A method, system, device, and medium for health management of an electromechanical coupling system

Abstract

The present disclosure relates to the technical field of electromechanical management, and particularly relates to a kind of electromechanical coupling system health management method, system, equipment and medium.The present disclosure adopts health state baseline modeling paradigm, does not rely on the mass historical fault data difficult to obtain, but utilizes sufficient normal operation data to model, fundamentally solves the technical problem that traditional data-driven method is difficult to be effectively deployed in industrial field due to the scarcity of fault sample.Meanwhile, through multilevel exception confirmation strategy, random noise and transient interference can be effectively filtered out, ensure that the alarm finally issued has very high confidence, significantly reduce the false alarm rate, thereby improving the overall reliability of system.The present disclosure upgrades the prediction of remaining useful life from the traditional single determined value to the complete probability distribution containing expectation value, standard deviation and confidence interval.Solve the problem that existing technology cannot quantify the uncertainty of prediction, provide a clear, probability-based risk assessment tool for managers.

Description

Technical Field

[0001] This disclosure relates to the field of electromechanical management technology, and in particular to a method, system, device and medium for health management of electromechanical coupling systems. Background Technology

[0002] In the field of live-line working robots for power distribution networks, the core energy transmission device—an electromechanical coupling system consisting of a motor, an insulated transmission rod, and a generator—is used to achieve critical power transmission and energy conversion. The long-term stable operation of this system directly determines the robot's operational efficiency and safety. However, this complex electromechanical coupling system inevitably faces performance degradation and potential failure risks due to various physical processes such as mechanical wear, material fatigue, and electrical aging during long-term operation. Therefore, advanced fault prediction and health management for this system have significant engineering value.

[0003] Digital twins, as high-fidelity, multi-dimensional, and dynamic virtual models of physical entities, construct a virtual copy that is synchronized with and accurately mapped to the physical entity in real time by deeply integrating multiphysics simulations, real-time sensor data, and historical operational information. In fault prediction and health management applications, their core advantages are reflected in two aspects: First, they allow virtual simulation experiments to be conducted without interfering with the normal operation of physical equipment, enabling the deduction and evaluation of various potential fault evolution paths; second, they can organically integrate mechanistic models with data-driven algorithms, thereby gaining in-depth insights into the internal physical state changes of the system and achieving accurate predictions of future health status.

[0004] However, existing digital twin-based fault diagnosis (PHM) technologies still suffer from several insurmountable limitations when addressing real-world industrial problems: First, the high accuracy of mainstream fault diagnosis models relies on large-scale historical datasets covering various fault modes. In reality, however, faults in sophisticated systems are low-probability events, and complete fault lifecycle data is extremely scarce. This constitutes a fundamental bottleneck hindering the effective deployment of traditional PHM methods. Second, the "black box" nature of the model's decision-making process contradicts the "credibility" requirement of high-risk decision-making. Even if existing models can identify faults, they cannot provide physically explainable causal paths, making it difficult for decision-makers to fully trust them. Third, existing models mostly predict remaining useful life (RUL) as single, deterministic values, ignoring the inherent uncertainties introduced by factors such as the model and data, thus failing to provide quantitative basis for risk management and maintenance planning. Finally, these models are typically static; once offline training is complete, their parameters are fixed, unable to adapt to the aging and parameter drift of physical systems due to long-term service, leading to a gradual decrease in the accuracy of their long-term predictions over time.

[0005] In summary, to solve the aforementioned technical challenges and fully leverage the potential of digital twins in the field of PHM, it is necessary to design a novel method for fault prediction and health management of electromechanical coupling systems that can break free from dependence on massive amounts of fault data, provide interpretable diagnostics, quantify and predict risks, and be self-optimizing, thereby improving the operational stability and safety of the energy transmission device of the live-line working robot in the distribution network. Summary of the Invention

[0006] To address the aforementioned issues, this disclosure provides a method, system, device, and medium for health management of electromechanical coupling systems. This disclosure integrates technologies such as baseline modeling based on health status, causal inference, uncertainty quantification, and closed-loop online learning to systematically solve problems such as sparse fault samples, insufficient model interpretability, unquantifiable predicted risks, and limited decision support capabilities of traditional data-driven methods.

[0007] A first aspect includes a health management method for an electromechanical coupling system, the method comprising:

[0008] Collect multidimensional time series data under healthy operating conditions, construct a multidimensional feature vector dataset under healthy conditions, and construct and train a health state baseline model to reflect the operating state of electromechanical coupling system based on the multidimensional feature vector dataset.

[0009] During the online operation of the electromechanical coupling system, the real-time sensor feature vector at the current moment is obtained and input into the health state baseline model to obtain the predicted feature vector. By calculating the difference between the real-time sensor feature vector and the predicted feature vector, a real-time residual sequence reflecting the deviation of the system state is generated.

[0010] Anomaly alarms are identified based on real-time residual sequences, and fault modes are obtained through analysis.

[0011] Based on failure modes, the remaining service life of the electromechanical coupling system under the current state is probabilistically predicted, and corresponding operation and maintenance guidance strategies are generated.

[0012] Furthermore, multidimensional time-series data under healthy working conditions are collected to construct a multidimensional feature vector dataset under healthy conditions, including:

[0013] Sensor arrays are deployed on key components of the electromechanical coupling system to collect multidimensional time series data under healthy operating conditions. By preprocessing and extracting features from the collected data, a multidimensional feature vector dataset under healthy conditions is constructed. The healthy operating conditions include normal operating conditions under different combinations of speed, load and ambient temperature.

[0014] Furthermore, the sensor array includes: an accelerometer for measuring mechanical vibration, a temperature sensor for monitoring the temperature of critical components, voltage and current transformers for measuring electrical parameters, and an encoder for acquiring rotor dynamic parameters.

[0015] Multidimensional time series data, including vibration acceleration, temperature, voltage, current, rotational speed, and torque measured by sensors.

[0016] Furthermore, the collected data undergoes preprocessing, including:

[0017] The collected time series data is cleaned, and outliers and missing values ​​are removed.

[0018] Digital filters are used to reduce noise in the signal, highlighting the core information related to the system state;

[0019] Normalization is performed to eliminate the influence of different physical dimensions on model training.

[0020] Further feature extraction includes:

[0021] Extracting time-domain statistical features and frequency-domain features from vibration acceleration;

[0022] Extracting power quality characteristics from voltage and current;

[0023] By combining time-domain statistical features, frequency-domain features, and power quality features with temperature and rotational speed, a multi-dimensional feature vector is formed.

