A motor operation state self-learning and fault early warning method and system

Abstract

This invention discloses a self-learning method and system for motor operation status and fault early warning, relating to the field of motor fault diagnosis technology. It includes the acquisition and standardized processing of multi-source operating data to generate a standardized feature set; a four-level self-learning model deployed at the edge and cloud levels, progressively completing sample mining, fault feature learning, early warning threshold optimization, and maintenance strategy generation, with bidirectional feedback iteration between each level; and master-slave collaboration between the edge master module and the cloud slave module to achieve real-time inference and accurate diagnosis, independent early warning at the edge in case of network anomalies, and synchronization of data and parameters after recovery. This self-learning method and system for motor operation status and fault early warning adopts a four-layer architecture: perception layer – edge master module – cloud slave module – application layer. Through the collaborative cooperation of real-time inference by the edge master module, accurate diagnosis by the cloud slave module, and dynamic optimization by the intelligent maintenance decision module, it achieves early warning of faults throughout the motor's entire lifecycle, prediction of remaining lifespan, and adaptive maintenance decision-making.

Description

Technical Field

[0001] This invention relates to the field of motor fault diagnosis technology, specifically to a method and system for self-learning motor operating status and fault early warning. Background Technology

[0002] As a core power equipment in modern industry and transportation, the reliability of electric motors directly determines the stable operation of the entire production line or equipment. Motor faults are complex and diverse, mainly including bearing wear, rotor imbalance, stator winding short circuit, air gap eccentricity, coupling asymmetry, insulation aging, etc. Different faults exhibit different characteristic patterns in vibration, current, temperature and other signals. In industrial scenarios, motor operating conditions are complex, fault characteristics are easily masked by noise, and fault samples are scarce and manual annotation is costly. More importantly, existing early warning technologies can only output fault warnings and RULs, lacking dynamic tracking of fault evolution speed. Maintenance strategies are mostly generated based on fixed rules, without considering actual scenario factors such as equipment operation priority, maintenance resource constraints, and production plan conflicts. This leads to poor adaptability of maintenance solutions, serious resource waste, or untimely maintenance, posing a huge challenge to the practical application of fault early warning technology.

[0003] Currently, motor fault early warning technologies are mainly divided into two categories: traditional signal processing methods and deep learning methods. Among them, deep learning methods have become a research hotspot due to their powerful feature extraction capabilities. However, existing technologies still have the following key shortcomings: the training of existing deep learning models largely depends on a large number of manually labeled fault samples. However, the incidence of fault events in industrial scenarios is low, fault samples are scarce, and the labeling process requires professional personnel to combine equipment manuals and testing experience, resulting in high labeling costs and long cycles. Existing early warning systems mostly adopt an architecture of edge-side acquisition + cloud-based diagnosis. The edge side only undertakes data acquisition and simple preprocessing functions, without participating in the model training and self-learning process, and the master-slave modules lack a collaborative decision-making mechanism. Fault early warning thresholds are mostly fixed values, set by engineers based on experience or historical data, which cannot adapt to dynamic operating conditions such as motor aging, load fluctuations, and changes in ambient temperature. During motor operation, both previously occurring fault types and new fault types may occur. Existing models only focus on learning the features of new faults, easily forgetting the features of old faults, leading to a decrease in the accuracy of old fault identification, and insufficient intelligence in maintenance decision-making. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for self-learning and fault early warning of motor operating status, in order to solve the problems mentioned in the background art. These problems include: the training of deep learning models relies heavily on a large number of manually labeled fault samples, which takes a long time; the architecture of the early warning system only undertakes data collection and simple preprocessing functions, without participating in the model training and self-learning process, lacking a collaborative decision-making mechanism; the fault early warning thresholds are mostly fixed values ​​set by engineers based on experience or historical data, which cannot adapt to dynamic operating conditions such as motor aging, load fluctuations, and changes in ambient temperature; and the existing models only focus on learning new fault features, easily forgetting old fault features, resulting in a decrease in the accuracy of old fault identification and insufficient intelligence in maintenance decision-making.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for self-learning and fault early warning of motor operating status, comprising the following steps: S1. Collection and standardization of multi-source operating data: Collect multi-dimensional operating data of the motor, perform noise reduction, feature extraction and standardization on the data, generate a standardized feature set, and provide a high-quality data foundation with a unified dimension for subsequent self-learning model training and fault early warning. S2. Training and Iteration of Hierarchical Self-Learning Model: Based on a standardized feature set, a four-level self-learning model with hierarchical deployment on the edge and cloud is used for progressive training. The previous level of self-learning provides data, parameters or results support for the next level, and the optimization results of the next level of self-learning are fed back to the previous level to realize model iteration. In sequence, sample mining, fault feature learning, dynamic optimization of early warning threshold, fault evolution modeling and maintenance strategy generation are completed. S3. Master-Slave Collaborative Fault Warning and Diagnosis: Based on the master-slave architecture of the edge master module and the cloud slave module, the edge master module realizes real-time reasoning and anomaly triggering based on the self-learning model, and the cloud slave module completes accurate diagnosis and decision output. When the network is abnormal, the edge side switches to an independent warning mode. After the network is restored, the data and model parameters of the edge and the cloud are synchronized, forming a closed-loop collaboration of warning-diagnosis-decision.

[0006] In this embodiment, in step S1, a suitable signal processing method is used for denoising and feature extraction based on the characteristics of the data in different dimensions. Only the extracted feature data is standardized to reduce the amount of data transmission, adapt to the collaborative transmission needs of the edge and the cloud, and eliminate the difference in feature dimensions to ensure the convergence efficiency of subsequent model training.

