Equipment full-life-cycle predictive maintenance method and system, equipment and medium

Abstract

The invention relates to an equipment full-life-cycle predictive maintenance method and system, equipment and a medium. The method comprises the steps that multi-device multi-source heterogeneous data is acquired and processed to obtain a multi-modal feature matrix and a preliminary device relation graph; equipment causal association is learned by combining causal inference and a graph neural network with a model, and a heterogeneous equipment causal graph is constructed; detecting abnormity and positioning root causes based on the time sequence diagram neural network; and optimizing a maintenance strategy through anti-factual reasoning in the digital twin, and generating an optimal maintenance sequence and report. By adopting the method, the data utilization rate and the association identification precision can be improved, abnormal accurate detection and root cause positioning are realized, a reliable basis is provided for operation and maintenance, the maintenance efficiency is improved, and the fault risk is reduced.

Description

Technical Field

[0001] This invention belongs to the field of Internet of Things (IoT) technology, and in particular relates to predictive maintenance methods, systems, devices, and media for the entire lifecycle of equipment. Background Technology

[0002] With the rapid development of IoT technology, the network of devices in urban infrastructure is becoming increasingly complex, including a variety of heterogeneous devices such as streetlights, fire hydrants, traffic lights, safety helmets, and handcarts. These devices are tightly coupled in terms of physical space and function. For example, a short circuit in a streetlight circuit may affect the water pressure sensor of a fire hydrant through the power supply network, or trigger a traffic light malfunction through environmental factors (such as a fire), leading to cascading failures. Existing predictive maintenance methods are mainly based on time series analysis, machine learning, or digital twin technology, which predict faults by monitoring device sensor data (such as current, water pressure, and flow rate).

[0003] However, most of these methods rely on correlation analysis, such as using recurrent neural networks (RNNs) or long short-term memory networks (LSTMs) to detect anomalous patterns, but they cannot distinguish between causal relationships and spurious correlations. Multi-source heterogeneous data fusion typically employs feature-level concatenation, such as merging sensor data from different devices into a feature vector for training a classification model, but this ignores the causal structure between devices, leading to a lack of interpretability in predictions and a high false alarm rate. Furthermore, while digital twin technology can simulate fault evolution, it lacks counterfactual reasoning and cannot quantify the causal impact of maintenance actions, such as "how much would the failure probability of fire equipment B decrease if street light A were maintained earlier?" Therefore, there is an urgent need for a method that can model causal fault propagation across devices and optimize counterfactual maintenance strategies to improve the reliability and maintenance efficiency throughout the equipment's lifecycle. Summary of the Invention

[0004] Therefore, it is necessary to provide predictive maintenance methods, systems, equipment, and media for the entire equipment lifecycle to address the aforementioned technical issues.

[0005] Firstly, this application provides a predictive maintenance method for the entire lifecycle of equipment, including:

[0006] S1. Acquire multi-source heterogeneous data from multiple devices, and perform feature processing on the multi-source heterogeneous data to obtain a multimodal feature matrix and a preliminary device relationship diagram; the multi-source heterogeneous data includes device historical data, usage status, location information, and maintenance records;

[0007] S2. Based on the multimodal feature matrix and the preliminary device relationship graph, the direction and intensity of causal influence between devices are learned by combining a pre-set causal inference and graph neural network model to obtain a heterogeneous device causal graph. The nodes of the heterogeneous device causal graph represent different types of devices, the edges represent the direction of causal influence between devices, and the edge weights represent the intensity of causal influence.

[0008] S3. Based on the heterogeneous device cause-effect graph, detect abnormal device states through a preset time-series neural network model, and perform cause-effect propagation backtracking based on the heterogeneous device cause-effect graph to locate the root cause fault device, thereby obtaining a list of root cause fault devices and the current abnormal state vector.

[0009] S4. Based on the cause-effect graph of heterogeneous equipment, the list of root cause failure equipment and historical equipment data, counterfactual reasoning is performed in the digital twin using a preset structural cause-effect model to simulate virtual maintenance operations, optimize maintenance strategies, and generate the optimal maintenance sequence and counterfactual impact report.

[0010] In one embodiment, based on a multimodal feature matrix and a preliminary device relationship graph, a heterogeneous device causal graph is obtained by learning the direction and intensity of causal influence between devices through a pre-defined causal inference and graph neural network combined model, including:

[0011] S11. Based on the multimodal feature matrix and the preliminary device relationship graph, node embedding learning is performed through the message passing mechanism of the graph neural network to generate the hidden embedding vector of each device. The message passing mechanism updates the node embedding through multi-layer graph convolution operations, and each layer uses a learnable weight matrix and activation function to process the neighbor node information.

[0012] S12. Based on the hidden embedding vector, the causal score of each pair of device nodes is calculated through the attention mechanism, and the causal direction is verified by combining the conditional independence test to generate a candidate causal edge set; the conditional independence test uses the partial correlation coefficient to test the independence of device nodes under a given confounding variable.

[0013] S13. Based on cross-device intervention experiment data, estimate the average causal effect of each edge in the candidate causal edge set through Bayesian linear regression, calibrate the causal influence strength, and generate a causal weight matrix; the cross-device intervention experiment includes applying random intervention to key equipment and recording the responses of other equipment;

[0014] S14. Based on the candidate causal edge set and the causal weight matrix, construct a directed weighted heterogeneous device causal graph; the node set of the heterogeneous device causal graph corresponds to all devices, the edge set represents the causal dependency relationship, and the weight matrix represents the causal strength.

[0015] In one embodiment, based on the hidden embedding vector, a causal score for each pair of device nodes is calculated using an attention mechanism, and the causal direction is verified by combining a conditional independence test to generate a candidate causal edge set, including:

[0016] S21. Based on the hidden embedding vector, calculate the initial causal score for each pair of device nodes using a learnable diagonal matrix; the formula for calculating the initial causal score is:

[0017]

[0018] in, Represents device node For device nodes The initial causal score, and Representing device nodes and The hidden embedding vector, Represents a learnable diagonal matrix;

[0019] S22. Using environmental data from the multimodal feature matrix as a confounding variable, the conditional independence between device nodes is tested using the partial correlation coefficient; the formula for calculating the partial correlation coefficient is:

[0020]

[0021] in, Represents device node and Given a set of confounding variables The partial correlation coefficient under the condition, Represents device node and Pearson correlation coefficient, and Representing device nodes and With the set of confusing variables The correlation coefficient;

[0022] S23. For each pair of device nodes, when the initial causal score is higher than the preset threshold and the absolute value of the partial correlation coefficient is less than the independence threshold, the corresponding causal edge is retained to obtain the result of retaining the causal edge.

[0023] S24. Based on the results of retaining causal edges, generate a set of candidate causal edges.

[0024] In one embodiment, based on the heterogeneous device causality graph, a preset time-series neural network model is used to detect abnormal device states. Then, based on the heterogeneous device causality graph, causal propagation backtracking is performed to locate the root cause fault device, resulting in a list of root cause fault devices and a current abnormal state vector, including:

[0025] S31. Based on real-time streaming data, extract the real-time feature vector of each device and calculate the device state embedding through a temporal graph neural network model. The temporal graph neural network combines a long short-term memory network and a heterogeneous device causal graph to generate state embedding vectors through temporal updates and graph convolution.

[0026] S32. Based on the state embedding vector, reconstruct the predicted embedding through an autoencoder, and calculate the reconstruction error as the anomaly score; the formula for calculating the anomaly score is:

[0027]

[0028] in, Indicates equipment In time Abnormal scores, Indicates equipment In time The state embedding vector, This represents the predicted embedding vector reconstructed by the autoencoder;

[0029] S33. When the abnormal score exceeds the dynamic threshold set based on historical data, the device is marked as abnormal and a set of abnormal devices is generated.

[0030] S34. Based on the cause-effect graph of heterogeneous devices, starting from the set of abnormal devices, calculate the causal influence score of each device through the causal propagation backtracking algorithm to locate the root cause fault device; the formula for calculating the causal influence score is:

[0031]

[0032] in, Represents device node The causal influence score Indicates pointing to the device node The neighborhood group, Indicates from device node arrive Causal weights Represents device node The degree of departure;

[0033] S35. Based on the causal influence score, select the device with the highest score as the root cause failure device, and generate a list of root cause failure devices and a current abnormal state vector.

[0034] In one embodiment, based on a heterogeneous device causality graph, a list of root cause failure devices, and historical device data, counterfactual reasoning is performed in a digital twin using a pre-defined structural causal model to simulate virtual maintenance operations, optimize maintenance strategies, and generate an optimal maintenance sequence and counterfactual impact report, including:

[0035] S41. Based on the cause-effect graph of heterogeneous devices, construct a structural cause-effect model; the structural cause-effect model includes the state variables, structural equations and external noise variables of each device, and the structural equations are learned from historical data;

[0036] S42. Based on the current abnormal state vector, simulate virtual maintenance operations using the Monte Carlo sampling method; the virtual maintenance operations intervene in the equipment state variables to normal values ​​and predict the counterfactual state distribution of downstream equipment based on a structural causal model;

[0037] S43. For each candidate maintenance sequence, calculate the expected cost using counterfactual reasoning; the formula for calculating the expected cost is:

[0038]

[0039] in, Represents the vector for performing maintenance actions. The expected cost afterward This indicates the maintenance action vector. Indicates maintenance equipment binary actions, Indicates maintenance equipment The cost, Indicates equipment Losses due to malfunction Indicating intervention and observation Lower device The probability of failure;

[0040] S44. Optimize the maintenance sequence through reinforcement learning algorithm, and select the sequence that minimizes the expected cost as the optimal maintenance sequence;

[0041] S45. Based on the counterfactual prediction results of the optimal maintenance sequence, generate a counterfactual impact report; the counterfactual impact report includes the amount by which the maintenance action reduces the probability of equipment failure.

