Method for constructing motor fault diagnosis model

Abstract

The invention belongs to the technical field of motor fault diagnosis, and particularly relates to a method for constructing a motor fault diagnosis model, which comprises the following steps: deploying a multi-modal sensor and carrying out data acquisition; preprocessing the collected data at the edge end; carrying out multi-modal data fusion and carrying out credibility estimation on the multi-modal fusion data; learning and identifying the depth representation of the fault mode according to the self-supervised time sequence comparison representation; incremental recognition and disastrous forgetting suppression are adopted; performing influence estimation and linear correction according to environmental causality; a non-linear compensation and edge-cloud cooperative triggering strategy of uncertainty gating is constructed; according to the method, the fault identification accuracy and the recall rate can be remarkably improved under the complex working condition, the false alarm rate and the missing report rate are remarkably reduced, the robustness to environment covariant and short-time interference is enhanced, and meanwhile, an end-cloud cooperative fault diagnosis system which supports online incremental updating, privacy protection, real-time alarm of a limited bandwidth field station and cloud continuous optimization is supported.

Description

Technical Field

[0001] This invention belongs to the field of motor fault diagnosis technology, and specifically relates to a method for constructing a motor fault diagnosis model. Background Technology

[0002] Chinese patent application number CN202411644708.0 discloses a method, apparatus, and approach for constructing a motor fault diagnosis model. This invention involves: constructing an initial model; acquiring multiple sets of health state data of a motor in a healthy state and using these sets as a training sample set; training the initial model using the training sample set based on adversarial training techniques to obtain a motor fault diagnosis model; the motor fault diagnosis model is used to output the encoding loss and classification result corresponding to the input state data, and fault diagnosis is performed jointly based on the encoding loss and classification result.

[0003] Although in the field of motor fault diagnosis technology, this invention, through its method, apparatus, and method for constructing motor fault diagnosis models, relies solely on health state data for training, and the generator, through optimization of multiple loss functions, makes the modeling of health state data more accurate, enabling precise detection and diagnosis of potential faults, the existing technology has the following problems in the context of wind power generation:

[0004] 1. Wind turbine generators are usually located in remote areas, which are far from cities and maintenance bases, resulting in low efficiency of manual inspections and high maintenance costs.

[0005] 2. Most existing wind turbine generator fault diagnosis technologies rely on single sensor data. Due to the lack of comprehensive analysis of multiple fault modes, the accuracy and reliability of fault diagnosis are insufficient, which in turn affects the stable operation of wind turbine generators.

[0006] Therefore, a method for constructing motor fault diagnosis models is needed that can integrate multimodal data and make adaptive adjustments to solve the aforementioned existing technical problems. Summary of the Invention

[0007] This invention provides a method for constructing a motor fault diagnosis model, aiming to solve the following problems: Wind turbine generators are usually located in remote areas, which leads to low efficiency and high maintenance costs due to their distance from cities and maintenance bases; Existing wind turbine generator fault diagnosis technologies mostly rely on single sensor data, and the lack of comprehensive analysis of multiple fault modes results in insufficient accuracy and reliability of fault diagnosis, thus affecting the stable operation of wind turbine generators.

[0008] The technical solution adopted by this invention to solve the above problems is as follows: a method for constructing a motor fault diagnosis model, comprising: deploying multimodal sensors and collecting data; preprocessing the collected data at the edge; multiplying the features of each mode by their corresponding confidence weights and introducing them into a self-attention cross-modal fusion layer; and incorporating temporal consistency constraints into the loss function of the fusion network.

[0009] A representation learning strategy combining self-supervised contrastive learning and weakly supervised clustering is adopted;

[0010] In the representation learning phase, the online fine-tuning process is introduced by combining the solidification of elastic weights with the experience playback mechanism, and a composite loss function is constructed.

[0011] An algorithm for constructing an environmental impact regression matrix identifies the interference paths of each modal signal and performs counterfactual correction on the signals.

[0012] A nonlinear compensation and edge-cloud collaborative triggering strategy for uncertainty gating is constructed. The edge node calculates the uncertainty index of the time-series diagnostic model for each inference output, compares the uncertainty index with the short-term moving average, obtains the judgment result of the current sample, and processes it according to the judgment result.

[0013] As a preferred embodiment, the specific steps for deploying the multimodal sensor and acquiring data are as follows:

[0014] By installing acceleration sensors, speed sensors, shaft current sensors, speed encoders, temperature sensors, current sensors, and wind speed, direction, and air pressure sensors at key locations such as the blade root, gearbox housing, bearing housing, generator end, and tower base of the wind turbine generator set, multi-source raw data covering structural vibration, thermal conditions, and environmental operating conditions can be obtained.

[0015] As a preferred embodiment, the specific steps for preprocessing the collected data at the edge are as follows:

[0016] At edge nodes, data collected by multiple sensors is debiased, pre-filtered with bandpass, and transmitted in packets using symmetric encryption. This allows transmission data with key frequency band characteristics to be preserved even in bandwidth-constrained sites. Sensors are calibrated regularly and a unified timestamp is used to obtain comparable sensor data sources.

[0017] As a preferred implementation, the specific steps of multiplying the features of each modality by their corresponding confidence weights and then introducing them into the self-attention cross-modal fusion layer are as follows:

[0018] For each modal signal, an intramodal reliability estimator is constructed to obtain the confidence score of each modality within the current time window. The self-attention cross-modal fusion layer projects each modal feature into a query, key, and value, calculates the intermodal attention score, normalizes it into weights, and then sums the value vectors according to the weights to output the fused features. The confidence fusion formula is as follows:

[0019] ,

[0020] B represents the fusion weight of modality a within time window c, where B is the fusion weight, a is the modality index, and c is the time window index. It is a natural exponential function. The scaling weights for the attention base score. This represents the importance score learned by mode a within time window c. The weights for the confidence term. Let D be the modal confidence score of mode a within time window c, and let D be the modal confidence score. For confidence sensitivity, The weights of the mutual information modulation terms, Let the mutual information modulation term of mode a in time window c be . Here, A represents the mutual information modulation term, and i represents the total number of modes. Let represent the importance score learned by modality i within time window c. Let be the mutual information modulation term of mode i within time window c, and let log be the logarithmic function. For summation, This is the dot product symbol.

