source heterogeneous information is performed to obtain multi-channel preprocessed signals.
[0012] The multi-channel preprocessed signals are subjected to time-frequency transformation to obtain multi-channel spectra;
[0013] The multiple spectrum channels are filtered to remove redundant spectrum to obtain the filtered spectrum;
[0014] The filtered spectra are aligned by timestamps and concatenated to obtain a spectral tensor.
[0015] In some embodiments, mapping the real-time operating parameters to a learnable spectral attention mask based on the physical mechanism of the motor fault includes:
[0016] Take the inherent parameters of the motor, and input the real-time operating parameters and the inherent parameters into a fault frequency encoder that encodes the fault characteristic frequency formula to obtain the theoretical fault characteristic frequency set.
[0017] A Gaussian spectral attention base mask containing a learnable center frequency, learnable bandwidth, and learnable attenuation coefficient is constructed based on the theoretical fault feature frequency set as the center frequency.
[0018] The real-time operating parameters are input into a lightweight adjustment network to predict the adaptive adjustment amount of the operating conditions for each learnable parameter.
[0019] The adaptive adjustment amount of the operating condition is superimposed with the corresponding learnable parameter in the Gaussian spectral attention base mask to obtain the dynamic spectral attention mask after nonlinear modulation of the operating condition parameters.
[0020] In some embodiments, the step of weighting the spectral tensor with the spectral attention mask to obtain the mechanism-weighted spectrum includes:
[0021] The primary weighted spectrum is obtained by multiplying the spectral attention mask element-wise with the spectral tensor.
[0022] The primary weighted spectrum is input into a physical constraint convolutional layer composed of fault physical laws encoding to enhance features and obtain a mechanism-enhanced feature map.
[0023] The mechanism-enhanced feature map and the primary weighted spectrum are concatenated along the channel dimension and then fused by attention to obtain the fused spectral feature as the mechanism-weighted spectrum.
[0024] In some embodiments, the process of processing the multi-source heterogeneous information and the real-time operating parameters through physical simulation and generative models to obtain targeted augmentation data for physical information enhancement includes:
[0025] The inherent parameters of the motor and the preset fault types are obtained, and the inherent parameters and the fault types are input into the dynamic simulation model based on digital twin to generate simulation signals containing fault characteristics.
[0026] The multi-source heterogeneous information is input into a pre-trained diagnostic model to identify fault categories with classification confidence scores below a threshold as challenging samples.
[0027] The simulation signal and the difficult sample are input into a diffusion model that uses the difficult sample as the generation condition to generate initial augmentation data for the difficult sample;
[0028] The initial augmentation data and real fault samples from the multi-source heterogeneous information are input into a recurrent generative adversarial network for style transfer so that the feature distribution of the initial augmentation data is aligned with the feature distribution of the real fault samples to output targeted augmentation data.
[0029] In some embodiments, constructing a hybrid diagnostic network including Transformers includes:
[0030] The mechanism-weighted spectrum is decomposed into time-step embedding vectors to form a token sequence. In the encoder layer of the Transformer, the self-attention score matrix of the token sequence is masked by the spectral attention mask to generate a physically enhanced hidden layer representation.
[0031] Fault feature maps are extracted layer by layer by cascading multiple encoder layers, and differentiable wavelet transform layers are introduced between encoder layers to perform multi-scale time-frequency decomposition on the hidden layer representation to extract residual features.
[0032] The residual features are skipped to the fault feature map to construct a hybrid diagnostic network that includes a physically enhanced Transformer.
[0033] In some embodiments, the step of extracting global temporal features in the Transformer using the spectral attention mask as a bias of the self-attention mechanism includes:
[0034] The hidden layer representation is linearly projected onto three different feature spaces to obtain the query matrix, key matrix, and value matrix;
[0035] The spectral attention mask is superimposed as a bias term onto the dot product of the query matrix and the key matrix to generate a physical bias attention score.
[0036] The physical bias attention score is normalized to obtain the attention weight matrix;
[0037] The attention weight matrix and the value matrix are weighted and aggregated to obtain global temporal features guided by fault mechanism priors.
[0038] In some embodiments, obtaining the diagnostic results based on the fault type probability, fault quantification parameters, and fault feature heatmap includes:
[0039] The confidence level of the fault type probability is calibrated to obtain the calibrated confidence level.
[0040] Based on the aforementioned fault quantification parameters, a damage quantification index is constructed;
[0041] The fault feature heatmap is superimposed onto the original signal waveform corresponding to the multi-source heterogeneous information to generate a visual diagnostic map.
[0042] The system integrates the post-calibration confidence level, the damage quantification index, and the visualized diagnostic atlas, and processes the data through an interpretable report generator to output a diagnostic result that includes explanations of the diagnostic criteria.
[0043] Secondly, a smart motor diagnostic system is provided, the system comprising: an embedded module, an enhancement module, a network module, and a diagnostic module; wherein,
[0044] The embedded module is used to acquire multi-source heterogeneous information of the motor and its real-time operating parameters, transform the multi-source heterogeneous information to the frequency domain to obtain a spectral tensor, and map the real-time operating parameters into a learnable spectral attention mask according to the physical mechanism of motor faults, and use the spectral attention mask to weight the spectral tensor to obtain a mechanism-weighted spectrum.
[0045] The enhancement module is used to process the multi-source heterogeneous information and the real-time operating parameters through physical simulation and generative modeling to obtain targeted enhancement data for physical information enhancement.
[0046] The network module is used to construct a hybrid diagnostic network containing a Transformer, taking the mechanism-weighted spectrum and the targeted enhancement data as inputs, and extracting global temporal features in the Transformer using the spectral attention mask as a bias of the self-attention mechanism.