[0042] Secondly, this application also provides a predictive maintenance system for the entire lifecycle of equipment, comprising:

[0043] The data acquisition and feature processing module is used to acquire multi-source heterogeneous data from multiple devices and perform feature processing on the multi-source heterogeneous data to obtain a multimodal feature matrix and a preliminary device relationship diagram; the multi-source heterogeneous data includes device historical data, usage status, location information and maintenance records;

[0044] The device causal graph construction module is used to learn the direction and intensity of causal influence between devices based on the multimodal feature matrix and the preliminary device relationship graph, through a pre-set causal inference and graph neural network combined model, to obtain a heterogeneous device causal graph. The nodes of the heterogeneous device causal graph represent different types of devices, the edges represent the direction of causal influence between devices, and the edge weights represent the intensity of causal influence.

[0045] The anomaly detection and root cause localization module is used to detect abnormal device states based on the heterogeneous device causal graph and a preset time-series neural network model, and to locate the root cause fault device by causal propagation back based on the heterogeneous device causal graph, thereby obtaining a list of root cause fault devices and the current abnormal state vector.

[0046] The maintenance strategy optimization module is used to perform counterfactual reasoning in a digital twin based on the cause-effect graph of heterogeneous equipment, the list of root cause failure equipment, and historical equipment data, through a preset structural cause-effect model, to simulate virtual maintenance operations, optimize maintenance strategies, and generate optimal maintenance sequences and counterfactual impact reports.

[0047] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect.

[0048] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect.

[0049] The aforementioned predictive maintenance methods, systems, equipment, and media throughout the entire equipment lifecycle overcome the challenges of fragmented and inefficient multi-source heterogeneous data by constructing a comprehensive data acquisition network and optimizing feature processing, providing a high-quality analytical foundation. Utilizing a fusion model of causal inference and graph neural networks, it accurately uncovers and quantifies the strength of causal relationships between devices, eliminating spurious correlations. Combined with temporal graph neural networks, it achieves accurate anomaly identification, quickly pinpointing root cause faulty devices through causal propagation backtracking, reducing false positives and false negatives. Based on digital twins and structural causal models, it conducts counterfactual reasoning to simulate maintenance effects and quantify costs. Through reinforcement learning optimization, it obtains a scientific maintenance sequence, which, along with counterfactual impact reports, provides decision support, effectively improving maintenance efficiency, reducing failure risks, and lowering overall maintenance costs. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying 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.

[0051] Figure 1 This is a flowchart illustrating a predictive maintenance method for the entire lifecycle of equipment in one embodiment;

[0052] Figure 2 This is a schematic diagram of the structure of a predictive maintenance system for the entire lifecycle of equipment in one embodiment. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0054] refer to Figure 1 The application presents a flowchart illustrating the predictive maintenance method for the entire lifecycle of equipment provided in this application, which includes the following steps:

[0055] S1. Acquire multi-source heterogeneous data from multiple devices, and perform feature processing on the multi-source heterogeneous data to obtain a multimodal feature matrix and a preliminary device relationship diagram.

[0056] The equipment includes, but is not limited to, streetlights, fire hydrants, traffic lights, safety helmets, and handcarts; the multi-source heterogeneous data includes historical data, usage status, location information, and maintenance records of the equipment. The core of acquiring multi-source heterogeneous data is to build a comprehensive data acquisition network covering all the aforementioned equipment, ensuring the integrity and temporal continuity of the data dimensions. Specifically, differentiated data acquisition schemes are required based on the physical characteristics and functional requirements of different equipment: streetlights are equipped with integrated sensing modules, including current sensors, voltage sensors, light sensors, and positioning modules, which can collect real-time data on the streetlight's operating current, supply voltage, ambient light intensity, and latitude and longitude of its location. The sampling frequency is set according to the equipment's operating characteristics and data requirements; fire hydrants are equipped with water pressure sensors, flow sensors, and tilt sensors, with the tilt sensor used to monitor the equipment's tilting state and simultaneously acquiring related water supply pressure data through the urban water supply network system, with the sampling frequency adapted to the frequency of changes in parameters such as water pressure; traffic lights use status sensors to capture real-time data. The operation status of traffic lights (red, yellow, and green) is typically distinguished by digital signals, indicating whether they are on or off. This is correlated with traffic flow detector data, and the sampling frequency is synchronized with the traffic light signal cycle. The safety helmet incorporates a triaxial accelerometer, heart rate sensor, and positioning module. The triaxial accelerometer collects the wearer's movement data, the heart rate sensor collects physiological data, and the positioning module determines the work area location. The sampling frequency meets the needs of personnel movement monitoring. The handcart is equipped with a weight sensor, wheel axle speed sensors, and an identification module. The weight sensor acquires the load weight, the wheel axle speed sensors calculate the movement speed, and the identification module records ownership information. The sampling frequency matches the handcart's usage frequency. Historical equipment data covers long-term operation records, and maintenance records are stored using blockchain technology. These records include structured information such as maintenance time, maintenance personnel, fault type, and replaced parts. The immutability of blockchain ensures the authenticity and traceability of these maintenance records.

[0057] Feature processing of multi-source heterogeneous data requires addressing data format differences and quality issues, with the process beginning with data cleaning. For sensor noise, wavelet thresholding is used to remove high-frequency interference from continuous data such as current and water pressure. The core threshold calculation formula is as follows: , in the formula Represents the standard deviation of noise. The threshold determined by this formula, representing the data length, can accurately separate valid signals from noise. For missing data, a differentiated imputation strategy is used based on the data type. Continuous data, such as current and illuminance, uses K-nearest neighbor imputation, selecting reference data based on time and location similarity. Discrete data, such as equipment status and maintenance type, uses pattern imputation, selecting the most frequently occurring values ​​in the dataset. Outlier identification is achieved using the 3σ criterion, marking data exceeding the mean ± 3 standard deviations as outliers. A secondary verification is then performed using equipment operating thresholds to ensure the accuracy of the outlier detection. After confirming an outlier, linear interpolation is used to replace the data. The feature extraction stage requires designing extraction schemes for different types of data. For continuous data, the focus is on extracting time-domain and frequency-domain features. Time-domain features include mean, variance, peak value, and kurtosis, while frequency-domain features are obtained through Fast Fourier Transform (FFT) to obtain the dominant frequency and spectral energy. Discrete data is converted into numerical features using one-hot encoding. Location information is calculated using a Gaussian kernel function to determine the spatial distance between devices, with the kernel bandwidth set according to the density of devices distributed along urban roads. Textual data such as maintenance records uses the BERT model to extract semantic features, selecting a basic version of the model to ensure efficiency and effectiveness in feature extraction. The construction of the multimodal feature matrix uses "device-timestamp" as a dual index, concatenating all features of a single device and a single timestamp into a feature vector, ultimately forming a matrix with dimensions of "number of devices × time series length × feature dimension." The feature dimension varies for different types of devices due to differences in monitoring parameters. The preliminary equipment relationship diagram is constructed based on physical, spatial, and functional associations. Physical associations are determined by the topology of the equipment power supply network, such as determining whether different devices share a power supply circuit. Spatial associations are determined by distance thresholds; devices within a set distance range are considered to have spatial associations. Functional associations are established based on preset rules; for example, traffic lights at an intersection and surrounding fire hydrants constitute a functional association. The relationship diagram is represented using an adjacency matrix, with matrix elements distinguishing the existence of associations using 0 and 1, ultimately forming an undirected graph structure containing nodes and edges. Nodes represent devices, and edges represent the associations between devices.

[0058] S2. Based on the multimodal feature matrix and the preliminary device relationship graph, the direction and intensity of causal influence between devices are learned by combining the preset causal inference with the graph neural network model, and a heterogeneous device causal graph is obtained.