[0021] As a preferred embodiment, the specific steps for incorporating temporal consistency constraints into the loss function of the fusion network are as follows:

[0022] After constructing the confidence fusion formula, by incorporating the temporal consistency constraint into the loss function of the fusion network, a fusion model is obtained that can accurately identify short-term transient faults and maintain the temporal continuity of features during training. Specifically, the temporal consistency constraint is refined into two parts and implemented step by step in the training pipeline: First, the L2 norm of the element-wise difference between the fusion features of two adjacent time windows is calculated and accumulated as a constraint term to obtain the velocity component that directly suppresses abrupt changes; Second, the temporal smoothing error of the fusion features is calculated within a fixed-length sliding window and added as a regularization term to the total loss to obtain a feature sequence that suppresses high-frequency jitter and retains long-term trends; By parallel computing of mutual modality and mutual information and suppressing redundant modal features with mutual information scoring, a joint representation with higher information content and lower redundancy is obtained; By performing temporal alignment and necessary interpolation on the modalities, an aligned sequence that can be used for parallel computing is obtained.

[0023] As a preferred implementation, the specific steps of the representation learning strategy that combines self-supervised contrastive learning and weakly supervised clustering are as follows:

[0024] By constructing positive and negative sample pairs within a sliding time window using the aligned temporal data of each modality and pre-training them with a contrastive learning objective, a discriminative feature representation capable of distinguishing between transient and steady-state states under unlabeled conditions is obtained. Center-based clustering is performed on the pre-trained features, and samples manually confirmed by the cluster centers are used as pseudo-labels to backfill the training set, resulting in a weakly supervised label set for targeted fine-tuning. By alternately executing contrastive learning and pseudo-label-based supervised fine-tuning, i.e., consolidating the representation through contrastive learning and then fine-tuning the classification boundary with pseudo-labels, and evaluating the cluster stability after each iteration to determine whether to retain it, a representation learning process that retains self-supervised discriminative ability and gradually forms reliable class division is obtained.

[0025] As a preferred implementation, the specific steps of introducing the online fine-tuning process by combining elastic weight solidification with an experience playback mechanism during the representation learning phase are as follows:

[0026] Elastic weight solidification constructs a quadratic penalty term by diagonally approximating the old model parameters with their corresponding Fisher information and incorporates this penalty term into the fine-tuning loss, thus limiting large updates to previously important parameters when introducing new samples. The experience replay mechanism saves representative historical samples in a replay pool of fixed capacity and uses the replay samples and new samples together for loss calculation and parameter updates in a predetermined proportion during each online fine-tuning, thus obtaining a training strategy that maintains the performance of the old task by retraining on historical data.

[0027] As a preferred implementation, the specific steps for constructing the composite loss function are as follows:

[0028] The time-series diagnostic model extracts low-dimensional features within a sliding time window using a one-dimensional convolutional layer and a recurrent neural network as a time encoder. A subsequent discriminator outputs the fault category, enabling the identification of transient and degenerative faults in real-time streaming data and supporting an end-to-end diagnostic model with online incremental updates. A composite loss function is constructed to obtain feature vectors that balance discriminability and signal detail preservation. The formula for the composite loss function is:

[0029] ,

[0030] For composite loss function, To compare learning loss, To control the relative contribution of the timing consistency term to the total loss, For timing consistency functions, To control the contribution of playback loss to the total loss, This represents the experience replay loss function. The regularization function is fixed for elastic weights.

[0031] As a preferred embodiment, the specific steps of the algorithm for constructing the environmental impact regression matrix, which identifies the interference paths of each modal signal and performs counterfactual correction on the signal, are as follows:

[0032] By using external environmental variables such as wind speed, wind direction, temperature, and air pressure as fusion inputs and employing an environmental condition gating mechanism in the fusion layer, the diagnostic output can dynamically adjust the modal fusion weights and decision thresholds under different environmental conditions. Through an algorithm that constructs an environmental impact regression matrix to analyze the interference paths of each modal signal and implements counterfactual correction, a corrected output that restores the true fault characteristic signal can be obtained when there are sudden changes in wind speed or climate disturbances. The algorithm formula for estimating the environmental impact regression matrix is ​​as follows:

[0033] ,

[0034] For the environmental impact regression matrix, To find the matrix that minimizes the cost function in the subsequent equation, O represents the regression matrix variable in the solution. The summation symbol is used, and T represents the time index. Let L be the square of the L2 norm of the vector. Let T be the fused feature vector, and E be the fused feature vector. Let F be the environmental variable vector corresponding to time window T. Let be the regularization coefficient for ridge regression. This indicates that L2 regularization is applied to the square of the 2-norm of O.

[0035] As a preferred implementation, the edge node simultaneously calculates the uncertainty index of the time-series diagnostic model for each inference output and compares this uncertainty index with the short-term moving average to determine whether the current sample is in a high-uncertainty state. A priority reporting packet is constructed for each sample determined to be in a high-uncertainty state, and this reporting packet is symmetrically encrypted and encapsulated with transport layer security at the edge. Upload priority and bandwidth adaptation strategies are set at the edge node. Model fusion and knowledge distillation are performed on incremental models from multiple sites in the cloud, enabling the sharing and optimization of fault knowledge without directly transmitting the original data. The lightweight inference model obtained through knowledge distillation is deployed at the edge node, and model update patches generated by the cloud are pushed according to plan.