[0007] In this embodiment, in step S2, the four-level self-learning model is implemented hierarchically according to function, and a two-way feedback of data and parameters is formed between each level, specifically: S21, Level 1 Self-Learning: Deployed on the edge side, based on unlabeled standardized feature sets, it automatically mines high-confidence samples and expands the training set through weakly supervised learning algorithms and lightweight feature recognition models, completes the training of the lightweight model on the edge side, provides model support for real-time edge inference, and uploads high-value samples to the cloud to provide a data foundation for cloud model training. S22, Level 2 Self-Learning: Deployed in the cloud, based on sample data uploaded from the edge side, it learns new fault features while memorizing historical fault features through the fusion feature learning model, outputs quantifiable indicators of motor health, provides core input parameters for subsequent early warning threshold optimization, and feeds back the optimized basic model parameters to the edge side to improve the edge model recognition accuracy. S23, Level 3 Self-Learning: Deployed in the cloud, based on the motor health and real-time operating condition changes, it dynamically adjusts the warning threshold through intelligent optimization algorithms to make the threshold adapt to the dynamic operating state of the motor. The generated dynamic threshold is sent to the edge side as the standard for real-time anomaly judgment on the edge side, and at the same time, it receives the warning effect feedback from the edge side to achieve continuous optimization of the threshold. S24, Level 4 Self-Learning: Deployed in the cloud, based on fault type, remaining lifespan and real-time operating data, it tracks the fault development pattern through a time-series evolution model, combines actual operating condition constraints, generates adaptive maintenance strategies through a multi-objective optimization algorithm, and receives feedback on the execution effect of the maintenance strategies to update the fault evolution model and health assessment system in reverse.

[0008] In this embodiment, in step S22, the fusion feature learning model achieves a balance between learning new fault features and remembering historical fault features through weighted fusion of loss functions, avoiding the model forgetting historical fault features and ensuring the stability of the model's long-term operation.

[0009] In this embodiment, in step S24, the actual operating condition constraints include one or more of the following: equipment operation priority, maintenance resource configuration, and production operation window. The generated adaptive maintenance strategy adapts to the actual execution requirements on site, realizing the transformation of fault warning results into executable maintenance solutions.

[0010] In this embodiment, the specific operational logic of the master-slave collaborative fault early warning and diagnosis in step S3 is as follows: S31. When the network is normal, the edge main module loads the first-level self-learning model and performs real-time inference based on the dynamic threshold issued by the third-level self-learning model. After identifying an anomaly, it triggers the priority upload of the abnormal data to the cloud. After receiving the abnormal data from the module, the cloud completes accurate diagnosis through the second, third and fourth-level self-learning models, outputs the fault type, remaining lifespan and maintenance priority, and generates an adaptive maintenance strategy. S32. When the network is abnormal, the edge main module automatically switches to independent early warning mode, enables locally cached model parameters, dynamic thresholds and basic maintenance strategy templates, maintains basic fault early warning and maintenance suggestion functions, and ensures continuous system operation. S33. After the network is restored, the edge master module and the cloud slave module complete the bidirectional synchronization of data, model parameters and optimized maintenance strategies within a preset time. The cloud integrates the running data of the edge side during the network outage to update the self-learning model, and the edge side loads the cloud-optimized model and thresholds to restore the master-slave collaborative operation mode.

[0011] In this embodiment, in S3, the core function of the edge master module is to realize real-time data processing, local model inference and rapid anomaly triggering, while the core function of the cloud slave module is to realize complex model training, accurate fault diagnosis and intelligent decision generation. Through the functional layering of the edge and the cloud, the real-time nature of fault warning and the accuracy of diagnostic decision are balanced.

[0012] A self-learning and fault early warning system for motor operation status is disclosed. The system adopts a four-layer architecture, consisting of a perception layer, an edge master module, a cloud slave module, and an application layer, from top to bottom. Each layer collaborates according to its function, realizing full-link support from data acquisition to real-time processing, intelligent training, and application deployment. Furthermore, a two-way flow of data, parameters, and decision results is established between layers, providing hardware and software support for self-learning model training and master-slave collaborative fault early warning. Sensing Layer: The sensing layer is responsible for the accurate acquisition and reliable transmission of multi-dimensional motor operation data. It supports multi-protocol data transmission and breakpoint resume transmission, ensuring the integrity and continuity of the original operation data and providing the data input foundation for the entire system. Edge main module: Equipped with a data preprocessing unit and a first-level self-learning model running environment, it undertakes the lightweight data preprocessing of S1, the training and iteration of the first-level self-learning model of S21, and the real-time inference, anomaly triggering and independent early warning functions of S3, realizing the rapid processing and response of local data, and the inference latency meets the real-time requirements of industry. Cloud-based module: Equipped with high-performance training nodes and massive data storage environment, it undertakes the training and optimization of the second, third, and fourth level self-learning models of S22-S24, and the precise diagnosis and intelligent decision generation functions of S3. At the same time, it establishes a communication and synchronization mechanism with the edge main module to achieve bidirectional synchronization of model parameters, running data, and decision results. Application Layer: This layer serves as the terminal for implementing system functions, handling tasks such as early warning information push, fault tracing, maintenance work order generation, maintenance strategy execution and feedback. It integrates an intelligent maintenance decision module, providing a runtime environment for S24's adaptive maintenance strategy generation. Simultaneously, it receives feedback on the execution effect of maintenance strategies and transmits the feedback data back to the cloud module, achieving closed-loop optimization of the entire system.

