An intelligent manufacturing production line fault prediction and health management method

Abstract

The application discloses a kind of intelligent manufacturing production line fault prediction and health management method, comprising: obtaining equipment state data and production operation data and time synchronization;Digital twin model is constructed and is dynamically calibrated based on adaptive transfer learning using synchronous data;Health index is generated based on synchronous data, and degradation trajectory data is generated in the calibrated digital twin model based on reinforcement learning to execute virtual fault injection;Based on health index, degradation trajectory data and synchronous data, root cause reasoning is carried out and fault mode library is updated;Risk level is determined based on health index and maintenance instruction is output;After maintenance, post-maintenance operation data is obtained, and at least one of digital twin model, health index, fault mode library or fault injection strategy is updated.After the above process, continuous calibration and non-stop degradation trajectory acquisition are realized with the change of working condition, supporting maintenance decision and knowledge base closed-loop update, and improving the adaptability of fault prediction and health management.

Description

Technical Field

[0001] This application relates to the field of production line monitoring technology, and in particular to a method for fault prediction and health management of intelligent manufacturing production lines. Background Technology

[0002] With the rapid development of intelligent manufacturing and the Industrial Internet, production lines in discrete manufacturing, process manufacturing, and other fields are increasingly characterized by diverse equipment types, tight process cycles, and continuously changing operating states. The health status of key equipment on the production line (such as motors, bearings, hydraulic actuators, and transmission mechanisms) directly affects production line uptime, product yield, and delivery stability. To reduce the risk of sudden downtime and improve operational efficiency, the industry is gradually shifting from post-maintenance and periodic inspections to condition-based prognostics and health management (PHM) and predictive maintenance. This involves analyzing equipment operating data to achieve early warnings, life assessments, and maintenance decisions.

[0003] In related technologies, production line fault prediction typically uses sensors to collect signals such as vibration, temperature, and current, combined with operating parameters from systems like PLCs and MES, for feature extraction and model training. This enables anomaly detection, fault identification, and remaining life prediction. Some solutions incorporate digital twin technology to construct virtual models corresponding to physical equipment for condition simulation and health assessment. However, existing methods still generally suffer from the following shortcomings: First, fault prediction models are mostly trained offline based on historical data, and the model parameters and thresholds are relatively fixed, making it difficult to calibrate in real time with changes in operating conditions such as load changes, ambient temperature fluctuations, and process cycle adjustments, resulting in a decrease in prediction accuracy and generalization ability. Second, fault mechanism verification and degradation law modeling often rely on methods such as shutdown testing, accelerated life testing, or fault reproduction, making it difficult to obtain high-quality samples covering multiple fault modes and continuous degradation processes under non-shutdown conditions, resulting in incomplete degradation trajectories and slow updates to the fault mode library. Third, prediction results that rely solely on data-driven approaches often lack interpretable root cause indications, and health indicators and risk classifications lack support for the continuous evolution from "minor degradation to severe failure," further making it difficult to dynamically optimize and improve maintenance actions (such as planned maintenance, high-frequency monitoring, and emergency shutdowns), affecting the reliability of alarms and the timeliness and economy of operation and maintenance decisions.

[0004] Therefore, in the fault prediction and health management of intelligent manufacturing production lines, the models are difficult to achieve highly consistent dynamic calibration with changes in operating conditions, difficult to obtain continuous degradation trajectories and improve the fault mode library under non-stop conditions, and maintenance decisions lack risk-oriented dynamic optimization and closed-loop feedback, which have become urgent technical problems to be solved. Summary of the Invention

[0005] This application provides a method for fault prediction and health management of intelligent manufacturing production lines, aiming to solve the problems in the existing technology of fault prediction and health management of intelligent manufacturing production lines, such as the difficulty in achieving highly consistent dynamic calibration of models with changes in operating conditions, the difficulty in obtaining continuous degradation trajectories and improving the fault mode library under non-stop conditions, and the lack of risk-oriented dynamic optimization and closed-loop feedback in maintenance decisions.

[0006] A method for fault prediction and health management of a smart manufacturing production line is provided, the method comprising: Acquire equipment status data and production operation data of the target equipment; The equipment status data and production operation data are synchronized in time to obtain time-synchronized equipment status data and time-synchronized production operation data. A digital twin model of the target device is constructed, and based on the time-synchronized device status data and the time-synchronized production operation data, adaptive transfer learning is used to perform dynamic calibration on the digital twin model to obtain a dynamically calibrated digital twin model. Based on the time-synchronized equipment status data and the time-synchronized production operation data, health indicators are generated; In the dynamically calibrated digital twin model, virtual fault injection is performed based on reinforcement learning to generate degradation trajectory data; Based on the health indicators, the degradation trajectory data, the time-synchronized equipment status data, and the time-synchronized production operation data, root cause reasoning is performed and the fault mode library is updated. Based on the aforementioned health indicators, the risk level is determined and maintenance instructions are output. After performing maintenance according to the maintenance instructions, the system obtains post-maintenance operation data and updates at least one of the following based on the post-maintenance operation data: the digital twin model, the health indicators, the fault mode library, and the fault injection strategy corresponding to the virtual fault injection. The post-maintenance operation data includes equipment status data and / or production operation data after maintenance.

[0007] Optionally, in the above scheme, the equipment status data includes at least two of the following: vibration data, temperature data, and current data. The production operation data includes at least one of the following: PLC operating parameters, MES cycle time parameters, and process parameters.

[0008] Optionally, in the above scheme, the step of synchronizing the equipment status data and production operation data to obtain time-synchronized equipment status data and time-synchronized production operation data includes: Obtain timestamps for the equipment status data and the production operation data respectively; The device status data and the production operation data are aligned based on the timestamp to obtain the time-synchronized device status data and the time-synchronized production operation data.

[0009] Optionally, in the above scheme, the step of performing dynamic calibration on the digital twin model using adaptive transfer learning to obtain a dynamically calibrated digital twin model includes: Based on the time-synchronized device status data, a source domain dataset is constructed; Based on the output of the digital twin model under the working conditions corresponding to the production operation data synchronized with the time, a target domain dataset is constructed. Based on the distribution difference between the source domain dataset and the target domain dataset, a domain adaptive weight matrix is ​​generated; The model parameters of the digital twin model are iteratively updated according to the domain adaptive weight matrix and using an adaptive learning rate. After each iteration update, the deviation between the output of the digital twin model and the time-synchronized device status data is calculated, and the current round of dynamic calibration is terminated when the deviation meets a preset condition.