[0036] The beneficial effects of this invention are as follows:

[0037] 1. By combining attention-based modal confidence estimation, parallel cross-modal mutual information suppression, and self-attention cross-modal fusion, and incorporating temporal consistency constraints such as minimizing the difference between consecutive frames and smoothing the sliding window into the training loss, robust and discriminative fusion features can still be generated under conditions of sudden wind speed changes and short-term interference. This can reduce the false alarm rate and false negative rate and improve the accuracy and recall of fault identification.

[0038] 2. By simultaneously employing a fixed-capacity experience replay strategy and Fisher-based elastic weight solidification constraints during the online incremental learning phase, the model can retain its ability to identify existing fault modes without significantly losing its ability to fine-tune the model when receiving new fault samples. This ensures the model's stability, maintainability, and continuous learning capability during long-term deployment.

[0039] 3. By implementing signal debiasing, bandpass pre-filtering, and hierarchical compression at edge nodes and triggering on-demand reporting with uncertainty gating, and by using a teacher group based on cross-site weighted fusion in the cloud to perform knowledge distillation on the student model and distribute updates with differential patches, it is possible to significantly reduce uplink bandwidth usage, reduce the number of full retraining times in the cloud, and reduce the computational and communication overhead of edge inference and operation and maintenance without directly transmitting a large amount of raw data. This reduces the total cost of ownership of the system and improves deployment feasibility and economy. Attached Figure Description

[0040] Figure 1 This is a flowchart of a method for constructing a motor fault diagnosis model.

[0041] Figure 2 This is a technical effect diagram of a method for constructing a motor fault diagnosis model. Detailed Implementation

[0042] To make the technical means, creative features, and achieved objectives and effects of this invention easier to understand, the invention is further described below with reference to specific embodiments. However, the following embodiments are merely preferred embodiments of this invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments described herein without creative effort are all within the protection scope of this invention. Unless otherwise specified, the experimental methods in the following embodiments are conventional methods, and the materials and reagents used in the following embodiments are commercially available unless otherwise specified.

[0043] Example 1 Figure 1A flowchart of a method for constructing a motor fault diagnosis model includes: deploying multimodal sensors and acquiring data; preprocessing the acquired data at the edge; performing multimodal data fusion and estimating the reliability of the fused multimodal data; learning to identify deep representations of fault modes based on self-supervised temporal contrastive representation; employing incremental identification and catastrophic forgetting suppression; estimating the impact based on environmental causality and performing linear correction; and constructing a nonlinear compensation and edge-cloud collaborative triggering strategy with uncertainty gating. The specific implementation steps are as follows:

[0044] Step 1: Deploy multimodal sensors and collect data; preprocess the collected data at the edge.

[0045] Specifically, the steps for deploying multimodal sensors for data acquisition are as follows: Accelerometers, speed sensors, shaft current sensors, speed encoders, temperature sensors, current sensors, and wind speed, direction, and air pressure sensors are installed at key locations such as the blade roots, gearbox housing, bearing housing, generator ends, and tower base of the wind turbine. This obtains multi-source raw data covering structural vibration, thermal conditions, and environmental operating conditions, providing a complete observation dimension for subsequent multimodal joint diagnosis. An event-triggered adaptive sampling strategy is adopted in the vibration channel. By setting energy judgment criteria and thresholds, when the energy judgment criteria exceed the threshold, the sampling rate is switched from a baseline value to a high sampling rate. A multi-timescale data acquisition scheme is available that can acquire high-resolution signals during transient impacts and reduce storage and transmission overhead during steady-state conditions. This allows for reduced bandwidth and storage costs while ensuring that fault details are captured. The energy determination is based on calculating short-time energy of the vibration channel within a predetermined time window. After calculation, the short-time energy is compared with a preset threshold. When the short-time energy is greater than the threshold, it is determined to be a transient impact. The transient impact is characterized by a sudden increase in vibration amplitude within a short period of time, with a duration less than a preset duration threshold, accompanied by broadband spectral components. The trigger edge node switches the sampling rate from the baseline sampling rate to a high sampling rate and saves the original waveform of that period for subsequent fault analysis and playback.

[0046] The specific steps for preprocessing the collected data at the edge are as follows: By performing bias correction, bandpass pre-filtering, and symmetric encryption on the data collected by multiple sensors at the edge nodes for packet transmission, the key frequency band characteristics of the transmitted data can still be preserved even in bandwidth-constrained wind farms, thereby reducing communication load and protecting on-site data privacy; By periodically calibrating the sensors and using a unified timestamp, comparable sensor data sources can be obtained, thereby ensuring the basic quality of the data and facilitating subsequent processing; The bandwidth-constrained wind farms refer to wind power farms in remote areas where network connections are intermittently interrupted and high-sampling-rate raw data cannot be continuously transmitted.

[0047] Step 2: Perform multimodal data fusion and estimate the credibility of the fused multimodal data;

[0048] Specifically, the steps for multimodal data fusion are as follows: An intramodal reliability estimator is constructed for each modal signal to obtain the confidence score of each modality within the current time window. This score is then used to assign weights for dynamic fusion and filter out anomalous modalities. The estimator includes an attention-based modal confidence network and a statistics-based noise power estimation submodule. The attention-based modal confidence network performs end-to-end mapping between modal feature vectors and known contextual information, including historical confidence scores and environmental variables. The network outputs the original confidence score as a single scalar value, which is then normalized and mapped to the 0-1 range. This yields a preliminary confidence score reflecting the modality's current information contribution capability, enabling the fusion of complex modal data. The contribution of the modes under the formula is made learnable and discriminable; the noise power estimation submodule of the statistics calculates the power spectral density of the multimodal signal and estimates the peak energy of noise and signal within a sliding time window, and derives the statistical confidence score from two quantitative indicators, noise power ratio and transient energy ratio, so as to obtain an objective numerical quantification of the degree of modal noise pollution, thereby providing physical quantity support for the confidence network and independently judging the modal quality in unlabeled scenarios; the unlabeled scenario refers to the sensor samples obtained during data acquisition that have not been labeled as fault categories by manual or automated methods, such as the initial operation of a wind farm, the lack of manual inspection records at remote stations, or the situation where historical data only exists in the original time series without fault labels.