[0013] In this embodiment, the edge master module and the cloud slave module establish a heartbeat mechanism through a standardized communication protocol to monitor the communication status in real time. This provides technical support for the independent early warning mode switching when the network is abnormal and for rapid synchronization after the network is restored, ensuring the robustness of the system in complex communication environments.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. The self-learning and fault early warning method and system for motor operation status, through a four-level self-learning mechanism—edge weak supervision pre-training—cloud feature inheritance—dynamic threshold optimization—fault evolution and maintenance strategy optimization—breaks through the dual bottlenecks of dependence on labeled data and rigid maintenance decisions. It relies on a master-slave collaborative architecture to balance real-time performance and diagnostic accuracy, combines dynamic threshold optimization to improve early warning reliability, and establishes a feature inheritance mechanism to avoid forgetting. Ultimately, it achieves early and accurate early warning of motor faults, prediction of remaining lifespan, and adaptive maintenance decisions, minimizing downtime losses and maintenance costs.

[0015] Furthermore, the edge-side first-level self-learning automatically mines high-confidence samples from massive amounts of unlabeled data using a weakly supervised learning algorithm, eliminating the need for extensive manual labeling. This effectively addresses the pain points of scarce fault samples and high labeling costs in industrial scenarios, enabling effective training even with a very low initial labeled sample ratio. This significantly lowers the barrier to industrial application of the technology. Through functional layering between the edge and cloud, the edge main module achieves millisecond-level real-time inference and anomaly triggering, meeting the low-latency requirements of fault warnings in industrial scenarios. The cloud slave module relies on high computing power to complete complex model training and accurate diagnosis, ensuring the accuracy of fault type and remaining life prediction. This resolves the contradiction that a single node cannot simultaneously achieve both real-time performance and accuracy. Based on the dynamic warning threshold of motor health and real-time operating conditions, it replaces the traditional fixed threshold, adapting to dynamic operating conditions such as motor aging, load fluctuations, and environmental changes. This significantly reduces the false alarm rate and false negative rate of fault warnings, improves the reliability of warnings, and reduces ineffective maintenance costs and the risk of equipment damage due to false alarms.

[0016] 2. Furthermore, the second-level self-learning, through a weighted fusion loss function, effectively memorizes historical fault features while learning new fault characteristics, avoiding catastrophic forgetting of the model. This ensures that the system can maintain high recognition accuracy for old faults even when facing new fault types during long-term operation, guaranteeing the long-term stability and reliability of the system. The fourth-level self-learning dynamically tracks fault development patterns through a time-series evolution model, and generates adaptive maintenance strategies by combining actual working conditions such as equipment priority and maintenance resource constraints. This transforms simple fault warning results into executable and implementable maintenance plans, while also enabling dynamic updates to maintenance strategies. This avoids over-maintenance or untimely maintenance, truly implementing predictive maintenance for motors and significantly reducing downtime losses and annual maintenance costs caused by motor failures.

[0017] 3. Furthermore, the heartbeat and data synchronization mechanisms established between the edge and the cloud enable the system to seamlessly switch to an independent edge warning mode when the network is abnormal. After the network is restored, data and model parameters are quickly synchronized to ensure continuous system operation. At the same time, the perception layer supports multi-protocol transmission and breakpoint resumption, and bidirectional feedback of data and parameters between layers makes the system adaptable to the complex communication environment and operating conditions in industrial sites, with strong robustness. From data acquisition, model training, early warning diagnosis to maintenance strategy execution, the entire system forms a closed-loop chain of data-model-decision-execution-feedback. The execution effect of the maintenance strategy will be fed back to the self-learning model to achieve continuous iterative upgrades of the model. This allows the fault identification accuracy, early warning reliability, and maintenance strategy adaptability to continuously improve with the system's running time. The method and system of this invention do not depend on a specific type of motor and can be adapted to various types of motors such as industrial asynchronous motors, new energy vehicle drive motors, and marine propulsion motors. At the same time, it can be implemented in multiple application scenarios such as industrial production, transportation, shipbuilding engineering, and new energy equipment, and has broad versatility. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the motor operating status self-learning and fault early warning method of the present invention; Figure 2 This is a schematic diagram of the motor operation status self-learning and fault early warning system of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] This invention provides a method and system for self-learning and fault early warning of motor operating status. Its core lies in addressing the technical shortcomings of existing technologies, such as reliance on labeled data, poor coordination between master and slave modules, fixed early warning thresholds, rigid maintenance decisions leading to false alarms and missed alarms, and high maintenance costs. This is achieved through three core steps: multi-source data acquisition and preprocessing, training of a four-level self-learning model (edge-cloud-threshold-maintenance strategy), and master-slave collaborative fault early warning and intelligent maintenance decision-making. The system adopts a four-layer architecture: perception layer – edge master module – cloud slave module – application layer. Through the collaborative cooperation of real-time inference by the edge master module, accurate diagnosis by the cloud slave module, and dynamic optimization by the intelligent maintenance decision-making module, it achieves early fault early warning, remaining life prediction, and adaptive maintenance decision-making throughout the motor's entire lifecycle. In industrial scenarios with insufficient labeled samples, it improves fault identification accuracy and reduces false alarm and missed alarm rates. It is adaptable to various equipment scenarios, including industrial asynchronous motors, new energy vehicle drive motors, and marine propulsion motors, significantly reducing downtime losses and maintenance costs.