[0047] The diagnostic module is used to input the global time-series features into the multi-task decoder to output the motor's fault type probability, fault quantification parameters, and fault feature heatmap used to characterize the diagnostic basis in parallel, and to obtain the diagnostic result based on the fault type probability, fault quantification parameters, and fault feature heatmap.
[0048] Thirdly, a computer-readable storage medium is provided having a computer program stored thereon that can run on a processor, wherein when the computer program is executed by the processor, it implements a motor intelligent diagnostic method as described in the first aspect.
[0049] Using the above method, this application obtains multi-source heterogeneous information and real-time operating parameters of the motor. The multi-source heterogeneous information is transformed to the frequency domain to obtain a spectral tensor. Based on the physical mechanism of the motor fault, the real-time operating parameters are mapped to a learnable spectral attention mask. The spectral tensor is then weighted using the spectral attention mask to obtain a mechanism-weighted spectrum. Physical simulation and generative models are used to process the multi-source heterogeneous information and real-time operating parameters to obtain targeted enhancement data for physical information enhancement. A hybrid diagnostic network including a Transformer is constructed, using the mechanism-weighted spectrum and targeted enhancement data as input. The spectral attention mask is used as a bias in the self-attention mechanism within the Transformer to extract global temporal features. Based on these global temporal features, the multi-task decoder outputs the motor fault type probability, fault quantification parameters, and a fault feature heatmap used to characterize the diagnostic basis in parallel. The diagnostic results are obtained based on the fault type probability, fault quantification parameters, and fault feature heatmap. This allows for more accurate motor diagnosis. Attached Figure Description
[0050] Figure 1 This is a block diagram of an intelligent diagnostic method for motors provided in this application.
[0051] Figure 2 This is a block diagram of the method provided in this application for mapping real-time operating parameters into learnable spectral attention masks based on the physical mechanism of motor faults.
[0052] Figure 3This is a block diagram of the method provided in this application for processing multi-source heterogeneous information and real-time operating parameters to obtain targeted augmentation data for physical information enhancement.
[0053] Figure 4 This is a schematic diagram of the overall architecture of the intelligent diagnostics for motors provided in this application.
[0054] Figure 5 This is a connection diagram of an intelligent motor diagnostic system provided in an embodiment of this application. Detailed Implementation
[0055] To better understand the purpose, technical solutions, and advantages of this application, it has been described and illustrated below with reference to the accompanying drawings and embodiments. However, those skilled in the art should understand that this application can be implemented without these details. It will be apparent to those skilled in the art that various modifications can be made to the embodiments disclosed in this application, and the general principles defined in this application can be applied to other embodiments and application scenarios without departing from the principles and scope of this application. Therefore, this application is not limited to the illustrated embodiments, but is consistent with the broadest scope claimed in this application.
[0056] The embodiments of this application will now be described in further detail with reference to the accompanying drawings.
[0057] Figure 1 This is a block diagram of a motor intelligent diagnostic method provided in this application. Figure 1 As shown, a method for intelligent diagnosis of motors includes the following steps:
[0058] Step S100: Obtain multi-source heterogeneous information of the motor and its real-time operating parameters, transform the multi-source heterogeneous information to the frequency domain to obtain a spectral tensor, and map the real-time operating parameters into a learnable spectral attention mask according to the physical mechanism of motor faults, and use the spectral attention mask to weight the spectral tensor to obtain the mechanism-weighted spectrum.
[0059] This application describes the process from a diagnostic perspective. Specifically, vibration, current, and temperature sensors are first placed in key parts of the motor (such as bearing housings and casing) to collect high-frequency vibration signals, three-phase current waveforms, and temperature time-series data, forming multi-source heterogeneous information covering both mechanical and electrical conditions. Simultaneously, the motor's current speed and load values are read in real-time via an encoder or control system as real-time operating parameters. To ensure data quality, all acquired signals undergo preprocessing operations such as noise reduction, filtering, and normalization to eliminate environmental interference and standardize dimensions, thereby constructing raw input data that accurately reflects the motor's condition. This achieves comprehensive perception of the motor's operating status, providing a high-quality, highly consistent data foundation for subsequent mechanism embedding and data augmentation, significantly improving the reliability and adaptability of diagnostic work.
[0060] To ensure that the representation of multi-source heterogeneous information in the frequency domain accurately reflects the fault characteristics of the motor under different physical fields, it needs to be refined to construct a high-quality spectral tensor. The process of transforming multi-source heterogeneous information to the frequency domain to obtain the spectral tensor includes the following steps:
[0061] Step S101: Perform customized hierarchical preprocessing on various signals in the multi-source heterogeneous information to adapt to their respective signal characteristics to obtain multi-channel preprocessed signals.
[0062] Step S102: Perform time-frequency transformation on the multiple preprocessed signals to obtain multiple spectra.
[0063] Step S103: Perform channel filtering on the multiple spectrums to remove redundant spectrums and obtain the filtered spectrum.
[0064] Step S104: Align the filtered spectra according to the timestamps and concatenate them to obtain the spectral tensor.