[0059] Specifically, in the heterogeneous device causal graph, nodes represent different types of devices, edges represent the direction of causal influence between devices, and edge weights represent the strength of causal influence. The pre-designed causal inference and graph neural network combined model adopts a two-level architecture of causal discovery and graph attention enhancement. The core objective is to eliminate spurious correlations and quantify causal strength based on the multimodal feature matrix and the initial device relationship graph. In the causal inference stage, an improved PC algorithm combined with Do-Calculus theory is used to identify the causal direction. First, the multimodal feature matrix is ​​divided into training and validation sets according to the time series. Using device feature vectors as variables, an initial undirected causal graph is constructed through the PC algorithm. In the PC algorithm, the conditional independence test uses partial correlation coefficient testing. By setting a significance level, conditionally independent edges are gradually deleted to narrow the range of variable associations. Subsequently, a dual constraint is introduced to optimize the causal direction judgment. The time sequence constraint clarifies that the collection time of the causal variable is earlier than that of the outcome variable. The device physical constraint determines the influence logic based on the operating principle of the device. For example, in a power supply network, upstream equipment can affect downstream equipment, but the reverse is not true. To further verify causal relationships, a Do-Calculus intervention is used. By simulating intervention in a specific device feature and observing changes in other device features, the causal direction is determined. A typical case involves intervening in a device parameter to cause it to malfunction, then monitoring subsequent parameter changes in related devices to clarify the causal direction. A graph neural network employs a graph attention network to learn causal strength. The input to the graph attention network is a multimodal feature matrix and an adjacency matrix of the initial device relationship graph. The model contains two attention layers. The first layer maps the device feature vectors from a high-dimensional space to a fixed-dimensional hidden space and calculates the attention coefficients between nodes using a self-attention mechanism. The calculation formula is as follows: ,in Representative equipment With equipment The original attention coefficients between them Represents the weight matrix. and respectively equipment With equipment eigenvectors, The first layer represents a vector concatenation operation. The second layer adjusts the attention coefficients based on causal direction constraints, retaining only edges with determined causal directions for calculation. The attention coefficients are normalized using the Softmax function to obtain the attention weights. The final output edge weights represent the causal influence strength, calculated as the product of the attention weight and the feature change rate in the intervention experiment. The result is normalized to a fixed interval before being used as the edge weight. During the generation of the heterogeneous device causal graph, node attributes include device ID, device type, and the current feature mean. Device types are labeled with concise symbols for easy differentiation. Edge attributes include causal direction, causal strength, and confidence. Causal direction is clearly indicated by arrows, and confidence is determined by the causal inference accuracy calculated from the validation set. Edges below a set threshold are removed. The final causal graph is stored in a graph database, supporting subsequent fast query and traversal operations.

[0060] S3. Based on the heterogeneous device cause-effect graph, detect abnormal device states through a preset time-series neural network model, and perform cause-effect propagation backtracking based on the heterogeneous device cause-effect graph to locate the root cause fault device, thereby obtaining a list of root cause fault devices and the current abnormal state vector.

[0061] Specifically, the pre-defined temporal graph neural network model employs a temporal graph convolutional network. Its core advantage lies in integrating the structural information of the causal graph of heterogeneous devices with the temporal dynamics of device features, simultaneously achieving two major functions: anomaly detection and root cause localization. The model input consists of two parts: first, a temporal feature sequence, where the multimodal feature matrix is ​​divided into temporal feature segments for each device by time windows. The size and step size of the time windows are set according to the device's operating cycle and fault warning requirements; second, the adjacency matrix and edge weight matrix of the causal graph of heterogeneous devices, used to accurately capture the causal relationships between devices. The temporal graph convolutional network model consists of graph convolutional layers and gated recurrent unit layers. The graph convolutional layers spatially aggregate the device features at each time step based on the causal graph structure, and the calculation method is as follows: ,in This represents the result after processing by the graph convolutional layer. Device feature matrix at time point, Represents the ReLU activation function. The normalized adjacency matrix is ​​calculated as follows: In the formula For degree matrix, It is the identity matrix. This is the original adjacency matrix. for The original device feature matrix at time step, The graph convolution weight matrix, This is the bias term. The gated recurrent unit layer is used to capture temporal dependencies. The spatially aggregated feature sequence is input into the gated recurrent unit, and the information transmission is precisely controlled by the update gate and reset gate, outputting the temporal feature representation of each time step.

[0062] Anomaly detection employs a reconstruction error method, calculating the mean squared error (MSE) between the output and input temporal feature sequences of the temporal graph convolutional network model. The distribution of the MSE is statistically analyzed using the training set data, and an anomaly threshold is set using the 3σ criterion. When the MSE at a certain time step exceeds the threshold, the corresponding device is determined to be in an abnormal state. Root cause fault device localization is based on a causal propagation backtracking algorithm. This algorithm starts with the abnormal device and, combined with the causal direction and strength of the causal graph of heterogeneous devices, traverses the causal path in reverse. The localization process consists of three steps: first, constructing an anomaly propagation tree, where the root node is the initial abnormal device, child nodes are devices affected by its causality, and edge weights represent the corresponding causal strengths; then, calculating the "fault propagation gain" of each node using the following formula: ,in Representative equipment Fault propagation gain, Representative equipment To the equipment The causal strength, Representative equipment Fault propagation gain, Representative equipment The degree of anomaly is represented by the ratio of mean squared error to a threshold. Finally, devices are sorted from largest to smallest by fault propagation gain, and those with gain values ​​exceeding a set threshold and no preceding causal node are identified as root cause fault devices, forming a list of root cause fault devices containing device ID, device type, anomaly occurrence time, and gain value. The current anomaly state vector is constructed with all devices as the dimension. The value of each dimension consists of an anomaly label and an anomaly degree. The anomaly label uses 0 and 1 to distinguish between normal and abnormal states, and the anomaly degree is the ratio of mean squared error to a threshold, normalized to a fixed range. For example, the numerical combination corresponding to a street light in the vector can intuitively reflect its anomaly state and severity.

[0063] S4. Based on the cause-effect graph of heterogeneous equipment, the list of root cause failure equipment and historical equipment data, counterfactual reasoning is performed in the digital twin using a preset structural cause-effect model to simulate virtual maintenance operations, optimize maintenance strategies, and generate the optimal maintenance sequence and counterfactual impact report.

[0064] Specifically, leveraging the high-fidelity mapping characteristics of digital twins, and combining them with structural causal models, precise optimization of maintenance strategies is achieved. The digital twin is a 1:1 virtual mapping of the physical entity of the equipment, integrating a 3D model of the equipment, real-time sensor data interfaces, causal relationship models, and a physical simulation engine. Multi-dimensional functional integration is achieved through the joint construction of multiple software programs. Before use, the digital twin needs to be calibrated. The least squares method is used to fit the equipment's operating status in historical data with the simulation results of the virtual model. Model parameters are adjusted to control the fitting error within a set range, ensuring consistency between the virtual model and the physical entity. The preset structural causal model includes exogenous variables, endogenous variables, and structural equations. Exogenous variables cover external influencing factors such as ambient temperature and power supply voltage fluctuations. Endogenous variables include equipment operating parameters, fault states, and other equipment-specific attributes. The structural equations are constructed based on regression equations trained using physical laws and historical data. For example, the power structural equation for a street light can be expressed as... In this formula, Streetlight power is a core parameter reflecting the operating status of streetlights; This refers to the power supply voltage, i.e., the input voltage when the streetlight is working; Operating current is the current consumed during the operation of the street light; Ambient temperature represents the external temperature environment in which the streetlight is located; This is a temperature influence coefficient obtained by fitting historical data, used to quantify the degree of influence of ambient temperature on street light power.

[0065] The counterfactual reasoning process uses devices in the root cause failure device list as intervention targets, setting up various counterfactual scenarios, such as replacing the core component of a root cause device in advance or increasing the monitoring frequency of a device. By intervening in the corresponding endogenous variables through a structural causal model while keeping other variables constant, the state changes of each device in the heterogeneous device causal graph after intervention are simulated. The effectiveness of different counterfactual scenarios is recorded; for example, after intervening in a root cause device to restore its parameters to normal, the abnormal states of related devices are observed to improve. Maintenance strategy optimization uses a genetic algorithm to construct a fitness function with the goal of "lowest maintenance cost, lowest failure risk, and shortest maintenance time," the function expression of which is... In this formula, The fitness value is used to comprehensively evaluate the quality of maintenance strategies; the higher the value, the better the strategy. , , These are weighting coefficients, set according to actual maintenance needs, used to reflect the importance of different optimization objectives; The maintenance cost for the current maintenance strategy includes labor costs, component costs, downtime losses, etc. The highest maintenance cost in historical maintenance data is used as the benchmark for cost comparison; To determine the failure risk of equipment after adopting the current maintenance strategy, a failure probability prediction value is determined based on counterfactual reasoning. This represents the upper limit of the equipment failure risk when no maintenance is performed; This represents the total maintenance time for the current maintenance strategy. The preset maximum allowable maintenance time is a constraint threshold for maintenance time. The execution flow of the genetic algorithm follows standard genetic operations. The initial population is a randomly generated set of maintenance strategies, each strategy containing key information such as maintenance equipment, maintenance method, and maintenance order. The selection operation uses roulette wheel selection, giving strategies with higher fitness values ​​a higher probability of being selected. The crossover operation uses single-point crossover, exchanging maintenance order segments of two strategies to achieve gene recombination. The mutation operation randomly adjusts the maintenance method or order, introducing appropriate mutations to avoid the algorithm getting trapped in local optima.

