Internet of Things monitoring system and method

Abstract

The invention relates to the technical field of Internet of Things monitoring, and provides an Internet of Things monitoring system and method, and the method comprises the steps: synchronously obtaining a valve position signal, a current signal and a vibration signal, extracting features, combining a valve position change rate as a weight, fusing the features of multiple signals, and carrying out the mapping output of a health feature vector; embedding the health feature vector into a causal knowledge graph, reasoning the causal knowledge graph through a graph neural network, and outputting a fault occurrence probability, residual life and a causal path; a part degradation model is integrated in digital twinning, different candidate maintenance strategies are simulated and deduced, and multi-target cooperation is realized through a target optimization algorithm, so that maintenance strategy one-sidedness caused by single dimension and fixed working conditions is avoided; after the maintenance strategy is executed, the actual performance track is generated and compared with the change track of the residual life to obtain the track deviation, parameters of the graph neural network and the component degradation model are adaptively corrected based on the track deviation, long-term stable operation of monitoring is guaranteed, and the manual operation and maintenance cost is greatly reduced.

Description

Technical Field

[0001] This application relates to the field of Internet of Things (IoT) monitoring technology, and in particular to an IoT monitoring system and method. Background Technology

[0002] With the popularization of Industrial Internet of Things (IIoT) technology, the real-time monitoring and fault early warning of actuators, as the core control units in industrial production processes, have become crucial for ensuring production continuity. Early methods relied on manual inspections or data collection from single sensors, using threshold judgments to achieve simple fault warnings. This stage could only identify obvious faults and could not capture early, subtle anomalies. With the development of deep learning technology, features of signals such as current and vibration were extracted using neural networks, combined with simple fault databases to achieve fault type matching. However, this stage did not consider the stage differences in the valve switching process of the actuator, and the fixed weights of multi-signal fusion resulted in insufficient accuracy in health feature representation.

[0003] In recent years, digital twin technology has been introduced into the monitoring of actuators. It simulates the degradation process and infers the maintenance effect by constructing a virtual model of the physical entity. However, in the existing technology, digital twins are mostly combined with independent component degradation models, without deep integration with the causal relationship reasoning between components. Moreover, the selection of maintenance strategies is mostly based on a single dimension and lacks multi-objective collaborative optimization. Furthermore, the existing technology cannot accurately correspond to specific mechanical actions when judging the health status of the actuator, and the basic data for subsequent reasoning and prediction is distorted.

[0004] Existing fault reasoning methods mostly target independent components, failing to consider the coupling characteristics of multiple components working collaboratively in the actuator, thus failing to locate fault propagation paths and resulting in insufficient interpretability of reasoning results. When deriving maintenance strategies in digital twins, they are mostly based on fixed operating conditions, failing to consider changes in operating load during actual operation, leading to significant deviations between the derived results and actual effects. Furthermore, the selection of maintenance strategies often focuses on a single dimension, failing to dynamically adjust evaluation weights based on the actuator's operating scenario and remaining lifespan, resulting in selected strategies that cannot adapt to actual needs.

[0005] Based on the shortcomings of the prior art, the technical problem solved by this application is how to monitor signals to infer actuator risks through component relationships and perform virtual maintenance to correct closed-loop monitoring. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this application provides an Internet of Things (IoT) monitoring system and method.

[0007] In a first aspect, this application provides an Internet of Things (IoT) monitoring method, which includes: acquiring valve position signals, current signals and vibration signals during the operation of an actuator; after synchronizing the signals, extracting the current harmonic amplitude of the current signal and the vibration harmonic components of the vibration signal during the valve switching process through a convolutional neural network; fusing them with the valve position change rate as a weight; and outputting a health feature vector characterizing the mechanical state. It receives health feature vectors and embeds them into a causal knowledge graph between pre-set mechanism components. It then uses a graph neural network to reason about the causal knowledge graph and outputs the failure probability, remaining lifespan, and causal path. The system generates multiple candidate maintenance strategies based on the inference results, simulates and infers different candidate maintenance strategies in a digital twin of the integrated component degradation model, and selects a maintenance strategy based on the inference results using an objective optimization algorithm. After implementing the maintenance strategy, the current and vibration signals of the actuator are continuously acquired and the actual performance trajectory is generated. The trajectory deviation is obtained by comparing the trajectory of the remaining life with the actual performance trajectory. The parameters of the graph neural network and the parameters of the component degradation model are adaptively corrected based on the trajectory deviation.

[0008] As an optional implementation, the output health feature vector includes: The valve position signal, current signal and vibration signal are acquired during the operation of the actuator. After synchronizing the signal, the execution stages of the valve switching process are divided based on the change characteristics of the valve opening. In each execution phase, the current harmonic amplitude of the current signal and the vibration harmonic component of the vibration signal are extracted by a convolutional neural network to determine the spatiotemporal correlation characteristics of the current signal and the vibration signal within the same execution phase. Using the valve position change rate as a weight, the current harmonic amplitude, vibration harmonic components, and spatiotemporal correlation characteristics of different execution stages are weighted and fused, and then mapped through a dual-channel residual network to output a health feature vector characterizing the mechanical state.

[0009] As an optional implementation, the reasoning on the causal knowledge graph includes: The health feature vector is decomposed into component state vectors corresponding to the causal knowledge graph nodes and embedded with node attributes. The causal strength weights of the edges of the causal knowledge graph are adjusted based on the temporal changes of the health feature vector. By using graph neural networks to perform component reasoning on the node attributes of the causal knowledge graph, the sub-failure probability and sub-remaining lifetime of different mechanism components are output. The reasoning results between mechanism components are aggregated based on the edge of the causal knowledge graph. The coupling influence coefficient of the related components is considered in combination with the causal strength weight, and the failure probability and remaining life of the actuator are output. During the reasoning process, edges in the causal knowledge graph that satisfy the preset causal confidence threshold are activated to form a directed causal path from the fault source node to the end influencing node, and the causal path is selected based on the product of the causal strength weights of each edge on the directed causal path.

[0010] As an optional implementation, the selected maintenance strategy includes: Based on the deduction results of the degradation trajectories of different mechanical components under candidate maintenance strategies, the evaluation dimensions of performance recovery time, total life cycle cost and disturbance degree of coupled components are determined. Based on the operating scenario and remaining life of the actuator, dynamic weights are assigned to different evaluation dimensions, and the deviation rate of the simulation results of different candidate maintenance strategies under the fluctuation of operating load is calculated. The weighted evaluation results of different candidate maintenance strategies are corrected based on the deviation rate. The revised weighted evaluation results are input into the target optimization algorithm to select a maintenance strategy based on the synergistic objectives of performance recovery time and total lifecycle cost.

[0011] As an optional implementation, the simulation of different candidate maintenance strategies includes: In response to the inference results, multiple candidate maintenance strategies are generated, and each candidate maintenance strategy is mapped to the parameter intervention rule of the component degradation model; The operating load parameters of the actuator are determined based on the fluctuation of current and vibration signals, and then the operating load parameters are input into the digital twin. Initiate a component degradation model in a digital twin and extrapolate the degradation trajectory of different components under different candidate maintenance strategies based on the coupling relationship between the components.

[0012] As an optional implementation, the parameters of the adaptive correction graph neural network include: Based on the causal strength weights of edges in the causal knowledge graph, the trajectory deviation is decomposed into component sub-deviations, and it is determined whether the component sub-deviations originate from state reasoning deviations or association reasoning deviations. If the component sub-bias originates from state reasoning bias, then the feature mapping coefficients of the corresponding component in the graph neural network are corrected; if the component sub-bias originates from association reasoning bias, then the aggregation weight coefficients between the corresponding components in the graph neural network are corrected. By combining the importance of the mechanism components and the degree of component sub-deviation, a correction magnitude coefficient is set to correct the parameters of the graph neural network, and a correction step size constraint is set based on the changing trend of the trajectory deviation after historical correction. After correcting the parameters, verify the deviation between the inference results and the actual performance trajectory. If the verification fails, backtrack the parameters of the graph neural network and adjust the correction magnitude coefficient.

[0013] As an optional implementation, generating the actual performance trajectory includes: Before and after implementing the maintenance strategy, the current and vibration signals of the actuator are continuously acquired, and the current harmonic amplitude, vibration harmonic components and spatiotemporal correlation features are extracted simultaneously to form baseline features and real-time feature sequences. Based on the importance of mechanism components in the causal knowledge graph, the real-time feature sequence is split into feature subsequences corresponding to the mechanism components, and the weight of each feature subsequence is dynamically adjusted according to the sub-remaining lifetime of the mechanism components. The real-time feature sequence is processed based on the environmental interference signal of the actuator, and the weighted feature sub-sequences are integrated according to different mechanism components to generate the actual performance trajectory including instantaneous performance fluctuations and trend degradation.

[0014] As an optional implementation, the parameters of the adaptive correction component degradation model include: Based on the parameter intervention rules of the component degradation model in the digital twin, the trajectory deviation is decomposed into the self-degradation deviation of a single mechanism component and the coupling degradation deviation between multiple mechanism components. If the deviation is self-degradation, then the basic degradation parameters of the corresponding mechanism component are corrected in a targeted manner; if the deviation is coupling degradation, then the coupling influence coefficient between related components is corrected in a targeted manner. By combining the degradation sensitivity and deviation duration of the mechanism components, a correction magnitude factor is set to correct the parameters of the component degradation model, and the correction results of the parameters of the graph neural network are synchronously associated. After correcting the parameters, the degenerate trajectory is re-enacted using a digital twin and the trajectory deviation is re-verified. If the verification fails, the correction magnitude factor is adjusted and the verification is repeated.