[0065] Specifically, the first step involves custom-designed, layered preprocessing of the acquired multi-source signals, such as vibration, current, and temperature, tailored to their respective characteristics. For example, bandpass filtering is applied to vibration signals to retain the frequency bands related to fault impact, fundamental wave removal is performed on current signals to highlight sideband components, and trend term removal is applied to temperature signals to enhance the response to abrupt changes. This yields multiple preprocessed signals, providing high-quality input for subsequent frequency domain transformation. Then, time-frequency transformations are performed on each preprocessed signal, such as short-time Fourier transform or wavelet transform, to obtain their corresponding spectrograms. This transforms fault features that are difficult to distinguish in the time domain into physically meaningful energy distributions in the frequency domain. Next, the generated multi-spectrum signals undergo channel filtering. The sensitivity of each spectrum channel is evaluated based on the fault mechanism, eliminating noise-dominated or redundant spectrum channels and retaining spectrum components highly correlated with typical fault modes such as bearing faults and rotor imbalance. This reduces the computational burden of subsequent models and improves feature effectiveness. Finally, the filtered multi-spectrum signals are aligned according to the acquisition timestamps and stitched together along the channel dimension to construct a spatiotemporally consistent spectral tensor. This tensor not only preserves the frequency domain structure of the original signal but also integrates the evolutionary patterns of multi-source information over time, providing a structurally unified and semantically rich input foundation for subsequent mechanism embedding and data fusion. Through customized hierarchical preprocessing, channel filtering, and alignment stitching operations, it achieves efficient fusion and structured expression of multi-source heterogeneous information, improving the physical consistency and diagnostic adaptability of frequency domain features. This lays a solid data foundation for subsequent mechanism-guided deep network modeling and enhances robustness and generalization ability under complex conditions.
[0066] After obtaining the multi-source heterogeneous information of the motor and its real-time operating parameters, it is necessary to further embed the physical mechanism into the deep learning model in a structured manner. Figure 2 This is a block diagram of the method provided in this application for mapping real-time operating parameters to learnable spectral attention masks based on the physical mechanism of motor faults. For example... Figure 2 As shown, mapping real-time operating parameters to learnable spectral attention masks based on the physical mechanism of motor faults includes the following steps:
[0067] Step S105: Obtain the inherent parameters of the motor, and input the real-time operating parameters and inherent parameters into the fault frequency encoder that encodes the fault characteristic frequency formula to obtain the theoretical fault characteristic frequency set.
[0068] Step S106: Construct a Gaussian spectral attention base mask based on the theoretical fault characteristic frequency set as the center frequency, which includes a learnable center frequency, a learnable bandwidth, and a learnable attenuation coefficient.
[0069] Step S107: Input the real-time operating condition parameters into the lightweight adjustment network to predict the adaptive adjustment amount of the operating condition for each learnable parameter.
[0070] Step S108: The adaptive adjustment amount of the operating condition is superimposed with the corresponding learnable parameter in the Gaussian spectral attention base mask to obtain the dynamic spectral attention mask after nonlinear modulation of the operating condition parameters.
[0071] Specifically, the encoder incorporates classical physical formulas, such as the failure frequency of the bearing outer race, and can dynamically calculate the theoretical frequency values corresponding to various faults based on current operating parameters such as speed and load, providing physical priors for the subsequent attention mechanism. Then, a Gaussian spectral attention basis mask is constructed based on the theoretical fault feature frequency set as the center frequency. Not only includes learnable center frequencies It also introduces learnable bandwidth. With attenuation coefficient Its form is This allows the model to adaptively adjust its focus around the physical frequency, adapting to frequency shifts or ambiguity during actual operation. Real-time operating parameters are then input into a lightweight adjustment network, such as an MLP, to predict the adaptive adjustment amount for each learnable parameter. This allows the mask parameters to be dynamically adjusted according to changes in operating conditions, enhancing the model's adaptability to varying operating scenarios. Finally, the adaptive adjustment amount is superimposed with the corresponding learnable parameters in the Gaussian spectral attention base mask to obtain the dynamic spectral attention mask after nonlinear modulation of operating conditions. This mask not only preserves the guiding role of physical frequencies but also flexibly adjusts the focus bands and sensitivity according to real-time operating conditions, achieving deep synergy between mechanistic knowledge and data-driven approaches at the algorithm level. This transforms traditionally discrete and fixed expert knowledge into differentiable, learnable, and dynamically adjustable network parameters, enabling the model to understand physical laws during diagnosis and maintain high robustness and interpretability under varying operating conditions and small sample scenarios. It enhances the ability to identify early, weak, and complex faults, laying a solid physical foundation for subsequent global temporal feature extraction and multi-task diagnosis.
[0072] After obtaining the dynamic spectral attention mask, it needs to be deeply fused with the spectral tension to achieve structured embedding of physical priors at the data level. The process of weighting the spectral tensor with the spectral attention mask to obtain the mechanism-weighted spectrum includes the following steps:
[0073] Step S109: Multiply the spectral attention mask element-wise with the spectral tensor to obtain the primary weighted spectrum.
[0074] Step S110: Input the primary weighted spectrum into the physical constraint convolutional layer composed of fault physical law encoding to enhance the features and obtain the mechanism enhancement feature map.
[0075] Step S111: The mechanism enhancement feature map and the primary weighted spectrum are spliced along the channel dimension and fused by attention to obtain the fused spectrum feature as the mechanism weighted spectrum.
[0076] Specifically, the generated dynamic spectral attention mask is first multiplied element-wise with the constructed spectral tensor on the corresponding frequency channels. This operation is equivalent to applying a soft gating mechanism driven by physical mechanisms to each frequency component in the frequency domain: frequency bands related to the theoretical fault frequency and its sidebands are preserved because their weights are close to 1, while irrelevant noise frequency bands are suppressed because their weights are close to 0. This significantly enhances the response intensity of fault-sensitive components while preserving the original spectral structure. The physically constrained convolutional layer is not a standard convolutional layer; its initial weights are parameterized by the physical morphology of the fault impact (such as an exponentially decaying oscillating waveform). During forward propagation, this convolutional layer performs a one-dimensional convolution operation on the primary weighted spectrum to extract local features that match the preset fault waveform. More importantly, the parameters of this convolutional layer are designed to be differentiable and learnable, allowing for fine-tuning based on the backpropagation gradient during network training. This enables the extracted feature waveforms to adaptively match subtle differences in real data while adhering to the basic physical morphology. Finally, the original weighted information and the convolutionally enhanced features are concatenated by skipping consecutively, preserving the original resolution information of the physical weighting and compensating for the detail loss that may be caused by the convolution operation. Subsequently, a lightweight attention fusion module is introduced to automatically learn the importance weights of each channel of the concatenated features, dynamically fusing the advantages of both, and ultimately outputting a fused spectral feature tensor that is rich in physical information and robust in features. This deeply embeds the physical mechanism into the extraction and enhancement process of frequency domain features in a differentiable and learnable manner. The spectral attention mask provides soft constraints based on physical priors, the physically constrained convolutional layer realizes feature re-extraction based on physical morphology, and the attention fusion ensures the integrity of information and the robustness of expression. This multi-layered coupled mechanism embedding method ensures that the final mechanism-weighted spectrum retains the strong interpretability of physical laws and possesses data-driven adaptive capabilities, improving the model's sensitivity to early minor faults and its adaptability to complex operating conditions, laying a solid physical information foundation for subsequent temporal feature extraction and accurate diagnosis.