[0010] Optionally, in the above scheme, generating a domain-adaptive weight matrix based on the distribution difference between the source domain dataset and the target domain dataset includes: The source domain dataset and the target domain dataset are mapped to the same feature space through a preset feature mapping; The maximum mean difference (MMD) is calculated in the feature space to obtain the distribution difference degree. Based on the distribution difference, the weights of different data modalities and / or different feature dimensions are determined, and the domain adaptive weight matrix is ​​generated.

[0011] Optionally, in the above scheme, generating health indicators based on the time-synchronized equipment status data and the time-synchronized production operation data includes: The time-synchronized equipment status data and the time-synchronized production operation data are segmented according to the preset analysis window; For each analysis window, multidimensional features are extracted, and the multidimensional features are weighted and fused to obtain fused features; wherein, the multidimensional features include at least two types of time-domain features, frequency-domain features, and time-frequency-domain features; The fused features are normalized to obtain the health indicators.

[0012] Optionally, in the above scheme, the step of performing virtual fault injection based on reinforcement learning to generate degradation trajectory data in the dynamically calibrated digital twin model includes: Construct a reinforcement learning agent and define the state of the reinforcement learning agent as including virtual state quantity and physical condition quantity, wherein the virtual state quantity is obtained by the output of the dynamically calibrated digital twin model, and the physical condition quantity is obtained by the time-synchronized equipment status data and / or the time-synchronized production operation data. The actions of the reinforcement learning agent are defined as fault type, fault injection strength, and fault injection timing. The fault injection strategy is obtained by training or reasoning the reinforcement learning agent based on a preset reward function. Virtual fault injection is performed in the dynamically calibrated digital twin model according to the fault injection strategy to obtain the degradation trajectory data.

[0013] Optionally, in the above scheme, the step of performing root cause reasoning and updating the fault mode library based on the health indicators, the degradation trajectory data, the time-synchronized equipment status data, and the time-synchronized production operation data includes: Based on the health indicators, the degradation trajectory data, the time-synchronized equipment status data, and the time-synchronized production operation data, the root cause reasoning is performed by calling the causal relationship graph and / or rule base to obtain fault cause information; The fault cause information is associated with the corresponding feature patterns and / or the degradation trajectory data, and written into the fault pattern library.

[0014] Optionally, in the above scheme, determining the risk level and outputting maintenance instructions based on the health indicators includes: Based on preset risk grading rules, the health indicators are mapped to the risk levels; The maintenance instruction is generated based on the risk level.

[0015] Optionally, in the above scheme, updating at least one of the digital twin model, the health indicators, the fault mode library, and the fault injection strategy corresponding to the virtual fault injection based on the maintenance-post-operational data includes: When updating the digital twin model, the model parameters of the digital twin model are iteratively updated again based on the maintained running data; When updating the health indicators, the feature set, weighted fusion parameters and / or normalization parameters used to generate the health indicators are adjusted based on the maintenance-post-operation data. When updating the fault mode library, feature patterns are extracted based on the post-maintenance operation data and then associated with the fault cause information obtained by root cause reasoning before being updated to the fault mode library. When the fault injection strategy is updated, the preset reward function parameters and / or policy parameters of the reinforcement learning agent are updated based on the post-maintenance running data.

[0016] Compared with the prior art, this application has at least the following beneficial effects: Based on further analysis and research of existing technical problems, this application recognizes that existing technologies in fault prediction and health management of intelligent manufacturing production lines suffer from several issues. First, existing models struggle to achieve highly consistent dynamic calibration with changing operating conditions. Second, they struggle to acquire continuous degradation trajectories and improve the fault mode library without downtime. Third, maintenance decisions lack risk-oriented dynamic optimization and closed-loop feedback. This application addresses these issues by acquiring and synchronizing the target equipment's status data and production operation data, ensuring subsequent analysis is based on the same time benchmark. This allows for a direct correlation between equipment status changes and operating condition changes within the same sequence. Furthermore, this application employs adaptive transfer learning to dynamically calibrate the digital twin model based on the time-synchronized equipment status data and production operation data. This ensures the digital twin model's output is continuously corrected as production conditions change, preventing model decoupling from the actual equipment state due to operating condition drift. Finally, this application generates health indicators based on the time-synchronized equipment status data and production operation data, enabling the target equipment's health status to be continuously represented in a unified quantitative form and used for subsequent judgment. Simultaneously, this application utilizes dynamically calibrated data... The digital twin model uses reinforcement learning to perform virtual fault injection and generate degradation trajectory data, enabling the acquisition of degradation evolution data consistent with the current operating conditions even without shutdown, thus supplementing the continuous degradation process that is difficult to cover by relying solely on historical samples. Furthermore, by combining health indicators, degradation trajectory data, and time-synchronized equipment status data and production operation data, root cause reasoning is performed to update the fault mode library, allowing abnormal states to be quantitatively represented and further linked to fault causes, forming continuously updated pattern knowledge. Finally, this application determines risk levels based on health indicators and outputs maintenance instructions. After maintenance, at least one of the digital twin model, health indicators, fault mode library, or fault injection strategy is updated based on post-maintenance operation data, enabling fault prediction and health management to form a closed-loop iteration. This achieves continuous updating and decision support for production line fault prediction and health management in scenarios with dynamically changing operating conditions, difficulty in obtaining degradation trajectories, and difficulty in dynamically optimizing maintenance decisions. This solves the problems in the background technology where models are difficult to dynamically calibrate with changing operating conditions, difficult to obtain continuous degradation trajectories and improve the fault mode library without shutdown, and lack of closed-loop optimization in maintenance decisions. Attached Figure Description

[0017] Figure 1 A schematic diagram illustrating the application environment of a smart manufacturing production line fault prediction and health management method provided in one embodiment of this application; Figure 2This is a flowchart illustrating a method for predicting and managing faults in a smart manufacturing production line, as provided in one embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] In one embodiment, such as Figure 1 As shown, a method for fault prediction and health management of intelligent manufacturing production lines is provided, including the following steps: Acquire equipment status data and production operation data of the target equipment; The equipment status data and production operation data are synchronized in time to obtain time-synchronized equipment status data and time-synchronized production operation data. A digital twin model of the target device is constructed, and based on the time-synchronized device status data and the time-synchronized production operation data, adaptive transfer learning is used to perform dynamic calibration on the digital twin model to obtain a dynamically calibrated digital twin model. Based on the time-synchronized equipment status data and the time-synchronized production operation data, health indicators are generated; In the dynamically calibrated digital twin model, virtual fault injection is performed based on reinforcement learning to generate degradation trajectory data; Based on the health indicators, the degradation trajectory data, the time-synchronized equipment status data, and the time-synchronized production operation data, root cause reasoning is performed and the fault mode library is updated. Based on the aforementioned health indicators, the risk level is determined and maintenance instructions are output. After performing maintenance according to the maintenance instructions, the system obtains post-maintenance operation data and updates at least one of the following based on the post-maintenance operation data: the digital twin model, the health indicators, the fault mode library, and the fault injection strategy corresponding to the virtual fault injection. The post-maintenance operation data includes equipment status data and / or production operation data after maintenance.