[0024] Furthermore, the health status baseline model is a physical information autoencoder, and its loss function is... Defined as data reconstruction loss With a physical law residual loss that includes multiple key failure physical processes Weighted sum:

[0025] ;

[0026] in, w recon Indicates the weight of the data reconstruction loss. w phys The weight represents the residual loss due to physical laws.

[0027] Furthermore, physical law residual loss This is the weighted sum of the mean square errors of the residuals generated by each physical model at all sampling points within the computational domain;

[0028] The physical models include: a nonlinear bearing dynamics model, a Paris crack propagation model, and a multi-factor electrical insulation aging model.

[0029] Furthermore, the real-time residual sequence is quantized by calculating the Mahalanobis distance between the real-time sensor feature vector and the predicted feature vector at time t.

[0030] Furthermore, anomaly alerts are identified based on real-time residual sequences, including:

[0031] A multi-level anomaly confirmation strategy is implemented for the real-time residual sequence to filter random disturbances and identify deterministic anomalies. A deviation is confirmed as a high-confidence anomaly and an anomaly alarm is triggered only if it passes the test at all levels.

[0032] Furthermore, a multi-level anomaly confirmation strategy includes:

[0033] The first layer is a dynamic time series test, used to confirm the persistence and trend of anomalies;

[0034] The second layer is the operating condition adaptive threshold evaluation, used to ensure the baseline threshold. The dynamic accuracy; where the threshold It is a function based on real-time operating condition vectors;

[0035] The third layer involves cross-verification of information from multiple sensors to confirm the physical correlation of anomalies.

[0036] Further analysis revealed the following failure modes:

[0037] When an abnormal alarm is triggered, causal tracing is performed based on the cause-effect graph, and reverse fact analysis is performed using the health status baseline model to identify and output the root cause event that caused the current abnormal state and its physical transmission path.

[0038] Among them, the cause-effect graph is a directed acyclic cause-effect graph, in which the nodes represent the key physical variables of the system and the directed edges represent the direct causal relationships between the variables.

[0039] Reverse fact analysis is based on causal graphs. When an anomaly is detected in a downstream node variable in the graph, its upstream parent node variable is selected as a potential cause. The upstream parent node variable is forced to be fixed to a normal value and the simulation is re-enhanced. The causal contribution of the upstream parent node variable is quantified by observing whether the downstream node variable recovers from the anomaly.

[0040] Furthermore, a probabilistic prediction is made of the remaining service life of the electromechanical coupling system in its current state, including:

[0041] Using a Bayesian neural network, the remaining useful life (RUL) of an electromechanical coupling system in its current state is probabilistically predicted. The prediction result is a probability density function characterizing the uncertainty of the RUL. .

[0042] Furthermore, corresponding operation and maintenance guidance strategies are generated, including:

[0043] Based on probabilistic predictions of remaining service life, corresponding operation and maintenance guidance strategies are generated, including dynamically adjusting the system's monitoring strategies and issuing a series of graded maintenance warnings with different levels of urgency to guide operation and maintenance activities.

[0044] The dynamic monitoring and tiered maintenance early warning system specifically includes:

[0045] The expected value of the remaining useful life (RUL) prediction is compared with a set of preset time thresholds. When the expected value of the remaining useful life (RUL) prediction is lower than the preset time thresholds, the sampling frequency of the sensor or the sensitivity of the anomaly detection algorithm is automatically increased.

[0046] Based on the RUL prediction results, maintenance instructions with different levels of urgency are generated, including planned, warning, and critical alerts.

[0047] Furthermore, it also includes:

[0048] Operation and maintenance are carried out according to the operation and maintenance guidance strategy. After the operation and maintenance activities are completed, the complete process data of this fault diagnosis and prediction, as well as the maintenance results confirmed by humans, are used as new knowledge samples and fed back to the health status baseline model and the process of probabilistic prediction of remaining service life, forming a closed-loop model for online optimization.

[0049] Furthermore, online optimization of the closed-loop model includes:

[0050] After the maintenance activity is completed, the sensor characteristic data from the entire process from the occurrence of the anomaly to the execution of the maintenance will be packaged into a data packet, and the root cause of the fault confirmed by humans will be used as the true label of the data packet.

[0051] Data packets with real labels are used as training samples to supplement the original training dataset, and the parameters in the health status baseline model and the probabilistic prediction process of remaining lifespan are updated by means of online learning or periodic fine-tuning.

[0052] Secondly, a health management system for an electromechanical coupling system includes:

[0053] State modeling unit, state prediction unit, fault identification unit, and operation and maintenance guidance unit;

[0054] The state modeling unit is used to collect multi-dimensional time series data under healthy operating conditions, construct a multi-dimensional feature vector dataset under healthy conditions, and construct and train a health state baseline model to reflect the operating state of the electromechanical coupling system based on the multi-dimensional feature vector dataset.

[0055] The state prediction unit is used to acquire the real-time sensor feature vector at the current moment during the online operation of the electromechanical coupling system, and input it into the health state baseline model to obtain the predicted feature vector. By calculating the difference between the real-time sensor feature vector and the predicted feature vector, a real-time residual sequence reflecting the deviation of the system state is generated.

[0056] The fault identification unit, based on the real-time residual sequence, is used to identify abnormal alarms and analyze them to obtain fault modes.

[0057] The operation and maintenance guidance unit, based on failure modes, is used to probabilistically predict the remaining service life of the electromechanical coupling system in its current state and generate corresponding operation and maintenance guidance strategies.

[0058] Thirdly, an electronic device includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;

[0059] Memory, which stores computer programs;

[0060] When a processor executes a computer program stored in memory, it implements the aforementioned electromechanical coupling system health management method.

[0061] Fourthly, a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned electromechanical coupling system health management method.

[0062] This disclosure includes at least the following beneficial effects:

[0063] This disclosure adopts a health status baseline modeling paradigm, which does not rely on the massive historical fault data that is difficult to obtain. Instead, it utilizes sufficient normal operation data for modeling, fundamentally solving the technical problem that traditional data-driven methods are difficult to effectively deploy in industrial fields due to the scarcity of fault samples. At the same time, through a multi-level anomaly confirmation strategy, random noise and transient interference can be effectively filtered out, ensuring that the final alarm has a very high confidence level, significantly reducing the false alarm rate, and thus improving the overall reliability of the system.