[0077] Step S200: Targeted augmentation data for physical information enhancement is obtained by processing multi-source heterogeneous information and real-time operating parameters through physical simulation and generative modeling.
[0078] After obtaining the mechanism-weighted spectrum, in order to further address the problem of scarce fault samples and difficulty in covering all fault modes in actual industrial scenarios, high-fidelity targeted augmentation data is generated in an active and controllable manner to improve the generalization ability of subsequent diagnostic models under small sample and complex working conditions. Figure 3 This is a block diagram of the method provided in this application for processing multi-source heterogeneous information and real-time operating parameters to obtain targeted augmentation data for physical information enhancement. (For example...) Figure 3As shown, the process of obtaining targeted augmentation data for physical information enhancement by processing multi-source heterogeneous information and actual operating parameters through physical simulation and generative modeling includes the following steps:
[0079] Step S201: Obtain the inherent parameters of the motor and the preset fault types, and input the inherent parameters and fault types into the dynamic simulation model built based on digital twin to generate simulation signals containing fault characteristics.
[0080] Step S202: Input multi-source heterogeneous information into a pre-trained diagnostic model to identify fault categories with classification confidence below a threshold as difficult samples.
[0081] Step S203: Input the simulation signal and the difficult sample into the diffusion model with the difficult sample as the generation condition to generate the initial augmentation data for the difficult sample.
[0082] Step S204: Input the initial augmentation data and real fault samples from multi-source heterogeneous information into a recurrent generative adversarial network to perform style transfer so that the feature distribution of the initial augmentation data is aligned with the feature distribution of the real fault samples to output targeted augmentation data.
[0083] Specifically, firstly, based on the inherent parameters of the target motor, such as its model, geometric dimensions, and material properties, as well as the types of faults to be simulated, such as pitting corrosion on the outer race and cracks on the inner race, a parameterized digital twin model is established in dynamic simulation software. This model can not only simulate the stable state of the fault, but also simulate the evolution process of the fault from initiation to expansion by adjusting fault parameters such as damage size and location, generating simulation signals containing clear physical characteristics. These simulation signals possess physical realism, but their distribution differs from real sensor data. Then, to achieve targeted enhancement of the model, a difficult sample focusing mechanism is introduced. The currently collected real fault samples are input into a pre-trained diagnostic model. By analyzing the classification probabilities output by the model, fault categories with confidence levels below a preset threshold or those easily confused are identified. These identified difficult samples represent the weak points in the current model's knowledge and will become the key targets for subsequent data augmentation. Next, for the identified difficult samples, the generated simulation signal is used as the physical prior, and the category label of the difficult samples is used as the generation condition, both input into a conditional diffusion model. This model progressively destroys the simulated signal into noise through a forward diffusion process, then learns to recover the data from the noise. During this recovery process, guided by challenging sample conditions, the model generates initial augmented data that statistically approximates these challenging samples while retaining the fault characteristics of the simulated signal in its physical structure. Finally, to eliminate the domain differences between the simulated and real signals, the generated initial augmented data and a small number of real fault samples are input into a recurrent generative adversarial network (GAN) such as CycleGAN. Through adversarial training, a nonlinear mapping function from the simulation domain to the real domain is learned, transferring the style of the initial augmented data (such as noise background and transmission path effects) to be consistent with the real data while retaining its core fault impact characteristics. The final output targeted augmented data is highly aligned with the real data in statistical distribution and strictly consistent with the simulated signal in physical mechanism. This constructs a simulation-generation-real ternary closed-loop physical information augmentation system, integrating the causality provided by physical simulation, the diversity generated by the diffusion model, and the realism of style transfer, achieving a shift from passively relying on data to actively creating data. The targeted augmentation data it generates can precisely strengthen the current weak links of the model. It not only effectively solves the problem of data scarcity in small sample scenarios, but also endows the model with the ability to learn the law of fault development by introducing simulation data of fault evolution process. This lays a solid data foundation for the subsequent realization of high-precision fault diagnosis and early warning, and improves the model's generalization ability and robustness in complex industrial environments.
[0084] Step S300: Construct a hybrid diagnostic network containing a Transformer, taking mechanism-weighted spectrum and target enhancement data as input, and extracting global temporal features in the Transformer using spectral attention mask as a bias of self-attention mechanism.
[0085] After obtaining the mechanism-weighted spectrum and targeted enhancement data, a hybrid diagnostic network needs to be constructed that deeply integrates physical priors into the feature learning process to achieve accurate extraction and representation of fault features. To this end, a hybrid diagnostic network incorporating Transformers is constructed, embedding physical laws into the core computational units of the network in a structured manner. The construction of the hybrid diagnostic network incorporating Transformers includes the following steps:
[0086] Step S301: Decompose the mechanism-weighted spectrum into time-step embedding vectors to form a token sequence. In the encoder layer of the Transformer, apply a masking constraint to the self-attention matrix of the token sequence with a spectral attention mask to generate a physically enhanced hidden layer representation.