[0049] The specific steps for confidence estimation of multimodal fusion data are as follows: First, the features of each modality are multiplied by their corresponding confidence weights, and then introduced into a self-attention cross-modal fusion layer to obtain a joint feature representation that adaptively adjusts according to the current working condition. This allows the input to remain stable even when some modalities are contaminated by noise. The self-attention cross-modal fusion layer projects each modal feature as a query, key, and value, calculates the attention score between modalities, normalizes it to weights, and then weights and sums the value vectors according to the weights to output the fusion features. This achieves adaptive enhancement of useful modal information and suppression of redundant modalities according to the current working condition. The value vector is a representation obtained by linearly mapping the feature vector of each modality, carrying the actual information of that modality, and is weighted and summed under the action of attention weights to form the final fusion output. The confidence fusion formula is:

[0050] ,

[0051] B represents the fusion weight of modality a within time window c, where B is the fusion weight, a is the modality index, and c is the time window index. It is a natural exponential function. The scaling weights for the attention base score. This represents the importance score learned by mode a within time window c. The weights for the confidence term. Let D be the modal confidence score of mode a within time window c, and let D be the modal confidence score. For confidence sensitivity, The weights of the mutual information modulation terms, Let the mutual information modulation term of mode a in time window c be . Here, A represents the mutual information modulation term, and i represents the total number of modes. Let represent the importance score learned by modality i within time window c. Let be the mutual information modulation term of mode i within time window c, and let log be the logarithmic function. For summation, The dot product symbol;

[0052] After constructing the confidence fusion formula, by incorporating the temporal consistency constraint into the loss function of the fusion network, a fusion model is obtained that can accurately identify short-term transient faults and maintain the temporal continuity of features during training. This improves the robustness of fusion features under conditions of sudden wind speed changes or short-term interference. The constraint includes a term that minimizes the feature difference between consecutive frames and a feature smoothing term within a sliding window. Specifically, the temporal consistency constraint is refined into two parts and implemented step by step in the training pipeline: First, the L2 norm of the element-wise difference between the fusion features of two adjacent time windows is calculated and accumulated as a constraint term, thereby obtaining a velocity component that directly suppresses sudden jumps and reduces false discrimination caused by transient noise; Second, the temporal smoothing error of the fusion features is calculated within a fixed-length sliding window and added as a regularization term to the total loss, thereby obtaining a feature sequence that suppresses high-frequency jitter and retains long-term variation trends, reducing short-term jitter while preserving degradation trends. Impact: By parallel computing of mutual modal information and suppressing redundant modal features with mutual information scoring, a joint representation with higher information content and lower redundancy is obtained, thereby reducing the computational redundancy of subsequent models and improving discrimination efficiency; by performing temporal alignment and necessary interpolation on the modalities, an aligned sequence that can be used for parallel computing is obtained, thereby ensuring the consistency of the fusion input format; the loss function of the fusion network consists of a task loss for supervising diagnostic performance, a temporal consistency constraint term for ensuring temporal continuity, and a parameter regularization term for preventing overfitting. The temporal consistency constraint term includes a term for minimizing the feature difference between consecutive frames and a feature smoothing term within a sliding window, which are weighted and added to the task loss according to preset weight coefficients to optimize network parameters; the fusion model is a temporal representation learning model based on self-supervised contrastive learning and incorporating temporal consistency constraints in the loss function, which is used to simultaneously consider instantaneous fault discrimination capability and temporal continuity during training.

[0053] Step 3: Learn and identify deep representations of failure modes based on self-supervised temporal contrastive representation; employ incremental identification and catastrophic forgetting suppression.

[0054] Specifically, the steps for identifying deep representations of fault modes using self-supervised temporal contrastive representation learning are as follows: By employing a representation learning strategy combining self-supervised contrastive learning and weakly supervised clustering, the system can distinguish the low-dimensional representation space of new fault modes from existing fault modes under unlabeled or poorly labeled conditions, thereby enabling the system to discover unknown fault modes. The representation learning strategy combining self-supervised contrastive learning and weakly supervised clustering constructs positive and negative sample pairs within a sliding time window using the aligned temporal data of each modality and pre-trains them with a contrastive learning objective. This yields discriminative feature representations capable of distinguishing transient and steady-state states under unlabeled conditions, thereby reducing reliance on manual annotation and improving early fault detection capabilities. Center-based clustering is applied to the features obtained from pre-training, and samples with manually confirmed cluster centers are used as pseudo-labels to backfill the training set, thereby obtaining a weakly supervised label set for targeted fine-tuning. This enhances the clarity of class boundaries and supports the identification of new fault clusters. By alternately performing contrastive learning and pseudo-label-based supervised fine-tuning, i.e., consolidating representations through contrastive learning and then fine-tuning classification boundaries with pseudo-labels, and evaluating cluster stability after each iteration to determine whether to retain them, a representation learning process is obtained that retains self-supervised discriminative ability and gradually forms reliable class division. This enables continuous improvement in recognition ability and reduces the risk of mislabeling in unlabeled scenarios. The positive sample pairs are samples from the same working condition or time-continuous samples; the negative sample pairs are samples randomly sampled from different working conditions.