[0064] To address the lack of transparency in the decision-making process of existing AI models due to their "black box" nature, this disclosure innovatively introduces a causal inference module. When an anomaly is detected, the system not only issues an alarm but also, through virtual intervention and causal verification, clearly traces and outputs the root cause of the fault and its physical transmission path. This interpretable diagnostic conclusion provides engineers with a clear fault logic chain, thereby greatly enhancing their trust in the system's decisions and supporting more accurate and efficient fault diagnosis.

[0065] This disclosure upgrades the prediction of remaining useful life (RUL) from a traditional single deterministic value to a complete probability distribution that includes the expected value, standard deviation, and confidence interval. This addresses the problem that existing technologies cannot quantify predictive uncertainty, providing managers with a clear, probability-based risk assessment tool. Decision-makers can use this probability distribution to conduct quantitative risk analysis, scientifically plan maintenance activities, and optimize spare parts inventory, thereby maximizing the total lifecycle value of equipment while ensuring safety.

[0066] This disclosure utilizes an adaptive monitoring and tiered early warning mechanism to enable the digital twin system to dynamically adjust its monitoring intensity and alarm strategies based on the predicted risk level, optimizing system resources while ensuring security. More importantly, through a closed-loop online model optimization mechanism, the system can iteratively update its model using newly acquired real data and labels after each maintenance activity. This endows the digital twin system with continuous learning and performance enhancement capabilities, ensuring its ability to adapt to the aging of the physical system and parameter drift, and guaranteeing the long-term effectiveness of prediction and diagnostic accuracy.

[0067] Other features and advantages of this disclosure will be set forth in the following description and will be apparent in part from the description or may be learned by practicing the disclosure. The objects and other advantages of this disclosure may be realized and obtained by means of the structures pointed out in the description and the accompanying drawings. Attached Figure Description

[0068] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0069] Figure 1 This is a schematic diagram of the management method flow according to an embodiment of the present disclosure;

[0070] Figure 2 This is a schematic diagram illustrating the principle of the management method according to an embodiment of this disclosure;

[0071] Figure 3 This is a schematic diagram of a multi-level anomaly confirmation strategy according to an embodiment of this disclosure;

[0072] Figure 4 This is a schematic diagram illustrating causal inference in an embodiment of the present disclosure;

[0073] Figure 5 This is a schematic diagram of the management system architecture according to an embodiment of the present disclosure;

[0074] Figure 6 This is a schematic diagram of the electronic device structure according to an embodiment of the present disclosure. Detailed Implementation

[0075] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0076] like Figure 1 As shown, a health management method for an electromechanical coupling system includes:

[0077] S101, Collect multi-dimensional time series data under healthy operating conditions, construct a multi-dimensional feature vector dataset under healthy conditions, and construct and train a health state baseline model to reflect the operating state of the electromechanical coupling system based on the multi-dimensional feature vector dataset.

[0078] S102: During the online operation of the electromechanical coupling system, the real-time sensor feature vector at the current moment is obtained and input into the health state baseline model to obtain the predicted feature vector. By calculating the difference between the real-time sensor feature vector and the predicted feature vector, a real-time residual sequence reflecting the deviation of the system state is generated.

[0079] S103, based on real-time residual sequence, identifies abnormal alarms and analyzes them to obtain fault modes;

[0080] S104, based on failure modes, makes a probabilistic prediction of the remaining service life of the electromechanical coupling system in the current state and generates corresponding operation and maintenance guidance strategies.

[0081] In practical implementation, the process and principles should be considered, such as... Figure 2 As shown, the introduction is as follows:

[0082] The electromechanical coupling system includes an electric motor, an insulated transmission rod, and a generator;

[0083] S101, a sensor array is deployed on the key components of the electromechanical coupling system to collect multi-dimensional time series data under typical healthy operating conditions of the system (e.g., operating conditions under different combinations of speed, load and ambient temperature). Then, through preprocessing and feature extraction of the collected data, a multi-dimensional feature vector dataset under healthy conditions is constructed for subsequent modeling.

[0084] By utilizing a feature vector dataset under healthy conditions and integrating it with a physical mechanism model used to describe the key failure processes of the system, a high-fidelity baseline model of the healthy state is constructed and trained offline as the core of the digital twin.

[0085] Specifically, the sensor array includes: an acceleration sensor for measuring mechanical vibration, a temperature sensor for monitoring the temperature of key components, voltage and current transformers for measuring electrical parameters, and an encoder for acquiring rotor dynamic parameters; the multidimensional time series data includes vibration acceleration measured by these sensors. ,temperature ,Voltage Current Rotation speed and torque .

[0086] Data preprocessing includes cleaning the collected time series data, removing outliers and missing values; using digital filters to reduce noise in the signal and highlight the core information related to the system state; and performing normalization to eliminate the influence of different physical dimensions on subsequent model training.

[0087] Feature extraction specifically includes: from vibration acceleration Extract time-domain statistical features and frequency-domain features from voltage. Current Extract power quality characteristics; and correlate these characteristics with temperature. Rotation speed The multidimensional feature vector is formed by combining direct physical quantities.

[0088] Specifically, the health status baseline model is a physical information autoencoder, and its loss function is... Defined as data reconstruction loss With a physical law residual loss that includes multiple key failure physical processes Weighted sum:

[0089]

[0090] Physical law residual loss It is calculated by substituting the output of the model into the following physical control equations; the physical control equations include:

[0091] Nonlinear bearing dynamics model: used to constrain the vibration behavior of a system, where the key bearing force is... Modeled as:

[0092]

[0093] in, This is the bearing reaction force; It is a time-varying stiffness coefficient, the value of which varies with the degree of bearing wear, and is a key identifiable parameter reflecting the health status of the bearing; This represents the relative displacement between the inner and outer rings of the bearing. denoted as ρ, where ρ is the relative displacement velocity; n is the Hertzian contact nonlinearity exponent, typically taken as 1.5; and C is the viscous damping coefficient.

[0094] Paris crack propagation model: used to constrain the fatigue damage accumulation process of the transmission rod. Its physical constraints follow the Paris crack propagation formula, in the form of:

[0095]

[0096] Where a is the instantaneous length of the crack; N is the number of stress cycles; This represents the crack propagation rate. The stress intensity factor range is calculated by the finite element module integrated within the digital twin model based on real-time dynamic loads; C and m are material constants characterizing the fatigue resistance of the transmission rod material.