[0087] Step S302 involves refining the fault feature map layer by layer by cascading multiple encoder layers, and introducing a differentiable wavelet transform layer between encoder layers to perform multi-scale time-frequency decomposition on the hidden layer representation to extract residual features.
[0088] Step S303: Connect the residual feature jump to the fault feature map to construct a hybrid diagnostic network containing a physically enhanced Transformer.
[0089] Specifically, the mechanism-weighted spectrum is first divided into several consecutive time windows along the time dimension. Each window serves as a time step, and is transformed into a fixed-dimensional embedding vector through linear mapping, thus forming the token sequence required by the Transformer network. Subsequently, in the encoder layer of the Transformer, the generated spectral attention mask is used as an explicit physical prior to actually mask and constrain the score matrix generated by the token sequence during self-attention computation. This masking constraint is not a simple filter, but rather a bias term superimposed on the dot product of the query matrix and the key matrix. This causes the attention mechanism to naturally focus on the time regions corresponding to those frequency bands marked as important by the physical mechanism when calculating the correlation weights between time steps in the sequence. Through this mechanism, the hidden layer representation generated by the network is injected with the genes of fault physics from the very beginning, achieving an initial fusion of mechanism and data.
[0090] To capture the rich time-frequency information contained in the transient impacts of the fault signal, this network does not simply stack multiple identical encoder layers. Instead, a differentiable wavelet transform layer is introduced between network layers during the process of cascading multiple standard encoder layers to progressively extract abstract fault feature maps. This layer performs parallel multi-scale wavelet decomposition on the hidden layer representation of the current layer, explicitly separating high-frequency details from low-frequency contours in the signal. Because this wavelet transform layer is differentiable, its decomposition scale, mother wavelet basis, and other parameters can be fine-tuned during end-to-end training of the entire network, allowing it to adaptively match the optimal time-frequency atoms for different fault types. Finally, the multi-scale time-frequency residual features extracted by the differentiable wavelet transform layer are directly added to the fault feature map output by subsequent encoder layers via skip connections. This residual connection structure, on the one hand, compensates for the loss of transient impact details that may occur during the abstraction process in deep networks, and on the other hand, forces the network to simultaneously include high-level semantic information learned through data-driven learning and low-level structural information extracted by physical transformations in the final feature representation. Thus, by constructing a hybrid diagnostic network, the physical mechanism (spectral attention mask) is embedded into the self-attention core of the Transformer, and the physical transformation (wavelet decomposition) is used as a differentiable component of the network. This achieves deep synergy between physical laws and data features throughout the forward propagation and backward training process, enabling the global temporal features extracted by the network to have both data sensitivity to distinguish different fault modes and strong interpretability that conforms to physical laws.
[0091] After obtaining the hidden layer representation extracted by the physically enhanced Transformer, this representation needs to be further transformed into global temporal features with physical semantics to support subsequent diagnostic tasks. To this end, this application uses the generated mechanism-weighted spectrum and targeted enhancement data together as input to a hybrid diagnostic network. This allows the network to learn statistical regularities from real data during training and continuously strengthen its understanding of physical fault modes through enhanced data, thereby improving its generalization ability under small sample sizes and complex operating conditions. The extraction of global temporal features in the Transformer using a spectral attention mask as a bias for the self-attention mechanism includes the following steps:
[0092] Step S304: The hidden layer representation is linearly projected onto three different feature spaces to obtain the query matrix, key matrix and value matrix.
[0093] Step S305: The spectral attention mask is superimposed as a bias term onto the dot product of the query matrix and the key matrix to generate the physical bias attention score.
[0094] Step S306: Normalize the physical bias attention score to obtain the attention weight matrix.
[0095] Step S307: Perform a weighted aggregation operation on the attention weight matrix and the value matrix to obtain global temporal features guided by the prior knowledge of the fault mechanism.
[0096] Introducing physical priors into the Transformer's self-attention mechanism, this involves incorporating a spectral attention mask as a bias term into the attention score calculation process, thereby explicitly guiding the feature extraction process with mechanistic knowledge. Specifically, let the input hidden layer representation be... , where n is the number of time steps and d is the feature dimension. This is achieved through three learnable linear transformation matrices. The query matrix is calculated separately. Key matrix Sum matrix This lays the foundation for subsequent calculations of attention mechanisms. Traditional self-attention score calculation is... Building upon this, this application introduces a spectral attention mask matrix generated by mechanistic structured embedding. Its elements represent the correlation strength between different time steps in the physical frequency band. The physical bias attention score is obtained by adding M as a bias term to the dot product result. This naturally leads the model to focus on time regions corresponding to frequency bands marked as important by physical mechanisms when calculating temporal dependencies, thus embedding the physical laws of faults into the calculation process of attention distribution. Next, a Softmax function is applied row-wise to the score matrix S to obtain the normalized attention weight matrix A = Softmax(S). This weight matrix not only reflects data-driven temporal correlations but also incorporates physical priors introduced by the spectral attention mask, ensuring that the model consistently follows the guidance of the physical laws of faults during global modeling. Finally, weighted aggregation is performed... , to obtain output features This feature, while preserving the original temporal information, incorporates the physical mechanisms' preference for key frequency bands, achieving deep coupling between physical knowledge and data features at the feature representation level. Thus, the spectral attention mask is explicitly embedded as a bias term into the Transformer's self-attention mechanism, enabling the model to always be guided by the physical laws of the fault when extracting global temporal features, enhancing its ability to focus on fault-sensitive components. This not only improves the physical interpretability of the feature representation but also allows the model to maintain high robustness under early weak faults and varying operating conditions, laying a solid feature foundation for subsequent accurate diagnosis.