[0055] In the representation learning phase, Elastic Weight Consolidation (EWC) is combined with an experience replay mechanism to introduce an online fine-tuning process. This ensures that the time-series diagnostic model does not forget its existing recognition capabilities when new fault samples are introduced, thus maintaining long-term stable diagnostic capabilities and gradually expanding the fault knowledge base. A composite loss function is constructed to obtain feature vectors that balance discriminability and signal detail preservation, thereby improving the separability of different fault categories. The Elastic Weight Consolidation constructs a quadratic penalty term by diagonally approximating the old model parameters with their corresponding Fisher information and incorporating this penalty term into the fine-tuning loss. This limits large updates to previously important parameters when new samples are introduced, thus suppressing catastrophic forgetting and maintaining existing recognition capabilities. The old model includes: a snapshot of the model parameters saved before fine-tuning or incremental learning begins, and the previously validated and distributed model. The model version is deployed at edge nodes or in the cloud; the experience replay mechanism saves representative historical samples in a fixed-capacity replay pool and uses the replay samples and new samples together for loss calculation and parameter update in a predetermined ratio during each online fine-tuning, thereby obtaining a training strategy to maintain the performance of the old task by retraining historical data, which can reduce the conflict between old and new samples and improve the stability of incremental learning; the temporal diagnostic model extracts low-dimensional features within a sliding time window by using a one-dimensional convolutional layer and a recurrent neural network as a temporal encoder, and outputs the fault category by a subsequent discriminator, thereby enabling the identification of transient and degenerate faults on real-time streaming data and supporting an end-to-end diagnostic model for online incremental updates, thus providing a computational basis for real-time on-site alarms and continuous learning; the formula for the composite loss function is:

[0056] ,

[0057] For composite loss function, To compare learning loss, To control the relative contribution of the timing consistency term to the total loss, For timing consistency functions, To control the contribution of playback loss to the total loss, This represents the experience replay loss function. For elastic weights, solidify the regularization function;

[0058] The specific steps of incremental identification and catastrophic forgetting suppression are as follows: By setting an uncertainty threshold in the incremental identification module and linking it with manual annotation in a closed loop, when the output confidence is lower than the preset threshold and the clustering algorithm detects a new cluster, expert review is triggered. Experts manually review the data contained in the new cluster and the corresponding environmental context one by one and make a judgment. If the expert confirms that the new cluster belongs to a new fault mode, the expert manually annotates the representative samples in the cluster and adds the annotated samples to the replay pool to update the fault database and cluster center. At the same time, the incremental fine-tuning process with elastic weight solidification constraints is triggered and the review results and evidence are recorded for traceability. If the expert determines that the new cluster is noise, occasional interference or misclustering, the expert marks the cluster as an anomaly and removes it from the training and replay process. At the same time, feedback is fed back to adjust the clustering parameters to reduce similar false alarms, and the review record is saved for subsequent analysis. This results in a closed-loop iterative mechanism that combines automatic learning and manual knowledge in an orderly manner, which plays a role in gradually accepting and confirming new fault modes while ensuring that the false alarm rate is controllable.

[0059] Step 4: Estimate the impact based on environmental causality and perform linear correction; construct a nonlinear compensation and edge-cloud collaborative triggering strategy for uncertainty gating;

[0060] Specifically, the steps for estimating the impact of environmental causality and performing linear correction are as follows: By using external environmental variables such as wind speed, wind direction, temperature, and air pressure as fusion inputs and employing an environmental condition gating mechanism at the fusion layer, the diagnostic output of the modal fusion weights and decision thresholds can be dynamically adjusted under different environmental conditions, thereby reducing false alarms and missed alarms caused by environmental covariance; by constructing an environmental impact regression matrix algorithm to address the interference paths of each modal signal and implementing counterfactual correction, the corrected output that restores the true fault characteristic signal can be obtained when there are sudden changes in wind speed or climate disturbances, thereby improving the causal interpretability of the diagnostic conclusions; the environmental condition gating mechanism calculates gating coefficients on the real-time collected environmental variable vectors and uses these gating coefficients to scale each modal feature proportionally, thereby enabling the estimation of the impact of environmental causality under different environmental conditions. The dynamic control unit that automatically adjusts the contribution of each modality information described below can avoid misleading fusion results due to single modality variations caused by environmental covariance and improve the robustness of diagnostic output. Environmental covariance refers to the phenomenon where external environmental variables, such as wind speed, wind direction, temperature, and air pressure, change over time or space, simultaneously causing systematic changes in sensor signals. This means that there is a common relationship between environmental changes and observed signals, resulting in observed signals containing both fault information and environmentally driven non-fault changes. Different environmental conditions refer to situations where wind speed, wind direction, temperature, air pressure, and other environmental quantities change significantly over time or space, including but not limited to sudden increases or decreases in wind speed, rapid changes in wind direction, rapid fluctuations in temperature or air pressure, and sudden environmental noise. The algorithm formula for estimating the environmental impact regression matrix is ​​as follows:

[0061] ,

[0062] For the environmental impact regression matrix, To find the matrix that minimizes the cost function in the subsequent equation, O represents the regression matrix variable in the solution. The summation symbol is used, and T represents the time index. Let L be the square of the L2 norm of the vector. Let T be the fused feature vector, and E be the fused feature vector. Let F be the environmental variable vector corresponding to time window T. Let be the regularization coefficient for ridge regression. This indicates that L2 regularization is applied to the square of the L2 norm of O;

[0063] The specific steps for constructing a nonlinear compensation and edge-cloud collaborative triggering strategy for uncertainty gating are as follows: Edge nodes simultaneously calculate the uncertainty index of the time-series diagnostic model for each inference output and compare this uncertainty index with the short-term moving average to determine whether the current sample is in a high-uncertainty state. This allows for the differentiation between low-uncertainty samples that can be processed locally by the edge and high-uncertainty samples that require cloud verification. The uncertainty index of the time-series diagnostic model includes the Shannon entropy of the prediction distribution used to measure the fuzziness of the output probability distribution, and the complement of the prediction confidence as a direct confidence index. The high-uncertainty state is defined as when any uncertainty index exceeds... When the corresponding preset high threshold is reached, and the duration of the anomaly exceeds the preset duration threshold within a short time window, the system determines it to be in a high uncertainty state and triggers priority reporting and manual review of the sample and context. For low uncertainty samples, if all uncertain indicators are below their corresponding preset low threshold within the same time window and the fluctuation range of the indicators is less than the preset jitter tolerance in the short term, the system will directly give a diagnostic conclusion for such samples at the edge without reporting to the cloud. For high uncertainty samples, if any uncertain indicator exceeds the high threshold, or if two or more uncertain indicators are abnormal at the same time and accompanied by low output confidence, the sample is marked as needing priority reporting and will enter the cloud review and labeling process.