[0066] After multiple iterations, the strategy with the highest fitness value is determined as the optimal maintenance strategy. The corresponding maintenance steps are sorted chronologically to generate the optimal maintenance sequence. The sequence includes detailed information such as maintenance task ID, target equipment, maintenance operation, estimated time, executor, and prerequisite tasks. For example, it is specified that the corresponding water supply valve must be closed before maintaining a fire hydrant. The counterfactual impact report is generated based on the counterfactual reasoning results. The content covers the causal impact range of the root cause failure equipment, the effect comparison of various virtual maintenance scenarios, the expected effect of the optimal maintenance sequence, and the confidence level. The causal impact range clearly defines the list of equipment affected by the root cause equipment and the degree of impact. The effect comparison is carried out from dimensions such as the reduction in failure probability and cost changes. The expected effect quantifies the optimization range of failure risk and maintenance cost. The confidence level is calculated based on the calibration error of the digital twin. The report is presented in a structured format and supports export to common document formats or integration with the maintenance management system, providing maintenance personnel with quantifiable decision-making basis.

[0067] The aforementioned predictive maintenance method for the entire equipment lifecycle overcomes the challenge of utilizing multi-source data by constructing a comprehensive data acquisition network and optimizing feature processing, providing a high-quality foundation for analysis. Combining causal inference and graph neural networks, it accurately mines causal relationships between devices and quantifies their strength, eliminating invalid correlations. Relying on temporal graph neural networks, it achieves accurate anomaly identification, quickly pinpointing the root cause of faulty devices through causal propagation and backtracking, reducing misjudgments. Utilizing digital twins and reinforcement learning, it simulates maintenance effects and optimizes strategies through counterfactual reasoning, generating optimal maintenance sequences and impact reports, providing reliable data for operation and maintenance, effectively improving maintenance efficiency, and reducing the cost waste caused by blind maintenance.

[0068] In an optional embodiment, based on a multimodal feature matrix and a preliminary device relationship graph, a heterogeneous device causal graph is obtained by learning the direction and intensity of causal influence between devices through a pre-defined causal inference and graph neural network combined model, including the following steps:

[0069] S11. Based on the multimodal feature matrix and the preliminary device relationship graph, node embedding learning is performed through the message passing mechanism of graph neural network to generate the hidden embedding vector of each device.

[0070] Specifically, the core of the message passing mechanism is to leverage the relationships between devices to fuse single-device features with those of neighboring devices, thereby enhancing and optimizing device features and providing a high-quality feature foundation for subsequent causal relationship mining. The graph neural network uses a graph convolutional network as its basic model, with inputs being a multimodal feature matrix and an adjacency matrix of the initial device relationship graph. The multimodal feature matrix has already undergone data cleaning and feature extraction, while the adjacency matrix clarifies the initial relationships between devices.

[0071] The message passing process iteratively updates node embeddings through multi-layer graph convolution operations. The core computation formula for each layer of graph convolution is: In this formula, Representing the In-layer equipment The node embedding vector is the device feature representation output after the graph convolution operation of this layer; The ReLU function is chosen as the activation function to enhance the model's nonlinear expressive power; It is the learnable weight matrix of the l-th layer, used for feature mapping and dimension transformation of the fused neighbor information; Indicates equipment The set of neighboring devices, that is, devices There is an initial group of associated devices; and respectively equipment With neighboring devices The number of neighbors, this ratio is used to normalize the neighbor information and avoid feature bias caused by differences in node degree; For the first In-layer neighbor devices The node embedding vector is the one that participates in the current device. Neighbor features fused from features; It is the first The bias term of each layer is used to adjust the overall offset of the feature output, improving the model's fitting accuracy. The graph convolutional network uses multiple hidden layers to balance expressive power and computational efficiency by progressively compressing the feature dimension. During model training, a cross-entropy loss function is used, with device fault labels as supervision signals. The learnable weight matrix and bias term of each layer are updated through backpropagation, enabling the hidden embedding vectors to accurately capture the correlation features between device operating status and fault risk. The final hidden embedding vector for each device contains both the device's own operating characteristics and the influence information of related devices, providing strong feature support for subsequent causal relationship analysis.

[0072] S12. Based on the hidden embedding vector, calculate the causal score of each pair of device nodes through the attention mechanism, and verify the causal direction by combining the conditional independence test to generate a candidate causal edge set.

[0073] Specifically, the introduction of the attention mechanism aims to quantify the importance of inter-device relationships, filter out relationships that significantly affect the state of the target device, and lay the foundation for causal relationship identification. The attention mechanism employs a scaled dot product attention structure to compute the relationship between each pair of device nodes. The attention score, which is the initial causal score, is calculated as follows: In this formula, It is equipment For equipment The causal score is used to measure the strength of the potential causal relationship between two devices; and These are the query weight matrix and the key weight matrix, respectively. Both are learnable parameters, and their dimensions are set according to the feature processing requirements. They are used to convert the device hidden embedding vector into the query vector and the key vector. and For equipment and The hidden embedding vector is the basic feature input for attention computation; Indicates equipment The product of the hidden embedding vector and the query weight matrix is ​​used to obtain the device. The query vector; Indicates equipment The transpose of the product of the hidden embedding vector and the key weight matrix is ​​used to obtain the device. The key vector is transposed; the product of the two is used to calculate the similarity between the query vector and the key vector. The dimension of the key vector is used in the denominator to scale the similarity results, preventing the gradient vanishing problem caused by excessively large scores. The Softmax function is applied to all pointing devices. The causal scores are normalized to obtain attention weights; higher weights indicate better device performance. For equipment The stronger the potential causal influence, the better. Subsequently, the Top-K device node pairs are selected based on attention weights to form an initial edge set, with the K value set according to the total number of devices and the association density. To eliminate spurious correlations in the initial edge set, a conditional independence test is needed to verify the causal direction. The partial correlation coefficient test is used as the testing method, the core of which is to determine the causal direction under a given confounding variable (i.e., one that may simultaneously affect devices). and Under conditions of third-party equipment or environmental factors, the equipment and Whether the hidden embedding vectors are independent. Set a significance level; if the absolute value of the partial correlation coefficient is less than the corresponding critical value, then the device is considered independent. and In the given If the conditions are independent, the edges between the two are considered spurious and are deleted; otherwise, a potential causal relationship is determined. This is combined with temporal constraints (equipment). The state change time is earlier than the device The causal direction is determined by physical constraints (such as upstream and downstream relationships in a power supply link), such as equipment. The upstream equipment supplies power, and its current anomaly occurs earlier than that of the downstream equipment. The voltage anomaly time determines the causal direction as follows: → Through the above process, pseudo-correlated edges in the initial edge set are removed and causal directions are clarified, ultimately generating a candidate causal edge set containing device node pairs and causal directions.

[0074] S13. Based on cross-device intervention experiment data, estimate the average causal effect of each edge in the candidate causal edge set through Bayesian linear regression, calibrate the causal influence strength, and generate a causal weight matrix.

[0075] Cross-device intervention experiments involve applying random interventions to critical devices and recording the responses of other devices.

[0076] Cross-device intervention experiment data is the core basis for quantifying causal strength. This data is obtained through a combination of physical experiments and simulations: physical experiments select key devices and manually control them to cause abnormal operating parameters, while simultaneously recording changes in the operating status of other related devices; simulations, based on digital twins, conduct batch intervention experiments on all devices to compensate for the limited coverage of physical experiments. Intervention methods include parameter adjustment and fault injection, and for each intervention scenario, the characteristic changes and time of change of the devices before and after the intervention are recorded. The experimental data must include structured information such as intervention device ID, intervention type, intervention time, response device ID, and change in response characteristics. A Bayesian linear regression model is used to establish the quantitative relationship between intervention and response variables, thereby estimating the average causal effect. The core formula of the model is... In this formula, The characteristic change of the responding device, that is, the change in the device status parameters caused by the intervention operation; The variable is the intervention variable, with a value of 1 indicating that intervention was applied to a certain device and 0 indicating that no intervention was applied. The intercept term represents the basic characteristic level of the responding device when no intervention is applied; The regression coefficient is the average causal effect, which directly reflects the degree of influence of the intervention variable on the response variable. The error term represents the influence of random factors that the model cannot explain, and it follows a normal distribution. Compared with traditional linear regression, Bayesian linear regression, by introducing a prior distribution and combining it with experimental data to calculate the posterior distribution, can more accurately estimate causal effects, especially suitable for scenarios with small sample experimental data. For each edge in the candidate causal edge set ( → ), to the equipment intervention for ,equipment The characteristic change is The Bayesian linear regression model is input for training, and the posterior mean of the regression coefficient β is obtained. This mean is the edge ( → The average causal effect. To ensure comparability of causal strengths, the average causal effect needs to be normalized, mapping it to a fixed interval. The normalization formula is: In this formula, The normalized causal strength is the edge weight ultimately used to construct the causal graph; For the edge ( → The average causal effect, i.e., the unnormalized original causal effect value; It is the minimum value of the average causal effect of all candidate causal edges; This represents the maximum average causal effect across all candidate causal edges. This formula eliminates the influence of differences in the original effect values ​​between different causal edges, ensuring that the causal strength of all edges is on the same comparable dimension. The normalized causal strengths of all edges are arranged in order of device node pairs, forming a causal weight matrix. The matrix elements... Corresponding edges ( → The causal strength of the device and If there are no candidate causal edges, then .