[0097] Step S400: Based on the global time-series features, input to the multi-task decoder to output the motor's fault type probability, fault quantification parameters, and fault feature heatmap used to characterize the diagnostic basis in parallel. The diagnostic result is obtained based on the fault type probability, fault quantification parameters, and fault feature heatmap.
[0098] After obtaining the global temporal features guided by prior fault mechanisms, they need to be mapped into diagnostic results with clear physical semantics to achieve a comprehensive characterization of the motor state. To this end, this application inputs the global temporal features into a multi-task decoder, which consists of three parallel sub-networks corresponding to fault classification, quantitative estimation, and interpretability generation tasks, respectively. Specifically, the global temporal features are first input into a classification sub-network composed of fully connected layers, and the Softmax function outputs the mode probability of the motor belonging to various fault types, forming fault type probabilities. Simultaneously, the features are input into a quantitative estimation sub-network composed of a multilayer perceptron regression network, directly regressing the physical parameters of the fault, such as bearing damage size and crack location angle, outputting quantitative fault parameters. To visualize the diagnostic basis, a heatmap generation sub-network based on an attention mechanism is also introduced. This sub-network upsamples and interpolates the attention weights of the global temporal features in the time dimension, generating a fault feature heatmap with the same length as the original signal, used to highlight impact sequences or spectral components that significantly contribute to the diagnostic decision. Thus, through the aforementioned parallel multi-task decoding architecture, classification, regression, and localization results are simultaneously output from the same feature space, forcing the network to learn more physically consistent shared representations. On one hand, the causal constraints between multiple tasks improve the physical and logical consistency of the diagnostic results. On the other hand, the fault feature map provides an intuitive decision-making basis for traditional black-box models, significantly enhancing the model's interpretability and engineering credibility. Furthermore, it enables the model not only to answer whether a fault exists, but also to explain where the fault is and how severe it is, providing comprehensive information support for accurate equipment maintenance and lifespan prediction.
[0099] After obtaining the fault type probabilities, fault quantification parameters, and fault feature heatmaps output in parallel by the multi-task decoder, these raw outputs need to be transformed into interpretable diagnostic conclusions with engineering guidance significance to achieve an effective closed loop from data to decision-making. To this end, the multi-task outputs are further post-processed and fused to generate structured diagnostic results. The diagnostic results obtained based on the fault type probabilities, fault quantification parameters, and fault feature heatmaps include the following steps:
[0100] Step S401: Perform confidence calibration on the probability of fault type to obtain the calibrated confidence level.
[0101] Step S402: Construct damage quantification indicators based on fault quantification parameters.
[0102] Step S403: Overlay the fault feature heatmap onto the original signal waveform corresponding to the multi-source heterogeneous information to generate a visual diagnostic map.
[0103] Step S404: The calibration confidence level, damage quantification index and visual diagnostic atlas are fused and processed by the interpretability report generator to output a diagnostic result containing explanations of the diagnostic basis.
[0104] Specifically, the original fault type probability output by the classification subnetwork is first input into a confidence calibrator based on ordinal regression or temperature scaling. This calibrator recalibrates the probability distribution predicted by the model by minimizing the negative log-likelihood loss, eliminating probability bias caused by data distribution shifts or model overconfidence, thus outputting a more realistic post-calibrated confidence level that reflects the accuracy and feasibility of the diagnosis, providing a reliable probabilistic basis for subsequent maintenance decisions. Then, for specific physical quantities output by the regression network, such as the estimated diameter of bearing outer ring pitting or the estimated mass eccentricity of rotor imbalance, they are compared with preset health thresholds or industry standards. Through a normalization mapping function, such as the sigmoid transform, they are converted into a dimensionless damage quantification index. This index, with a value ranging from 0 to 1, is used to intuitively characterize the severity of the fault, thereby transforming abstract physical parameters into easily understandable health status levels. Next, the attention weight sequence output by the heatmap generation subnetwork, which is the same length as the origin signal, is upsampled and smoothed to correspond one-to-one with the sampling points of the original vibration or current waveform. Subsequently, the weight sequence is overlaid in color mapping below or in the corresponding area of the original time-domain waveform, generating a visual diagnostic atlas. The highlighted areas in the atlas represent the key impact sequences or modulation sidebands upon which the model bases its diagnostic decisions, providing intuitive visual evidence for the diagnostic conclusions. Finally, the generated post-calibration confidence level, damage quantification indicators, and visual diagnostic atlas, along with basic fault type information, are input into a natural language generator based on a preset template. This generator automatically combines the input numerical values and atlas features into a structured diagnostic report text, ultimately outputting a complete diagnostic result containing numerical conclusions, quantification levels, and textual evidence. This transforms the raw output of the black-box model into multi-dimensional, visual diagnostic conclusions, enhancing the model's credibility and interpretability in industrial applications. It enables maintenance personnel to quickly understand the nature of the fault, locate its position, and assess its severity, thus providing comprehensive and transparent information support for developing precise maintenance strategies.