[0020] This implementation method targets any target device (e.g., motor, bearing, hydraulic actuator, or transmission mechanism) in a smart manufacturing production line, acquiring device status data and production operation data during production. Device status data can be collected from various sensors, such as vibration data from vibration sensors, temperature data from temperature sensors, and current data from current sensors. Production operation data can be acquired from the production line's control and management system, such as PLC operating parameters, MES cycle time parameters, and process parameters. The collected data can first be cached and preprocessed at edge nodes or data acquisition gateways before being uploaded to a server or local industrial computing node.

[0021] When synchronizing equipment status data and production operation data, timestamps (such as acquisition terminal timestamps, control system record timestamps, or unified clock source timestamps) can be appended or read separately for the two types of data, and alignment can be performed based on these timestamps. Alignment methods can include: nearest neighbor alignment, linear interpolation alignment, or fixed-period resampling alignment, thereby obtaining time-synchronized equipment status data and time-synchronized production operation data respectively, for subsequent unified window segmentation, feature extraction, and model calibration.

[0022] When constructing a digital twin model of a target device, a physical mechanism model, a data-driven model, or a hybrid model can be used. Taking a motor as an example, a digital twin model can include an electrical sub-model (such as the relationship between windings, current, and voltage), a mechanical sub-model (such as the relationship between rotational inertia and load torque), and a thermal sub-model (such as the relationship between heat conduction and temperature rise). Based on the time-synchronized equipment status data and the time-synchronized production operation data, adaptive transfer learning is used to perform dynamic calibration on the digital twin model: during operation, the deviation between the output of the digital twin model and the time-synchronized equipment status data is periodically compared, and the model parameters are updated according to the transfer learning mechanism until a preset termination condition is met, resulting in a dynamically calibrated digital twin model.

[0023] When generating health indicators, the analysis window is segmented based on time-synchronized equipment status data and time-synchronized production operation data. Features are extracted, fused, and normalized within each window to generate health indicators. Health indicators can be single scalars or multi-dimensional vectors. When outputting risk levels and maintenance instructions, health indicators can be converted into risk levels according to preset mapping rules, and maintenance instructions (such as planned maintenance, high-frequency monitoring, emergency shutdown, etc.) can be output in combination with the risk levels.

[0024] When performing virtual fault injection based on reinforcement learning in a dynamically calibrated digital twin model, the digital twin model can be used as a reinforcement learning environment. The parameters or states of the twin model can be perturbed according to the fault type, fault injection strength, and fault injection timing to generate degradation trajectory data. This degradation trajectory data can include time-varying sequences of key state variables, feature sequences, and health indicator changes, which, together with real data, support root cause reasoning and fault mode library updates.

[0025] When performing root cause reasoning and updating the fault mode library, health indicators, degradation trajectory data, time-synchronized equipment status data, and time-synchronized production operation data can be used as inputs. The causal relationship graph and / or rule base are invoked for reasoning to obtain fault cause information. This fault cause information, along with its corresponding feature patterns and degradation trajectory data, is then written to or updated to the fault mode library. The fault mode library can be organized by fields such as "fault type—trigger condition—feature pattern—degradation trajectory segment—related operating condition," supporting subsequent matching and retrieval.

[0026] After performing maintenance according to the maintenance instructions, acquire the post-maintenance operational data; the post-maintenance operational data includes the equipment status data and / or the production operation data after maintenance. Based on the post-maintenance operational data, update at least one of the following: digital twin model, health indicators, fault mode library, and fault injection strategy corresponding to virtual fault injection; for example, recalibrate model parameters with post-maintenance data, adjust health indicator construction parameters, supplement / revise fault mode library entries, or update reinforcement learning strategy parameters, thereby forming a closed loop.

[0027] This embodiment integrates equipment status data and production operation data under a unified time reference, achieves dynamic calibration of the digital twin model based on adaptive transfer learning, and generates degradation trajectory data by injecting virtual faults through reinforcement learning within the calibrated digital twin model. Combined with root cause reasoning, risk classification, and closed-loop updates of post-maintenance operation data, fault prediction and health management can maintain sustainable operation under conditions of changing operating conditions and scarce fault samples, and support maintenance decisions and knowledge base iteration.

[0028] In this embodiment, the device status data includes at least two of the following: vibration data, temperature data, and current data. The production operation data includes at least one of the following: PLC operating parameters, MES cycle time parameters, and process parameters.

[0029] This embodiment specifies the composition of "equipment status data" and "production operation data". Equipment status data includes at least two of the following: vibration data, temperature data, and current data. In actual deployment, vibration sensors (e.g., bearing housings or casings), temperature sensors (e.g., winding or casing temperature points), and current sampling units (e.g., motor drive circuits) can be placed at key parts of the target equipment, and time-series data streams can be generated respectively.

[0030] Production operation data must include at least one of the following: PLC operating parameters, MES cycle time parameters, or process parameters. PLC operating parameters may include speed settings, load settings, start / stop status, valve positions, or actuator control values; MES cycle time parameters may include workstation cycle time, output count, and shift information; process parameters may include processing recipe parameters, temperature control settings, pressure settings, material batches, or process stage identifiers. Production operation data can be obtained through industrial protocols or system interface calls and will be processed synchronously with equipment status data in subsequent time synchronization processes.

[0031] This embodiment clarifies the typical composition of equipment status data and production operation data, enabling the method to cover common multi-source monitoring and process / cycle time data sources in production lines. This provides an achievable data foundation for subsequent time synchronization, dynamic calibration, health indicator generation, root cause reasoning, and maintenance closed-loop.

[0032] In this embodiment, the step of synchronizing the equipment status data and production operation data to obtain time-synchronized equipment status data and time-synchronized production operation data includes: Obtain timestamps for the equipment status data and the production operation data respectively; The device status data and the production operation data are aligned based on the timestamp to obtain the time-synchronized device status data and the time-synchronized production operation data.

[0033] This implementation details the "time synchronization" process. First, timestamps are obtained for both equipment status data and production operation data. These timestamps can originate from the local clock at the data acquisition terminal, a unified time synchronization server, the PLC system clock, or the time recorded by the MES. To improve consistency, a unified clock synchronization mechanism (e.g., NTP / PTP or unified time synchronization via an industrial gateway) can be configured between the data acquisition terminal and the control system, and a timestamp field can be included in each data record.