[0021] Example 1: To better understand the above technical solution, the following will provide a detailed description of the technical solution in conjunction with the accompanying drawings and specific implementation methods. (Refer to...) Figure 1 As shown in the figure, this is a schematic diagram of the self-learning and fault early warning method for motor operating status. The self-learning and fault early warning method for motor operating status includes the following steps: S1. Collection and standardization of multi-source operating data: Collect multi-dimensional operating data of the motor, perform noise reduction, feature extraction and standardization on the data, generate a standardized feature set, and provide a high-quality data foundation with a unified dimension for subsequent self-learning model training and fault early warning. Multi-dimensional operational data are the core physical quantities characterizing the motor's operating status, including four types of data: motor vibration, current, temperature, and load. These are the basic data reflecting mechanical faults, electrical faults, thermal faults, and changes in operating conditions. Data is collected from key monitoring points of the motor. Vibration data is collected from vibration-sensitive parts such as the motor bearing end cover and frame. Current data is collected from the stator power supply circuit. Temperature data is collected from easily heated parts such as the stator winding, bearings, and frame. Load data is collected from the motor control circuit. During the collection process, time synchronization of multi-source data is achieved to ensure the matching of data in different dimensions in the time dimension, providing a time-consistent basic data for subsequent feature fusion learning. To address the characteristics of operational data across different dimensions, appropriate signal processing methods are employed for denoising and feature extraction. For vibration data, wavelet transform is used for denoising, effectively separating environmental mechanical noise, electromagnetic interference, and fault impact features, extracting core features in the time and frequency domains to reflect mechanical faults such as motor bearing wear and rotor imbalance. For current data, Fourier transform is used for feature extraction, separating grid harmonic interference and extracting motor-related harmonic components and fundamental wave features to reflect electrical faults such as stator winding short circuits and air gap eccentricity. For temperature data, moving average filtering is used for denoising, eliminating instantaneous interference from environmental temperature fluctuations, and extracting features such as temperature deviation and temperature change rate to reflect thermal faults such as motor overload, overheating, and cooling system failure. For load data, sliding window statistics are used for feature extraction, generating features such as average load value and load fluctuation coefficient to reflect the real-time operating conditions of the motor, providing a basis for subsequent threshold optimization for operating condition adaptation and fault evolution modeling. Only the extracted feature data is standardized using Z-score normalization. This maps feature data of different dimensions and magnitudes to a standard interval with a mean of 0 and a variance of 1, generating a fixed-dimensional standardized feature set. This reduces data transmission volume, adapts to the collaborative transmission needs between the edge and the cloud, and eliminates differences in feature dimensions, ensuring the convergence efficiency of subsequent model training. The amount of extracted feature data is much smaller than the original collected data. Only the standardized feature set is transmitted between the edge and the cloud, significantly reducing network bandwidth usage, adapting to the collaborative transmission needs between the edge and the cloud, and avoiding transmission delays caused by transmitting the entire original data.

[0022] S2. Training and Iteration of Hierarchical Self-Learning Model: Based on a standardized feature set, a four-level self-learning model with hierarchical deployment on the edge and cloud is used for progressive training. The previous level of self-learning provides data, parameters or results support for the next level, and the optimization results of the next level of self-learning are fed back to the previous level to realize model iteration. The model successively completes sample mining, fault feature learning, dynamic optimization of early warning threshold, fault evolution modeling and maintenance strategy generation. The four-level self-learning model is implemented hierarchically according to function, and a two-way feedback of data and parameters is formed between each level. Specifically, after the standardized feature set is generated, it is first pushed to the edge to provide basic data for Level 1 self-learning. After the edge completes sample mining and lightweight model training, only high-value labeled / pseudo-labeled samples are uploaded to the cloud to provide training data for Level 2 self-learning. The motor health quantification indicators output by Level 2 self-learning serve as the core input for Level 3 self-learning. The dynamic early warning thresholds generated by Level 3 self-learning are sent to the edge to become the anomaly judgment criteria for real-time inference of the Level 1 self-learning model. The fault type, remaining lifespan, health status, and dynamic thresholds output by Level 2 and Level 3 self-learning together provide the decision-making basis for Level 4 self-learning. The execution effect of the maintenance strategy generated by Level 4 self-learning is fed back to the Level 2, 3, and 4 self-learning models to achieve iterative optimization of the health assessment system, early warning thresholds, and fault evolution model. At the same time, the cloud sends the optimized basic model parameters to the edge to update the Level 1 self-learning lightweight model, forming a closed loop of model iteration and performance improvement across the entire edge-cloud chain.

[0023] S21, Level 1 Self-Learning: Deployed at the edge, based on an unlabeled standardized feature set, it uses a weakly supervised learning algorithm and a lightweight feature recognition model. Specifically, the weakly supervised learning algorithm is a consistency regularization algorithm. It addresses the pain point of having many unlabeled samples and few labeled samples in industrial scenarios. By applying different data augmentation processes to the same unlabeled standardized feature set, a dual-model inference architecture is constructed. This allows the model to automatically learn the feature patterns of samples without manual annotation, achieving accurate mining of high-confidence samples. Specifically, it automatically mines high-confidence samples and expands the training set. After the model infers the state of the unlabeled samples, it outputs the confidence value of the sample's state (healthy / faulty / specific fault type). Only samples with a confidence value ≥ a preset threshold are selected as high-confidence samples. Simultaneously, a category difference check is added to ensure that the state determination of high-confidence samples is unique and the features are clear, avoiding the inclusion of ambiguous samples. The input of data into the training set leads to model bias; the training of a lightweight model on the edge side is completed; specifically, the model adopts a lightweight network architecture, retaining only the core feature extraction and classification layers, discarding complex deep network structures. The core is adapted to the low computing power and low latency requirements of the edge side, ensuring that the inference speed of the model on the edge side meets the real-time requirements of industry. At the same time, the number of model parameters can be flexibly adapted to the computing power level of different edge hardware, providing model support for real-time edge inference. Meanwhile, high-value samples are uploaded to the cloud to provide a data foundation for cloud model training. On the edge side, based on the standardized feature set collected in real time, incremental training is carried out according to a preset cycle. Each incremental training only targets the newly collected feature data, quickly mining new high-confidence samples and expanding the training set, while updating the lightweight model. This ensures that the model can adapt to small changes in motor operating conditions without the need for full cloud participation, reducing the collaborative cost between the edge and the cloud.