[0015] Secondly, this application provides an Internet of Things (IoT) monitoring system, which includes: a data acquisition module, a fault diagnosis module, and a maintenance execution module; The data acquisition module is used to acquire valve position signals, current signals and vibration signals during the operation of the actuator. After synchronizing the signals, the current harmonic amplitude of the current signal and the vibration harmonic components of the vibration signal are extracted through a convolutional neural network during the valve switching process. The data is then fused with the valve position change rate as a weight to output a health feature vector representing the mechanical state. The fault diagnosis module is used to receive health feature vectors and embed them into a causal knowledge graph between pre-set mechanism components. It performs reasoning on the causal knowledge graph based on a graph neural network, outputs the probability of fault occurrence, remaining lifespan and causal path, generates multiple candidate maintenance strategies in response to the reasoning results, simulates and deduces different candidate maintenance strategies in a digital twin of the integrated component degradation model, and selects a maintenance strategy based on the deduction results through a target optimization algorithm. The maintenance execution module is used to continuously acquire the current signal and vibration signal of the actuator after executing the maintenance strategy and generate the actual performance trajectory. The trajectory deviation is obtained by comparing the change trajectory of the remaining life with the actual performance trajectory. The parameters of the graph neural network and the parameters of the component degradation model are adaptively corrected based on the trajectory deviation.

[0016] Compared with existing technologies, the beneficial effects of this application are as follows: by synchronously acquiring valve position signals, current signals, and vibration signals and extracting features, combining the valve position change rate as a weight to fuse the features of multiple signals, and then mapping and outputting a health feature vector, the mechanical state of the actuator can be truly represented, providing data assurance for high accuracy in subsequent fault reasoning; by embedding the health feature vector into a causal knowledge graph, and then using a graph neural network to reason about the causal knowledge graph, the probability of fault occurrence, remaining lifespan, and causal path are output. This reasoning process avoids the one-sidedness of modeling independent components and is interpretable, making it easy for maintenance personnel to quickly locate the root cause of the fault and block its propagation.

[0017] By integrating a component degradation model into a digital twin to simulate and deduce different candidate maintenance strategies, and then using a target optimization algorithm to achieve multi-objective collaboration, the one-sidedness of maintenance strategies caused by a single dimension and fixed operating conditions can be avoided. The selected maintenance strategy can not only ensure the reliability of the actuator, but also reduce the total life cycle maintenance cost. After the maintenance strategy is executed, the actual performance trajectory is generated and compared with the trajectory of the remaining life to obtain the trajectory deviation. Based on the trajectory deviation, the parameters of the graph neural network and the component degradation model are adaptively corrected. Without manual intervention, the long-term stable operation of the monitoring system is guaranteed, and the manual maintenance cost is greatly reduced. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart of an Internet of Things (IoT) monitoring method provided in an embodiment of this application; Figure 2 This is a logical flowchart of the reasoning causal knowledge graph of an IoT monitoring method provided in an embodiment of this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application more apparent and understandable, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0020] Example 1 like Figure 1 The diagram shown is a flowchart of an IoT monitoring method provided in this application embodiment. The method includes: S1. Acquire valve position signal, current signal and vibration signal during the operation of the actuator. After synchronizing the signal, extract the current harmonic amplitude of the current signal and the vibration harmonic component of the vibration signal during the valve switching process through a convolutional neural network. Weigh the components with the valve position change rate as the weight and output a health feature vector representing the mechanical state.

[0021] Specifically, the output health feature vector includes: The valve position signal, current signal and vibration signal are acquired during the operation of the actuator. After synchronizing the signal, the execution stages of the valve switching process are divided based on the change characteristics of the valve opening. In each execution phase, the current harmonic amplitude of the current signal and the vibration harmonic component of the vibration signal are extracted by a convolutional neural network to determine the spatiotemporal correlation characteristics of the current signal and the vibration signal within the same execution phase. Using the valve position change rate as a weight, the current harmonic amplitude, vibration harmonic components, and spatiotemporal correlation characteristics of different execution stages are weighted and fused, and then mapped through a dual-channel residual network to output a health feature vector characterizing the mechanical state.

[0022] During valve operation, the mechanical actions of the actuator exhibit significant stage differences. For example, during the valve opening stage, the static friction of the valve must be overcome, and the motor load fluctuates greatly. During the valve holding stage, the valve opening is stable, and the mechanical action is smooth. During the valve closing stage, the valve needs to be controlled to close at a uniform speed to avoid impact. If feature extraction is directly performed on the entire process signal, the signal features of different stages will be mixed, making it impossible to accurately correspond to the specific mechanical action state, thus affecting the representation accuracy of the subsequent health feature vector. First, the valve position signal is obtained through a displacement sensor installed on the valve stem side to avoid interference from mechanical contact with valve position detection. Then, the current signal is obtained through a Hall current sensor connected in series with the actuator motor power supply circuit to capture changes in motor current in real time. Next, the vibration signal is obtained through a vibration sensor attached to the valve body and the motor end cover to cover the vibration transmission path of key mechanical components. Finally, the three signals are processed synchronously using a unified timestamp to ensure that the valve position signal, current signal, and vibration signal correspond at the same time, avoiding deviations in the execution stage division caused by signal delays.

[0023] The classification rules are set based on the changing characteristics of valve opening. When the valve opening continuously rises from 0, it is determined to enter the valve opening stage; when the valve opening stabilizes, it is determined to enter the valve holding stage, which corresponds to the stage where the valve remains stable and the mechanical components operate under low load; when the valve opening continuously drops to 0, it is determined to enter the valve closing stage. During the classification process, the timing of the execution stage switching is dynamically adjusted by monitoring the changing characteristics of the valve opening in real time. For example, if the vibration signal suddenly increases during the valve closing stage, the execution stage switching can be temporarily delayed to avoid misjudgment. This achieves precise binding between signal characteristics and mechanical action stages, avoids the mixing of signal characteristics in different execution stages, and ensures that the features extracted later can directly reflect the state under specific mechanical actions. For example, the current characteristics of the valve opening stage only correspond to the load state when the motor starts, and reduces interference from invalid signals, improving the targeting of subsequent feature extraction and providing execution stage basis for abnormal state positioning of the actuator.

[0024] Abnormal mechanical states of the actuator are directly reflected in the harmonics of the current and vibration signals. For example, bearing wear can cause the current to generate harmonics of a specific frequency, which in turn can amplify the harmonic components of the vibration signal at the corresponding frequency. Furthermore, the changes in the current and vibration signals within the same execution phase are correlated in both time and space. Extracting only single signal features cannot capture this cross-signal correlation and is insufficient to comprehensively characterize the mechanical state. In each execution phase, a convolutional neural network is used to extract the current harmonic amplitude of the current signal and the vibration harmonic components of the vibration signal. The synchronized current signal is divided into sub-signals according to a time window. Convolution operations are used to capture the current harmonic amplitudes of different frequencies. Then, a pooling layer is used to filter out the current harmonic amplitude with the largest amplitude, which corresponds to the key harmonic that best reflects the mechanical state. Similarly, vibration harmonic components are extracted. During the training of the convolutional neural network, the parameters of the convolutional neural network are optimized using the harmonics during normal operation of the actuator as samples to ensure accurate separation of abnormal harmonics from normal harmonics.

[0025] Within the same execution phase, the correlation between current harmonic amplitude and vibration harmonic components is analyzed from both temporal and spatial dimensions. In the temporal dimension, a sliding time window is used to calculate the time difference between the peak values ​​of the current harmonic amplitude and the peak values ​​of the vibration harmonic components. If the time difference remains stable within a preset range, a temporal correlation is identified. In the spatial dimension, considering the sensor's installation location, the transmission relationship between changes in the motor-end current harmonic amplitude and changes in the valve stem-end vibration harmonic components is analyzed. For example, when the motor-end current harmonic amplitude increases, does the valve stem-end vibration harmonic component increase synchronously? An attention mechanism is used to assign higher weight to the correlation along the spatial transmission path, ultimately integrating the temporal and spatial correlation features into a spatiotemporal correlation feature. This allows for the accurate extraction of weak abnormal harmonics, avoiding the difficulty in capturing subtle anomalies and improving the sensitivity of mechanical state characterization. The spatiotemporal correlation feature reflects the cross-signal correlation between current and vibration signals, compensating for the limitations of single-signal harmonic features and comprehensively characterizing the collaborative state between mechanical components. Based on the harmonic features extracted during the execution phase and the spatiotemporal correlation feature, abnormal patterns in different phases can be distinguished.

[0026] The valve position change rate varies in different execution stages, and the corresponding mechanical activity level and state influence weight also differ. For example, the valve position change rate is large in the valve opening and closing stages, and the mechanical components are in a dynamic stress state. The features of this execution stage have a higher weight in representing the mechanical health state. However, the valve position change rate is small in the valve holding stage, and the mechanical components are in a stable state. The feature weight is lower. If equal weight fusion is used, the contribution of key stage features will be weakened, resulting in the health feature vector failing to highlight key state information. Without mapping through a dual-channel residual network, the gradient vanishing problem is likely to occur, making it impossible to effectively integrate the three types of features, resulting in low accuracy of the output health feature vector. First, the valve position change rate of each execution stage is calculated, and the valve position change rate of each execution stage is compared with the sum of the valve position change rates of all execution stages to obtain the weight of that execution stage. The valve opening and closing stages have higher weights than the valve holding stage. The current harmonic amplitude, vibration harmonic components, and spatiotemporal correlation features of different execution stages are weighted and fused. Then, the weighted results of all execution stages are summarized according to feature type to form three types of global weighted results.

[0027] A dual-channel residual network is constructed, dividing the global weighted result into two inputs. The first channel inputs the weighted current harmonic amplitude and the temporal correlation component of the spatiotemporal correlation feature, as both are directly related to the motor's operating state. The second channel inputs the weighted vibration harmonic component and the spatial correlation component of the spatiotemporal correlation feature, as both are directly related to the vibration transmission of mechanical components. Each channel includes two residual blocks, which are used to compress the dimensions of the features and integrate information. Finally, a fusion layer concatenates the output features of the two channels, and then a fully connected layer maps them into a fixed-dimensional vector, which is the health feature vector representing the mechanical state. This is achieved during network training. At that time, the known health status of the actuator is used as a label to optimize the network parameters, ensuring that the output health feature vector can accurately distinguish different health statuses, including normal, light wear, and heavy wear. This allows the health feature vector to preferentially reflect the health information under the dynamic stress state of the machinery, improving the ability to represent early dynamic anomalies. The dual-channel residual network avoids the gradient vanishing problem and can effectively integrate features across signals and stages, ensuring that the health feature vector has a moderate dimension and complete information, avoiding feature redundancy or information loss. The output health feature vector has a unified dimension and clear state representation meaning, which facilitates standardized processing and inference in subsequent steps.