[0034] Then, the equipment status data and production operation data are aligned based on timestamps. Alignment methods may include: Resample according to the same sampling period and match the nearest neighbor by timestamp; For low-frequency production operation data, retain the original sampling points; for high-frequency equipment status data, aggregate by window (e.g., calculate mean / variance / energy) and then align. When the sampling frequencies of the two types of data differ significantly, linear interpolation or spline interpolation is used to fill in the missing time points.

[0035] After alignment, we obtain time-synchronized equipment status data and time-synchronized production operation data.

[0036] This embodiment acquires timestamps separately and aligns them based on the timestamps, enabling data from different sampling frequencies and sources to be organized on a unified timeline. This facilitates unified window analysis, model calibration, and inference calculations, and reduces analytical bias caused by data asynchrony.

[0037] In this embodiment, the step of performing dynamic calibration on the digital twin model using adaptive transfer learning to obtain a dynamically calibrated digital twin model includes: Based on the time-synchronized device status data, a source domain dataset is constructed; Based on the output of the digital twin model under the working conditions corresponding to the production operation data synchronized with the time, a target domain dataset is constructed. Based on the distribution difference between the source domain dataset and the target domain dataset, a domain adaptive weight matrix is ​​generated; The model parameters of the digital twin model are iteratively updated according to the domain adaptive weight matrix and using an adaptive learning rate. After each iteration update, the deviation between the output of the digital twin model and the time-synchronized device status data is calculated, and the current round of dynamic calibration is terminated when the deviation meets a preset condition.

[0038] This implementation details the process of "dynamically calibrating the digital twin model using adaptive transfer learning". First, a source domain dataset is constructed based on the time-synchronized equipment status data. The source domain dataset can include multimodal monitoring quantities and their sample sets under different production operation data conditions. Then, a target domain dataset is constructed based on the output of the digital twin model under the operating conditions corresponding to the time-synchronized production operation data. The target domain dataset can be the corresponding output sequence obtained from the twin model simulation.

[0039] Next, a domain-adaptive weight matrix is ​​generated based on the distribution dissimilarity between the source and target domain datasets. The distribution dissimilarity characterizes the degree of difference between the real data distribution and the twinned output distribution; the domain-adaptive weight matrix is ​​used to assign weights to different data modalities or different feature dimensions for subsequent parameter update process to measure the dissimilarity and adjust the update strategy.

[0040] Subsequently, the model parameters of the digital twin model are iteratively updated according to the domain adaptive weight matrix and using an adaptive learning rate. Model parameters can include at least one of the following: electrical parameters (e.g., equivalent resistance, equivalent inductance), mechanical parameters (e.g., moment of inertia, coefficient of friction), and thermal parameters (e.g., thermal conductivity, heat capacity). The adaptive learning rate can be dynamically adjusted based on the distribution difference, deviation, or number of iterations to meet the update requirements of different operating conditions. After each iteration, the deviation between the digital twin model output and the time-synchronized equipment status data is calculated. The dynamic calibration is terminated when the deviation meets preset conditions (e.g., below a threshold, meeting the threshold multiple times consecutively, or reaching the maximum number of iterations), resulting in a dynamically calibrated digital twin model.

[0041] This embodiment constructs source domain datasets and target domain datasets, drives the generation of domain adaptive weight matrices with distribution difference, and iteratively updates the parameters of the digital twin model with adaptive learning rate, enabling the digital twin model to be dynamically calibrated as production conditions change.

[0042] In this embodiment, generating a domain-adaptive weight matrix based on the distribution difference between the source domain dataset and the target domain dataset includes: The source domain dataset and the target domain dataset are mapped to the same feature space through a preset feature mapping; The maximum mean difference (MMD) is calculated in the feature space to obtain the distribution difference degree. Based on the distribution difference, the weights of different data modalities and / or different feature dimensions are determined, and the domain adaptive weight matrix is ​​generated.

[0043] This embodiment describes the steps of "generating an adaptive weight matrix for the target domain". First, the source domain dataset and the target domain dataset are mapped to the same feature space through a preset feature mapping. The preset feature mapping can be a linear mapping, a kernel mapping, or a mapping implemented by a feature extraction network; the purpose of the mapping is to eliminate the representational differences caused by different data sources, different units, or different sampling resolutions, so that the source domain and the target domain can be compared in a unified space.

[0044] Subsequently, the maximum mean difference (MMD) is calculated in this feature space to obtain the distribution dissimilarity. The calculation of MMD can be based on a kernel function (such as a Gaussian kernel or a multinomial kernel) to measure the difference in mean embeddings between the source domain samples and the target domain samples, thereby obtaining a dissimilarity index in scalar or vector form, which is used to characterize the distribution difference between the two domains.

[0045] After obtaining the distribution dissimilarity, weights for different data modes and / or different feature dimensions are determined based on the distribution dissimilarity, and a domain-adaptive weight matrix is ​​generated. The weights can be determined by assigning higher weights to modes / dimensions that contribute more to the dissimilarity and lower weights to modes / dimensions with stronger noise or poorer stability; alternatively, the weights can be normalized to satisfy a weight sum of 1 or matrix norm constraints. The generated domain-adaptive weight matrix is ​​used in subsequent model parameter iteration and update processes.

[0046] This embodiment maps the source domain and target domain to the same feature space, calculates the distribution difference degree (MMD), and generates a domain adaptive weight matrix based on the difference degree. This enables the dynamic calibration process to adaptively weight different data modes and feature dimensions, thereby supporting the feasibility and consistency control of digital twin model parameter updates.

[0047] In this embodiment, generating health indicators based on the time-synchronized equipment status data and the time-synchronized production operation data includes: The time-synchronized equipment status data and the time-synchronized production operation data are segmented according to the preset analysis window; For each analysis window, multidimensional features are extracted, and the multidimensional features are weighted and fused to obtain fused features; wherein, the multidimensional features include at least two types of time-domain features, frequency-domain features, and time-frequency-domain features; The fused features are normalized to obtain the health indicators.

[0048] This implementation method describes the specific process of "generating health indicators". First, the time-synchronized equipment status data and time-synchronized production operation data are segmented according to the preset analysis window. The analysis window can be a fixed-length sliding window. The window length and step size can be set according to the equipment response speed and process cycle time. For example, the window length can be several seconds to several minutes, and the step size is a part of the window length, so as to balance real-time performance and stability.