[0064] By constructing a priority reporting packet for each sample identified as having high uncertainty, and performing symmetric encryption and transport layer security encapsulation on the edge, a high-confidence evidence packet can be directly used by the cloud for verification and retraining. This enables priority reporting and automated cloud verification triggering while ensuring data privacy and link security. By setting upload priority and bandwidth adaptation strategies at the edge nodes, such as uploading only compressed waveforms and complete environmental context when the uplink bandwidth is lower than a preset value, a reporting mechanism that balances real-time performance and verification requirements can be achieved even in bandwidth-constrained sites. The reporting packet includes: the original sensor waveform at the trigger time, the corresponding fused feature vector, the model output and uncertainty metric, environmental context such as wind speed, wind direction, temperature, air pressure, and device identifier and timestamp.

[0065] By fusing incremental and implementation models from multiple sites in the cloud and performing knowledge distillation, fault knowledge can be shared and optimized without directly transmitting raw data. This expands the model's applicability and reduces training costs per site. The incremental model is a model instance obtained during online learning by receiving new samples and fine-tuning the model parameters with small steps while retaining replay samples and elastic weights. Its input is the newly arrived fused features and replay samples, and its output is the updated fault discrimination results and a snapshot of model parameters. This is used to gradually accept and identify newly emerging fault modes while suppressing catastrophic forgetting. The implementation model is a validated model deployed on edge nodes. Its input is a preprocessed and aligned multi-mode model. The system uses time-series data, outputting fault category probabilities, uncertainty metrics, and necessary feature embeddings, while meeting predetermined latency, memory, and security constraints to support online alarms and local rollback. By deploying lightweight inference models obtained through knowledge distillation on edge nodes and pushing scheduled update patches generated from the cloud, a deployment architecture that balances real-time alarms with long-term model updates can be achieved, ensuring system availability and ease of maintenance. The incremental models from multiple sites undergo model fusion and knowledge distillation by collecting online fine-tuned candidate models from multiple sites in the cloud, assigning weights to the candidate models based on site reliability, sample quality, and validation performance, and then constructing a teacher model group using a weighted ensemble method. By leveraging the teacher group to generate probabilistic soft labels on unlabeled data and using these soft labels as supervisory signals to train a smaller student model through knowledge distillation, a teacher-student collaborative system is obtained that integrates the discrimination capabilities of various stations and performs stably under diverse operating conditions. This enables the sharing of cross-station knowledge without directly transmitting large amounts of raw time-series data and reduces the training overhead of each edge node. The lightweight inference model restricts the distilled student model to a parameter-sparse neural network and ensures that the single inference latency at edge nodes is below a preset threshold. The output includes fault category probability, confidence level, and necessary embedding vectors, thus enabling real-time alarms and supporting differential patching in environments with limited computing power and bandwidth. The deployable inference model with hot updates and rollback ensures real-time performance and maintainability while preserving the discriminative ability of the teacher model as much as possible. The candidate models are a set of selectable models obtained by initializing and independently training on data from different sites in the cloud. Each member is evaluated based on validation set performance, uncertainty distribution, and sample coverage, and serves as an option for subsequent weighted integration, teacher construction, or differential update screening. The teacher model obtains the ability to integrate the candidate models from multiple sites in the cloud according to the reliability of the sites and the sample quality, and generates probabilistic outputs on the validation set or unlabeled data. This allows for the aggregation of cross-site experience and provides supervision signals to the student model to improve its generalization ability.The student model is trained by distillation using the temperature-based probability distribution output by the teacher model as a soft label and combining it with the original hard labels to obtain a network with significantly reduced parameters and computational cost. This results in a lightweight inference model that can perform inference with low latency on edge devices and achieve performance close to that of the teacher model, enabling real-time alerts and supporting differential patch hot updates.

[0066] like Figure 2 The figure shows the technical effect of a method for constructing a motor fault diagnosis model. The black bars in the figure represent the technical effect achieved by using the present invention, and the gray bars represent the technical effect achieved by using the prior art. It can be seen from the figure that the technical effect of the present invention is better than that of the prior art.

[0067] Example 2, a method for constructing a motor fault diagnosis model based on Example 1, the specific scheme is as follows:

[0068] Step 1: Deploy sensors and preprocess them at the edge;

[0069] Specifically, the steps for deploying sensors and preprocessing them at the edge are as follows: Acceleration sensors, shaft current sensors, temperature sensors, and wind speed and direction sensors are installed at the blade roots, gearbox housing, bearing housing, and generator end of the wind turbine generator, respectively. The sampling rate of the vibration channel is set to no less than 18kHz, and the sampling rate of the electrical channel is set to 2kHz to 8kHz. This obtains multimodal raw data covering structural vibration, electromagnetic characteristics, and environmental conditions, thus providing complete time-frequency information for diagnosis. By performing timestamp synchronization, bandpass filtering, de-polarization, and wavelet denoising on the raw data at the edge device, and by performing layered compression of segmented signals according to the energy preservation strategy and using encrypted transmission, edge-processed data with key fault characteristics can still be preserved in bandwidth-constrained sites. This reduces communication load and ensures data integrity.

