Intelligent operation and maintenance decision-making method and system based on multivariate heterogeneous data fusion

Abstract

The invention relates to the field of data processing and information technology operation and maintenance, and discloses an intelligent operation and maintenance decision-making method and system based on multivariate heterogeneous data fusion. Comprising the following steps: constructing a dynamic causal graph reflecting a system state and an environmental factor causal relationship based on a causal structure learning algorithm; constructing a multi-agent game decision model based on the dynamic causal diagram, and respectively setting a global operation and maintenance target and a component local demand as a leader strategy and a follower strategy of the game; and solving game equilibrium by using multi-agent reinforcement learning, and outputting an optimal joint operation and maintenance decision instruction. The system is composed of a data acquisition preprocessing module, a multi-modal feature fusion module, a causal structure learning module, a game decision solving module and an execution monitoring module. According to the method, the causal inference and the game theory are fused, so that the causal logic can be accurately extracted, the decision robustness and interpretability are improved, and the global optimal collaborative configuration is realized.

Description

Technical Field

[0001] This invention belongs to the field of data processing and information technology operation and maintenance, specifically involving an intelligent operation and maintenance decision-making method and system based on the fusion of diverse heterogeneous data. Background Technology

[0002] With the deep integration of the Industrial Internet and big data technologies, intelligent operation and maintenance has become the core foundation for ensuring the efficient and stable operation of large and complex systems such as smart factories, modern data centers, and smart city infrastructure. These systems, by integrating massive amounts of industrial sensors, equipment logs, and business data, have achieved a leapfrog development from traditional manual inspection to automated monitoring and digital management, greatly improving the ability to perceive the operating status of equipment in complex environments and the efficiency of management.

[0003] Intelligent operation and maintenance decision-making technology based on the fusion of diverse and heterogeneous data is currently a core focus of research in this field. This technology aims to establish a comprehensive information model reflecting the overall operation of the system by deeply integrating and correlating data from different dimensions and formats, such as equipment operating parameters, time-series performance indicators, alarm texts, and historical work order records. Its basic principle is to use data-driven algorithms to identify abnormal patterns and predict failure risks from massive data streams, thereby driving the generation of automated operation and maintenance strategies to reduce unplanned downtime and optimize system resource allocation.

[0004] Existing intelligent operation and maintenance technologies still face many challenges in practical applications. Current systems generally rely excessively on data-driven correlation analysis models, which essentially only capture statistical correlations between data points. This makes them highly susceptible to external environmental variables or random human interference, leading to spurious correlations. Furthermore, they fail to fully integrate multi-dimensional operation and maintenance correlation data output by computer vision and biometric recognition software, resulting in significant limitations in data sources and feature dimensions. This leads to the system frequently outputting erroneous decision commands.

[0005] Due to a lack of in-depth exploration of the real causal relationships in the physical world, models often fall into the trap of knowing the "what" but not the "why" when facing complex system failures, resulting in extremely poor robustness and a lack of necessary interpretability in decision recommendations. Existing decision-making mechanisms often ignore the dynamic collaboration and interest game among components, making it difficult to achieve multi-objective optimal configurations in dimensions such as energy consumption and lifespan while ensuring overall service levels. These problems collectively lead to a severe deficiency in the ability of existing operation and maintenance systems to accurately isolate causal logic, handle unseen failure scenarios, and achieve globally optimal decisions in complex environments, making them unsuitable for the implementation requirements of artificial intelligence application software development. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent operation and maintenance decision-making method and system based on the fusion of diverse heterogeneous data, which can solve the problems mentioned in the background technology. Addressing the core technical shortcomings of existing intelligent operation and maintenance systems when dealing with complex industrial environments or large-scale computing clusters—namely, their over-reliance on statistical correlation analysis making it difficult to isolate environmental interference and spurious correlations, their lack of deep adaptation to AI-optimized operating systems, AI middleware and function libraries, and their failure to fully integrate multi-dimensional operation and maintenance data from computer vision and audiovisual software and biometric recognition software—leading to insufficient robustness and poor interpretability in the face of complex faults, this invention proposes a novel technical framework that combines causal inference deep graph learning with Stackelberg game decision-making. This is an innovative solution for AI application software development, aiming to achieve accurate extraction from data to causal logic in complex dynamic environments and to construct a globally optimal collaborative decision-making system.

[0007] To achieve the above objectives, the technical solution adopted by this invention is as follows: 1. Acquire multi-dimensional heterogeneous data streams from the system to be monitored, wherein the data streams include operation and maintenance related data output by computer audiovisual software and biometric recognition software, and transform them into a unified-dimensional embedding space vector using multimodal coding technology adapted to artificial intelligence-optimized operating systems, artificial intelligence middleware, and function libraries; 2. Mine the embedding space vectors based on causal structure learning algorithms to construct a dynamic causal graph reflecting the causal relationship between equipment operating status and environmental factors; 3. Construct a multi-agent game decision-making model based on the dynamic causal graph, and set the global operation and maintenance goal and the local requirements of each component as the leader strategy and follower strategy of the game, respectively; 4. Solve the Nash equilibrium state of the game model using a multi-agent reinforcement learning algorithm and output the optimal joint operation and maintenance decision instructions.