[0049] Multidimensional features are extracted within each analysis window, and these features are then weighted and fused to obtain the fused features. The multidimensional features include at least two of the following: time-domain features, frequency-domain features, and time-frequency-domain features. Temporal characteristics include root mean square, peak value, kurtosis, peak factor, etc. Frequency domain characteristics include characteristic frequency band energy, harmonic parameters, and sideband energy ratio. Time-frequency domain features include wavelet energy distribution, wavelet entropy, or short-time Fourier spectrum statistics.

[0050] Weighted fusion can linearly weight various features according to preset weights, concatenate them and then reduce their dimensionality, or obtain a fused feature vector according to an attention mechanism.

[0051] Finally, the fused features are normalized to obtain health indicators. Normalization can be min-max normalization, z-score standardization followed by mapping to a fixed interval, or the fused features can be input into a health indicator mapping model to obtain scalar health indicators.

[0052] This embodiment obtains health indicators through window segmentation, multi-dimensional feature extraction, weighted fusion, and normalization, enabling the production line operating status to be represented in a unified indicator form, providing a calculable health metric for risk level mapping, root cause inference input, and maintenance decision-making.

[0053] In this embodiment, the step of generating degradation trajectory data by performing virtual fault injection based on reinforcement learning in the dynamically calibrated digital twin model includes: Construct a reinforcement learning agent and define the state of the reinforcement learning agent as including virtual state quantity and physical condition quantity, wherein the virtual state quantity is obtained by the output of the dynamically calibrated digital twin model, and the physical condition quantity is obtained by the time-synchronized equipment status data and / or the time-synchronized production operation data. The actions of the reinforcement learning agent are defined as fault type, fault injection strength, and fault injection timing. The fault injection strategy is obtained by training or reasoning the reinforcement learning agent based on a preset reward function. Virtual fault injection is performed in the dynamically calibrated digital twin model according to the fault injection strategy to obtain the degradation trajectory data.

[0054] This implementation details the process of "generating degradation trajectory data by performing virtual fault injection based on reinforcement learning". First, a reinforcement learning agent is constructed, and its state is defined to include virtual state variables and physical operating condition variables: virtual state variables are obtained from the output of a dynamically calibrated digital twin model, such as speed, torque, temperature rise, and current; physical operating condition variables are obtained from time-synchronized equipment state data and / or time-synchronized production operation data, such as load rate, process stage identifier, ambient temperature, or cycle time intensity. By using both virtual state variables and physical operating condition variables as state inputs, the reinforcement learning agent can select fault injection actions under constraints consistent with real operating conditions.

[0055] The actions of the reinforcement learning agent are then defined, including fault type, fault injection strength, and fault injection timing. Fault type can cover typical degradation / fault modes of the target equipment, such as winding aging, bearing wear, insufficient lubrication, and sensor drift; fault injection strength can be the disturbance amplitude, parameter offset ratio, or noise injection intensity; fault injection timing can correspond to idle, steady-state, or high-load sections of the production line.

[0056] The reinforcement learning agent is trained or inferred based on a preset reward function to obtain a fault injection strategy. The reward function can consist of multiple metrics, such as fault mode matching degree, time cost of reaching a preset fault state, and penalty term for production operation constraints; the training algorithm can be DQN, DDPG, PPO, etc. Finally, virtual fault injection is performed in the dynamically calibrated digital twin model according to the fault injection strategy, and the state variables and feature sequences of the twin output changing over time are recorded to form degradation trajectory data.

[0057] This embodiment constructs a reinforcement learning agent within a dynamically calibrated digital twin model and uses fault type, fault injection intensity, and fault injection timing as actions to implement virtual fault injection, enabling degradation trajectory data to be generated under controllable conditions and used for subsequent root cause reasoning and fault mode library construction.

[0058] In this embodiment, the step of performing root cause reasoning and updating the fault mode library based on the health indicators, the degradation trajectory data, the time-synchronized equipment status data, and the time-synchronized production operation data includes: Based on the health indicators, the degradation trajectory data, the time-synchronized equipment status data, and the time-synchronized production operation data, the root cause reasoning is performed by calling the causal relationship graph and / or rule base to obtain fault cause information; The fault cause information is associated with the corresponding feature patterns and / or the degradation trajectory data, and written into the fault pattern library.

[0059] This implementation details the steps of "performing root cause reasoning and updating the fault mode library". Root cause reasoning inputs include health indicators, degradation trajectory data, time-synchronized equipment status data, and time-synchronized production operation data. Root cause reasoning can be executed by calling causal relationship graphs and / or rule bases. Causal relationship diagrams can establish causal edges for "operating conditions, state variables, characteristic patterns, and fault causes", and obtain fault cause information through evidence propagation, causal inference, or confidence updates; The rule base can consist of expert rules, threshold rules, or statistical rules. For example, "when a certain frequency domain index exceeds the threshold and the temperature rise rate exceeds the threshold and the system is under high load, output the corresponding fault cause candidate."

[0060] The inference output is fault cause information, which may include fault type identifier, associated component identifier, and triggering conditions.

[0061] After associating the fault cause information with its corresponding feature patterns and / or degradation trajectory data, the information is written into the fault pattern library. The feature pattern can be a window feature vector, feature statistical distribution, or feature template; the degradation trajectory data can be a health indicator / feature sequence fragment that evolves over time. If similar fault cause information entries already exist in the fault pattern library, the entry version can be maintained during implementation using a strategy of "updating with the same key and adding with different keys" to support subsequent matching and tracing.

[0062] This embodiment performs root cause reasoning on health indicators, degradation trajectory data, and multi-source data after time synchronization based on causal relationship graphs and / or rule bases, and writes the fault cause information and feature patterns / degradation trajectories into the fault mode library, so that the fault mode library can be updated in conjunction with the operating data and support subsequent fault identification and maintenance decisions.

[0063] In this embodiment, determining the risk level and outputting maintenance instructions based on the health indicators includes: Based on preset risk grading rules, the health indicators are mapped to the risk levels; The maintenance instruction is generated based on the risk level.

[0064] This implementation method describes the process of "determining the risk level and outputting maintenance instructions". First, based on the preset risk grading rules, health indicators are mapped to risk levels. The risk grading rules can adopt threshold segmentation, interval mapping, or multi-dimensional discrimination methods: for example, dividing health indicators into multiple intervals, each interval corresponding to a risk level; or modifying the mapping based on health indicators and working condition constraints.

[0065] Maintenance instructions are then generated based on the risk level. These instructions can be generated according to preset risk level strategies: for example, lower risk levels correspond to continued operation and periodic inspections, medium risk levels to high-frequency monitoring and planned maintenance, and high risk levels to emergency shutdowns or load-limited operation. Maintenance instructions can include fields such as maintenance type, suggested time window, target components, precautions, and required resource information, and can be pushed to the operation and maintenance platform or industrial control system.