[0028] S2. Receive health feature vectors and embed a pre-set causal knowledge graph between components. Perform reasoning on the causal knowledge graph based on a graph neural network, outputting the failure probability, remaining lifespan, and causal path. In response to the reasoning results, generate multiple candidate maintenance strategies. Simulate and deduce different candidate maintenance strategies in a digital twin of the integrated component degradation model. Select a maintenance strategy based on the deduction results using a target optimization algorithm.

[0029] It needs to be explained that the causal knowledge graph is first determined based on the mechanical structure and operating principle of the actuator to establish the nodes and initial causal edges of the graph. Nodes include mechanism component nodes and state nodes, while edges represent causal relationships of faults, clarifying the direction of the causal relationship. Then, the graph is supplemented and adjusted using historical fault data of the actuator. For example, if historical fault data shows that valve stem jamming indirectly leads to increased bearing vibration, then a new causal edge of valve stem jamming → increased bearing vibration is added. At the same time, based on the probability that the effect node occurs after the cause node occurs in the historical fault data, an initial causal strength weight is set for each causal edge. The higher the probability, the more the initial causal strength weight reflects the tightness of the causal relationship.

[0030] After receiving the health feature vector and decomposing it into component state vectors, an update is triggered. If the state vector of a component node continues to change, all associated causal edges of that node are located. If the state vector of the corresponding result node also changes synchronously, the causal strength weight of that edge is increased, and vice versa. This ensures that the weight can reflect the tightness of the causal relationship between components in real time and avoids inference bias caused by static weights.

[0031] Similarly, the component degradation model is based on the degradation characteristics of different components of the actuator, and a separate degradation logic is built for each core component. For example, the motor degradation logic is designed around the aging coefficient and the acceleration effect of load on aging, and the bearing degradation logic is designed around the wear amount and the impact of vibration on wear. This ensures that each component degradation model can accurately reflect the real degradation law of the corresponding component and avoid the one-sidedness of the general model. Based on the mechanical coupling relationship of the actuator, coupling effects are added to each component degradation model. For example, when the motor load increases, it will transmit the wear acceleration signal to the bearing degradation model and adjust the wear-related parameters of the bearing. When the bearing vibration intensifies, it will transmit the aging interference signal to the motor degradation model and adjust the aging-related parameters of the motor. This ensures that the model can simulate the mutual influence between components, rather than independent degradation.

[0032] The degradation calculation logic of the component degradation model is based on basic variables such as running time and number of actions, combined with core parameters to calculate the degradation state of the component. For example, the degree of motor aging = initial aging state + aging coefficient × running time + acceleration related to the working load; where the initial state parameter is the initial degradation state of the component, such as the wear amount and initial friction coefficient at the time of leaving the factory, which is the starting point of the degradation calculation.

[0033] Coupling effects are based on the mutual influence between components to achieve coordinated degradation of multiple components. Coupling parameters are parameters that quantify the degree of influence between components. They receive degradation state signals from other components and convert them into adjustment instructions for the degradation parameters of the current component. They also clarify the effective boundaries of coupling effects. For example, a motor only affects the directly connected bearings and does not affect the remote seals, avoiding meaningless cross-component interference and ensuring the rationality of the simulation.

[0034] It receives operating load parameters and converts them into adjustments to basic degradation parameters. For example, it increases the aging acceleration parameters of the motor under continuous high load and increases the fatigue-related parameters of the bearing under impact load. It dynamically updates the calculation results of the load effect as the operating conditions change, avoiding the inference distortion caused by fixed operating conditions.

[0035] Specifically, such as Figure 2 As shown, reasoning on a causal knowledge graph includes: The health feature vector is decomposed into component state vectors corresponding to the causal knowledge graph nodes and embedded with node attributes. The causal strength weights of the edges of the causal knowledge graph are adjusted based on the temporal changes of the health feature vector. By using graph neural networks to perform component reasoning on the node attributes of the causal knowledge graph, the sub-failure probability and sub-remaining lifetime of different mechanism components are output. The reasoning results between mechanism components are aggregated based on the edge of the causal knowledge graph. The coupling influence coefficient of the related components is considered in combination with the causal strength weight, and the failure probability and remaining life of the actuator are output. During the reasoning process, edges in the causal knowledge graph that satisfy the preset causal confidence threshold are activated to form a directed causal path from the fault source node to the end influencing node, and the causal path is selected based on the product of the causal strength weights of each edge on the directed causal path.

[0036] A health feature vector is a comprehensive description of the mechanical state of an actuator. A causal knowledge graph, with its nodes being components such as motors, bearings, and valve stems, requires decomposing the overall features into the state information of the corresponding components to enable component-based reasoning, such as determining whether the abnormality occurs in the motor or the bearing. Furthermore, the health state of the actuator changes dynamically over time, and the correlation strength between components in the causal knowledge graph must be adjusted synchronously with these changes to avoid static weights causing lag in correlation reasoning. Node attributes of the causal knowledge graph are pre-defined; for example, motor node attributes include current harmonic adaptability and load stability, while bearing node attributes include vibration harmonic adaptability and wear correlation. The health feature vector is then decomposed according to the dimensions of these node attributes. For instance, the current harmonic amplitude in the health feature vector corresponds to the current harmonic adaptability attribute of the motor node, and the vibration harmonic components correspond to the vibration harmonic adaptability attribute of the bearing node, forming the component state vector corresponding to each node. This component state vector is then embedded into the corresponding node, replacing the original node attributes.

[0037] The system monitors the temporal trends of health feature vectors in real time. For example, if the vibration harmonic components of a bearing continuously increase over multiple time windows, and the temporal change rate of the component state vector of a certain component exceeds the normal threshold set based on historical normal states, the associated edges of that component node are located, such as the edge between the bearing and the motor, or the edge between the bearing and the valve stem. The causal strength weights are adjusted according to the rule that the temporal change trend is positively correlated with the weight adjustment magnitude. For example, if the temporal change rate of the bearing vibration feature is greater than the normal threshold, the causal strength weight of its edge with the motor is adjusted upward accordingly. If the temporal change rate is less than the normal threshold, the causal strength weights remain unchanged to avoid fluctuations in the inference results due to frequent adjustments. This achieves accurate mapping from overall features to component features, enabling the node attributes of the causal knowledge graph to reflect the real state of the corresponding component in real time, providing accurate input for subsequent component inference. The causal strength weights of the edges are dynamically adjusted with the state to ensure that the association representation between components is synchronized with the actual health changes, avoiding deviations in association inference caused by static weights. This injects dynamic adaptability into the causal knowledge graph, allowing it to adapt to the state changes of the actuator in the long term without the need for frequent manual updates of the graph.

[0038] Node attributes alone cannot directly determine the failure risk and remaining lifespan of a component. Graph neural networks can utilize the relationships between edges in a causal knowledge graph and combine node attributes to infer the potential state of a component. Simultaneously, the component's own failure risk and the duration of its continued normal operation provide a basis for subsequent aggregation, avoiding the problem of ambiguous abnormal component location caused by directly inferring the overall state. The graph attention network within the graph neural network performs component inference based on the node attributes of the causal knowledge graph. The causal strength weights of the edges serve as the initial values ​​for the attention weights. Using historical failure data of the actuator as training samples, the attention weights and output layer parameters of the graph attention network are optimized, enabling the graph neural network to learn the mapping pattern from node attributes to sub-failure probabilities / sub-remaining lifespans. This allows it to output the sub-failure probabilities and sub-remaining lifespans of different mechanism components. For example, when the current harmonic adaptation degree in a motor node is less than the normal threshold, the sub-failure probability increases significantly.

[0039] The causal knowledge graph embedded with real-time component state vectors is input into the trained graph attention network. The graph attention network performs independent reasoning for each node through multi-layer attention propagation. First, it calculates the deviation degree between the node attribute and the normal threshold. For example, if the bearing vibration harmonic fit degree is less than the normal threshold, the deviation degree increases accordingly. Then, it combines the deviation degree of associated nodes and edge weights to output the sub-fault probability of the component. The larger the deviation degree and the more severe the abnormality of the associated node, the higher the sub-fault probability. At the same time, based on the trend of deviation degree changing over time, the remaining life of the sub-fault is calculated, that is, the time required for the deviation degree to reach the fault threshold. This allows the fusion of state information of associated components to assist in reasoning and avoids the one-sidedness of reasoning based on single node attributes. For example, when the motor is abnormal, it can distinguish whether it is a problem of its own or an indirect problem caused by bearing wear. The output sub-fault probability and sub-remaining life accurately correspond to a single component, providing a direct basis for subsequent location of abnormal components. The reasoning process is highly interpretable and facilitates subsequent analysis of the root cause of the fault.

[0040] An actuator is a system in which multiple components work collaboratively. An anomaly in a single component can affect other components through coupling relationships. For example, bearing wear can increase the motor load and accelerate motor aging. Simply treating the sub-fault probability ignores this coupling amplification effect. Therefore, a coupling influence coefficient is introduced to quantify the degree of mutual influence between components. Combined with the causal strength weight of the edges, a scientific aggregation of component and system indicators is achieved, ensuring that the failure probability and remaining lifespan reflect the true risk of the system. First, a coupling influence coefficient is set for each component. For example, the coupling influence coefficient between the motor and bearing is determined based on the probability that bearing anomalies cause motor anomalies in historical fault data; the higher the probability, the larger the coupling influence coefficient. Then, the overall failure probability is calculated through weighted aggregation and coupling correction. First, the sub-fault probability of each component is multiplied by its importance weight to obtain a weighted sub-fault probability. The importance weight is preset based on the degree of influence of the component on the actuator's function. Then, for each pair of related components, the weighted sub-fault probabilities of both are multiplied, and then multiplied by the corresponding coupling influence coefficient and the causal strength weight of the edges to obtain a coupling correction value. Finally, all weighted sub-fault probabilities and coupling correction values ​​are summed to obtain the overall failure probability of the actuator.