[0024] S22, Level 2 Self-Learning: Deployed in the cloud, based on sample data uploaded from the edge, it uses a fusion feature learning model. This model achieves a balance between learning new fault features and remembering historical fault features through weighted fusion of loss functions, avoiding the model's forgetting of historical fault features and ensuring the model's long-term stability. The model integrates spatial feature extraction and temporal feature modeling capabilities, enabling it to mine spatial correlation features of different dimensions of data in the standardized feature set, such as the coupling relationship between vibration features and current features, and to learn the temporal evolution features of motor operating states, such as the feature change pattern of faults from inception to development, achieving in-depth mining and accurate classification of complex fault features. The execution logic of weighted fusion of loss functions: cross-entropy loss and inheritance loss are weighted and fused according to a preset ratio. The core function of cross-entropy loss is to enable the model to learn new fault features and improve the recognition accuracy of new fault types. The core function of inheritance loss is to constrain the correlation between new fault feature parameters and historical fault feature parameters. By describing the intrinsic relationship between new and old fault features through the feature topology relationship matrix, the model retains the ability to remember historical fault features when learning new features, avoiding catastrophic forgetting, and realizing the memorization of historical fault features while learning new fault features. The output is a quantified indicator of motor health. Specifically, the health quantified indicator is generated through a multi-feature fusion evaluation model. The input is the fault feature parameters, motor running time, cumulative maintenance times, historical fault records, etc., output by the fusion feature learning model. The output is a value in the range of 0-1, where 0 represents complete failure and 1 represents complete health. This indicator is not only the core input of the three-level self-learning, but also serves as the core quantitative basis for the motor's operating status. It is pushed to the application layer to realize the visualization of the motor's health status. This provides core input parameters for subsequent early warning threshold optimization, and simultaneously feeds back the optimized basic model parameters to the edge side to improve the edge model's recognition accuracy. Specifically, the cloud distributes the common fault feature extraction layer parameters from the fusion feature learning model to the edge side, which then updates these parameters into the first-level self-learning lightweight feature recognition model, replacing the model's original feature extraction layer parameters. This allows the edge model to obtain the cloud's deep learning feature extraction capabilities without retraining, significantly improving the edge model's accuracy in recognizing fault features.

[0025] S23, Level 3 Self-Learning: Deployed in the cloud, based on motor health and real-time operating condition changes. Specifically, the real-time operating condition is based on the load factor as the core input, combined with auxiliary operating condition parameters such as ambient temperature and motor operating speed. The load factor is the ratio of the motor's real-time load to its rated load, which reflects the motor's operating load status and ensures that the warning threshold can adapt to the changes in fault characteristics under different operating conditions such as light load / heavy load / variable load. The warning threshold is dynamically adjusted through intelligent optimization algorithms. Specifically, the cloud sends the dynamic warning threshold to the edge side at a real-time / near real-time frequency. The edge side caches the threshold locally as an anomaly judgment standard for real-time inference of the first-level self-learning model. When the motor health drops significantly or the operating conditions change abruptly, the cloud will send the updated threshold first to ensure that the edge side can quickly adapt to the changes in motor status and improve the timeliness and accuracy of the warning. To adapt the threshold to the dynamic operating state of the motor, the generated dynamic threshold is sent to the edge side as a standard for real-time anomaly judgment. At the same time, it receives feedback on the early warning effect from the edge side to achieve continuous optimization of the threshold. Specifically, the edge side uploads early warning records, actual fault verification results, and false alarm / missed alarm cases to the cloud in real time. The cloud uses this data as feedback samples for reinforcement learning algorithms to retrain and optimize algorithm parameters and adjust the threshold generation logic, so that the dynamic threshold can continuously adapt to changes in the operating state of the motor, achieving closed-loop optimization of the threshold.

[0026] S24, Level 4 Self-Learning: Deployed in the cloud, based on fault type, remaining lifespan, and real-time operational data, it tracks fault development patterns through a time-series evolution model. Combined with actual operating condition constraints (including equipment operation priority, maintenance resource configuration, and one or more production operation windows), the generated adaptive maintenance strategy adapts to actual on-site execution needs, transforming fault warning results into executable maintenance plans. Specifically, equipment operation priority is quantified into a weight value of 0-1 based on production importance; the higher the weight, the higher the maintenance priority. Maintenance resource allocation: quantified as the number of maintenance personnel, spare parts inventory, maintenance equipment, etc., to clarify the upper limit of resources; Production operation window: quantified as a specific time interval and duration, to clarify the time range in which maintenance operations can be carried out; The model combines the above quantitative constraints with the fault evolution rate and remaining life prediction value as input conditions for multi-objective optimization. An adaptive maintenance strategy is generated through a multi-objective optimization algorithm. Specifically, the optimization algorithm takes minimizing maintenance costs, minimizing downtime, and maximizing maintenance resource utilization as its core optimization objectives. It generates multiple candidate maintenance strategies by solving for the multi-objective Pareto optimal solution, and sorts the candidate strategies according to equipment operation priority and fault evolution speed, outputting the optimal maintenance strategy. When a single objective cannot be satisfied, the algorithm can make objective trade-offs according to preset weights to ensure the actual executability of the maintenance strategy.