[0105] Figure 4 This is a schematic diagram of the overall architecture of the intelligent motor diagnostics provided in this application. Figure 4As shown, the overall framework for intelligent motor diagnosis is an end-to-end mechanism-guided deep network framework, consisting of five execution units: a signal acquisition layer, a preprocessing layer, a mechanism-structured embedding network, a physical information enhancement data generation system, and a network training, adaptive fusion output, and diagnostic layer. The signal acquisition layer collects heterogeneous signals from multiple sources, such as vibration, current, and temperature. The preprocessing layer performs noise reduction, filtering, and other preprocessing on the information collected by the signal acquisition layer, and performs time-frequency transformation operations to provide high-quality input for subsequent units. The mechanism-structured embedding network encodes the physical laws of the fault into a learnable spectral attention mask, and weights the frequency domain spectral tensor to obtain a mechanism-weighted spectrum. The physical information enhancement data generation unit generates targeted enhancement data based on digital twin dynamics simulation and generative models. The Transformer hybrid network is constructed by integrating network training, adaptive fusion output, and diagnostic layers. It integrates mechanism-weighted spectrum and targeted enhancement data to extract global temporal features. The fusion weights of mechanism and data modules are dynamically adjusted according to real-time operating parameters. Online incremental learning is achieved through knowledge distillation. Fault types and quantitative parameters are output through a multi-task decoder. Finally, interpretable diagnostic results with visualized graphs are generated. Each unit works in a progressive and collaborative manner to achieve deep fusion diagnosis of mechanism and data.
[0106] Figure 5 This is a connection diagram of a motor intelligent diagnostic system provided in an embodiment of this application. Figure 5 As shown, a motor intelligent diagnostic system includes: an embedded module, an enhancement module, a network module, and a diagnostic module.
[0107] The system comprises the following modules: An embedding module acquires multi-source heterogeneous information about the motor and its real-time operating parameters. It transforms this information to the frequency domain to obtain a spectral tensor and maps the real-time operating parameters to a learnable spectral attention mask based on the physical mechanism of the motor fault. The spectral tensor is then weighted using the spectral attention mask to obtain a mechanism-weighted spectrum. An enhancement module processes the multi-source heterogeneous information and real-time operating parameters through physical simulation and generative models to obtain targeted enhancement data for physical information enhancement. A network module constructs a hybrid diagnostic network including a Transformer. It takes the mechanism-weighted spectrum and targeted enhancement data as input and extracts global temporal features within the Transformer using a spectral attention mask as a bias for its self-attention mechanism. A diagnostic module inputs these global temporal features into a multi-task decoder to output the motor's fault type probability, fault quantification parameters, and a fault feature heatmap used to characterize the diagnostic criteria. The diagnostic results are obtained based on the fault type probability, fault quantification parameters, and fault feature heatmap.
[0108] The other functions performed by the aforementioned embedded module, enhancement module, network module, and diagnostic module, as well as the technical details of each function, are the same as or similar to the corresponding features in the previously described intelligent motor diagnostic method, and therefore will not be repeated here.
[0109] This application also provides a computer storage medium storing a computer program that, when run on a computer, enables the computer to execute the steps in the aforementioned intelligent motor diagnostic method.
[0110] It should be understood that although the steps in the flowcharts in the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily performed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order requirement for the execution of these steps, and they can be performed in other orders.
[0111] The above description is only a partial embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.
Claims
1. A method for intelligent diagnosis of motors, characterized in that, The method includes: The process involves acquiring multi-source heterogeneous information of the motor and its real-time operating parameters, transforming the multi-source heterogeneous information to the frequency domain to obtain a spectral tensor, mapping the real-time operating parameters to a learnable spectral attention mask based on the physical mechanism of the motor fault, and weighting the spectral tensor with the spectral attention mask to obtain a mechanism-weighted spectrum. The multi-source heterogeneous information and the real-time operating parameters are processed by physical simulation and generative modeling to obtain targeted enhancement data for physical information enhancement; A hybrid diagnostic network containing a Transformer is constructed, taking the mechanism-weighted spectrum and the target enhancement data as input, and extracting global temporal features in the Transformer using the spectral attention mask as a bias of the self-attention mechanism; The global time-series features are input into a multi-task decoder to output the motor's fault type probability, fault quantification parameters, and fault feature heatmap used to characterize the diagnostic basis in parallel. The diagnostic result is obtained based on the fault type probability, fault quantification parameters, and fault feature heatmap. The step of mapping the real-time operating parameters into a learnable spectral attention mask based on the physical mechanism of motor faults includes: The inherent parameters of the motor are obtained, and the real-time operating parameters and the inherent parameters are input into a fault frequency encoder that encodes a fault characteristic frequency formula to obtain a theoretical fault characteristic frequency set. A Gaussian spectral attention base mask containing a learnable center frequency, learnable bandwidth, and learnable attenuation coefficient is constructed based on the theoretical fault feature frequency set as the center frequency. The real-time operating parameters are input into a lightweight adjustment network to predict the adaptive adjustment amount of the operating conditions for each learnable parameter. The adaptive adjustment amount of the working condition is superimposed with the corresponding learnable parameter in the Gaussian spectral attention base mask to obtain the dynamic spectral attention mask after nonlinear modulation of the working condition parameters. The process of processing the multi-source heterogeneous information and the real-time operating parameters through physical simulation and generative models to obtain targeted enhancement data for physical information enhancement includes: The inherent parameters of the motor and the preset fault types are obtained, and the inherent parameters and the fault types are input into the dynamic simulation model based on digital twin to generate simulation signals containing fault characteristics. The multi-source heterogeneous information is input into a pre-trained diagnostic model to identify fault categories with classification confidence scores below a threshold as challenging samples. The simulation signal and the difficult sample are input into a diffusion model that uses the difficult sample as the generation condition to generate initial augmentation data for the difficult sample; The initial augmentation data and real fault samples from the multi-source heterogeneous information are input into a recurrent generative adversarial network for style transfer so that the feature distribution of the initial augmentation data is aligned with the feature distribution of the real fault samples to output targeted augmentation data.
2. The method according to claim 1, characterized in that, The step of transforming the multi-source heterogeneous information to the frequency domain to obtain the spectral tensor includes: Customized hierarchical preprocessing adapted to the characteristics of each type of signal in the multi-source heterogeneous information is performed to obtain multi-channel preprocessed signals. The multi-channel preprocessed signals are subjected to time-frequency transformation to obtain multi-channel spectra; The multiple spectrum channels are filtered to remove redundant spectrum to obtain the filtered spectrum; The filtered spectra are aligned by timestamps and concatenated to obtain a spectral tensor.