[0066] This embodiment maps health indicators to risk levels by pre-setting risk grading rules, and generates maintenance instructions based on the risk levels. This enables maintenance decisions to form structured outputs with health indicators as the entry point, facilitating the linkage between production line operation and maintenance execution and management system.

[0067] In this embodiment, updating at least one of the digital twin model, the health indicators, the fault mode library, and the fault injection strategy corresponding to the virtual fault injection based on the post-maintenance operational data includes: When updating the digital twin model, the model parameters of the digital twin model are iteratively updated again based on the maintained running data; When updating the health indicators, the feature set, weighted fusion parameters and / or normalization parameters used to generate the health indicators are adjusted based on the maintenance-post-operation data. When updating the fault mode library, feature patterns are extracted based on the post-maintenance operation data and then associated with the fault cause information obtained by root cause reasoning before being updated to the fault mode library. When the fault injection strategy is updated, the preset reward function parameters and / or policy parameters of the reinforcement learning agent are updated based on the post-maintenance running data.

[0068] This implementation details the process of "updating at least one item based on the post-maintenance operational data". When updating the digital twin model, the model parameters of the digital twin model are iteratively updated again based on the post-maintenance operational data. In practice, the post-maintenance operational data can be used as a new round of source domain dataset input into the dynamic calibration process, so that the model parameters reconverge with the post-maintenance state baseline.

[0069] When updating health metrics, the feature set, weighted fusion parameters, and / or normalization parameters used to generate the health metrics are adjusted based on the post-maintenance operational data. For example, the feature distribution can be recalculated in the post-maintenance data to remove unstable features, or the fusion weights of different modal features can be re-estimated; normalization parameters (such as maximum and minimum values, mean and variance, etc.) can be updated to adapt to the changes in status after maintenance.

[0070] When updating the fault mode library, feature patterns are extracted based on post-maintenance operational data and associated with fault cause information obtained through root cause reasoning before being updated to the fault mode library. For example, feature patterns within a short period after maintenance are used as "post-maintenance verification samples" and associated with fault cause entries to record post-maintenance status change characteristics or post-repair feature regression results.

[0071] When updating the fault injection strategy, the preset reward function parameters and / or policy parameters of the reinforcement learning agent are updated based on the post-maintenance operational data. In practice, the operating conditions, constraints, and target state reflected in the post-maintenance operational data can be used as training or fine-tuning data to update the reward function weight coefficients or policy network parameters, ensuring that subsequent virtual fault injections are consistent with the state boundaries after actual maintenance.

[0072] This embodiment uses the post-maintenance operational data to update at least one of the parameters of the digital twin model, the parameters for generating health indicators, the entries in the fault mode library, and the fault injection strategy, thereby enabling the fault prediction and health management process to form a post-maintenance feedback loop, which supports the continuous iteration and consistent maintenance of the model, indicators, and knowledge base.

[0073] In one embodiment, a method for fault prediction and health management of a smart manufacturing production line based on digital twins is provided, such as... Figure 2 As shown, the method includes the following steps: S1: First, deploy a sensor network for key equipment: motors, bearings, and hydraulic systems to collect vibration, temperature, and current operating data. At the same time, synchronize data with PLC operating parameters, MES production cycle time, and process parameters. S2: Based on the real-time data from S1, an adaptive transfer learning algorithm is used to adjust the model parameters to achieve state mapping, and the virtual model synchronously displays the device temperature and speed. S3: Extract time-domain and frequency-domain features from the multi-source data in S1; combine process knowledge to fuse the features into a quantitative health indicator HI. S4: Virtual Fault Injection and Root Cause Reverse Reasoning: Using the calibrated model in step S2, virtual faults are actively injected to observe the degradation trajectory and fault performance; then, combined with real-time physical line data, the root cause is inferred through the cause-effect graph to generate a fault mode library to supplement the knowledge base. S5: Based on the health index HI from S3 and the prediction results from S4, a health-risk level matrix is ​​constructed: the status is divided into normal (HI≥90, risk level 0), mild degradation (70≤HI<90, risk level 1), moderate warning (50≤HI<70, risk level 2), and severe failure (HI<50, risk level 3). Risk level 1 requires planned maintenance, risk level 2 requires high-frequency monitoring, and risk level 3 requires emergency shutdown. S6: Integrate the early warning data from step S5, the root cause data from step S4, and the historical data from step S1, and then provide feedback on the repair effect after maintenance to form a closed loop.

[0074] This embodiment discloses a method for fault prediction and health management of intelligent manufacturing production lines based on digital twins. By dynamically calibrating the digital twin model through an adaptive transfer learning algorithm, the deviation between the virtual model and the physical equipment is controlled within an extremely low range, providing a high-precision foundation for fault injection simulation. The intelligent agent selects a fault injection strategy based on actual operating conditions, and the simulated fault degradation trajectory highly matches the characteristics of real faults. It can completely present the entire process of equipment from a healthy state to the occurrence of a fault, providing real and effective data support for the construction of health indicators and early warning systems. This reduces production interruptions caused by sudden faults. By using fault simulation in a virtual environment, the testing risks in actual production are avoided. Verification of various fault scenarios can be completed without downtime, significantly reducing the cost of fault testing and its impact on production, and significantly improving the stability and operating efficiency of the intelligent manufacturing system.

[0075] In a preferred embodiment, in step S1, the vibration acceleration is sampled at a frequency of 2kHz, and the sensor is installed at the motor bearing end cover; the stator three-phase current is sampled at a frequency of 1kHz, and the sensor is deployed at the stator output terminal; the winding temperature and ambient temperature are sampled at a frequency of 1Hz, and the temperature sensor is embedded inside the winding and the motor housing. During the data preprocessing stage, the vibration signal is filtered through a low-pass filter with a cutoff frequency of 500Hz to remove high-frequency noise, and the current signal is denoised using wavelet transform. All data is normalized to the 0-1 range and then packaged into CSV format every minute for storage, with the synchronization latency to the edge computing node controlled within 100ms. Simultaneously, the data pool needs to integrate over 100 sets of historical fault cases and over 500 hours of normal operating condition data to provide a foundation for subsequent model calibration and feature extraction.

[0076] In a preferred embodiment, in step S2, the real-time physical motor data collected in S1—vibration acceleration, stator three-phase current, and winding temperature—are used as the source domain dataset Ds, and the simulated data output by the virtual model—virtual current and virtual temperature—are used as the target domain dataset Dt.