[0097] Multi-factor electrical insulation aging model: used to constrain the insulation degradation process of motors, generators, or transmission rods. It models the characteristic life L of the insulation material as a function of the combined effects of thermal and electrical stresses, following the Arrhenius-InversePower hybrid model, and its form is as follows:

[0098]

[0099] Where L is the predicted lifetime of the insulating material under specific operating conditions; T is the thermodynamic temperature of the insulating material; V is the electric field strength or equivalent voltage applied to the insulating material; A is the lifetime constant; B is a constant related to the activation energy of the material; and n is the voltage withstand index, which is a dimensionless constant.

[0100] The physical law residual loss It is the weighted sum of the mean square errors of the residuals (i.e., the differences between the left and right sides of the equation) generated by the above physical models at all sampling points within the computational domain.

[0101] S102, during system online operation, acquire the feature vector at the current moment in real time. And input it into the health status baseline model to obtain the predicted feature vector. By calculating the difference between the true vector and the predicted vector, a real-time residual sequence that reflects the deviation of the system state is generated. .

[0102] Specifically, the residual sequence By calculating the real-time sensor feature vector at time t The feature vector predicted by the model The Mahalanobis distance between them is used for quantification.

[0103] S103, for real-time residual sequences Implement a multi-level verification strategy to filter random disturbances and identify deterministic anomalies. A deviation is only identified as a high-confidence anomaly and a system alarm is triggered if and only if it passes the verification at all levels.

[0104] When the system triggers a high-confidence anomaly alarm, the causal inference module built into the digital twin is activated. This module initiates a causal tracing process, uses the health status baseline model to perform reverse fact analysis, and identifies and outputs the root cause event that led to the current abnormal state and its physical transmission path.

[0105] Specifically, such as Figure 3 As shown, the multi-level anomaly confirmation strategy aims to filter out instantaneous residual fluctuations caused by random noise or non-persistent disturbances layer by layer to ensure the reliability of alarms.

[0106] The first layer of the strategy is time series dynamic testing, which is used to confirm the persistence and trend of anomalies.

[0107] The second layer of the strategy is condition-adaptive threshold evaluation, used to ensure the baseline threshold. Dynamic accuracy; the threshold It is a vector based on real-time operating conditions function .

[0108] The third layer of the strategy involves cross-verification of information from multiple sensors to confirm the physical correlation of anomalies; a high-confidence alarm is only ultimately triggered when multiple physically correlated residual flows are simultaneously abnormal.

[0109] Specifically, such as Figure 4 As shown, the causal inference module has a built-in directed acyclic causal graph, where nodes represent key physical variables of the system and directed edges represent direct causal relationships between variables. The reverse fact analysis is based on this causal graph. When an anomaly is detected in a "downstream" node variable in the graph, the module automatically selects its "upstream" parent node variable as a potential cause. It then forces the upstream variable to be fixed to a normal value in the digital twin model and resimulates it. The causal contribution of the upstream variable is quantified by observing whether the downstream abnormal variable returns to normal.

[0110] S104, after identifying the failure mode, the digital twin model is used in conjunction with uncertainty quantification technology to perform a probabilistic prediction of the system's remaining useful life (RUL) in the current state. The prediction result is a probability density function that can characterize the uncertainty of the RUL. ;

[0111] Based on the RUL probability distribution, corresponding operation and maintenance guidance strategies are generated, including dynamically adjusting the system's monitoring strategy and issuing a series of graded maintenance warnings with different levels of urgency to guide specific operation and maintenance activities.

[0112] After the maintenance activity is completed, the complete process data of this fault diagnosis and prediction, as well as the maintenance results confirmed by humans, are used as new knowledge samples and fed back to the health status baseline model and RUL prediction model. Based on this, online updates or periodic retraining are implemented, thereby forming a closed-loop, continuously iterative self-optimization mechanism.

[0113] Specifically, the uncertainty quantification technique is a Bayesian neural network.

[0114] Dynamic monitoring and tiered maintenance early warning specifically include:

[0115] Adaptive monitoring: The expected value of the RUL prediction is compared with a set of preset time thresholds. When the expected value of the remaining service life RUL prediction is lower than the preset time thresholds, the system automatically increases the sampling frequency of the sensor or the sensitivity of the anomaly detection algorithm.

[0116] Tiered early warning: Based on the RUL prediction results, maintenance instructions with different levels of urgency are generated, including planned, warning and critical warnings.

[0117] Online optimization of the closed-loop model specifically includes:

[0118] Data and Tag Acquisition: After the maintenance activity is completed, the sensor feature data from the entire process from the occurrence of the anomaly to the execution of the maintenance will be packaged, and the root cause of the fault confirmed by the human will be used as the true tag of the data packet;

[0119] Incremental learning and model iteration: Data packets with real labels are added to the original training dataset as high-quality training samples, and the parameters of the health status baseline model and RUL prediction model are updated by means of online learning or periodic fine-tuning.

[0120] like Figure 5 As shown, a health management system for an electromechanical coupling system includes:

[0121] State modeling unit 501, state prediction unit 502, fault identification unit 503, and operation and maintenance guidance unit 504;

[0122] State modeling unit 501 is used to collect multi-dimensional time series data under healthy operating conditions, construct a multi-dimensional feature vector dataset under healthy conditions, and construct and train a health state baseline model to reflect the operating state of electromechanical coupling system based on the multi-dimensional feature vector dataset.

[0123] The state prediction unit 502 is used to acquire the real-time sensor feature vector at the current moment during the online operation of the electromechanical coupling system, and input it into the health state baseline model to obtain the predicted feature vector. By calculating the difference between the real-time sensor feature vector and the predicted feature vector, a real-time residual sequence reflecting the deviation of the system state is generated.

[0124] The fault identification unit 503, based on the real-time residual sequence, is used to identify abnormal alarms and analyze the fault modes.

[0125] The operation and maintenance guidance unit 504, based on the failure mode, is used to probabilistically predict the remaining service life of the electromechanical coupling system in the current state and generate corresponding operation and maintenance guidance strategies.

[0126] like Figure 6 As shown, this disclosure provides an electronic device, including a processor 601, a communication interface 602, a memory 603, and a communication bus 604, wherein the processor 601, the communication interface 602, and the memory 603 communicate with each other through the communication bus 604;

[0127] Memory 603 stores computer programs;

[0128] The processor 601 implements the above method when executing a computer program stored in the memory 603.