[0027] Simultaneously, it receives feedback on the execution effect of maintenance strategies and updates the fault evolution model and health assessment system in reverse. Specifically, the application layer transmits execution effect data such as execution time, execution cost, fault repair effect, resource consumption, and downtime to the cloud in real time. The cloud uses this data as feedback samples to iteratively optimize the time-series evolution model and multi-objective optimization algorithm. For the time-series evolution model: the reasoning logic for fault evolution speed is corrected to improve the accuracy of remaining life prediction; for the multi-objective optimization algorithm: the quantification logic of the weights and constraints of the optimization objectives is adjusted to make the generated maintenance strategy more in line with the actual needs on site; at the same time, the maintenance effect data will also update the secondary self-learning motor health assessment system in reverse, so that the health calculation is more in line with the actual operating state of the motor.

[0028] S3. Master-Slave Collaborative Fault Warning and Diagnosis: Relying on a master-slave architecture of an edge master module and a cloud slave module, the edge master module uses a self-learning model to achieve real-time inference and anomaly triggering, while the cloud slave module completes accurate diagnosis and decision output. When the network is abnormal, the edge side switches to an independent warning mode. Specifically, the operational logic of the master-slave collaborative fault warning and diagnosis is as follows: When the network is normal, the edge master module loads the first-level self-learning model and performs real-time inference based on the dynamic thresholds issued by the third-level self-learning model. After identifying an anomaly, it triggers the priority upload of abnormal data to the cloud. After receiving the abnormal data, the cloud slave module completes accurate diagnosis through the second, third, and fourth-level self-learning models, outputs the fault type, remaining lifespan, and maintenance priority, and generates an adaptive maintenance strategy. When the network is abnormal, the edge master module automatically switches to an independent warning mode, enables locally cached model parameters, dynamic thresholds, and basic maintenance strategy templates, maintains basic fault warning and maintenance suggestion functions, and ensures continuous system operation. After network recovery, data and model parameters are synchronized between the edge and the cloud, forming a closed-loop collaboration of early warning, diagnosis, and decision-making. The core function of the edge main module is to realize real-time data processing, local model inference, and rapid anomaly triggering, while the core function of the cloud slave module is to realize complex model training, accurate fault diagnosis, and intelligent decision generation. Through the functional layering of the edge and the cloud, the real-time nature of fault early warning and the accuracy of diagnostic decision-making are balanced. Next, after the network is restored, the edge master module and the cloud slave module complete the bidirectional synchronization of data, model parameters and optimized maintenance strategies within a preset time. The cloud integrates the running data of the edge side during the network outage to update the self-learning model, and the edge side loads the cloud-optimized model and thresholds to restore the master-slave collaborative operation mode.

[0029] Example 2: To better understand the above technical solution, the following will provide a detailed description of the technical solution in conjunction with the accompanying drawings and specific implementation methods. (Refer to...) Figure 2As shown in the diagram, this is a schematic diagram of a motor operation status self-learning and fault early warning system. The system adopts a four-layer architecture: from top to bottom, it consists of a perception layer, an edge master module, a cloud slave module, and an application layer. Each layer collaborates according to its function, realizing full-link support from data acquisition to real-time processing to intelligent training to application deployment. Furthermore, a two-way flow of data, parameters, and decision results is established between layers, providing hardware and software support for self-learning model training and master-slave collaborative fault early warning. (1) Perception layer: The perception layer is responsible for the accurate acquisition and reliable transmission of multi-dimensional motor operation data, supports multi-protocol data transmission and breakpoint resume transmission, ensures the integrity and continuity of the original operation data, and provides the data input basis for the entire system; Specifically, the accuracy parameters of the sensing layer sensors, such as ±0.01g for vibration sensors and ±0.1℃ for temperature sensors, determine the reliability of the preprocessed data. Their sampling capabilities, such as 1kHz for vibration and 50Hz for current, match the sampling rate requirements of the preprocessing. The 4G / 5G / Ethernet transmission and breakpoint resume function of the perception layer ensure that the original data is delivered completely to the edge main module, providing a continuous data source for preprocessing, wavelet transform, FFT, and moving average filtering. (2) Edge main module: It is equipped with a data preprocessing unit and a first-level self-learning model running environment. It undertakes the lightweight data preprocessing of S1, the training and iteration of the first-level self-learning model of S21, and the real-time inference, anomaly triggering and independent early warning functions of S3, so as to realize the rapid processing and response of local data and the inference latency meets the real-time requirements of industry. Specifically, the edge main module has 128GB SSD storage and computing resources to support the local lightweight operation of the preprocessing unit, avoiding the latency caused by uploading all data. The processor and memory configuration of the edge main module is adapted to the training and inference of a first-level self-learning lightweight CNN model.