[0077] S14. Based on the candidate causal edge set and causal weight matrix, construct a directed weighted heterogeneous device causal graph.

[0078] Specifically, the construction of a heterogeneous device causal graph requires clarifying three core elements: node set, edge set, and weight matrix. This ensures that the graphical structure can completely and accurately reflect the causal relationships and influence strength between devices. The node set consists of all participating devices. Each node contains basic attributes and characteristic attributes: basic attributes include a unique device ID, device type, and installation location coordinates. The device type uses standardized coding for differentiation. The characteristic attribute is a hidden embedding vector used for subsequent anomaly detection and maintenance strategy optimization. The edge set directly uses a candidate causal edge set. Each edge is directed, with the arrow direction representing the causal influence direction, i.e., from the cause device to the result device. The edge attributes include the causal relationship type, which is determined based on the causal association mechanism between devices. For example, a causal edge formed by streetlights and fire hydrants sharing a power supply circuit is labeled "power supply association type," and a causal edge between traffic lights at an intersection and nearby fire hydrants is labeled "functional coupling type." The causal weight matrix is ​​a normalized matrix. The non-zero element values ​​in the matrix are directly used as the weights of the corresponding edges. The larger the weight value, the stronger the causal influence of the cause device on the result device. During the graph construction process, a graph data structure is used for storage. A graph database is used to associate and store nodes and edges, and an indexing mechanism is established to improve the traversal and query efficiency of the causal graph. To ensure the reliability of the causal graph, validity verification is also required: a subset of causal edges are randomly selected, and cross-device intervention experiments are performed again. The deviation between the experimentally measured causal strength and the weight values ​​in the graph is compared. If the deviation is less than a set threshold, the causal graph is considered valid; if the deviation exceeds the threshold, the conditional independence test parameters and the Bayesian linear regression model are re-optimized until the causal graph meets the validity requirements. The final heterogeneous device causal graph clearly presents the causal dependencies between different types of devices, quantifies the strength of causal influence, and provides core structural support for subsequent anomaly detection, root cause localization, and maintenance strategy optimization.

[0079] In an optional embodiment, based on the hidden embedding vector, a causal score for each pair of device nodes is calculated using an attention mechanism, and the causal direction is verified by combining a conditional independence test to generate a candidate causal edge set, including the following steps:

[0080] S21. Based on the hidden embedding vector, calculate the initial causal score for each pair of device nodes using a learnable diagonal matrix; the formula for calculating the initial causal score is:

[0081]

[0082] in, Represents device node For device nodes The initial causal score, and Representing device nodes and The hidden embedding vector, This represents a learnable diagonal matrix.

[0083] Specifically, the initial causal score is used to preliminarily screen potential correlations between devices. Compared to traditional fully connected weight matrices, learnable diagonal matrices can reduce computational complexity while maintaining correlation capture capabilities. In the above formula, Represents device node For device nodes The initial causal score indicates that the higher the score, the greater the likelihood of a potential causal relationship between the two nodes. and Representing device nodes and The hidden embedding vectors are fused together with the device's own operational characteristics and the association information of neighboring devices; Represents device node Hidden embedding vector transpose, using transpose operation to calculate the inner product between vectors; This represents a learnable diagonal matrix, where the elements on the main diagonal are trainable parameters, and the elements on the off-diagonal are fixed at 0. This structure allows for focusing on the importance weights of the device's own features, reducing redundant computation. During model training... It will be continuously updated as the entire network propagates backward, gradually optimizing the calculation accuracy of the initial causal score.

[0084] S22. Using environmental data from the multimodal feature matrix as a confounding variable, the conditional independence between device nodes is tested using the partial correlation coefficient; the formula for calculating the partial correlation coefficient is:

[0085]

[0086] in, Represents device node and Given a set of confounding variables The partial correlation coefficient under the condition, Represents device node and Pearson correlation coefficient, and Representing device nodes and With the set of confusing variables The correlation coefficient.

[0087] Specifically, confounding variables are a key factor leading to spurious correlations between devices. Selecting environmental data (such as ambient temperature, humidity, and power supply voltage fluctuations) from the multimodal feature matrix as confounding variables can accurately eliminate interference from the external environment in determining device correlations. In the formula for calculating the partial correlation coefficient mentioned above, Represents device node and Given a set of confounding variables The partial correlation coefficient below is used to measure the exclusion After the impact, and The strength of the real correlation; Represents device node and The Pearson correlation coefficient reflects the apparent correlation between two nodes without considering confounding variables; and Representing device nodes and With the set of confusing variables The correlation coefficient quantifies the degree of association between the equipment and external environmental factors; the numerator is obtained by subtracting the association components introduced by confounding variables. and The net correlation value; the denominator is a standardization factor used to constrain the partial correlation coefficient within the range of [-1,1] to ensure that the calculation results under different scenarios are comparable.

[0088] S23. For each pair of device nodes, when the initial causal score is higher than the preset threshold and the absolute value of the partial correlation coefficient is less than the independence threshold, the corresponding causal edge is retained, and the result of retaining the causal edge is obtained.

[0089] Specifically, the setting of two thresholds forms a dual screening mechanism, ensuring that the selected causal edges have strong potential correlations while also preventing interference from spurious correlations. The preset threshold is determined based on the causal score distribution of devices in the training set during normal operation, typically selecting the high percentile value of the score distribution. The aim is to screen device-node pairs with sufficient correlation strength to constitute a potential causal relationship. The independence threshold is set based on the statistical significance level, referring to the critical value table of partial correlation coefficients. When the absolute value of the partial correlation coefficient is less than this threshold, it indicates that, given the set of confounding variables... Under these conditions, device nodes and The probability of them being independent is extremely high, and the correlation between them is not a spurious correlation. Only device node pairs that simultaneously satisfy both the conditions of "high initial causal score" and "low conditional independence probability" will have their corresponding causal edges retained, thereby effectively narrowing the candidate range and improving the efficiency of subsequent processing.

[0090] S24. Based on the results of retaining causal edges, generate a set of candidate causal edges.

[0091] Specifically, while preserving causal edges, it is also necessary to clarify the causal direction of each edge to ensure that the candidate causal edge set is a directed edge set. The determination of the causal direction combines two constraints: a time sequence constraint based on the acquisition time of the device status data, and a constraint that if the device... The characteristic changes occurred earlier than the equipment. If the characteristic change time is considered, then the causal direction may be preliminarily determined to be... → The physical constraints of the equipment are based on the operating principles and topological relationships of the equipment. For example, changes in the parameters of upstream equipment in a power supply network may affect downstream equipment, while downstream equipment usually does not affect upstream equipment in the reverse. Based on this, the causal direction can be further calibrated. All retained edges after the causal direction is clarified are integrated to form a structured candidate causal edge set containing "starting device - ending device - causal direction", providing an accurate edge set basis for subsequent causal strength calibration.

[0092] In an optional embodiment, based on the heterogeneous device causality graph, abnormal device states are detected using a preset temporal graph neural network model, and causal propagation backtracking is performed based on the heterogeneous device causality graph to locate the root cause fault device, obtaining a list of root cause fault devices and a current abnormal state vector, including the following steps:

[0093] S31. Based on real-time streaming data, extract the real-time feature vector of each device and calculate the device state embedding through a time-series graph neural network model.

[0094] Specifically, real-time streaming data directly reflects the current operating status of equipment, including continuous data such as current, water pressure, and temperature collected by sensors in real time, as well as discrete data such as equipment on / off status. This data needs to be received and preprocessed in real time using a streaming computing framework. The feature extraction process follows the same logic as historical data processing. Considering the time-sensitive nature of real-time data, a sliding window is used to extract temporal features, ensuring that the features promptly reflect changes in equipment status. The time-series graph neural network model innovatively combines a long short-term memory network (LSTM) and a graph convolutional network. The former captures temporal dynamic dependencies, while the latter integrates causal relationships between devices. The model input consists of two parts: a sequence of real-time feature vectors from the devices and the adjacency matrix and causal weight matrix of the heterogeneous device causal graph. During computation, the LSM first updates the real-time feature sequence of a single device temporally, outputting a device feature vector containing temporal information. The graph convolutional network then aggregates the feature information of neighboring devices into the current device based on the heterogeneous device causal graph, generating the final state embedding vector through spatial information fusion. This "temporal-spatial" dual-dimensional fusion approach allows the state embedding vector to include both the real-time operational characteristics of the device itself and the causal influence information of related devices, providing high-quality feature support for subsequent anomaly detection.

[0095] S32. Based on the state embedding vector, reconstruct the predicted embedding through an autoencoder, and calculate the reconstruction error as the anomaly score; the formula for calculating the anomaly score is:

[0096]

[0097] in, Indicates equipment In time Abnormal scores, Indicates equipment In time The state embedding vector, This represents the predicted embedding vector reconstructed by the autoencoder.