[0129] This disclosure provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0130] The computer-readable storage medium may be included in the device / apparatus described in the above embodiments; or it may exist independently and not assembled into the device / apparatus. The computer-readable storage medium carries one or more programs that, when executed, implement the method according to the embodiments of this disclosure.

[0131] According to embodiments of this disclosure, the computer-readable storage medium can be a non-volatile computer-readable storage medium, such as including, but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this disclosure, the computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0132] To enable those skilled in the art to better understand this disclosure, the principles of this disclosure are explained below in conjunction with the accompanying drawings:

[0133] For an electromechanical coupling system consisting of a motor, an insulated transmission rod, and a generator connected in series, a self-optimizing intelligent digital twin can be constructed to achieve accurate diagnosis of the system's health status, probabilistic prediction of its future lifespan, and provide dynamic operation and maintenance guidance.

[0134] The core of this disclosure lies in the fact that, without relying on massive amounts of historical fault data, it constructs a high-fidelity baseline model of the system's health status and combines it with multi-level anomaly confirmation strategies, causal inference, probabilistic prediction, and closed-loop optimization techniques to achieve reliable capture of early minor faults, interpretable tracing of the root causes of faults, quantitative risk assessment of remaining lifetime, and continuous iterative optimization of the model itself based on real-time monitoring data.

[0135] The electromechanical coupling system addressed in this embodiment is specifically constructed as follows: an AC asynchronous motor driven by a frequency converter serves as the power source, and its power is transmitted through a transmission rod with specific insulation properties to drive a synchronous generator to generate electricity.

[0136] like Figure 2 As shown, a baseline model of the health status is constructed during the offline phase. The system collects a large amount of sensor data from the electromechanical coupling system under all design-allowed healthy operating conditions (e.g., various combinations of different speeds and loads). The deployed sensor network includes: accelerometers, which collect vibration signals at high sampling frequencies covering the characteristic frequencies of all key components of the system (such as bearings, gears, etc.); temperature sensors, which monitor the thermal state of key components; voltage and current transformers, which assess power quality and electrical load; and encoders and torque sensors, which monitor real-time speed and torque data.

[0137] The raw data is cleaned, filtered, and normalized. Then, deep features are extracted from the preprocessed data, for example, from vibration signals. The calculation includes a series of time-domain and frequency-domain characteristics, including root mean square (RMS) and kurtosis; from electrical signals , The calculation of power quality characteristics such as total harmonic distortion (THD) is performed, and these characteristics are correlated with temperature. Rotation speed The direct physical quantities are combined to form a multidimensional health status feature vector dataset.

[0138] Based on this dataset and by deeply integrating physical mechanism models used to describe key failure processes of the system, a physical information autoencoder is constructed as the core of the digital twin. The loss function of this model... Designed as data reconstruction loss Residual loss in accordance with physical laws Weighted sum:

[0139]

[0140] in, To ensure the model's predictions adhere to recognized physical laws in the field, such as nonlinear bearing dynamics models, Paris crack propagation models, and multi-factor electrical insulation aging models, the model is trained by minimizing this composite loss function, ultimately yielding a baseline health model capable of accurately reproducing any health state of the system.

[0141] During online operation, the system performs real-time monitoring and diagnostics. The system collects sensor data in real time and constructs the current real-time feature vector. This data is then input into the health status baseline model to obtain the model's prediction of the current health status (i.e., health-oriented reconstruction) feature vector. The Mahalanobis distance between the two was calculated as a real-time residual sequence. .

[0142] The residual sequence This then triggers a rigorous, multi-level anomaly confirmation process, such as... Figure 2 As shown. The first layer uses time series analysis to examine the persistence and growth trend of the residuals, filtering out transient interference. The second layer compares the residuals that pass the test with an adaptive threshold dynamically generated based on the current real-time operating conditions (speed, load). The system compares and excludes fluctuations caused by changes in normal operating conditions. In the third layer, the system queries a preset physical correlation matrix; only when the residuals of multiple physically related sensors are simultaneously confirmed as abnormal will a high-confidence anomaly alarm be triggered.

[0143] Once an alarm is triggered, the causal inference module starts immediately. For example... Figure 3 As shown, this module, based on a preset system cause-effect graph, quickly and accurately locates and outputs the root cause of the fault and its transmission path by performing reverse fact analysis in a digital twin model.

[0144] Subsequently, the system shifts to prediction and operational guidance. After identifying the failure modes, the digital twin model uses uncertainty quantification techniques such as Bayesian neural networks to probabilistically predict the system's remaining useful life (RUL), outputting a probability density function. Based on this RUL distribution, the system performs adaptive monitoring and tiered maintenance early warning, and issues "planned," "warning," or "critical" level maintenance early warnings to maintenance personnel according to the distance of the RUL.

[0145] Finally, the system achieves closed-loop self-optimization. After the maintenance activity is completed, the complete data from the occurrence of the fault to its repair, along with the manually confirmed root cause of the fault, is used as a real label and fed back to the system as a high-quality knowledge sample. Using these new samples, the system iteratively updates the health status baseline model and the RUL prediction model through online learning or periodic fine-tuning, thereby continuously improving its prediction accuracy and generalization ability over time.

[0146] The advantages of this disclosure will be explained below in specific operational scenarios:

[0147] Scenario 1: Diagnosis and Life Prediction of Early Fatigue Cracks in Insulated Transmission Rods

[0148] In this scenario, it is assumed that after long-term operation, microcracks will develop inside the transmission rod due to alternating stress.

[0149] In the initial stage, the cracks are extremely small, and the system uses real-time generated residual sequences. An anomaly was detected; the residual components associated with high-frequency vibrations began to show a continuous and slow increase in deviation.

[0150] The signal enters the first level of the multi-level anomaly confirmation process, which is a dynamic time series test. Through time series analysis, the system confirms that the residual sequence is not only consistently higher than the benchmark, but also shows a statistically significant growth trend.

[0151] The signal enters the second-level adaptive threshold evaluation, and the system confirms that the residual level has significantly exceeded the exclusive normal threshold dynamically calculated based on the current high load and stable speed conditions.

[0152] In the third-layer multi-sensor information cross-verification, if the system finds that other residual flows such as temperature and electrical parameters remain normal, according to the built-in physical correlation matrix, this characteristic matches the pattern of a "purely mechanical structural" fault. Therefore, the system ultimately triggers an anomaly alarm with a medium confidence level.