[0070] Step two: Perform multimodal fusion and conduct credibility assessment;

[0071] Specifically, the steps for multimodal fusion and credibility assessment are as follows: For each modality, the noise power ratio and sample variance are evaluated within a sliding time window to determine modality confidence. The correlation between modalities is estimated based on historical sample statistics to measure redundancy, thus obtaining confidence and redundancy metrics that can be used for dynamic weighting, providing an objective weighting basis for weighted fusion. In the fusion layer, weighted adaptive parallel projection is used, and the weighted summation is performed. The modal features are adjusted according to confidence and linearly mapped, then summarized into a joint feature vector. This yields a fusion representation that still has discriminative power even when some modalities fail, providing stable input for subsequent representation learning. The weighted adaptive parallel projection projects the feature vectors of each modality in parallel to a common low-dimensional subspace. Weights are adaptively assigned to each projection result based on quantitative indicators such as modality confidence and mutual information calculated within the same time window. The weighted projection results are then element-wise summed and normalized to obtain a joint fusion representation that balances the discriminative power and robustness of each modality.

[0072] Step 3: Perform self-supervised temporal representation learning and incremental recognition;

[0073] Specifically, the steps for self-supervised temporal representation learning and incremental identification are as follows: Positive and negative samples are constructed within a time window using the joint features obtained in step two, and a contrastive learning paradigm is used to train the representation network using joint temporal smoothing. This enables a discriminative low-dimensional representation that can distinguish between transient faults and steady-state conditions under label-less conditions, thereby reducing reliance on manual annotation and improving the ability to identify unknown faults. In the online update process, representative old samples are stored in a fixed-capacity empirical replay pool, and fine-tuning is performed using parameter retention measures. An uncertainty threshold is set, and samples exceeding this threshold trigger a manual review process. This ensures that accepting new fault samples does not significantly reduce the existing recognition capability. The incremental learning system, which boasts high performance and the ability to gradually expand the fault database, can suppress catastrophic forgetting and ensure long-term model stability. The fixed-capacity experience replay pool maintains the historical sample set by reserving a fixed-capacity storage buffer at the edge and following a predetermined replacement strategy, such as a priority replay strategy. This allows for the sampling of historical sample storage units for retraining at a set ratio during the incremental update phase, thereby preventing catastrophic forgetting and providing replayable data. The representative old samples are selected from the replay pool based on information content measures, such as prediction uncertainty, feature space diversity, cluster centrality, or class balance, and samples that can cover the historical fault distribution and have high discriminative power are preferentially retained.

[0074] Step four: Perform environmental causal correction and construct an uncertainty-gated edge-cloud collaborative strategy;

[0075] Specifically, the steps for performing environmental causal correction and constructing an uncertainty-gated edge-cloud collaborative strategy are as follows: Regularized linear regression is used to periodically estimate the linear impact of environmental variables on the fused features, with the residuals serving as the basic features after linear correction. This eliminates the diagnostic input after removing major linear environmental interference, thereby reducing false alarms caused by environmental covariance. Uncertainty measures, such as model prediction variance or entropy, are calculated on the model inference output. Gating rules are set to introduce a context-based nonlinear compensation model under high uncertainty, and the sample and complete context are prioritized and uploaded to the cloud for expert review. This results in an operational mechanism that enables on-demand collaboration between real-time edge response and precise cloud review, achieving both low-latency early warning and high-precision review capabilities. The nonlinear compensation model is a learning-based mapping module used to further correct the linear correction residuals. Its inputs are the linear residual vector, fused features, and environmental variables, and its output is a nonlinear correction term or corrected features for the fused features.

[0076] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for constructing a motor fault diagnosis model, characterized in that: Deploy multimodal sensors and collect data; preprocess the collected data at the edge; multiply the features of each modality by their corresponding confidence weights and introduce them into a self-attention cross-modal fusion layer; incorporate temporal consistency constraints into the loss function of the fusion network; A representation learning strategy combining self-supervised contrastive learning and weakly supervised clustering is adopted; In the representation learning phase, the online fine-tuning process is introduced by combining the solidification of elastic weights with the experience playback mechanism, and a composite loss function is constructed. An algorithm for constructing an environmental impact regression matrix identifies the interference paths of each modal signal and performs counterfactual correction on the signals. A nonlinear compensation and edge-cloud collaborative triggering strategy for uncertainty gating is constructed. The edge node calculates the uncertainty index of the time-series diagnostic model for each inference output, compares the uncertainty index with the short-term moving average, obtains the judgment result of the current sample, and processes it according to the judgment result.

2. The method for constructing a motor fault diagnosis model according to claim 1, characterized in that: The specific steps for deploying multimodal sensors and acquiring data are as follows: By installing acceleration sensors, speed sensors, shaft current sensors, speed encoders, temperature sensors, current sensors, and wind speed, direction, and air pressure sensors at key locations such as the blade root, gearbox housing, bearing housing, generator end, and tower base of the wind turbine generator set, multi-source raw data covering structural vibration, thermal conditions, and environmental operating conditions can be obtained.

3. The method for constructing a motor fault diagnosis model according to claim 1, characterized in that: The specific steps for preprocessing the collected data at the edge are as follows: By performing bias correction, bandpass pre-filtering, and symmetric encryption on the data collected by multiple sensors at the edge nodes for packet transmission, the transmission data with key frequency band characteristics can still be preserved even in bandwidth-constrained field stations. By periodically calibrating the sensors and using a unified timestamp, comparable sensor data sources can be obtained.

4. The method for constructing a motor fault diagnosis model according to claim 1, characterized in that: The specific steps for multiplying the features of each modality by their corresponding confidence weights and then introducing them into the self-attention cross-modal fusion layer are as follows: For each modal signal, an intramodal reliability estimator is constructed to obtain the confidence score of each modality within the current time window. The self-attention cross-modal fusion layer projects each modal feature into a query, key, and value, calculates the intermodal attention score, normalizes it into weights, and then sums the value vectors according to the weights to output the fused features. The confidence fusion formula is as follows: ， B represents the fusion weight of modality a within time window c, where B is the fusion weight, a is the modality index, and c is the time window index. It is a natural exponential function. The scaling weights for the attention base score. This represents the importance score learned by mode a within time window c. The weights for the confidence term. Let D be the modal confidence score of mode a within time window c, and let D be the modal confidence score. For confidence sensitivity, The weights of the mutual information modulation terms, Let the mutual information modulation term of mode a in time window c be . Here, A represents the mutual information modulation term, and i represents the total number of modes. Let represent the importance score learned by modality i within time window c. Let be the mutual information modulation term of mode i within time window c, and let log be the logarithmic function. For summation, This is the dot product symbol.