[0030] (3) Cloud Slave Module: Equipped with high computing power training nodes and massive data storage environment, it undertakes the training and optimization of the second, third and fourth level self-learning models of S22-S24, the accurate diagnosis and intelligent decision generation function of S3, and establishes a communication and synchronization mechanism with the edge master module to realize the two-way synchronization of model parameters, running data and decision results; Specifically, the edge master module and the cloud slave module establish a heartbeat mechanism through a standardized communication protocol to monitor the communication status in real time. This provides technical support for switching to an independent early warning mode when the network is abnormal and for rapid synchronization after the network is restored, ensuring the robustness of the system in complex communication environments. (4) Application layer: It is the terminal for the implementation of system functions, and undertakes the functions of early warning information push, fault tracing, maintenance work order generation, maintenance strategy execution and feedback. It integrates intelligent maintenance decision module, provides operating environment for S24's adaptive maintenance strategy generation, and receives feedback on the execution effect of maintenance strategy. It then sends the feedback data back to the cloud module to realize the closed-loop optimization of the entire system. Specifically, the Intel Xeon processor and PyTorch framework of the application-layer intelligent maintenance decision module directly support four-level self-learning fault evolution modeling and multi-objective optimization, enabling rapid generation and dynamic updating of maintenance strategies.

[0031] The core role of the four-layer architecture in the solution: A complete technical support system is built: The four-layer architecture, from the data acquisition and perception layer, the real-time processing edge main module, the intelligent training cloud module, to the application layer, forms a complete technology stack of hardware-software-algorithm-application. It provides solid hardware computing power and software environment support for multi-source data acquisition, four-level self-learning model operation, and master-slave collaborative early warning, and is the basic carrier for the implementation of system functions.

[0032] Ensuring high-quality data flow throughout the entire process: The perception layer ensures the accuracy of raw data through high-precision sensors, and multiple transmission protocols and breakpoint resume function ensure that data is not lost; the hardware configuration of the edge and cloud matches the computing power requirements of data preprocessing, model inference and training, and the time-series database meets the long-term storage of massive amounts of operational data, realizing controllable data quality throughout the entire chain of acquisition, transmission, storage and processing.

[0033] Balancing real-time and intelligence requirements: The lightweight hardware configuration of the edge main module is adapted to the real-time inference of the first-level self-learning model with a latency of ≤100ms, meeting the rapid response requirements of industrial scenarios; the high-performance GPU nodes in the cloud support the training of complex second / third / fourth-level self-learning models, achieving high-precision diagnosis and decision-making, thus taking into account both real-time and intelligence requirements.

[0034] Achieving a balance between system robustness and ease of use: edge-side local caching and independent operation capabilities during network outages, and rapid synchronization after network recovery, ensure stable 24 / 7 system operation; application-layer multi-terminal adaptation and user-friendly feature design, such as multi-mode early warning and QR code inspection, lower the maintenance threshold, and promote the system from technical feasibility to practical ease of use, realizing the implementation of technical and business value.

[0035] Inter-layer collaborative reinforcement system closed loop: Perception layer → Edge main module: Accurate data → Real-time preprocessing and inference, supporting immediate early warning; Edge main module → Cloud slave module: Abnormal data + labeled samples → Model training and optimization, supporting intelligent upgrades; From the module to the application layer in the cloud: diagnostic results + maintenance strategies → multi-terminal output, supporting business implementation; Application layer → Cloud / Edge: Maintain records + feedback data → Model iteration, forming a closed-loop optimization of the entire chain of data-model-decision-feedback.

[0036] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

[0037] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A motor operation state self-learning and fault early warning method, characterized in that, Includes the following steps: S1. Collection and standardization of multi-source operating data: Collect multi-dimensional operating data of the motor, perform noise reduction, feature extraction and standardization on the data, generate a standardized feature set, and provide a high-quality data foundation with a unified dimension for subsequent self-learning model training and fault early warning. S2. Training and Iteration of Hierarchical Self-Learning Model: Based on a standardized feature set, a four-level self-learning model with hierarchical deployment on the edge and cloud is used for progressive training. The previous level of self-learning provides data, parameters or results support for the next level, and the optimization results of the next level of self-learning are fed back to the previous level to realize model iteration. In sequence, sample mining, fault feature learning, dynamic optimization of early warning threshold, fault evolution modeling and maintenance strategy generation are completed. S3. Master-Slave Collaborative Fault Warning and Diagnosis: Based on the master-slave architecture of the edge master module and the cloud slave module, the edge master module realizes real-time reasoning and anomaly triggering based on the self-learning model, and the cloud slave module completes accurate diagnosis and decision output. When the network is abnormal, the edge side switches to an independent warning mode. After the network is restored, the data and model parameters of the edge and the cloud are synchronized, forming a closed-loop collaboration of warning-diagnosis-decision.

2. The method of claim 1, wherein the method further comprises: In S1, appropriate signal processing methods are used for denoising and feature extraction based on the characteristics of data from different dimensions. Only the extracted feature data is standardized to reduce the amount of data transmission, adapt to the collaborative transmission needs of edge and cloud, and eliminate the differences in feature dimensions to ensure the convergence efficiency of subsequent model training.