[0153] In response to this alarm, the causal inference module is activated to trace the root cause. Based on the built-in causal graph, the module identifies "external load fluctuations" and "internal material degradation" as the main upstream causes of the vibration anomaly. To distinguish between the two, the module prioritizes examining external factors and performs virtual intervention in the digital twin model, that is, forcibly fixing the input torque variable to a stable value without fluctuations.

[0154] After resimulating under this intervention, the module found that the virtual vibration residual did not decrease significantly. Based on this, the system ruled out external factors and finally output a diagnostic conclusion: the root cause has a high confidence level as "degradation of the internal material properties of the transmission rod," manifesting as early fatigue characteristics.

[0155] After obtaining a definitive diagnostic conclusion, the remaining useful life (RUL) probabilistic prediction module is activated. Based on the diagnostic conclusion, this module primarily utilizes the Paris crack propagation model built into the digital twin model. Combining historical load spectrum data and quantifying uncertainty through a Bayesian neural network, the system ultimately outputs a complete RUL probability density function. The function takes the form of a probability distribution, and its expected value is... It indicates the most likely remaining lifetime, while the width of the distribution quantifies the uncertainty of the prediction.

[0156] When the RUL probability density function After generation, the system immediately executes tiered early warning and adaptive monitoring logic. The system first calculates its key statistics based on the probability distribution, including the expected value. with standard deviation Therefore, its 95% confidence interval is determined as [ 1.96 , +1.96 ].

[0157] The system compares the lower limit of the confidence interval with the preset "planning level" threshold. Compare. When the lower limit of the confidence interval ( 1.96 (lower than) If this occurs, a Level 3 warning will be triggered. The system will then generate and output a warning message, the content of which includes:

[0158] Warning level: Level 3 (planned level);

[0159] Diagnostic conclusion: Early degradation related to material fatigue;

[0160] Predicted RUL expected value: ;

[0161] Predict the 95% confidence interval for RUL: [ 1.96 , +1.96 ];

[0162] Risk assessment: Low;

[0163] Maintenance recommendation: Include the "drive rod" in the observation or replacement list for the next planned overhaul, and it is recommended to plan spare parts inventory and procurement process.

[0164] At the same time, the triggering of the Level 3 warning will automatically activate the adaptive monitoring function. The system will increase the monitoring frequency and data analysis accuracy of sensors related to the fault mode (such as vibration and torque sensors) to achieve more refined tracking of the degradation process.

[0165] After the maintenance personnel replaced the transmission rod and confirmed through non-destructive testing that there were fatigue cracks inside that matched the model's predicted trend, this confirmation result, along with previous operating data, was used to optimize the RUL model, making its fatigue life prediction for the transmission rod more accurate.

[0166] Scenario 2: Accelerated aging of generator bearings due to poor lubrication

[0167] In this scenario, a slight blockage in the lubrication system leads to a decrease in the lubrication effect of the generator bearings.

[0168] In the initial stage, the physical characteristics of the fault were first manifested as a continuous and slow increase in the residual error of the temperature sensor in the generator bearing housing. Subsequently, the residual error of high-frequency vibration energy, which is related to the impact characteristics of the bearing rolling elements, also began to show a similar abnormal increase.

[0169] Both residual sequences were sent to the first level of the multi-level anomaly confirmation process for dynamic time series analysis. Through time series analysis, the system confirmed that both sequences not only remained consistently above the benchmark but also exhibited a statistically significant upward trend.

[0170] Subsequently, the signal enters the second-level adaptive threshold evaluation. The system confirms that the temperature residual level has significantly exceeded the normal thermal state threshold dynamically calculated based on the current operating conditions, and the vibration residual has also exceeded its corresponding dynamic threshold.

[0171] In the third layer of multi-sensor information cross-verification, the system queries the built-in physical correlation matrix and confirms that "temperature rise" and "increased high-frequency vibration" are typical concurrent characteristics of bearing failure. Due to the existence of strong circumstantial evidence, the system ultimately triggers a high-confidence anomaly alarm.

[0172] In response to this alarm, the causal inference module is activated to trace the root cause. Based on the built-in causal graph, the module identifies "lubrication status" and "radial load" as the common upstream causes of the abnormal temperature and vibration.

[0173] The module prioritizes the "lubrication status" (reflected by temperature residual) that first exhibits an anomaly based on the chronological order of the anomalies, setting it as the highest priority candidate cause for investigation.

[0174] The module performs a virtual intervention on the "lubrication state" (represented as the heat dissipation coefficient in the model) within the digital twin. After resimulating under this intervention, the module found that the virtual temperature and vibration residuals both returned to normal baseline levels. Based on this, the system outputs a diagnostic conclusion with high confidence: the root cause is "poor lubrication leading to bearing overheating, which in turn causes accelerated wear on the rolling element surface."

[0175] After obtaining a clear diagnostic conclusion, the remaining useful life (RUL) probabilistic prediction module is activated. Based on the diagnostic conclusion, this module primarily utilizes the nonlinear bearing dynamics model built into the digital twin model to predict the time-varying stiffness coefficients. The predicted rate of decline is used as the core health indicator. The system ultimately outputs a complete RUL probability density function. .

[0176] When the RUL probability density function After generation, the system immediately executes tiered early warning and adaptive monitoring logic. The system first calculates its key statistics based on the probability distribution, including the expected value. with standard deviation Therefore, its 95% confidence interval is determined as [ 1.96 , +1.96 ].

[0177] The system will set the expected value of RUL. Compared to the preset "warning level" threshold Compare. When the expected value Below If this occurs, a Level 2 warning is triggered. The system then generates and outputs a warning message, the content of which includes:

[0178] Warning level: Level 2 (Warning);

[0179] Diagnostic conclusion: Accelerated degradation related to bearing wear;

[0180] Predicted expected RUL value: μ;

[0181] Predicted 95% confidence interval for RUL: [μ 1.96σ,μ+1.96σ];

[0182] Risk assessment: Medium;

[0183] Maintenance recommendation: Specific downtime maintenance tasks must be scheduled within the expected lifespan.

[0184] At the same time, the triggering of the Level 2 warning will automatically activate the adaptive monitoring function. The system will increase the monitoring frequency and analysis accuracy of the temperature and vibration sensors of the bearing to the highest level, and tighten the judgment threshold for abnormal confirmation in order to capture any signs of accelerated deterioration.

[0185] After maintenance personnel complete the inspection based on the warning and confirm that the fault is a blockage in the lubrication oil circuit, this confirmation result serves as a true label. Together with the previous operating data, it is used to optimize the causal inference model and the RUL model, making their diagnosis and prediction of this type of thermo-mechanical coupling fault more accurate.