[0041] Based on the remaining life of key components, and combined with adjustments to the coupling influence coefficient, key components refer to components with high importance weights, including motors, bearings, and valve stems. For example, if the remaining life of a certain key component is short and its coupling influence coefficient on another key component is large, then the actual remaining life of the other key component needs to be adjusted accordingly, because the short life of the former will shorten the actual life of the latter through the coupling relationship. The minimum value of the remaining life of the key components after correction is taken as the remaining life of the actuator, ensuring that the remaining life is based on the shortest-lived key component to avoid overestimation. Through weighted aggregation and coupling correction, the coupling amplification effect between components is accurately quantified, avoiding underestimation or overestimation of the probability of failure due to simple aggregation. The remaining life is based on key components, combined with coupling relationship correction, to ensure that the calculation results are consistent with the actual law that the key components of the actuator determine the overall remaining life. The aggregation process is deeply bound to the edge weights and coupling influence coefficients of the causal knowledge graph, so that the overall indicator can reflect the associated risks, rather than the risk superposition of isolated components.

[0042] Simply outputting the probability of failure and remaining lifetime is insufficient to pinpoint the source of the failure and the causal path, leading to a lack of targeted maintenance in subsequent operations. By identifying the fault source node and the chain of influence through the causal path, and by setting a causal confidence threshold, weakly related edges can be filtered out to avoid generating redundant paths. By filtering through the product of the causal strength weights of each edge, causal paths with a high probability of propagation can be prioritized. Based on historical failure data, the minimum causal strength weight of the edges in the real failure path is statistically analyzed and set as the preset causal confidence threshold. During the inference process, all edges of the causal knowledge graph are traversed. If the causal strength weight of an edge is greater than or equal to the preset causal confidence threshold, the edge is determined to meet the confidence requirement and is activated. If it is less than the preset causal confidence threshold, it is temporarily not activated.

[0043] Component nodes with higher sub-fault probabilities are selected as candidate fault source nodes. Starting from each candidate source node, a directed causal path is formed by tracing along the activated edges to the associated nodes, such as motor → bearing → valve stem, with the arrow direction indicating the fault propagation direction. The causal association direction based on the edges is preset. The product of the causal strength weights of all edges on the directed causal path is calculated, and the paths are sorted from largest to smallest to select several paths as causal paths. The larger the product, the higher the probability of the causal path's actual propagation. The fault source node and the terminal influencing node of each causal path are marked. This clarifies the source node and influence chain of fault propagation, enabling subsequent maintenance strategies to accurately locate the initial faulty component and block the propagation path, avoiding blind maintenance. Through threshold filtering and product screening, redundant weakly associated paths are eliminated, reducing interference information in the formulation of subsequent maintenance strategies and providing a basis for fault root cause analysis.

[0044] Furthermore, the simulations of different candidate maintenance strategies include: In response to the inference results, multiple candidate maintenance strategies are generated, and each candidate maintenance strategy is mapped to the parameter intervention rule of the component degradation model; The operating load parameters of the actuator are determined based on the fluctuation of current and vibration signals, and then the operating load parameters are input into the digital twin. Initiate a component degradation model in a digital twin and extrapolate the degradation trajectory of different components under different candidate maintenance strategies based on the coupling relationship between the components.

[0045] The inference results clearly identify the current risks and failure propagation patterns of the executing mechanism. A feasible maintenance plan needs to be generated to address these risks. However, the component degradation model in the digital twin cannot directly identify abstract strategies such as component replacement and parameter adjustment. These strategies must be transformed into parameter intervention rules that the component degradation model can understand to simulate the effectiveness of the maintenance strategies and avoid a disconnect between the maintenance strategies and the component degradation model, which would prevent simulations from being conducted. Based on the failure probability and causal path identified in the inference results, rules for generating candidate maintenance strategies are formulated. If the failure probability of the executing mechanism is significantly higher than expected, the remaining lifespan is significantly shorter than expected, and the failure source node in the causal path is a single component, then a new candidate maintenance strategy is generated. Strategies include emergency replacement of source node components and status monitoring of impact chain components, as well as comprehensive shutdown and overhaul of source nodes and propagation path components. If the probability of failure and remaining lifespan are at a moderate level, strategies such as adjusting the operating parameters of source node components, regular monitoring, lubrication / cleaning maintenance of source node components, and tracking the degradation trend of impact chain components are generated. If the probability of failure is low and the remaining lifespan is long, strategies such as enhancing the signal sampling frequency of source node components and optimizing the component inspection sequence based on causal paths are generated. It is ensured that candidate maintenance strategies cover different scenarios such as emergency handling, routine maintenance, and risk prevention, and the number is controlled within a reasonable range to avoid excessive strategies that lead to low inference efficiency.

[0046] The mapping relationship between pre-set maintenance actions and parameter intervention rules is established. Based on the specific maintenance actions of candidate maintenance strategies, corresponding parameter intervention rules are matched. For example, replacing a bearing corresponds to resetting the cumulative fatigue damage coefficient of the bearing component to its initial value and clearing the wear parameter to zero; adjusting motor operating parameters corresponds to adjusting the aging rate coefficient and load loss coefficient of the motor component according to a set ratio; cleaning the valve stem corresponds to correcting the friction coefficient of the valve stem component to a standard value. During the mapping process, if a candidate maintenance strategy contains multiple maintenance actions, the parameter intervention rules for each maintenance action are mapped separately, forming the parameter intervention rules for that candidate maintenance strategy. This achieves a deep binding between candidate maintenance strategies and inference results, ensuring that candidate maintenance strategies can specifically address current risks and avoiding the problem of general strategies failing to match actual risks. Through the mapping between candidate maintenance strategies and parameter intervention rules, abstract maintenance actions are transformed into operable quantitative parameter adjustments for the component degradation model, providing executable input instructions for the digital twin's deduction, preventing the deduction from failing due to the inability of candidate maintenance strategies to be implemented, and facilitating subsequent comparison of the intervention intensity of different candidate maintenance strategies.

[0047] The component degradation rate of the actuator is directly related to the operating load. If the digital twin only simulates maintenance strategies based on fixed operating conditions, it will be out of touch with the actual operating environment. The fluctuation of current and vibration signals can directly reflect changes in operating load. It is necessary to extract operating load parameters through the fluctuation and input them into the digital twin to ensure that the simulation environment is consistent with reality and avoid distortion of simulation results. The current and vibration signals of the actuator are acquired in real time, and the fluctuation of the two signals is analyzed. The frequency and amplitude of peak fluctuations of the current signal per unit time are calculated. If the fluctuation is frequent and the amplitude is large, it is judged as an instantaneous impact load, and the operating load parameters of impact frequency and impact intensity are extracted. If the fluctuation is infrequent but the amplitude is stable and high, it is judged as a continuous high load, and the operating load coefficient of continuous load is extracted. The stability of the vibration amplitude in the vibration signal, i.e., the standard deviation, is calculated. If the stability is low and synchronized with the current fluctuation, the load type judgment result is verified. Finally, the operating load parameters of impact frequency, impact intensity, and continuous load coefficient are integrated to obtain the operating load parameters.

[0048] The operating load parameters are input into the digital twin according to the correspondence between components and loads. For example, instantaneous impact loads mainly affect motors and bearings, so the impact frequency and impact intensity are input into the component degradation model of the motor and bearings; while continuous high loads mainly affect the valve stem, so the continuous load coefficient is input into the component degradation model of the valve stem. After receiving the operating load parameters, the digital twin automatically adjusts the load settings of the internal environment simulation to ensure that the load environment of the digital twin is consistent with the actual operating conditions of the actuator. This avoids misjudgments of effects caused by fixed operating condition extrapolation. The operating load parameters are extracted based on the fluctuations of current and vibration signals, eliminating the need for additional load sensors, reducing hardware costs, and enabling real-time updates of the operating load parameters. Inputting parameters according to the correspondence between components and loads ensures that the impact of the load on the components accurately matches the actual situation, avoiding deviations in component degradation model extrapolation caused by incorrect input.

[0049] Component degradation models are tools for simulating maintenance strategies. By extrapolating the changes in component status over time after maintenance, the advantages and disadvantages of different maintenance strategies can be compared. However, there are coupling relationships between actuator components; ignoring these couplings during simulation will lead to inaccurate degradation trajectories. A dedicated component degradation model is constructed for each core component. The motor degradation model uses aging coefficient and load loss as core parameters to simulate the degradation process of motor insulation aging and winding losses. The bearing degradation model uses wear amount and accumulated fatigue damage as core parameters to simulate the process of bearing rolling element wear and fatigue failure. The valve stem degradation model uses friction and corrosion degree as core parameters to simulate the process of valve stem friction wear and media corrosion. Coupling relationships are also embedded in the component degradation models. For example, when the load loss of the motor increases, a wear acceleration signal is simultaneously output to the bearing degradation model, causing the bearing wear amount to increase accordingly. When the bearing wear amount exceeds the wear threshold, a vibration interference signal is output to the motor degradation model, causing the motor aging coefficient to increase accordingly.

[0050] Load the parameter intervention rules and operating load parameters corresponding to each candidate maintenance strategy, and start the deduction process of the digital twin. First, initialize the degradation parameters of each component, and then perform deduction at fixed time intervals as deduction steps. Within each step, first calculate the basic degradation amount of each component based on the operating load parameters, and then calculate the degradation impact amount of related components based on the coupling relationship. The total degradation amount of that step is obtained by summing them up. Record the degradation parameter values ​​of each component at each step to form a degradation trajectory of time and degradation parameters. Repeat this process to complete the deduction of all candidate maintenance strategies and obtain the multi-component degradation trajectory corresponding to each candidate maintenance strategy. Based on the collaborative deduction of coupling relationships, the degradation trajectory can reflect the indirect impact of maintenance components on other components, avoiding the one-sidedness of independent deduction. The degradation trajectory is presented in the form of a curve of time and degradation parameters, which allows for intuitive comparison of the effect differences of different maintenance strategies. The deduction process facilitates the subsequent optimization of maintenance strategy details.