3. The method of claim 1, wherein the method further comprises: In S2, the four-level self-learning model is implemented hierarchically according to function, and a two-way feedback of data and parameters is formed between each level, specifically: S21, Level 1 Self-Learning: Deployed on the edge side, based on unlabeled standardized feature sets, it automatically mines high-confidence samples and expands the training set through weakly supervised learning algorithms and lightweight feature recognition models, completes the training of the lightweight model on the edge side, provides model support for real-time edge inference, and uploads high-value samples to the cloud to provide a data foundation for cloud model training. S22, Level 2 Self-Learning: Deployed in the cloud, based on sample data uploaded from the edge side, it learns new fault features while memorizing historical fault features through the fusion feature learning model, outputs quantifiable indicators of motor health, provides core input parameters for subsequent early warning threshold optimization, and feeds back the optimized basic model parameters to the edge side to improve the edge model recognition accuracy. S23, Level 3 Self-Learning: Deployed in the cloud, based on the motor health and real-time operating condition changes, it dynamically adjusts the warning threshold through intelligent optimization algorithms to make the threshold adapt to the dynamic operating state of the motor. The generated dynamic threshold is sent to the edge side as the standard for real-time anomaly judgment on the edge side, and at the same time, it receives the warning effect feedback from the edge side to achieve continuous optimization of the threshold. S24, Level 4 Self-Learning: Deployed in the cloud, based on fault type, remaining lifespan and real-time operating data, it tracks the fault development pattern through a time-series evolution model, combines actual operating condition constraints, generates adaptive maintenance strategies through a multi-objective optimization algorithm, and receives feedback on the execution effect of the maintenance strategies to update the fault evolution model and health assessment system in reverse.

4. The method for self-learning and fault early warning of motor operating status according to claim 3, characterized in that: In step S22, the fusion feature learning model achieves a balance between learning new fault features and remembering historical fault features through weighted fusion of loss functions, avoiding the model's forgetting of historical fault features and ensuring the stability of the model's long-term operation.

5. The method and system for self-learning and fault early warning of motor operating status according to claim 4, characterized in that: In S24, the actual operating condition constraints include one or more of the following: equipment operation priority, maintenance resource configuration, and production operation window. The generated adaptive maintenance strategy adapts to the actual execution needs on site, realizing the transformation of fault warning results into executable maintenance solutions.

6. The method and system for self-learning and fault early warning of motor operating status according to claim 5, characterized in that: In S3, the specific operational logic of the master-slave collaborative fault early warning and diagnosis is as follows: S31. When the network is normal, the edge main module loads the first-level self-learning model, performs real-time inference based on the dynamic threshold issued by the third-level self-learning, and triggers the abnormal data to be uploaded to the cloud first after an anomaly is detected. After receiving abnormal data from the module, the cloud performs accurate diagnosis through the second, third and fourth level self-learning models, outputs the fault type, remaining lifespan and maintenance priority, and generates an adaptive maintenance strategy. S32. When the network is abnormal, the edge main module automatically switches to independent early warning mode, enables locally cached model parameters, dynamic thresholds and basic maintenance strategy templates, maintains basic fault early warning and maintenance suggestion functions, and ensures continuous system operation. S33. After the network is restored, the edge master module and the cloud slave module complete the bidirectional synchronization of data, model parameters and optimized maintenance strategies within a preset time. The cloud integrates the running data of the edge side during the network outage to update the self-learning model, and the edge side loads the cloud-optimized model and thresholds to restore the master-slave collaborative operation mode.

7. The method and system for self-learning and fault early warning of motor operating status according to claim 6, characterized in that: In S3, the core function of the edge master module is to realize real-time data processing, local model inference and rapid anomaly triggering, while the core function of the cloud slave module is to realize complex model training, accurate fault diagnosis and intelligent decision generation. Through the functional layering of the edge and the cloud, the real-time nature of fault warning and the accuracy of diagnostic decision are balanced.

8. A motor operating status self-learning and fault early warning system, applied to the motor operating status self-learning and fault early warning method described in any one of claims 1-7, characterized in that, The system adopts a four-layer architecture, from top to bottom: the perception layer, the edge master module, the cloud slave module, and the application layer. Each layer collaborates according to its function, realizing full-link support from data acquisition to real-time processing to intelligent training to application deployment. Furthermore, a two-way flow of data, parameters, and decision results is established between layers, providing hardware and software support for self-learning model training and master-slave collaborative fault early warning. Sensing Layer: The sensing layer is responsible for the accurate acquisition and reliable transmission of multi-dimensional motor operation data. It supports multi-protocol data transmission and breakpoint resume transmission, ensuring the integrity and continuity of the original operation data and providing the data input foundation for the entire system. Edge main module: Equipped with a data preprocessing unit and a first-level self-learning model running environment, it undertakes the lightweight data preprocessing of S1, the training and iteration of the first-level self-learning model of S21, and the real-time inference, anomaly triggering and independent early warning functions of S3, realizing the rapid processing and response of local data, and the inference latency meets the real-time requirements of industry. Cloud-based module: Equipped with high-performance training nodes and massive data storage environment, it undertakes the training and optimization of the second, third, and fourth level self-learning models of S22-S24, and the precise diagnosis and intelligent decision generation functions of S3. At the same time, it establishes a communication and synchronization mechanism with the edge main module to achieve bidirectional synchronization of model parameters, running data, and decision results. Application Layer: This layer serves as the terminal for implementing system functions, handling tasks such as early warning information push, fault tracing, maintenance work order generation, maintenance strategy execution and feedback. It integrates an intelligent maintenance decision module, providing a runtime environment for S24's adaptive maintenance strategy generation. Simultaneously, it receives feedback on the execution effect of maintenance strategies and transmits the feedback data back to the cloud module, achieving closed-loop optimization of the entire system.

9. The motor operating status self-learning and fault early warning system according to claim 8, characterized in that: The edge master module and the cloud slave module establish a heartbeat mechanism through a standardized communication protocol to monitor the communication status in real time. This provides technical support for switching to an independent early warning mode when the network is abnormal and for rapid synchronization after the network is restored, ensuring the robustness of the system in complex communication environments.

Images