[0098] Specifically, an autoencoder consists of two parts: an encoder and a decoder. The encoder compresses the high-dimensional state embedding vector into a low-dimensional latent vector, while the decoder reconstructs the low-dimensional latent vector into a predicted embedding vector with the same dimension as the input. When the device is in a normal state, the state embedding vector has a stable distribution, and the autoencoder can achieve accurate reconstruction. When the device malfunctions, the state embedding vector deviates from the normal distribution, and the reconstruction error increases significantly. Therefore, the degree of device malfunction can be quantified by the reconstruction error. In the above formula for calculating the malfunction score, Indicates equipment In time The anomaly score indicates the severity of the anomaly at that moment; a higher score indicates a more severe anomaly. Indicates equipment In time The state embedding vector is the output of the temporal graph neural network model; This represents the predicted embedding vector reconstructed by the autoencoder, which is the result of the decoder restoring the low-dimensional latent vector; L2 norm (Euclidean norm) is used to calculate the Euclidean distance between two vectors. This distance value directly reflects the difference between the state embedding vector and the predicted embedding vector, i.e., the reconstruction error.

[0099] S33. When the abnormal score exceeds the dynamic threshold set based on historical data, the device is marked as abnormal, and a set of abnormal devices is generated.

[0100] Specifically, the dynamism of the threshold is key to improving the accuracy of anomaly detection. Fixed thresholds are prone to misjudgment or missed detection when equipment operating status fluctuates. Dynamic thresholds are set based on the distribution of anomaly scores in historical normal operating data. The initial threshold range is determined by statistically analyzing the mean and standard deviation of anomaly scores in historical data, using the 3σ criterion, and then dynamically adjusted in conjunction with equipment type and operating scenario. For example, streetlights experience greater fluctuations in anomaly scores during high-load periods at night, so the threshold is adjusted upwards accordingly; fire hydrants operate stably during off-peak hours, so the threshold is lowered. The update cycle of the dynamic threshold is synchronized with the sampling cycle of real-time data to ensure timely adaptation to changes in equipment operating status. When the anomaly score of a device at a certain time step exceeds the corresponding dynamic threshold, the device is immediately marked as an anomaly, and the anomaly occurrence time and score are recorded. All marked anomaly devices are then integrated to form an anomaly device set.

[0101] S34. Based on the cause-effect graph of heterogeneous devices, starting from the set of abnormal devices, calculate the causal influence score of each device through the causal propagation backtracking algorithm to locate the root cause fault device; the formula for calculating the causal influence score is:

[0102]

[0103] in, Represents device node The causal influence score Indicates pointing to the device node The neighborhood group, Indicates from device node arrive Causal weights Represents device node The degree of departure.

[0104] Specifically, the causal influence score is used to quantify the likelihood that a device is the root cause of a failure. The core logic is that a failure propagates along causal relationships from the root device to related devices, and the root device typically has the highest causal influence score. In the formula for calculating the causal influence score above, Represents device node The causal influence score indicates that the higher the score, the better the device. The greater the probability that the faulty device is the root cause; Indicates pointing to the device node The set of neighbors, i.e., those in causal relationships that may affect the device. The collection of upstream equipment; Indicates from device node To device node The causal weights are derived from the causal weight matrix of the heterogeneous device causal graph, reflecting... right The strength of the causal influence; Represents device node The causal influence score enables the iterative transmission of influence; Represents device node The output degree, i.e., the equipment The number of causal edges pointing to other devices, used to... The influence is normalized to avoid excessive amplification of influence by highly connected devices. The calculation process starts with the set of anomalous devices, initializes the causal influence score of anomalous devices as their anomalous score, and sets the initial score of non-anomalous devices to 0. Then, iterative calculation is performed along the reverse causal path of the heterogeneous device causal graph (from the result device to the cause device) until the scores of all devices converge.

[0105] S35. Based on the causal influence score, select the device with the highest score as the root cause failure device, and generate a list of root cause failure devices and a current abnormal state vector.

[0106] Specifically, to ensure the accuracy of root cause localization, a score screening threshold needs to be set. Only devices with causal influence scores exceeding the threshold are included in the root cause candidate range. The threshold is set based on the score distribution of historical root cause localization data, typically selecting a score value that covers 95% of the actual root cause devices. If multiple devices exist within the candidate range, further judgment is needed based on the topological structure of the heterogeneous device causal graph. Devices with preceding causal nodes are removed. That is, if device A has a high causal influence score, but device B is an upstream causal device of A and its score also exceeds the threshold, then A is the fault propagation node, and B is the root cause fault device. The finally selected root cause fault devices are sorted in descending order of causal influence score to generate a root cause fault device list. The list includes information such as device ID, device type, anomaly occurrence time, anomaly score, and causal influence score. The current anomaly state vector uses all devices as dimensions. Each dimension stores a "anomaly label - anomaly degree" tuple. The anomaly label uses 1 to represent anomaly and 0 to represent normal. The anomaly degree is the ratio of the anomaly score to the dynamic threshold and normalized to the [0,1] interval, intuitively presenting the anomaly status of all devices.

[0107] In an optional embodiment, based on the heterogeneous device causal graph, the root cause failure device list, and device historical data, counterfactual reasoning is performed in the digital twin using a preset structural causal model to simulate virtual maintenance operations, optimize maintenance strategies, and generate an optimal maintenance sequence and counterfactual impact report, including the following steps:

[0108] S41. Based on the cause-effect graph of heterogeneous devices, construct a structural cause-effect model.

[0109] Specifically, structural causal models are the core carriers of counterfactual reasoning. Their construction must closely align with the causal relationships between devices to ensure accurate mapping of the transmission patterns of device states. A structural causal model includes state variables, structural equations, and external noise variables for each device. State variables encompass core attributes such as device operating parameters and fault states, directly reflecting the real-time operating conditions of the device. External noise variables characterize the impact of uncontrollable factors such as environmental fluctuations and measurement errors on device states, typically following a normal distribution. The structural equation is the core of the model, used to quantify the causal relationships between variables. Its form is learned from historical device data through regression analysis and causal verification. For example, for power supply-related devices, the structural equation can reflect the impact of upstream device voltage changes on downstream device current. During construction, the nodes and edges of the heterogeneous device causal graph serve as the framework, correspondingly associating state variables with causal edges to ensure the input-output relationship of the structural equation aligns with the causal direction, guaranteeing that the model structure conforms to the causal logic of the physical devices.

[0110] S42. Based on the current abnormal state vector, simulate virtual maintenance operations using the Monte Carlo sampling method; the virtual maintenance operations intervene in the equipment state variables to normal values ​​and predict the counterfactual state distribution of downstream equipment based on the structural causal model.

[0111] Specifically, the core of virtual maintenance operations is to intervene in the state of abnormal equipment, predict the maintenance effect, and provide data support for strategy optimization. The current abnormal state vector clearly identifies the abnormality markers and degrees of abnormality for each device, thereby determining the target devices requiring intervention, primarily those in the root cause failure device list and the abnormal devices affected by them. Virtual maintenance operations are achieved by intervening in device state variables to normal values, such as adjusting the power supply voltage variable of an abnormal street light to the standard voltage range. Monte Carlo sampling is used to handle uncertainties in the intervention process. Different intervention scenarios are generated through multiple random samplings, and in each scenario, a structural causal model is used to predict the counterfactual state distribution of downstream devices. The counterfactual state distribution reflects the possible range of changes in the state of downstream related devices after implementing the maintenance operation, and the concentration of the distribution reflects the reliability of the prediction results, providing a comprehensive scenario basis for subsequent cost calculations.

[0112] S43. For each candidate maintenance sequence, calculate the expected cost using counterfactual reasoning; the formula for calculating the expected cost is:

[0113]

[0114] in, Represents the vector for performing maintenance actions. The expected cost afterward This indicates the maintenance action vector. Indicates maintenance equipment binary actions, Indicates maintenance equipment The cost, Indicates equipment Losses due to malfunction Indicating intervention and observation Lower device The probability of failure.

[0115] Specifically, expected cost is a core indicator for evaluating the merits of candidate maintenance sequences. It comprehensively considers the cost of maintenance implementation and the potential losses from failures due to lack of maintenance, achieving a balance between cost and risk. The formula for calculating expected cost above... Represents the vector for performing maintenance actions. The smaller the expected cost, the better the maintenance sequence. This represents a maintenance action vector, where each element corresponds to a maintenance action for a device. Indicates maintenance equipment The binary action, when the value is 1, indicates an action on the device. Perform maintenance; a value of 0 indicates no maintenance. Indicates maintenance equipment The cost covers all direct and indirect costs, including labor, parts, and downtime losses. Indicates equipment Losses from malfunctions include repair costs and cascading losses caused by the malfunction. Indicating intervention and observation Lower device The probability of failure, where Represents the vector of maintenance actions. Intervention operations, This represents the currently observed equipment state (i.e., the observed value corresponding to the current abnormal state vector). This probability is obtained through counterfactual reasoning of the structural causal model and directly reflects the effect of maintenance operations on reducing the risk of equipment failure.