[0077] First, the feature distribution difference between Ds and Dt is calculated using a kernel function (using an improved maximum mean difference, MMD), and a domain adaptive weight matrix W is dynamically generated (high weights are assigned to features with similar distributions, such as current (0.6), temperature (0.3), and vibration (0.1). Next, a loss function is constructed based on the weight matrix, and the key model parameters—stator winding resistance Rs, rotor moment of inertia J, and thermal conductivity k—are updated using an adaptive learning rate η (which decreases linearly with decreasing distribution difference, initially η=0.001). Parameter updates are performed every 5 minutes, and the root mean square error (RMSE) is used to evaluate the state deviation between the model and the physical line. The current adjustment round stops when the deviation is ≤5%, achieving a high-precision state mapping between the virtual model and the physical motor. During this process, the parameter adjustment logs are synchronized to the S1 data pool for subsequent analysis, and the adjusted high-precision model directly serves as the core carrier for virtual fault injection in S4.

[0078] In a preferred embodiment, in step S2, the weighted maximum mean difference formula is: ; In the formula: W-MMD represents the weighted maximum mean difference; wi represents the weight coefficient of the i-th feature; n represents the number of feature dimensions; m represents the number of data samples in the source domain; k represents the number of data samples in the target domain; (·) indicates kernel function mapping (using a Gaussian kernel); xi represents the i-th feature sample in the source domain; yi represents the i-th feature sample in the target domain; The formula for adaptive learning rate is: formula:

[0079] In the formula: η(t) represents the adaptive learning rate at time t; η0 represents the initial learning rate; α represents the attenuation coefficient; W-MMD(t) represents the weighted maximum mean difference at time t; The formula for evaluating model state deviation is:

[0080] In the formula: Dev represents the percentage of model state deviation; T represents the number of samples within the evaluation period; V(t) represents the output value of the virtual model at time t; P(t) represents the actual measured value of the physical line at time t.

[0081] In a preferred embodiment, in step S3, a 10-second sliding window and a 5-second step size are used to process real-time data. The root mean square and peak factor are calculated for time-domain features, and the current harmonic distortion rate and vibration sideband energy ratio are obtained through Fast Fourier Transform for frequency-domain features. Wavelet entropy quantization is used to quantize the signal complexity for time-frequency domain features. The weights for health indicators are 0.4 for time-domain features, 0.3 for time-domain features, and 0.3 for time-frequency features. After weighted summation, the weights are normalized to the 0-1 range. A warning signal is triggered when the health indicator threshold is set to 0.8. The feature library is updated weekly, integrating new fault case features and health indicator change trajectories to ensure the timeliness and accuracy of feature extraction.

[0082] In a preferred embodiment, in step S4, an agent is constructed based on the motor digital twin model calibrated in S2 (as the DRL environment) to actively decide on the fault injection strategy. The agent's state space encompasses the real-time parameters of the virtual model (speed, stator current, output torque) and the physical motor operating condition data (load rate, ambient temperature) collected in S1; the action space is defined as fault type (stator winding aging, rotor bar breakage, bearing wear, etc.), injection intensity (resistance increase ratio, inertia change rate), and injection timing (idle period / high load period of production cycle). The algorithm learns the optimal strategy through the Q-value function of DQN: after the agent selects an action, the environment outputs a new state (such as current increase, torque fluctuation), and the reward value that integrates fault matching degree, simulation efficiency, and production impact is calculated through the fault injection reward function formula. The Q-network parameters are adjusted according to the DQN target Q-value update formula to continuously optimize the injection logic. This process relies on the high-precision mapping of the S2 model, and the action selection must match the actual operating conditions of the physical data in S1.

[0083] In a preferred embodiment, during step S4, when observing the degradation trajectory, the agent preferentially selects to inject a stator winding aging fault (action: resistance increases by 15%, timing: load rate 85%), and the trajectory is displayed by the virtual model output. When the injection intensity increases to 25%, the model simulates a winding overheating breakdown state (temperature exceeds 150°C, protection device triggers shutdown), and the virtual geometric model displays cracks in the winding insulation layer. These trajectories (current-time curves, efficiency-load curves) are quantified into feature sequences and synchronized to S3 for health indicator construction, providing a complete path of fault evolution for the S5 early warning.

[0084] In a preferred embodiment, in step S4, The formula for the fault injection reward function is:

[0085] In the formula: R represents the single-step reward value; λ1, λ2, and λ3 represent weighting coefficients, which are 0.6, 0.3, and 0.1, respectively. M represents the fault feature matching degree (cosine similarity between the virtual and actual fault databases, ranging from 0 to 1). T represents the time (in minutes) required to simulate reaching the fault state; the smaller the better. I represents the production impact (the percentage of downtime to cycle time, ranging from 0 to 1). The formula for harmonic distortion rate (THD) is:

[0086] In the formula: THD represents the total harmonic distortion rate of current. In represents the effective value (A) of the nth harmonic current. I1 represents the effective value of the fundamental current (A). The formula for updating the target Q value in DQN is:

[0087] In the formula: yj represents the target Q value; rj represents the reward value at step j (from the fault injection reward function formula); γ represents the discount factor (taken as 0.95); Qtarget represents the target Q-network; sj+1 represents the state at step j+1; a′ indicates the next action; Indicates the target network parameters; The formula for efficiency degradation rate is:

[0088] In the formula: : Efficiency degradation rate at time t; ηt represents the virtual motor efficiency at time t; η0 represents the motor efficiency under normal conditions.

[0089] In a preferred embodiment, in step S5, the threshold optimization cycle is performed every two weeks; the root cause reasoning relies on a rule base, and when the current harmonic distortion rate is greater than 5% and the winding temperature exceeds 120°C, the root cause is determined to be stator winding aging. The warning information is pushed to the maintenance personnel via SMS, and the fault location and processing time are displayed on the monitoring platform.

[0090] In a preferred embodiment, in step S6, planned maintenance is recommended when the health index is less than 0.7, and emergency shutdown is recommended when it is greater than 0.9; maintenance resources are prioritized for allocation to the nearest maintenance team, with a response time of no more than 30 minutes.

[0091] In summary, by employing the above technical solutions, this embodiment dynamically calibrates the digital twin model using an adaptive transfer learning algorithm, keeping the state deviation between the virtual model and the physical equipment within an extremely low range, thus providing a high-precision foundation for fault injection simulation. The intelligent agent selects a fault injection strategy based on actual operating conditions, and the simulated fault degradation trajectory highly matches the characteristics of real faults. It can completely present the entire process of equipment from a healthy state to the occurrence of a fault, providing real and effective data support for the construction of health indicators and early warning systems. This helps to identify potential equipment problems in advance and reduce production interruptions caused by sudden failures.