[0051] Specifically, the maintenance strategy selection includes: Based on the deduction results of the degradation trajectories of different mechanical components under candidate maintenance strategies, the evaluation dimensions of performance recovery time, total life cycle cost and disturbance degree of coupled components are determined. Based on the operating scenario and remaining life of the actuator, dynamic weights are assigned to different evaluation dimensions, and the deviation rate of the simulation results of different candidate maintenance strategies under the fluctuation of operating load is calculated. The weighted evaluation results of different candidate maintenance strategies are corrected based on the deviation rate. The revised weighted evaluation results are input into the target optimization algorithm to select a maintenance strategy based on the synergistic objectives of performance recovery time and total lifecycle cost.

[0052] The results of different candidate maintenance strategies need to be evaluated from multiple dimensions to comprehensively determine their merits. Three core dimensions need to be identified: performance recovery time, total lifecycle cost, and disturbance degree of coupled components. Candidate maintenance strategies should be evaluated from the perspectives of effectiveness, cost, and side effects to avoid short-sighted decisions caused by selecting a single dimension. The time required for the degradation parameters of the maintained component to recover to a safe threshold range should be extracted from the degradation trajectory. For example, the safe threshold for bearing wear is a certain standard value. If the degradation trajectory of a candidate maintenance strategy shows that the wear drops to this standard value within a certain period after maintenance, then this time is the performance recovery time. If there are multiple components, the maximum recovery time of the critical component should be taken to ensure that all critical components recover to a safe state. The maintenance action cost and the subsequent repair cost caused by degradation should be calculated based on the degradation trajectory. The maintenance action cost, as preset according to industry standards, includes the procurement cost of replacement components, labor costs, and the debugging cost of adjusting parameters. The subsequent degradation cost is calculated based on the time required from the maintenance to the next maintenance, predicted by the degradation trajectory, combined with the component replacement probability within this period. The sum of these two is the total lifecycle cost.

[0053] The analysis focuses on the degradation parameter changes of non-target components in the degradation trajectory. If the candidate maintenance strategy targets the bearing, and the non-target components are the motor and valve stem, the fluctuation range of the degradation parameters of the motor and valve stem within several steps after maintenance is calculated. The larger the fluctuation range, the higher the disturbance of the coupled components, indicating that the maintenance has a negative impact on the non-target components. The data for each dimension is validated. For example, if the performance recovery time is shorter than the time required for the maintenance action itself, the degradation trajectory is re-examined. If the total life cycle cost is lower than the cost of the basic maintenance action, the subsequent degradation cost estimation is revised to ensure that the dimensional data can truly reflect the candidate maintenance strategy. The three dimensions cover short-term effects, long-term costs, and system side effects, respectively, avoiding the one-sidedness of single-dimensional evaluation. The data for each dimension is extracted from the degradation trajectory to ensure the objectivity of the evaluation basis, reduce the interference of human factors on the selection of maintenance strategies, and provide a basis for adjusting the priority of maintenance strategies in different scenarios.

[0054] Different operating scenarios have different requirements for evaluation dimensions. If fixed weights are used, the maintenance strategy selection will be out of touch with the scenario requirements. At the same time, the operating load fluctuates, and the maintenance strategy will be biased under the fluctuation. The evaluation results are corrected by the deviation rate to avoid selecting the maintenance strategy that is only effective under ideal load. The operating scenarios are divided into continuous production scenarios and intermittent operation scenarios. The remaining lifespan is divided into emergency state, normal state and low risk state. Dynamic weights are assigned to different evaluation dimensions according to the operating scenario and remaining lifespan of the actuator. For example, in continuous production scenarios and emergency state, performance recovery time is assigned higher weight, while total life cycle cost and coupling component disturbance degree are assigned lower weight. In intermittent operation scenarios and low risk state, total life cycle cost is assigned higher weight, while performance recovery time and coupling component disturbance degree are assigned lower weight.

[0055] For each candidate maintenance strategy, the load under simulated conditions fluctuates within a reasonable range, and its degradation trajectory is re-deduced. The dimensional data after fluctuation is extracted, and the percentage difference between the fluctuated data and the original deduced data is calculated. This is the deviation rate of the candidate maintenance strategy. The larger the deviation rate, the more significantly the effect of the candidate maintenance strategy is affected by load fluctuations. The evaluation result after dynamic weighting is corrected based on the deviation rate. The higher the deviation rate, the lower the corrected evaluation value, ensuring that candidate maintenance strategies with high deviation rates are not prioritized due to their high original weighted evaluation values. Thus, the dynamic weights are bound to the operating scenario and remaining lifetime, making the maintenance strategy selection fit the actual needs and avoiding strategy mismatch caused by general weights. The deviation rate correction takes into account the impact of load fluctuations, ensuring that the selected maintenance strategy can maintain stable performance under non-ideal loads. The correction process is transparent and traceable, facilitating subsequent maintenance strategy adjustments and verification.

[0056] While the revised weighted evaluation results quantify the overall performance of the maintenance strategies, a multi-objective balance problem exists. An optimal balance between performance recovery time and total lifecycle cost needs to be found through an objective optimization algorithm, rather than simply selecting the optimal maintenance strategy for one dimension. Furthermore, the algorithm should be able to handle quantitative data from multiple maintenance strategies, avoiding subjective bias from manual selection and ensuring the objectivity and optimality of the selection results. A non-dominated sorting genetic algorithm is chosen as the objective optimization algorithm because it excels at handling multi-objective optimization problems and can generate Pareto optimal solution sets. This means it cannot improve one objective without decreasing another. Minimizing performance recovery time and minimizing total lifecycle cost are set as the two optimization objectives. The revised weighted evaluation results are used as the input fitness value for the non-dominated sorting genetic algorithm. Algorithm parameters are set to ensure efficient search for the optimal solution and avoid getting trapped in local optima.

[0057] After running, the non-dominated sorting genetic algorithm generates a Pareto optimal solution set. Based on the final decision preference of the executing agency, i.e., extracted from scenario requirements (e.g., recovery time is prioritized over cost in continuous production scenarios, while cost is prioritized over recovery time in warehousing scenarios), the optimal strategy is selected from the solution set. For example, in continuous production scenarios, the strategy with the shortest performance recovery time and cost within budget is selected; in warehousing scenarios, the strategy with the lowest total lifecycle cost and acceptable recovery time is selected. If multiple maintenance strategies satisfy the decision preference exist in the solution set, the maintenance strategy with the lowest disturbance to coupled components is selected. This ensures a balance among multiple objectives, avoiding the disadvantages caused by single-objective optimization. Combining decision preference selection ensures that the final maintenance strategy meets the core needs of the executing agency, preventing the algorithm output from deviating from actual needs, reducing subjective bias, and ensuring high processing efficiency, objectivity, and timeliness of the selection results.

[0058] S3. After implementing the maintenance strategy, continuously acquire the current signal and vibration signal of the actuator and generate the actual performance trajectory. Compare the trajectory of the remaining life with the actual performance trajectory to obtain the trajectory deviation. Based on the trajectory deviation, adaptively correct the parameters of the graph neural network and the parameters of the component degradation model.

[0059] Furthermore, generating actual performance trajectories includes: Before and after implementing the maintenance strategy, the current and vibration signals of the actuator are continuously acquired, and the current harmonic amplitude, vibration harmonic components and spatiotemporal correlation features are extracted simultaneously to form baseline features and real-time feature sequences. Based on the importance of mechanism components in the causal knowledge graph, the real-time feature sequence is split into feature subsequences corresponding to the mechanism components, and the weight of each feature subsequence is dynamically adjusted according to the sub-remaining lifetime of the mechanism components. The real-time feature sequence is processed based on the environmental interference signal of the actuator, and the weighted feature sub-sequences are integrated according to different mechanism components to generate the actual performance trajectory including instantaneous performance fluctuations and trend degradation.

[0060] The effectiveness of the maintenance strategy needs to be verified by comparing the state before and after maintenance. Simply obtaining the post-maintenance signal is insufficient to determine whether the performance change is due to maintenance or natural fluctuations. It is necessary to extract identical features to ensure consistent feature dimensions and avoid comparison failures caused by differences in feature types. Simultaneously, the pre-maintenance features are used as a baseline, and the post-maintenance features as a real-time sequence, providing a feature basis for subsequent generation of actual performance trajectories. Current and vibration signals of the actuator are continuously acquired before and after the implementation of the maintenance strategy, and the average values ​​of current harmonic amplitude, vibration harmonic components, and spatiotemporal correlation features are extracted simultaneously to form baseline features and a real-time feature sequence. The baseline features represent the stable mechanical state before maintenance, while the real-time feature sequence reflects the change in state over time after maintenance. This ensures the objectivity of the performance change comparison and avoids distortion caused by different feature dimensions or extraction methods. The baseline features reflect the stable state before maintenance, and the real-time sequence reflects the dynamic changes after maintenance, providing data support for the generation of actual performance trajectories for subsequent instantaneous fluctuations and trend degradation.

[0061] Real-time feature sequences provide a comprehensive representation of the actuator's state, while actual performance trajectories need to reflect component-level performance changes. This requires feature decomposition according to component nodes in the causal knowledge graph. Furthermore, different components have varying importance and sub-remaining lifetimes, necessitating dynamic weight adjustment to ensure the actual performance trajectory prioritizes the state changes of critical and high-risk components, avoiding the obscuring of crucial information by average weights. The real-time feature sequence is broken down into feature sub-sequences corresponding to the actuator's components, ensuring each component's sub-sequence only includes features directly related to its state. Then, based on the importance of the actuator components in the causal knowledge graph, basic weights are assigned to different feature sub-sequences. Simultaneously, the weights of each feature sub-sequence are dynamically adjusted in conjunction with the sub-remaining lifetime of the actuator components. If a component's sub-remaining lifetime indicates a high-risk state, the weight of its feature sub-sequence is further increased on top of the basic weight; if the sub-remaining lifetime is within a safe range, the basic weight remains unchanged. During weight adjustment, the component inference results are updated in real-time. This ensures that subsequent trajectories accurately reflect the performance changes of individual components, avoiding ambiguity in component states caused by overall features, and ensuring that the trajectory prioritizes the states of critical and high-risk components, facilitating subsequent deviation analysis to focus on core issues.