[0116] S44. Optimize the maintenance sequence through reinforcement learning algorithm, and select the sequence that minimizes the expected cost as the optimal maintenance sequence.

[0117] Specifically, reinforcement learning algorithms can learn optimal maintenance decision-making strategies through iterative interaction with the environment, making them particularly suitable for dynamically changing equipment maintenance scenarios. The algorithm uses the "current equipment state (based on anomaly state vectors and causal graphs)" as the state space, "all possible combinations of maintenance actions" as the action space, and the "negative value of expected cost" as the reward signal; that is, the lower the expected cost, the higher the reward. During training, the agent starts from the initial state and selects a maintenance action vector (corresponding to a candidate maintenance sequence) each time. It obtains the expected cost corresponding to the action through counterfactual reasoning and converts it into a reward value. Simultaneously, it obtains the new equipment state after implementing the action based on a structural causal model. Through multiple iterative training iterations, the agent gradually masters the optimal maintenance action selection rules under different states, ultimately outputting the maintenance sequence that minimizes the expected cost in the current state—the optimal maintenance sequence. This sequence clearly defines the order of equipment maintenance, specific maintenance actions, and implementation timing, ensuring a balance between maintenance efficiency and cost control.

[0118] S45. Based on the counterfactual prediction results of the optimal maintenance sequence, generate a counterfactual impact report.

[0119] Specifically, the counterfactual impact report is a crucial basis for maintenance decisions. It must clearly present the implementation effects and quantifiable value of the optimal maintenance sequence, providing maintenance personnel with intuitive and actionable reference information. The core content of the report includes the reduction in equipment failure probability caused by maintenance actions. This reduction is obtained by comparing the difference in equipment failure probability between "performing maintenance" and "not performing maintenance." For example, it could show the average reduction in failure probability of three downstream related devices after maintenance of a root cause device. Furthermore, the report must include specific details of the optimal maintenance sequence (maintenance equipment, actions, sequence, and estimated time), the expected cost composition (a detailed breakdown of maintenance costs and risk losses), a comparison of the effects of different maintenance scenarios, and the confidence level of the prediction results. The confidence level is calculated based on the distribution concentration of Monte Carlo sampling to demonstrate the reliability of the report's conclusions. The report is presented in a structured format, supports export and integration with maintenance management systems, and directly serves the implementation of actual maintenance work.

[0120] The aforementioned predictive maintenance method for the entire equipment lifecycle acquires multi-source data, effectively integrating and extracting features from equipment-wide data to provide a high-quality data foundation for subsequent analysis. It employs a modeling approach combining causal inference and graph neural networks to accurately uncover causal relationships and influence strength between equipment, resolving the issue of spurious correlation interference in traditional correlation analysis. A complete technical chain is formed, encompassing anomaly detection, root cause localization, and maintenance optimization. Relying on temporal graph neural networks and digital twin counterfactual reasoning, it achieves accurate anomaly identification, rapid root cause localization, and quantitative optimization of maintenance strategies. Finally, it outputs a practical optimal maintenance sequence and counterfactual impact report, providing data support for operational decisions, improving maintenance efficiency, and reducing failure risks and costs.

[0121] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0122] Based on the same inventive concept, this application also provides a system for implementing the aforementioned predictive maintenance method for the entire equipment lifecycle. The solution provided by this system is similar to the implementation described in the above method. Therefore, the specific limitations of one or more embodiments of the predictive maintenance system for the entire equipment lifecycle provided below can be found in the limitations of the predictive maintenance method for the entire equipment lifecycle described above, and will not be repeated here.

[0123] In one exemplary embodiment, such as Figure 2 As shown, a predictive maintenance system 200 for the entire equipment lifecycle is provided, comprising:

[0124] The data acquisition and feature processing module 201 is used to acquire multi-source heterogeneous data from multiple devices and perform feature processing on the multi-source heterogeneous data to obtain a multimodal feature matrix and a preliminary device relationship diagram; the multi-source heterogeneous data includes device historical data, usage status, location information and maintenance records;

[0125] The device causal graph construction module 202 is used to learn the direction and intensity of causal influence between devices based on the multimodal feature matrix and the preliminary device relationship graph, through a preset causal inference and graph neural network combined model, to obtain a heterogeneous device causal graph; the nodes of the heterogeneous device causal graph represent different types of devices, the edges represent the direction of causal influence between devices, and the edge weights represent the intensity of causal influence.

[0126] The anomaly detection and root cause localization module 203 is used to detect abnormal device states based on the heterogeneous device causal graph and through a preset time-series neural network model, and to locate the root cause fault device by causal propagation back based on the heterogeneous device causal graph, thereby obtaining a list of root cause fault devices and the current abnormal state vector.

[0127] The maintenance strategy optimization module 204 is used to perform counterfactual reasoning in a digital twin based on the heterogeneous equipment cause-effect graph, the root cause failure equipment list and equipment historical data, through a preset structural cause-effect model, to simulate virtual maintenance operations, optimize maintenance strategies, and generate the optimal maintenance sequence and counterfactual impact report.

[0128] Furthermore, the device cause-effect graph construction module 202 is also used for:

[0129] S11. Based on the multimodal feature matrix and the preliminary device relationship graph, node embedding learning is performed through the message passing mechanism of the graph neural network to generate the hidden embedding vector of each device. The message passing mechanism updates the node embedding through multi-layer graph convolution operations, and each layer uses a learnable weight matrix and activation function to process the neighbor node information.

[0130] S12. Based on the hidden embedding vector, the causal score of each pair of device nodes is calculated through the attention mechanism, and the causal direction is verified by combining the conditional independence test to generate a candidate causal edge set; the conditional independence test uses the partial correlation coefficient to test the independence of device nodes under a given confounding variable.

[0131] S13. Based on cross-device intervention experiment data, estimate the average causal effect of each edge in the candidate causal edge set through Bayesian linear regression, calibrate the causal influence strength, and generate a causal weight matrix; the cross-device intervention experiment includes applying random intervention to key equipment and recording the responses of other equipment;

[0132] S14. Based on the candidate causal edge set and the causal weight matrix, construct a directed weighted heterogeneous device causal graph; the node set of the heterogeneous device causal graph corresponds to all devices, the edge set represents the causal dependency relationship, and the weight matrix represents the causal strength.

[0133] Furthermore, the device cause-effect graph construction module 202 is also used for:

[0134] S21. Based on the hidden embedding vector, calculate the initial causal score for each pair of device nodes using a learnable diagonal matrix; the formula for calculating the initial causal score is:

[0135]

[0136] in, Represents device node For device nodes The initial causal score, and Representing device nodes and The hidden embedding vector, Represents a learnable diagonal matrix;

[0137] S22. Using environmental data from the multimodal feature matrix as a confounding variable, the conditional independence between device nodes is tested using the partial correlation coefficient; the formula for calculating the partial correlation coefficient is:

[0138]

[0139] in, Represents device node and Given a set of confounding variables The partial correlation coefficient under the condition, Represents device node and Pearson correlation coefficient, and Representing device nodes and With the set of confusing variables The correlation coefficient;

[0140] S23. For each pair of device nodes, when the initial causal score is higher than the preset threshold and the absolute value of the partial correlation coefficient is less than the independence threshold, the corresponding causal edge is retained to obtain the result of retaining the causal edge.

[0141] S24. Based on the results of retaining causal edges, generate a set of candidate causal edges.

[0142] Furthermore, the anomaly detection and root cause localization module 203 is also used for:

[0143] S31. Based on real-time streaming data, extract the real-time feature vector of each device and calculate the device state embedding through a temporal graph neural network model. The temporal graph neural network combines a long short-term memory network and a heterogeneous device causal graph to generate state embedding vectors through temporal updates and graph convolution.

[0144] S32. Based on the state embedding vector, reconstruct the predicted embedding through an autoencoder, and calculate the reconstruction error as the anomaly score; the formula for calculating the anomaly score is:

[0145]

[0146] in, Indicates equipment In time Abnormal scores, Indicates equipment In time The state embedding vector, This represents the predicted embedding vector reconstructed by the autoencoder;

[0147] S33. When the abnormal score exceeds the dynamic threshold set based on historical data, the device is marked as abnormal and a set of abnormal devices is generated.

[0148] S34. Based on the cause-effect graph of heterogeneous devices, starting from the set of abnormal devices, calculate the causal influence score of each device through the causal propagation backtracking algorithm to locate the root cause fault device; the formula for calculating the causal influence score is:

[0149]

[0150] in, Represents device node The causal influence score Indicates pointing to the device node The neighborhood group, Indicates from device node arrive Causal weights Represents device node The degree of departure;

[0151] S35. Based on the causal influence score, select the device with the highest score as the root cause failure device, and generate a list of root cause failure devices and a current abnormal state vector.