[0092] By using fault simulation in a virtual environment, the testing risks in actual production are avoided. Verification of various fault scenarios can be completed without downtime, significantly reducing the cost of fault testing and its impact on production. The generated fault evolution paths and health indicators can guide maintenance personnel to develop targeted maintenance plans and optimize the allocation and scheduling of maintenance resources. Simultaneously, the closed-loop optimization mechanism continuously updates models and strategies, constantly improving the accuracy of fault prediction and the scientific nature of maintenance decisions. This drives the transformation of production line maintenance from passive response to proactive prediction, significantly improving the stability and operational efficiency of the intelligent manufacturing system.

[0093] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for fault prediction and health management of an intelligent manufacturing production line, characterized in that, The method includes: Acquire equipment status data and production operation data of the target equipment; The equipment status data and production operation data are synchronized in time to obtain time-synchronized equipment status data and time-synchronized production operation data. A digital twin model of the target device is constructed, and based on the time-synchronized device status data and the time-synchronized production operation data, adaptive transfer learning is used to perform dynamic calibration on the digital twin model to obtain a dynamically calibrated digital twin model. Based on the time-synchronized equipment status data and the time-synchronized production operation data, health indicators are generated; In the dynamically calibrated digital twin model, virtual fault injection is performed based on reinforcement learning to generate degradation trajectory data; Based on the health indicators, the degradation trajectory data, the time-synchronized equipment status data, and the time-synchronized production operation data, root cause reasoning is performed and the fault mode library is updated. Based on the aforementioned health indicators, the risk level is determined and maintenance instructions are output. After performing maintenance according to the maintenance instructions, the system obtains post-maintenance operation data and updates at least one of the following based on the post-maintenance operation data: the digital twin model, the health indicators, the fault mode library, and the fault injection strategy corresponding to the virtual fault injection. The post-maintenance operation data includes equipment status data and / or production operation data after maintenance.

2. The method according to claim 1, characterized in that, The device status data includes at least two of the following: vibration data, temperature data, and current data. The production operation data includes at least one of the following: PLC operating parameters, MES cycle time parameters, and process parameters.

3. The method according to claim 1, characterized in that, The step of synchronizing the equipment status data and production operation data in time to obtain time-synchronized equipment status data and time-synchronized production operation data includes: Obtain timestamps for the equipment status data and the production operation data respectively; The device status data and the production operation data are aligned based on the timestamp to obtain the time-synchronized device status data and the time-synchronized production operation data.

4. The method according to claim 1, characterized in that, The step of performing dynamic calibration on the digital twin model using adaptive transfer learning to obtain a dynamically calibrated digital twin model includes: Based on the time-synchronized device status data, a source domain dataset is constructed; Based on the output of the digital twin model under the working conditions corresponding to the production operation data synchronized with the time, a target domain dataset is constructed. Based on the distribution difference between the source domain dataset and the target domain dataset, a domain adaptive weight matrix is ​​generated; The model parameters of the digital twin model are iteratively updated according to the domain adaptive weight matrix and using an adaptive learning rate. After each iteration update, the deviation between the output of the digital twin model and the time-synchronized device status data is calculated, and the current round of dynamic calibration is terminated when the deviation meets a preset condition.

5. The method according to claim 4, characterized in that, The step of generating a domain-adaptive weight matrix based on the distribution difference between the source domain dataset and the target domain dataset includes: The source domain dataset and the target domain dataset are mapped to the same feature space through a preset feature mapping; The maximum mean difference (MMD) is calculated in the feature space to obtain the distribution difference degree. Based on the distribution difference, the weights of different data modalities and / or different feature dimensions are determined, and the domain adaptive weight matrix is ​​generated.

6. The method according to claim 1, characterized in that, The generation of health indicators based on the time-synchronized equipment status data and the time-synchronized production operation data includes: The time-synchronized equipment status data and the time-synchronized production operation data are segmented according to the preset analysis window; For each analysis window, multidimensional features are extracted, and the multidimensional features are weighted and fused to obtain fused features; wherein, the multidimensional features include at least two types of time-domain features, frequency-domain features, and time-frequency-domain features; The fused features are normalized to obtain the health indicators.

7. The method according to claim 1, characterized in that, In the dynamically calibrated digital twin model, virtual fault injection is performed based on reinforcement learning to generate degradation trajectory data, including: Construct a reinforcement learning agent and define the state of the reinforcement learning agent as including virtual state quantity and physical condition quantity, wherein the virtual state quantity is obtained by the output of the dynamically calibrated digital twin model, and the physical condition quantity is obtained by the time-synchronized equipment status data and / or the time-synchronized production operation data. The actions of the reinforcement learning agent are defined as fault type, fault injection strength, and fault injection timing. The fault injection strategy is obtained by training or reasoning the reinforcement learning agent based on a preset reward function. Virtual fault injection is performed in the dynamically calibrated digital twin model according to the fault injection strategy to obtain the degradation trajectory data.

8. The method according to claim 1, characterized in that, The step of performing root cause reasoning and updating the fault mode library based on the health indicators, the degradation trajectory data, the time-synchronized equipment status data, and the time-synchronized production operation data includes: Based on the health indicators, the degradation trajectory data, the time-synchronized equipment status data, and the time-synchronized production operation data, the root cause reasoning is performed by calling the causal relationship graph and / or rule base to obtain fault cause information; The fault cause information is associated with the corresponding feature patterns and / or the degradation trajectory data, and written into the fault pattern library.

9. The method according to claim 1, characterized in that, The process of determining the risk level and outputting maintenance instructions based on the health indicators includes: Based on preset risk grading rules, the health indicators are mapped to the risk levels; The maintenance instruction is generated based on the risk level.

10. The method according to claim 1, characterized in that, The step of updating at least one of the following based on the maintained operational data: the digital twin model, the health indicators, the fault mode library, and the fault injection strategy corresponding to the virtual fault injection, includes: When updating the digital twin model, the model parameters of the digital twin model are iteratively updated again based on the maintained running data; When updating the health indicators, the feature set, weighted fusion parameters and / or normalization parameters used to generate the health indicators are adjusted based on the maintenance-post-operation data. When updating the fault mode library, feature patterns are extracted based on the post-maintenance operation data and then associated with the fault cause information obtained by root cause reasoning before being updated to the fault mode library. When the fault injection strategy is updated, the preset reward function parameters and / or policy parameters of the reinforcement learning agent are updated based on the post-maintenance running data.

Images