[0186] As can be seen from the detailed analysis of the two specific scenarios above, the method proposed in this disclosure can achieve sensitive capture of early and minor faults by constructing a high-fidelity health status baseline under the condition of scarce fault samples; and ensure the reliability of alarms through a multi-level confirmation strategy; its core advantage lies in providing interpretable diagnosis by using causal inference, providing quantitative risk assessment by using probabilistic prediction, and finally achieving continuous self-optimization of the model through closed-loop learning.

[0187] This disclosure differs from traditional digital twin methods that rely on supervised training using historical fault data. It proposes a novel approach based on health status baseline modeling. Its innovation lies in constructing a high-fidelity model of an "absolutely normal" system using only sufficient health operation data and physical mechanisms, and confirming real-time deviations through a multi-level strategy. This not only fundamentally solves the bottleneck of traditional models being ineffective in real-world applications due to the scarcity of fault samples, but also expands the detection capability from "identifying known faults" to "perceiving unknown anomalies."

[0188] This disclosure elevates fault diagnosis from traditional "correlation" analysis to "causal" inference, innovatively integrating causal inference with digital twins. By performing virtual intervention and causal verification within the digital twin model, the root cause of the fault and its physical transmission path can be clearly and interpretably traced and output, thereby breaking through the "black box" limitation of existing AI models and significantly improving the credibility and practicality of diagnostic conclusions.

[0189] The forecast output of this disclosure is a complete probability distribution including expected value, standard deviation, and confidence interval, rather than a traditional single deterministic value. This probabilistic forecast result enables managers to make tolerance-based quantitative assessments of future failure risks, thereby supporting more scientific and robust maintenance planning and resource allocation decisions.

[0190] Unlike traditional digital twins, which are used as static simulation or monitoring tools, this disclosure constructs a dynamic, self-evolving technological closed loop. Its innovation lies in transforming the digital twin system from a passive analysis tool into a proactive model capable of dynamically adjusting monitoring strategies based on risks and continuously learning and improving performance from operational events through adaptive monitoring and closed-loop online optimization. This ensures its long-term effectiveness and accuracy throughout the entire lifecycle of the physical system.

[0191] Although the present disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure.

Claims

1. A method of health management of an electromechanical coupling system, characterized by, The method comprises: Collecting multi-dimensional time series data under healthy working conditions, constructing a multi-dimensional feature vector data set under healthy conditions, and constructing and training a health state baseline model for reflecting the operating state of the electromechanical coupling system based on the multi-dimensional feature vector data set; wherein the electromechanical coupling system comprises an electric motor, an insulated transmission rod and a generator; During the online operation of the electromechanical coupling system, the real-time sensor feature vector at the current time is obtained and input into the health state baseline model to obtain a predicted feature vector, and a real-time residual sequence reflecting the state deviation of the system is generated by calculating the difference between the real-time sensor feature vector and the predicted feature vector; Based on the real-time residual sequence, an abnormal alarm is identified, and a fault mode is analyzed; Based on the fault mode, the remaining useful life of the electromechanical coupling system under the current state is probabilistically predicted, and a corresponding operation and maintenance guidance strategy is generated; a health state baseline model, a physical information autoencoder, whose loss function is defined as a data reconstruction loss plus a physical law residual loss containing multiple critical failure physical processes weighted sum: ； wherein, w recon denotes a data reconstruction loss weight, w phys denotes a physical law residual loss weight; Physical law residual loss is the weighted sum of the mean squared error of the residuals produced by each physical model over all the sampling points in the computational domain. Each physical model includes: a nonlinear bearing dynamics model, a Paris crack propagation model and a multi-factor electrical insulation aging model; The nonlinear bearing dynamics model is used to constrain the vibration behavior of the system; The Paris crack propagation model is used to constrain the fatigue damage accumulation process of the transmission rod; The multi-factor electrical insulation aging model is used to constrain the insulation degradation process of the motor, generator or transmission rod; Based on the real-time residual sequence, an abnormal alarm is identified, including: A multi-level abnormality confirmation strategy is performed on the real-time residual sequence to filter random disturbances and identify deterministic abnormal events, and only when a deviation passes all levels of testing, it is confirmed as a high-confidence abnormality and triggers an abnormal alarm; The multi-level abnormality confirmation strategy includes: The first layer is a time series dynamic test for confirming the persistence and trend of the abnormality; The second layer is a working condition adaptive threshold evaluation, used to ensure the dynamic accuracy of the reference threshold ; wherein the threshold is a function based on a real-time working condition vector; The third layer is a multi-sensor information cross-verification for confirming the physical relevance of the abnormality; The fault mode is analyzed, including: When an abnormal alarm is triggered, a causal graph is used for causal tracing, and a reverse fact analysis is performed using the health state baseline model to identify and output the root cause event and its physical transmission path that leads to the current abnormal state; The causal graph is a directed acyclic causal graph, and the nodes represent key physical variables of the system, and the directed edges represent the direct causal relationship between the variables; The reverse fact analysis is based on the causal graph, when a downstream node variable in the graph is detected to be abnormal, the upstream parent node variable is selected as a potential cause, by forcing the upstream parent node variable to be normal and re-simulating, the causal contribution of the upstream parent node variable is quantified by observing whether the downstream node variable returns to normal.

2. The electromechanical coupling system health management method of claim 1, wherein Collecting multi-dimensional time series data under healthy working conditions, constructing a multi-dimensional feature vector data set under healthy conditions, comprises: A sensor array is arranged on the key components of the electromechanical coupling system to collect multi-dimensional time series data under healthy working conditions, and a multi-dimensional feature vector data set under healthy conditions is constructed by preprocessing and feature extraction of the collected data; wherein the healthy working conditions include normal operating conditions under different combinations of speed, load and environmental temperature.

3. The electromechanical coupling system health management method of claim 2, wherein a sensor array comprising an acceleration sensor for measuring mechanical vibration, a temperature sensor for monitoring key component temperature, a voltage and current transformer for measuring electrical parameters, and an encoder for acquiring rotor dynamic parameters; multi-dimensional time series data comprising vibration acceleration, temperature, voltage, current, speed and torque measured by sensors.

4. The electromechanical coupling system health management method of claim 2, wherein preprocessing the collected data, including cleaning the collected time series data to handle outliers and missing values; using a digital filter to denoise the signal and highlight core information related to system status; normalizing to eliminate the influence of different physical dimensions on model training.