[0062] The real-time feature sequence includes environmental interference signals, such as current noise caused by power grid voltage fluctuations and vibration signal drift caused by ambient temperature changes. These interferences are not performance changes of the machinery itself. If left untreated, they will lead to trajectory distortion. It is necessary to integrate weighted feature subsequences by component to ensure that the actual performance trajectory can reflect the performance changes at the component level. At the same time, the actual performance trajectory must reflect both instantaneous fluctuations and trend degradation to comprehensively reflect the mechanical state and avoid misjudgment of the state due to a single dimension. The signal of the standby actuator operating without load in the same environment is obtained as a pure environmental interference signal, or the environmental background signal is obtained when the current actuator is stopped. The components that are highly similar to the environmental interference signals in the real-time feature sequence are identified by signal comparison. The component is removed from the real-time feature sequence by a filtering algorithm. During the processing, it is ensured that only environmental interference is removed and the feature fluctuations caused by changes in mechanical state are retained.

[0063] The processed weighted feature subsequences of each component are integrated along the time axis, with time as the horizontal axis and the component weighted feature subsequences as the vertical axis, generating an independent component performance sub-trajectory for each component. Based on this, the weights of each component are superimposed, with higher-importance component sub-trajectories having a higher proportion in the overall trajectory, forming the overall actual performance trajectory. During trajectory generation, short-term time windows capture instantaneous performance fluctuations, while long-term time windows capture trend degradation, ensuring the trajectory simultaneously covers both instantaneous and trend features. After processing environmental interference signals, the actual performance trajectory only reflects the performance changes of the machine itself, avoiding trajectory distortion caused by environmental factors and improving the accuracy of subsequent deviation comparisons. Integrating the component weighted feature subsequences along the time axis allows the trajectory to intuitively present instantaneous fluctuations and trend degradation, comprehensively reflecting the performance recovery and long-term degradation state after maintenance. Simultaneously generating component sub-trajectories and the overall trajectory facilitates subsequent deviation analysis to locate specific components while also reflecting the overall performance of the actuator, taking into account both local and overall state presentation.

[0064] Specifically, the parameters of the adaptive correction graph neural network include: Based on the causal strength weights of edges in the causal knowledge graph, the trajectory deviation is decomposed into component sub-deviations, and it is determined whether the component sub-deviations originate from state reasoning deviations or association reasoning deviations. If the component sub-bias originates from state reasoning bias, then the feature mapping coefficients of the corresponding component in the graph neural network are corrected; if the component sub-bias originates from association reasoning bias, then the aggregation weight coefficients between the corresponding components in the graph neural network are corrected. By combining the importance of the mechanism components and the degree of component sub-deviation, a correction magnitude coefficient is set to correct the parameters of the graph neural network, and a correction step size constraint is set based on the changing trend of the trajectory deviation after historical correction. After correcting the parameters, verify the deviation between the inference results and the actual performance trajectory. If the verification fails, backtrack the parameters of the graph neural network and adjust the correction magnitude coefficient.

[0065] Trajectory deviation represents overall performance differences. Without breaking it down into component sub-deviations, it's impossible to pinpoint which component's reasoning is flawed. The causal strength weights of edges in the causal knowledge graph reflect the tightness of the connections between components. We can prioritize breaking down closely related component sub-deviations based on these weights, avoiding blind splitting. Simultaneously, we need to determine whether the deviation originates from state reasoning deviation or association reasoning deviation in order to specifically correct the graph neural network parameters and avoid blind parameter correction. After calculating the trajectory deviation between the remaining lifetime change trajectory and the actual performance trajectory, we prioritize breaking down component pairs corresponding to edges with high causal strength weights based on the causal strength weights of the causal knowledge graph edges. We compare the performance data of each component pair in the two trajectories, calculate the deviation value of that component pair, and then break it down into individual components to obtain component sub-deviations. During the decomposition process, if the proportion of a component's sub-deviation to the trajectory deviation exceeds a preset value, it is marked as a major deviation component.

[0066] For major deviation components, the deviation type is determined by combining the reasoning process. If the sub-failure probability and sub-remaining lifetime of the component do not match the state in the actual performance trajectory, but the reasoning results of its related components are consistent (e.g., the motor reasoning deviation is large, but the bearing reasoning is consistent with the actual situation), it is determined to be a state reasoning deviation, which indicates that the state reasoning of only a single component is inaccurate. If the reasoning results of the component and its related components are both inconsistent with the actual situation, and the causal path of the two does not match the actual fault propagation path (e.g., the reasoning path is motor → bearing, but the actual path is bearing → motor), it is determined to be an association reasoning deviation, which indicates that the association reasoning between components is inaccurate. Based on the causal strength weight, the trajectory deviation is decomposed, prioritizing closely related components to reduce the blindness of decomposition, improve the efficiency of deviation location, clarify the deviation type, provide a basis for the targeted correction of the parameters of the subsequent graph neural network, avoid parameter chaos caused by indiscriminate correction, and mark the major deviation components so that subsequent correction can focus on the core issues and avoid wasting resources on minor deviation components.

[0067] In a graph neural network, feature mapping coefficients are responsible for mapping component state vectors to node attributes, directly affecting the state inference of a single component. If there is a state inference deviation, correcting this coefficient can directly improve the inference accuracy of a single component. Meanwhile, aggregation weight coefficients are responsible for aggregating the correlation information between components, directly affecting the correlation inference between components. If there is a correlation inference deviation, correcting this coefficient can improve the causal path inference between components. Different deviation types correspond to different graph neural network parameters. Mixing these parameters can lead to parameter dysfunction, requiring targeted matching and correction. For components originating from state inference deviations, the feature mapping coefficients of the component's nodes in the graph neural network are located, such as the mapping parameter from current harmonic amplitude to node attributes in a motor node. If the actual performance characteristics of the component are lower than the inference expectation, the feature mapping coefficients are adjusted to enhance the influence of actual characteristics on node attributes, such as increasing the mapping weight corresponding to the current harmonic amplitude. If the actual characteristics are higher than the inference expectation, the coefficients are adjusted in the opposite direction. During the correction process, the feature extraction logic of the component is referenced, such as the feature weights of a convolutional neural network, to ensure that the correction direction is consistent with the physical meaning of the features.

[0068] For component pairs with inference biases, the aggregation weight coefficients of the edges between these component pairs in the localization graph neural network are used. For example, the aggregation parameters of the association information between the motor and bearing edges are used. If the causal path of the inference does not match the reality (e.g., the inference states that the motor affects the bearing, but the actual result is that the bearing affects the motor), the direction of the aggregation weight coefficients is adjusted. If the association strength of the inference does not match the reality, the aggregation weight coefficients are increased. During the correction process, the causal strength weights of the edges in the causal knowledge graph are referenced to ensure that the aggregation weight coefficients match the actual degree of association. The type of bias corresponds to the correction parameters to ensure the targeted nature of the correction, avoid the disconnect between parameter correction and the root cause of the bias, and improve the correction efficiency. The correction process references the feature extraction logic and the edge weights of the causal knowledge graph to maintain the coherence of the technical solution and avoid parameter logic confusion caused by independent correction. The correction directly targets the parameters of the graph neural network without adjusting the entire network, reducing the impact of correction on network stability.

[0069] Different importance of mechanical components necessitates different correction magnitudes, and different degrees of component sub-deviation also require different correction magnitudes. Setting the correction magnitude solely based on importance or deviation degree can lead to overcorrection. Furthermore, it's necessary to consider historical correction trends when setting step size constraints to prevent excessively large single correction magnitudes from disrupting the stability of the graph neural network and ensuring a smooth correction process. Based on the importance of mechanical components in the causal knowledge graph, a base magnitude coefficient is assigned to different levels. Then, the base magnitude coefficient is adjusted according to the degree of component sub-deviation, i.e., the proportion of deviation to trajectory deviation. Higher deviation degrees result in an upward adjustment of the base coefficient, while lower deviation degrees result in a downward adjustment. For example, if the base magnitude coefficient for a first-level component like a motor is a certain value, and its deviation degree is high, the final correction magnitude coefficient will be the result of adjusting the base value upwards. Similarly, if the base magnitude coefficient for a second-level component like a valve stem is low, even with a high deviation degree, the final correction magnitude coefficient will not exceed the upper limit of the first-level component.

[0070] Record historical correction data, including the magnitude coefficients of the past three corrections and the changes in trajectory deviation after correction. If a certain type of parameter had a large correction magnitude in the last time, and the trajectory deviation first decreased and then rebounded after correction, it indicates overcorrection. In this case, set an upper limit for the correction step size constraint for this type of parameter. If the historical correction magnitude is moderate and the deviation continues to decrease, it indicates reasonable correction, and the correction step size constraint is kept loose. After setting the correction step size constraint, ensure that the parameter change in a single correction does not exceed the upper limit of the correction step size constraint to avoid parameter mutations that cause oscillations in graph neural network inference. Thus, the correction magnitude coefficient is set by combining the importance and deviation degree, taking into account the priority of core components and the matching of deviation size, avoiding correction imbalance caused by a single factor. The correction step size constraint is set based on the historical correction trend to avoid overcorrection in a single correction from damaging the stability of the graph neural network, ensuring smooth convergence of the correction process. The correction magnitude coefficient and correction step size constraint can be dynamically adjusted to adapt to the deviation correction needs in different scenarios and improve the flexibility of correction.