[0152] Furthermore, the maintenance strategy optimization module 204 is also used for:

[0153] S41. Based on the cause-effect graph of heterogeneous devices, construct a structural cause-effect model; the structural cause-effect model includes the state variables, structural equations and external noise variables of each device, and the structural equations are learned from historical data;

[0154] S42. Based on the current abnormal state vector, simulate virtual maintenance operations using the Monte Carlo sampling method; the virtual maintenance operations intervene in the equipment state variables to normal values ​​and predict the counterfactual state distribution of downstream equipment based on a structural causal model;

[0155] S43. For each candidate maintenance sequence, calculate the expected cost using counterfactual reasoning; the formula for calculating the expected cost is:

[0156]

[0157] in, Represents the vector for performing maintenance actions. The expected cost afterward This indicates the maintenance action vector. Indicates maintenance equipment binary actions, Indicates maintenance equipment The cost, Indicates equipment Losses due to malfunction Indicating intervention and observation Lower device The probability of failure;

[0158] S44. Optimize the maintenance sequence through reinforcement learning algorithm, and select the sequence that minimizes the expected cost as the optimal maintenance sequence;

[0159] S45. Based on the counterfactual prediction results of the optimal maintenance sequence, generate a counterfactual impact report; the counterfactual impact report includes the amount by which the maintenance action reduces the probability of equipment failure.

[0160] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the device lifecycle predictive maintenance method as described above.

[0161] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0162] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0163] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A predictive maintenance method for the entire lifecycle of equipment, characterized in that, The method includes: S1. Acquire multi-source heterogeneous data from multiple devices, and perform feature processing on the multi-source heterogeneous data to obtain a multimodal feature matrix and a preliminary device relationship diagram; the multi-source heterogeneous data includes device historical data, usage status, location information, and maintenance records; S2. Based on the multimodal feature matrix and the preliminary device relationship graph, the direction and intensity of causal influence between devices are learned through a preset causal inference and graph neural network combination model to obtain a heterogeneous device causal graph; the nodes of the heterogeneous device causal graph represent different types of devices, the edges represent the direction of causal influence between devices, and the edge weights represent the intensity of causal influence. S3. Based on the heterogeneous device cause-effect graph, detect abnormal device states through a preset time-series neural network model, and perform cause-effect propagation backtracking based on the heterogeneous device cause-effect graph to locate the root cause fault device, thereby obtaining a list of root cause fault devices and the current abnormal state vector. S4. Based on the heterogeneous device causal graph, the root cause failure device list, and the device historical data, counterfactual reasoning is performed in the digital twin using a preset structural causal model to simulate virtual maintenance operations, optimize maintenance strategies, and generate the optimal maintenance sequence and counterfactual impact report.

2. The method according to claim 1, characterized in that, Based on the multimodal feature matrix and the preliminary device relationship graph, a heterogeneous device causal graph is obtained by learning the direction and intensity of causal influence between devices through a preset causal inference and graph neural network combined model, including: S11. Based on the multimodal feature matrix and the preliminary device relationship graph, node embedding learning is performed through the message passing mechanism of a graph neural network to generate a hidden embedding vector for each device; the message passing mechanism updates the node embedding through multi-layer graph convolution operations, and each layer uses a learnable weight matrix and activation function to process neighbor node information; S12. Based on the hidden embedding vector, calculate the causal score of each pair of device nodes through an attention mechanism, and verify the causal direction by combining the conditional independence test to generate a candidate causal edge set; the conditional independence test uses the partial correlation coefficient to test the independence of device nodes under a given confounding variable. S13. Based on cross-device intervention experiment data, estimate the average causal effect of each edge in the candidate causal edge set using Bayesian linear regression, calibrate the causal influence strength, and generate a causal weight matrix; the cross-device intervention experiment includes applying random intervention to key equipment and recording the responses of other equipment; S14. Based on the candidate causal edge set and the causal weight matrix, construct a directed weighted heterogeneous device causal graph; the node set of the heterogeneous device causal graph corresponds to all devices, the edge set represents causal dependencies, and the weight matrix represents causal strength.

3. The method according to claim 2, characterized in that, Based on the hidden embedding vector, the causal score of each pair of device nodes is calculated through an attention mechanism, and the causal direction is verified by combining a conditional independence test to generate a candidate causal edge set, including: S21. Based on the hidden embedding vector, calculate the initial causal score for each pair of device nodes using a learnable diagonal matrix; the formula for calculating the initial causal score is: in, Represents device node For device nodes The initial causal score, and Representing device nodes and The hidden embedding vector, Represents a learnable diagonal matrix; S22. Using the environmental data in the multimodal feature matrix as a confounding variable, the conditional independence between device nodes is tested using the partial correlation coefficient; the formula for calculating the partial correlation coefficient is: in, Represents device node and Given a set of confounding variables The partial correlation coefficient under the condition, Represents device node and Pearson correlation coefficient, and Representing device nodes and With the set of confusing variables The correlation coefficient; S23. For each pair of device nodes, when the initial causal score is higher than a preset threshold and the absolute value of the partial correlation coefficient is less than the independence threshold, the corresponding causal edge is retained to obtain the result of retaining causal edges. S24. Based on the results of retaining causal edges, generate the candidate causal edge set.

4. The method according to claim 1, characterized in that, The process involves detecting abnormal device states using a preset time-series neural network model based on the heterogeneous device causality graph, and locating the root cause fault device through causal propagation backtracking based on the heterogeneous device causality graph, thereby obtaining a list of root cause fault devices and a current abnormal state vector, including: S31. Based on real-time streaming data, extract the real-time feature vector of each device, and calculate the device state embedding through a temporal graph neural network model; the temporal graph neural network combines a long short-term memory network and the heterogeneous device causal graph to generate the state embedding vector through temporal updates and graph convolution. S32. Based on the state embedding vector, reconstruct the predicted embedding using an autoencoder, and calculate the reconstruction error as the anomaly score; the formula for calculating the anomaly score is: in, Indicates equipment In time Abnormal scores, Indicates equipment In time The state embedding vector, This represents the predicted embedding vector reconstructed by the autoencoder; S33. When the abnormal score exceeds the dynamic threshold set based on historical data, the device is marked as abnormal, and a set of abnormal devices is generated. S34. Based on the heterogeneous device causality graph, starting from the set of abnormal devices, calculate the causal influence score of each device using a causal propagation backtracking algorithm to locate the root cause fault device; the formula for calculating the causal influence score is: in, Represents device node The causal influence score Indicates pointing to the device node The neighborhood group, Indicates from device node arrive Causal weights Represents device node The degree of departure; S35. Based on the causal influence score, select the device with the highest score as the root cause failure device, and generate the root cause failure device list and the current abnormal state vector.

5. The method according to claim 1, characterized in that, Based on the heterogeneous device causal graph, the root cause failure device list, and the device historical data, counterfactual reasoning is performed in the digital twin using a preset structural causal model to simulate virtual maintenance operations, optimize maintenance strategies, and generate an optimal maintenance sequence and counterfactual impact report, including: S41. Based on the heterogeneous device causal graph, construct a structural causal model; the structural causal model includes the state variables, structural equations, and external noise variables of each device, and the structural equations are learned from historical data; S42. Based on the current abnormal state vector, simulate virtual maintenance operations using the Monte Carlo sampling method; the virtual maintenance operations intervene in the equipment state variables to normal values ​​and predict the counterfactual state distribution of downstream equipment based on the structural causal model; S43. For each candidate maintenance sequence, the expected cost is calculated using counterfactual reasoning; the formula for calculating the expected cost is: in, Represents the vector for performing maintenance actions. The expected cost afterward This indicates the maintenance action vector. Indicates maintenance equipment binary actions, Indicates maintenance equipment The cost, Indicates equipment Losses due to malfunction Indicating intervention and observation Lower device The probability of failure; S44. Optimize the maintenance sequence using a reinforcement learning algorithm, and select the sequence that minimizes the expected cost as the optimal maintenance sequence; S45. Based on the counterfactual prediction results of the optimal maintenance sequence, generate the counterfactual impact report; the counterfactual impact report includes the amount by which the maintenance action reduces the probability of equipment failure.

6. A predictive maintenance system for the entire lifecycle of equipment, characterized in that, The system includes: The data acquisition and feature processing module is used to acquire multi-source heterogeneous data from multiple devices and perform feature processing on the multi-source heterogeneous data to obtain a multimodal feature matrix and a preliminary device relationship diagram; the multi-source heterogeneous data includes device historical data, usage status, location information and maintenance records; The device causal graph construction module is used to learn the direction and intensity of causal influence between devices based on the multimodal feature matrix and the preliminary device relationship graph, through a preset causal inference and graph neural network combined model, to obtain a heterogeneous device causal graph; the nodes of the heterogeneous device causal graph represent different types of devices, the edges represent the direction of causal influence between devices, and the edge weights represent the intensity of causal influence. The anomaly detection and root cause localization module is used to detect abnormal device states based on the heterogeneous device causal graph using a preset time-series neural network model, and to perform causal propagation backtracking based on the heterogeneous device causal graph to locate the root cause fault device, thereby obtaining a list of root cause fault devices and a current abnormal state vector. The maintenance strategy optimization module is used to perform counterfactual reasoning in a digital twin based on the heterogeneous device causal graph, the root cause failure device list, and the device historical data, through a preset structural causal model, to simulate virtual maintenance operations, optimize maintenance strategies, and generate an optimal maintenance sequence and counterfactual impact report.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 5.

Images