5. The electromechanical coupling system health management method of claim 2, wherein feature extraction, including extracting time-domain statistical features and frequency-domain features from vibration acceleration; extracting power quality features from voltage and current; combining time-domain statistical features, frequency-domain features and power quality features with temperature and speed to form a multi-dimensional feature vector.

6. The electromechanical coupling system health management method of claim 1, wherein a real-time residual sequence is quantified by calculating the Mahalanobis distance between the real-time sensor feature vector and the predicted feature vector at time t.

7. The electromechanical coupling system health management method of claim 1, wherein probabilistic prediction of the remaining useful life of the electromechanical coupling system under the current state, including The remaining useful life of the mechatronic coupling system in the current state is probabilistically predicted by a Bayesian neural network, and the prediction result is a probability density function representing uncertainty of the remaining useful life RUL .

8. The electromechanical coupling system health management method of claim 1, wherein generating a corresponding operation and maintenance guidance strategy, including based on the probabilistic prediction of the remaining useful life, generating a corresponding operation and maintenance guidance strategy, including dynamically adjusting the monitoring strategy of the system, and issuing a series of hierarchical maintenance warnings with different levels of urgency to guide operation and maintenance activities; wherein dynamic monitoring and hierarchical maintenance warning specifically includes comparing the expected value of the remaining useful life RUL prediction with a set of preset time thresholds, and when the expected value of the remaining useful life RUL prediction is lower than the preset time threshold, automatically increasing the sampling frequency of the sensor or the sensitivity of the anomaly detection algorithm; based on the RUL prediction result, generating maintenance instructions with different levels of urgency, including planned, warning and critical level warnings.

9. The electromechanical coupling system health management method of claim 1, further comprising based on the operation and maintenance guidance strategy, performing operation and maintenance, and after the operation and maintenance activity is completed, feeding the complete process data of this fault diagnosis and prediction, as well as the maintenance result confirmed by human into the health state baseline model and the process of probabilistic prediction of the remaining useful life, forming a closed-loop model for online optimization.

10. The electromechanical coupling system health management method of claim 9, wherein closed-loop model online optimization, including ​ After the maintenance activity is completed, sensor feature data from the entire process from the occurrence of the anomaly to the execution of the maintenance is packaged into a data packet, and the manually confirmed root cause of the failure is taken as the true label of the data packet; The data packet with the true label is supplemented to the original training data set as a training sample, and the health state baseline model and the parameters in the probabilistic prediction process of the remaining useful life are updated in an online learning or periodic fine-tuning manner.

11. A health management system for an electromechanical coupling system, the system comprising: Comprise: a state modeling unit, a state prediction unit, a fault identification unit, and an operation and maintenance guidance unit; The state modeling unit is configured to collect multi-dimensional time series data in a healthy working condition, construct a multi-dimensional feature vector data set in a healthy state, and construct and train a health state baseline model for reflecting the operating state of the electromechanical coupling system based on the multi-dimensional feature vector data set; wherein the electromechanical coupling system comprises an electric motor, an insulated transmission rod, and a generator; The state prediction unit is configured to obtain a real-time sensor feature vector at the current time during online operation of the electromechanical coupling system, input the real-time sensor feature vector into the health state baseline model to obtain a predicted feature vector, and generate a real-time residual sequence reflecting the state deviation of the system by calculating the difference between the real-time sensor feature vector and the predicted feature vector; The fault identification unit is configured to identify an abnormal alarm based on the real-time residual sequence and analyze a fault mode; The operation and maintenance guidance unit is configured to probabilistically predict the remaining useful life of the electromechanical coupling system in the current state based on the fault mode and generate a corresponding operation and maintenance guidance strategy; a health state baseline model, a physical information autoencoder, whose loss function is defined as a data reconstruction loss plus a physical law residual loss containing multiple critical failure physical processes weighted sum: ； wherein, w recon denotes a data reconstruction loss weight, w phys denotes a physical law residual loss weight; Physical law residual loss is the weighted sum of the mean squared error of the residuals produced by each physical model over all the sampling points in the computational domain. Each physical model comprises: a nonlinear bearing dynamics model, a Paris crack propagation model, and a multi-factor electrical insulation aging model; The nonlinear bearing dynamics model is configured to constrain the vibration behavior of the system; The Paris crack propagation model is configured to constrain the fatigue damage accumulation process of the transmission rod; The multi-factor electrical insulation aging model is configured to constrain the insulation degradation process of the motor, generator, or transmission rod; Identifying an abnormal alarm based on the real-time residual sequence comprises: Performing a multi-level abnormality confirmation strategy on the real-time residual sequence to filter random disturbances and identify deterministic abnormal events, and only when a deviation passes all levels of inspection, it is confirmed as a high-confidence abnormality and triggers an abnormal alarm; The multi-level abnormality confirmation strategy comprises: The first level is a time series dynamic inspection for confirming the persistence and trend of the abnormality; The second layer is a working condition adaptive threshold evaluation, used to ensure the dynamic accuracy of the reference threshold ; wherein the threshold is a function based on a real-time working condition vector; The third level is a multi-sensor information cross-verification for confirming the physical correlation of the abnormality; Analyzing the fault mode comprises: When an abnormal alarm is triggered, a causal graph is used for causal tracing, and a reverse fact analysis is performed on the health state baseline model to identify and output the root cause event and its physical transmission path that leads to the current abnormal state; The causal graph is a directed acyclic causal graph, and the nodes represent key physical variables of the system, and the directed edges represent the direct causal relationship between the variables. The reverse fact analysis is based on a cause-effect diagram, when an abnormality of a lower node variable in the diagram is detected, a parent node variable of the lower node variable is selected as a potential cause, the parent node variable is forced to be a normal value and is simulated again, and whether the lower node variable returns to normal from abnormal is observed to quantify a cause-effect contribution degree of the parent node variable.

12. An electronic device, comprising: The device comprises a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory complete mutual communication through the communication bus; The memory stores a computer program; The processor is used for executing the computer program stored on the memory, and a health management method of the electromechanical coupling system in any one of claims 1-10 is realized.

13. A computer readable storage medium storing a computer program, wherein the computer program comprises program instructions configured to cause a processor to perform the method according to any one of claims 1 to 12. The computer program is executed by the processor, and a health management method of the electromechanical coupling system in any one of claims 1-10 is realized.

Images