[0071] The validity of the corrected graph neural network parameters needs to be verified by comparing the deviation between the inference results and the actual performance trajectory. For example, does the sub-remaining lifetime of the corrected inference more closely match the actual trajectory? If applied directly without verification, erroneous parameters will continue to affect the inference. Parameter backtracking needs to be set up. If the verification fails, the parameters should be restored to those before the correction to avoid error accumulation. At the same time, the correction magnitude coefficient needs to be adjusted to provide optimization direction for the next correction. The corrected graph neural network parameters are then applied to the current causal knowledge graph, and component inference and causal path inference are re-performed. Based on the new inference results, a new trajectory of remaining lifetime changes is generated, and the trajectory deviation is recalculated with the actual performance trajectory. If the new trajectory deviation is lower than before the correction, or falls within a preset threshold, the verification is considered successful, and the corrected parameters are retained. If the new trajectory deviation does not decrease, or even increases, the verification is considered unsuccessful.

[0072] When validation fails, the parameters of the graph neural network are backtracked to their state before correction. The reasons for the failure are analyzed. If the increased deviation is due to the correction magnitude coefficient being too large, for example, the parameter change exceeding the correction step size constraint, the correction magnitude coefficient for that type of parameter will be lowered for the next iteration. If the deviation is due to the correction magnitude coefficient being too small, i.e., the deviation does not change significantly, the correction magnitude coefficient will be appropriately increased. The new correction magnitude coefficient will be recorded for the next correction. Validation ensures that the corrected parameters are effective, avoids erroneous parameters from continuously affecting inference, and reduces system risk. Backtracking avoids error accumulation and ensures that the parameters of the graph neural network do not fall into an abnormal state due to a single correction failure. Based on the validation results, the correction magnitude coefficient is adjusted to improve the success rate of subsequent corrections, forming a closed-loop optimization.

[0073] Specifically, the parameters of the adaptive correction component degradation model include: Based on the parameter intervention rules of the component degradation model in the digital twin, the trajectory deviation is decomposed into the self-degradation deviation of a single mechanism component and the coupling degradation deviation between multiple mechanism components. If the deviation is self-degradation, then the basic degradation parameters of the corresponding mechanism component are corrected in a targeted manner; if the deviation is coupling degradation, then the coupling influence coefficient between related components is corrected in a targeted manner. By combining the degradation sensitivity and deviation duration of the mechanism components, a correction magnitude factor is set to correct the parameters of the component degradation model, and the correction results of the parameters of the graph neural network are synchronously associated. After correcting the parameters, the degenerate trajectory is re-enacted using a digital twin and the trajectory deviation is re-verified. If the verification fails, the correction magnitude factor is adjusted and the verification is repeated.

[0074] The parameters of the component degradation model include intrinsic degradation parameters and coupled degradation parameters. Trajectory deviations originate from inaccurate intrinsic degradation parameters and coupled degradation parameters. It is necessary to reverse locate the parameter type corresponding to the deviation based on parameter intervention rules to avoid blindly decomposing the deviation. Extract the parameter intervention rules mapped by the candidate maintenance strategy. For example, if the rule mapped by a certain maintenance strategy is to adjust the motor aging rate coefficient and correct the coupling influence coefficient between the bearing and the motor, then the trajectory deviation should be associated with the component state corresponding to these two types of parameters first. If the maintenance strategy does not involve a certain type of parameter, the deviation should not be attributed to that type of parameter for the time being to reduce invalid decomposition.

[0075] By comparing the projected degradation trajectory with the actual performance trajectory, if the projected trajectory of a single component deviates significantly from the actual trajectory, while the deviation of its associated components is small, it is determined to be a self-degradation deviation. If the projected deviation of a certain component is large, and the deviations of its associated components are also large, and the trends of the deviation changes are consistent, it is determined to be a coupled degradation deviation, that is, the coupling influence coefficient between the two is inaccurate. During the decomposition process, the parameter correlation of the component degradation model in the digital twin is referenced. For example, the self-parameter only affects a single component, while the coupled parameters affect multiple components, ensuring that the decomposition type matches the parameter function. Based on the parameter intervention rules, the deviation is decomposed in reverse, prioritizing the parameters involved in the maintenance strategy to reduce the decomposition range, improve the efficiency of deviation location, and clarify the deviation type to provide a basis for the targeted correction of subsequent degradation model parameters, avoiding the disconnect between parameter correction and the root cause of deviation. The parameter correlation of the degradation model is referenced to ensure that the deviation decomposition is consistent with the model design and to avoid logical contradictions caused by independent decomposition.

[0076] Basic degradation parameters are the core parameters of a component's own degradation law, such as the aging coefficient of a motor and the wear amount of a bearing. They directly determine the degradation trajectory of a single component. If it is a self-degradation deviation, correcting this parameter can directly match the actual degradation law. The coupling influence coefficient is the core parameter of the mutual influence between components, which directly determines the collaborative degradation trajectory of multiple components. For example, a large coupling coefficient means a strong acceleration effect of the motor on the bearing. If it is a coupling degradation deviation, correcting this parameter can match the actual coupling relationship. Different deviation types correspond to different parameters. Mixed corrections will cause the component state and coupling relationship in the component degradation model to be distorted. For components with self-degradation deviations, locate their basic degradation parameters. For example, the self-degradation deviation of a motor comes from the aging coefficient. If the degradation speed of the motor predicted by the component degradation model is faster than the actual speed, then the aging coefficient is lowered; if the predicted speed is slower than the actual speed, then the aging coefficient is raised. The self-degradation deviation of a bearing comes from the wear amount. If the predicted wear amount is greater than the actual wear amount, then the wear amount is lowered. During the correction process, refer to the initial parameter settings of the component degradation model to ensure that the corrected parameters are within a reasonable physical range.

[0077] For component pairs with coupling degradation deviations, their coupling influence coefficients are located. For example, the coupling degradation deviation between a motor and a bearing originates from the acceleration coefficient of bearing wear caused by the motor load. If the acceleration effect of increased motor load on bearing wear in the component degradation model is weaker than the actual effect, the coupling influence coefficient is increased; if the acceleration effect is stronger than the actual effect, the coupling influence coefficient is decreased. During the correction process, the degradation trajectory derived from the component coupling relationship is referenced to ensure that the corrected coupling influence coefficient matches the degree of mutual influence between actual components. This ensures that the deviation type corresponds one-to-one with the correction parameter, ensuring the targeted nature of the correction, avoiding the disconnect between parameter correction and the root cause of the deviation, and quickly improving the accuracy of the degradation trajectory. The correction process references the initial parameter settings and coupling derivation logic to maintain the consistency of the technical solution, avoid the confusion of component degradation model logic caused by independent correction, eliminate the need for overall model structure adjustment, reduce the impact of correction on model stability, and reduce system complexity.

[0078] Different degradation sensitivities of mechanical components necessitate differentiated correction amplitude factors. Components with high degradation sensitivity should have smaller correction amplitude factors to avoid trajectory oscillations. Different deviation durations also require different correction amplitudes; longer-lasting deviations indicate larger parameter deviations, requiring larger amplitudes to avoid long-term deviations. Setting amplitudes based on a single factor can lead to over-correction of sensitive components or under-correction of long-term deviations. Furthermore, since both the component degradation model and the graph neural network infer the same component state, the correction results of the graph neural network must be synchronously correlated to avoid contradictory judgments on the same component state. The degradation sensitivity of mechanical components is preset based on historical fault data, and a base amplitude factor is assigned. This base amplitude factor is then adjusted according to the duration of the deviation, increasing for longer durations and decreasing for shorter durations. For example, the base amplitude factor for a high-sensitivity component is a relatively small value; if the deviation lasts a long time, it is slightly increased from the base value. Conversely, a low-sensitivity component has a large base amplitude factor, allowing the correction amplitude factor to remain high even with short-lasting deviations. By extracting the correction results of the graph neural network parameters, the aging coefficient of the motor is simultaneously increased when correcting the basic degradation parameters of the motor in the component degradation model. If the aggregate weight coefficient of the motor and bearing in the graph neural network is increased, the coupling influence coefficient of the motor and bearing in the degradation model is also simultaneously increased. During the collaborative process, it is ensured that the two models have the same trend of change in the state of the same component. Thus, the amplitude factor is set by combining the two factors of degradation sensitivity and deviation duration, taking into account both the smooth correction of sensitive components and the rapid correction of long-term deviations, avoiding the imbalance of correction caused by a single factor. It is coordinated with the parameter correction results of the graph neural network to ensure that the judgment trends of the two core models on the state of the same component are consistent, avoiding logical contradictions between models, improving the consistency of the overall scheme, and dynamically adapting the correction amplitude factor to the component characteristics and deviation duration, improving the flexibility and accuracy of correction, and reducing the probability of subsequent verification failure.

[0079] The validity of the corrected degradation model parameters needs to be verified by re-enacting the degradation trajectory using a digital twin. Failure to do so will result in erroneous parameters continuously affecting the deduction, necessitating repeated verification. If the initial verification fails, the correction magnitude factor should be adjusted before re-verification to avoid failure due to a single inappropriate magnitude. Furthermore, digital twin re-enactment can reproduce the corrected degradation process, facilitating analysis of why the deviation has not been reduced and providing a basis for subsequent adjustments. The parameters of the corrected component degradation model are input into the digital twin, and the same operating load parameters as the actual conditions are loaded. The maintained degradation trajectory is then re-enacted. The deviation between the re-enacted degradation trajectory and the actual performance trajectory is compared. If the deviation decreases, the verification is considered successful, and the corrected parameters are retained. If the deviation does not decrease, the verification is considered unsuccessful.