3. The method according to claim 2, characterized in that, The step of weighting the spectral tensor with the spectral attention mask to obtain the mechanism-weighted spectrum includes: The primary weighted spectrum is obtained by multiplying the spectral attention mask element-wise with the spectral tensor. The primary weighted spectrum is input into a physical constraint convolutional layer composed of fault physical laws encoding to enhance features and obtain a mechanism-enhanced feature map. The mechanism-enhanced feature map and the primary weighted spectrum are concatenated along the channel dimension and then fused by attention to obtain the fused spectral feature as the mechanism-weighted spectrum.
4. The method according to claim 1, characterized in that, The construction of the hybrid diagnostic network including Transformer includes: The mechanism-weighted spectrum is decomposed into time-step embedding vectors to form a token sequence. In the encoder layer of the Transformer, the self-attention score matrix of the token sequence is masked by the spectral attention mask to generate a physically enhanced hidden layer representation. Fault feature maps are extracted layer by layer by cascading multiple encoder layers, and differentiable wavelet transform layers are introduced between encoder layers to perform multi-scale time-frequency decomposition on the hidden layer representation to extract residual features. The residual features are skipped to the fault feature map to construct a hybrid diagnostic network that includes a physically enhanced Transformer.
5. The method according to claim 4, characterized in that, The step of extracting global temporal features in the Transformer using the spectral attention mask as a bias for the self-attention mechanism includes: The hidden layer representation is linearly projected onto three different feature spaces to obtain the query matrix, key matrix, and value matrix; The spectral attention mask is superimposed as a bias term onto the dot product of the query matrix and the key matrix to generate a physical bias attention score. The physical bias attention score is normalized to obtain the attention weight matrix; The attention weight matrix and the value matrix are weighted and aggregated to obtain global temporal features guided by fault mechanism priors.
6. The method according to claim 1, characterized in that, The diagnostic results obtained based on the fault type probability, fault quantification parameters, and fault feature heatmap include: The confidence level of the fault type probability is calibrated to obtain the calibrated confidence level. Based on the aforementioned fault quantification parameters, a damage quantification index is constructed; The fault feature heatmap is superimposed onto the original signal waveform corresponding to the multi-source heterogeneous information to generate a visual diagnostic map. The system integrates the post-calibration confidence level, the damage quantification index, and the visualized diagnostic atlas, and processes the data through an interpretable report generator to output a diagnostic result that includes explanations of the diagnostic criteria.
7. A motor intelligent diagnostic system, characterized in that, The system includes: an embedding module, an enhancement module, a network module, and a diagnostic module; wherein, The embedded module is used to acquire multi-source heterogeneous information of the motor and its real-time operating parameters, transform the multi-source heterogeneous information to the frequency domain to obtain a spectral tensor, and map the real-time operating parameters into a learnable spectral attention mask according to the physical mechanism of motor faults, and use the spectral attention mask to weight the spectral tensor to obtain a mechanism-weighted spectrum. The enhancement module is used to process the multi-source heterogeneous information and the real-time operating parameters through physical simulation and generative modeling to obtain targeted enhancement data for physical information enhancement. The network module is used to construct a hybrid diagnostic network containing a Transformer, taking the mechanism-weighted spectrum and the targeted enhancement data as inputs, and extracting global temporal features in the Transformer using the spectral attention mask as a bias of the self-attention mechanism. The diagnostic module is used to input the global time-series features into the multi-task decoder to output the motor's fault type probability, fault quantification parameters, and fault feature heatmap used to characterize the diagnostic basis in parallel, and to obtain the diagnostic result based on the fault type probability, fault quantification parameters, and fault feature heatmap. The step of mapping the real-time operating parameters into a learnable spectral attention mask based on the physical mechanism of motor faults includes: The inherent parameters of the motor are obtained, and the real-time operating parameters and the inherent parameters are input into a fault frequency encoder that encodes a fault characteristic frequency formula to obtain a theoretical fault characteristic frequency set. A Gaussian spectral attention base mask containing a learnable center frequency, learnable bandwidth, and learnable attenuation coefficient is constructed based on the theoretical fault feature frequency set as the center frequency. The real-time operating parameters are input into a lightweight adjustment network to predict the adaptive adjustment amount of the operating conditions for each learnable parameter. The adaptive adjustment amount of the working condition is superimposed with the corresponding learnable parameter in the Gaussian spectral attention base mask to obtain the dynamic spectral attention mask after nonlinear modulation of the working condition parameters. The process of processing the multi-source heterogeneous information and the real-time operating parameters through physical simulation and generative models to obtain targeted enhancement data for physical information enhancement includes: The inherent parameters of the motor and the preset fault types are obtained, and the inherent parameters and the fault types are input into the dynamic simulation model based on digital twin to generate simulation signals containing fault characteristics. The multi-source heterogeneous information is input into a pre-trained diagnostic model to identify fault categories with classification confidence scores below a threshold as challenging samples. The simulation signal and the difficult sample are input into a diffusion model that uses the difficult sample as the generation condition to generate initial augmentation data for the difficult sample; The initial augmentation data and real fault samples from the multi-source heterogeneous information are input into a recurrent generative adversarial network for style transfer so that the feature distribution of the initial augmentation data is aligned with the feature distribution of the real fault samples to output targeted augmentation data.
8. A computer-readable storage medium having a computer program stored thereon that can run on a processor, characterized in that, When the computer program is executed by the processor, it implements a motor intelligent diagnostic method as described in any one of claims 1 to 6.