5. The method for constructing a motor fault diagnosis model according to claim 1, characterized in that: The specific steps for incorporating temporal consistency constraints into the loss function of the fusion network are as follows: After constructing the confidence fusion formula, by incorporating the temporal consistency constraint into the loss function of the fusion network, a fusion model is obtained that can accurately identify short-term transient faults and maintain the temporal continuity of features during training. Specifically, the temporal consistency constraint is refined into two parts and implemented step by step in the training pipeline: First, the L2 norm of the element-wise difference between the fusion features of two adjacent time windows is calculated and accumulated as a constraint term to obtain the velocity component that directly suppresses abrupt changes; Second, the temporal smoothing error of the fusion features is calculated within a fixed-length sliding window and added as a regularization term to the total loss to obtain a feature sequence that suppresses high-frequency jitter and retains long-term trends; By parallel computing of mutual modality and mutual information and suppressing redundant modal features with mutual information scoring, a joint representation with higher information content and lower redundancy is obtained; By performing temporal alignment and necessary interpolation on the modalities, an aligned sequence that can be used for parallel computing is obtained.

6. The method for constructing a motor fault diagnosis model according to claim 1, characterized in that: The specific steps of the representation learning strategy that combines self-supervised contrastive learning and weakly supervised clustering are as follows: By constructing positive and negative sample pairs within a sliding time window using the aligned temporal data of each modality and pre-training them with a contrastive learning objective, a discriminative feature representation capable of distinguishing between transient and steady-state states under unlabeled conditions is obtained. Center-based clustering is performed on the pre-trained features, and samples manually confirmed by the cluster centers are used as pseudo-labels to backfill the training set, resulting in a weakly supervised label set for targeted fine-tuning. By alternately executing contrastive learning and pseudo-label-based supervised fine-tuning, i.e., consolidating the representation through contrastive learning and then fine-tuning the classification boundary with pseudo-labels, and evaluating the cluster stability after each iteration to determine whether to retain it, a representation learning process that retains self-supervised discriminative ability and gradually forms reliable class division is obtained.

7. The method for constructing a motor fault diagnosis model according to claim 1, characterized in that: The specific steps of introducing the online fine-tuning process by combining elastic weight solidification with an experience playback mechanism during the representation learning phase are as follows: Elastic weight solidification constructs a quadratic penalty term by diagonally approximating the old model parameters with their corresponding Fisher information and incorporates this penalty term into the fine-tuning loss, thus limiting large updates to previously important parameters when introducing new samples. The experience replay mechanism saves representative historical samples in a replay pool of fixed capacity and uses the replay samples and new samples together for loss calculation and parameter updates in a predetermined proportion during each online fine-tuning, thus obtaining a training strategy that maintains the performance of the old task by retraining on historical data.

8. The method for constructing a motor fault diagnosis model according to claim 1, characterized in that: The specific steps for constructing the composite loss function are as follows: The time-series diagnostic model extracts low-dimensional features within a sliding time window using a one-dimensional convolutional layer and a recurrent neural network as a time encoder. A subsequent discriminator outputs the fault category, enabling the identification of transient and degenerative faults in real-time streaming data and supporting an end-to-end diagnostic model with online incremental updates. A composite loss function is constructed to obtain feature vectors that balance discriminability and signal detail preservation. The formula for the composite loss function is: ， For composite loss function, To compare learning loss, To control the relative contribution of the timing consistency term to the total loss, For timing consistency functions, To control the contribution of playback loss to the total loss, This represents the experience replay loss function. The regularization function is fixed for elastic weights.

9. The method for constructing a motor fault diagnosis model according to claim 1, characterized in that: The specific steps of the algorithm for constructing the environmental impact regression matrix, which identifies the interference paths of each modal signal and performs counterfactual correction on the signal, are as follows: By using external environmental variables such as wind speed, wind direction, temperature, and air pressure as fusion inputs and employing an environmental condition gating mechanism in the fusion layer, the diagnostic output can dynamically adjust the modal fusion weights and decision thresholds under different environmental conditions. Through an algorithm that constructs an environmental impact regression matrix to analyze the interference paths of each modal signal and implements counterfactual correction, a corrected output that restores the true fault characteristic signal can be obtained when there are sudden changes in wind speed or climate disturbances. The algorithm formula for estimating the environmental impact regression matrix is ​​as follows: ， For the environmental impact regression matrix, To find the matrix that minimizes the cost function in the subsequent equation, O represents the regression matrix variable in the solution. The summation symbol is used, and T represents the time index. Let L be the square of the L2 norm of the vector. Let T be the fused feature vector, and E be the fused feature vector. Let F be the environmental variable vector corresponding to time window T. Let be the regularization coefficient for ridge regression. This indicates that L2 regularization is applied to the square of the 2-norm of O.

10. The method for constructing a motor fault diagnosis model according to claim 1, characterized in that: The specific steps for constructing the nonlinear compensation and edge-cloud collaborative triggering strategy for uncertainty gating are as follows: Edge nodes simultaneously calculate the uncertainty index of the time-series diagnostic model for each inference output and compare the uncertainty index with the short-term moving average to determine whether the current sample is in a high uncertainty state. A priority reporting packet is constructed for each sample determined to be in a high uncertainty state, and the reporting packet is symmetrically encrypted and encapsulated with transport layer security at the edge end. Configure upload priority and bandwidth adaptation strategies at edge nodes; perform model fusion and knowledge distillation on incremental models from multiple sites in the cloud, enabling the sharing and optimization of fault knowledge without directly transmitting the original data; deploy lightweight inference models obtained through knowledge distillation at edge nodes and push model update patches generated by the cloud according to plan.

Images