[0080] When verification fails, the reasons for the failure to reduce the deviation are analyzed. If the deviation increases due to an excessively large correction factor, the correction factor for that component will be lowered in the next iteration. If the correction factor is too small, it will be appropriately increased. After adjustment, the parameters of the component degradation model are re-corrected, and the trajectory is re-enacted and verified using a digital twin until the deviation meets the standard or the preset number of repetitions is reached. This digital twin re-enactment verification ensures that the corrected parameters accurately reflect the actual degradation pattern, preventing the application of incorrect parameters to subsequent deductions, improving the reliability of the component degradation model. Repeated verification and amplitude adjustment prevent correction failure due to improper amplitude in a single instance, increasing the success rate of correction and forming a closed-loop optimization. The re-enactment process can reproduce degradation details, facilitating the location of the root cause of the deviation and providing a more detailed basis for subsequent optimization. The parameters of the verified component degradation model and the parameters of the graph neural network can be used for subsequent maintenance strategy optimization of the actuator, while providing a foundation for continuous adaptive optimization of IoT monitoring methods, ensuring model accuracy in long-term applications.

[0081] Example 2 This application provides an Internet of Things (IoT) monitoring system, which includes a data acquisition module, a fault diagnosis module, and a maintenance execution module.

[0082] The data acquisition module is used to acquire valve position signals, current signals and vibration signals during the operation of the actuator. After synchronizing the signals, the current harmonic amplitude of the current signal and the vibration harmonic components of the vibration signal are extracted through a convolutional neural network during the valve switching process. The data is then fused with the valve position change rate as a weight to output a health feature vector that characterizes the mechanical state. The fault diagnosis module receives health feature vectors and embeds them into a causal knowledge graph between pre-set mechanism components. It then performs reasoning on the causal knowledge graph based on a graph neural network, outputting the probability of fault occurrence, remaining lifespan, and causal path. In response to the reasoning results, it generates multiple candidate maintenance strategies, simulates and infers different candidate maintenance strategies in a digital twin of the integrated component degradation model, and selects a maintenance strategy based on the inference results using a target optimization algorithm. The maintenance execution module is used to continuously acquire the current and vibration signals of the actuator after executing the maintenance strategy and generate the actual performance trajectory. The trajectory deviation is obtained by comparing the change trajectory of the remaining life with the actual performance trajectory. Based on the trajectory deviation, the parameters of the graph neural network and the parameters of the component degradation model are adaptively corrected.

[0083] Since the principle of the system in this application embodiment is similar to the method described above in this application embodiment, the implementation of the system is the same as the implementation of the method, and the repeated parts will not be described again.

Claims

1. An Internet of Things (IoT) monitoring method, characterized in that, include: The system acquires valve position signals, current signals, and vibration signals during the operation of the actuator. After synchronizing the signals, it extracts the current harmonic amplitude of the current signal and the vibration harmonic components of the vibration signal during the valve switching process through a convolutional neural network. The system then fuses these components with the valve position change rate as a weight and outputs a health feature vector that characterizes the mechanical state. It receives health feature vectors and embeds them into a causal knowledge graph between pre-set mechanism components. It then uses a graph neural network to reason about the causal knowledge graph and outputs the failure probability, remaining lifespan, and causal path. The system generates multiple candidate maintenance strategies based on the inference results, simulates and infers different candidate maintenance strategies in a digital twin of the integrated component degradation model, and selects the maintenance strategy based on the inference results using an objective optimization algorithm. After implementing the maintenance strategy, the current and vibration signals of the actuator are continuously acquired and the actual performance trajectory is generated. The trajectory deviation is obtained by comparing the trajectory of the remaining life with the actual performance trajectory. The parameters of the graph neural network and the parameters of the component degradation model are adaptively corrected based on the trajectory deviation.

2. The IoT monitoring method as described in claim 1, characterized in that, The output health feature vector includes: The valve position signal, current signal and vibration signal are acquired during the operation of the actuator. After synchronizing the signal, the execution stages of the valve switching process are divided based on the change characteristics of the valve opening. In each execution phase, the current harmonic amplitude of the current signal and the vibration harmonic component of the vibration signal are extracted by a convolutional neural network to determine the spatiotemporal correlation characteristics of the current signal and the vibration signal within the same execution phase. Using the valve position change rate as a weight, the current harmonic amplitude, vibration harmonic components, and spatiotemporal correlation characteristics of different execution stages are weighted and fused, and then mapped through a dual-channel residual network to output a health feature vector characterizing the mechanical state.

3. The IoT monitoring method as described in claim 2, characterized in that, The reasoning based on the causal knowledge graph includes: The health feature vector is decomposed into component state vectors corresponding to the causal knowledge graph nodes and embedded with node attributes. The causal strength weights of the edges of the causal knowledge graph are adjusted based on the temporal changes of the health feature vector. By using graph neural networks to perform component reasoning on the node attributes of the causal knowledge graph, the sub-failure probability and sub-remaining lifetime of different mechanism components are output. The reasoning results between mechanism components are aggregated based on the edge of the causal knowledge graph. The coupling influence coefficient of the related components is considered in combination with the causal strength weight, and the failure probability and remaining life of the actuator are output. During the reasoning process, edges in the causal knowledge graph that satisfy the preset causal confidence threshold are activated to form a directed causal path from the fault source node to the end influencing node, and the causal path is selected based on the product of the causal strength weights of each edge on the directed causal path.

4. The IoT monitoring method as described in claim 3, characterized in that, The selected maintenance strategy includes: Based on the deduction results of the degradation trajectories of different mechanical components under candidate maintenance strategies, the evaluation dimensions of performance recovery time, total life cycle cost and disturbance degree of coupled components are determined. Based on the operating scenario and remaining life of the actuator, dynamic weights are assigned to different evaluation dimensions, and the deviation rate of the simulation results of different candidate maintenance strategies under the fluctuation of operating load is calculated. The weighted evaluation results of different candidate maintenance strategies are corrected based on the deviation rate. The revised weighted evaluation results are input into the target optimization algorithm to select a maintenance strategy based on the synergistic objectives of performance recovery time and total lifecycle cost.

5. The IoT monitoring method as described in claim 4, characterized in that, The simulation of different candidate maintenance strategies includes: In response to the inference results, multiple candidate maintenance strategies are generated, and each candidate maintenance strategy is mapped to the parameter intervention rule of the component degradation model; The operating load parameters of the actuator are determined based on the fluctuation of current and vibration signals, and then the operating load parameters are input into the digital twin. Initiate a component degradation model in a digital twin and extrapolate the degradation trajectory of different components under different candidate maintenance strategies based on the coupling relationship between the components.

6. The IoT monitoring method as described in claim 5, characterized in that, The parameters of the adaptive correction graph neural network include: Based on the causal strength weights of edges in the causal knowledge graph, the trajectory deviation is decomposed into component sub-deviations, and it is determined whether the component sub-deviations originate from state reasoning deviations or association reasoning deviations. If the component sub-bias originates from state reasoning bias, then the feature mapping coefficients of the corresponding component in the graph neural network are corrected; if the component sub-bias originates from association reasoning bias, then the aggregation weight coefficients between the corresponding components in the graph neural network are corrected. By combining the importance of the mechanism components and the degree of component sub-deviation, a correction magnitude coefficient is set to correct the parameters of the graph neural network, and a correction step size constraint is set based on the changing trend of the trajectory deviation after historical correction. After correcting the parameters, verify the deviation between the inference results and the actual performance trajectory. If the verification fails, backtrack the parameters of the graph neural network and adjust the correction magnitude coefficient.

7. The IoT monitoring method as described in claim 6, characterized in that, The generation of the actual performance trajectory includes: Before and after implementing the maintenance strategy, the current and vibration signals of the actuator are continuously acquired, and the current harmonic amplitude, vibration harmonic components and spatiotemporal correlation features are extracted simultaneously to form baseline features and real-time feature sequences. Based on the importance of mechanism components in the causal knowledge graph, the real-time feature sequence is split into feature subsequences corresponding to the mechanism components, and the weight of each feature subsequence is dynamically adjusted according to the sub-remaining lifetime of the mechanism components. The real-time feature sequence is processed based on the environmental interference signal of the actuator, and the weighted feature sub-sequences are integrated according to different mechanism components to generate the actual performance trajectory including instantaneous performance fluctuations and trend degradation.

8. The IoT monitoring method as described in claim 7, characterized in that, The parameters of the adaptive correction component degradation model include: Based on the parameter intervention rules of the component degradation model in the digital twin, the trajectory deviation is decomposed into the self-degradation deviation of a single mechanism component and the coupling degradation deviation between multiple mechanism components. If the deviation is self-degradation, then the basic degradation parameters of the corresponding mechanism component are corrected in a targeted manner; if the deviation is coupling degradation, then the coupling influence coefficient between related components is corrected in a targeted manner. By combining the degradation sensitivity and deviation duration of the mechanism components, a correction magnitude factor is set to correct the parameters of the component degradation model, and the correction results of the parameters of the graph neural network are synchronously associated. After correcting the parameters, the degenerate trajectory is re-enacted using a digital twin and the trajectory deviation is re-verified. If the verification fails, the correction magnitude factor is adjusted and the verification is repeated.

9. An Internet of Things (IoT) monitoring system for implementing the IoT monitoring method according to any one of claims 1-8, characterized in that, include: Data acquisition module, fault diagnosis module, and maintenance execution module; The data acquisition module is used to acquire valve position signals, current signals and vibration signals during the operation of the actuator. After synchronizing the signals, the current harmonic amplitude of the current signal and the vibration harmonic components of the vibration signal are extracted through a convolutional neural network during the valve switching process. The data is then fused with the valve position change rate as a weight to output a health feature vector. The fault diagnosis module receives health feature vectors and embeds them into a causal knowledge graph between pre-set mechanism components. It then performs reasoning on the causal knowledge graph based on a graph neural network, outputting the probability of fault occurrence, remaining lifespan, and causal path. In response to the reasoning results, it generates multiple candidate maintenance strategies, simulates and infers different candidate maintenance strategies in a digital twin of an integrated component degradation model, and selects a maintenance strategy based on the inference results using a target optimization algorithm. The maintenance execution module is used to continuously acquire the current and vibration signals of the actuator after executing the maintenance strategy and generate the actual performance trajectory. It compares the trajectory of the remaining lifespan change with the actual performance trajectory to obtain the trajectory deviation, and adaptively corrects the parameters of the graph neural network and the parameters of the component degradation model based on the trajectory deviation.

Images