[0008] Preferably, step 1 specifically includes the following steps: 11. Synchronously acquire structured time-series data and unstructured text data through sensor nodes and log acquisition modules deployed in the system to be monitored; 12. Normalize and extract sliding window features from the structured time-series data to generate a continuous state vector reflecting the performance characteristics of the equipment; 13. Use natural language processing technology to perform entity recognition and event extraction on the unstructured text data, and transform it into discrete intervention signal features; 14. Input the continuous state vector and intervention signal features into a cross-modal feature fusion network, and output multi-dimensional heterogeneous fusion features representing the real-time operating status of the system.

[0009] Preferably, the structured time-series data includes CPU utilization, memory usage, input / output latency, real-time operating temperature of the device, structured data of device defect detection output by computer audiovisual software, and structured data of operation and maintenance permission verification output by biometric recognition software; the unstructured text data includes device operation log information, historical maintenance work order records, system status descriptions manually entered by operators, image recognition annotation text output by computer audiovisual software, and identity recognition event text output by biometric recognition software.

[0010] Preferably, step 2 specifically includes the following steps: 21. Using a constraint-based path constraint algorithm to perform preliminary screening of multi-variable heterogeneous fusion features, determine the conditional independence relationship between nodes, and generate an initial undirected graph; 22. Applying a score-based greedy equivalence search algorithm to score and optimize the initial undirected graph, and determining the directed edge direction between nodes through the maximization Bayesian information criterion scoring, forming an initial causal structure; 23. Introducing a graph neural network to dynamically update the initial causal structure, using an attention mechanism to capture the nonlinear evolution law of node features changing over time, and generating a dynamic causal graph that can reflect the causal logic of the system in real time.

[0011] Preferably, the nodes in the dynamic causal graph represent different monitoring indicators or external interference events, the directed edges represent the causal driving directions between nodes, and the edge weights represent the metric values ​​of causal strength.

[0012] Preferably, the step of dynamically updating the initial causal structure using a graph neural network specifically includes: inputting multi-heterogeneous fusion features as node attributes into a spatiotemporal graph convolutional network; using a temporal convolutional layer to capture the evolution trend of a single indicator within a preset time window; using a spatial graph convolutional layer to aggregate the feature information of neighboring nodes based on the connection strength of causal edges; and using residual connections and normalization layers to prevent the gradient vanishing problem during deep network training, outputting a dynamic node feature representation containing prediction performance.

[0013] Preferably, the application of the score-based greedy equivalence search algorithm includes: starting from an empty graph, attempting to add a directed edge that can maximize the improvement of the model score in each iteration; when the score no longer increases, entering the pruning phase, attempting to remove redundant edges that can improve the score; and by continuously searching the graph space until it converges to a locally optimal equivalence class graph structure, the automatic discovery of causal relationships is achieved.

[0014] Preferably, step 3 specifically includes the following steps: 31. Define each core hardware component or functional module in the operation and maintenance system as an independent controlled intelligent agent and assign it a specific local utility function; 32. Define the global operation and maintenance optimization goal as a leader intelligent agent, which guides the behavior of the controlled intelligent agents by formulating global incentive strategies or resource allocation schemes; 33. Establish state transition constraints between controlled intelligent agents based on dynamic causal graphs to ensure that the decision-making behavior of each intelligent agent conforms to the causal logic evolution law within the physical system.

[0015] Preferably, the local utility function is configured to comprehensively consider the component's operating energy consumption, remaining lifespan, and task execution efficiency; the global objective of the leader agent is configured to minimize the total operating cost of the entire system while satisfying the preset business service level agreement.

[0016] Preferably, step 4 specifically includes the following steps: 41. Initialize the policy network and value network parameters of each agent; 42. The leader agent outputs a global policy vector of a preset dimension based on the current dynamic causal graph state; 43. Each controlled agent, as a follower, responds based on its own local utility function after receiving the global policy vector, and outputs its own action selection; 44. The system outputs a reward signal based on the joint actions of each controlled agent, and iteratively updates the policy network parameters using the gradient descent algorithm until the policies of all agents reach Nash equilibrium, that is, no agent can obtain higher returns by changing its own policy without changing the policies of other agents.

[0017] Preferably, the multi-agent reinforcement learning algorithm employs a deep deterministic policy gradient algorithm with a centralized training and distributed execution architecture; during the training phase, a global commentator network is used to aggregate the observation and action information of all agents to overcome the non-stationarity problem in a multi-agent environment; during the execution phase, each agent makes real-time decisions based solely on its own local observations and the received leader signals.

[0018] Preferably, the intervention signal features are introduced as specific nodes in the dynamic causal graph to simulate the impact of operator intervention on the system state; by calculating the change in the conditional probability distribution of each node after intervention, the consequences of the decision can be predicted in advance, avoiding decision-making errors caused by statistical correlation of data.

[0019] Preferably, an intelligent operation and maintenance decision system based on multi-source heterogeneous data fusion is used to implement the above method, including: a data acquisition and preprocessing module configured to acquire and synchronize multi-source heterogeneous data streams of the system to be monitored; and a multimodal feature fusion module whose input end is connected to the output end of the data acquisition and preprocessing module, used to convert data of different formats into a unified embedding vector. The causal structure learning module is configured to construct and maintain a dynamic causal graph based on a causal inference algorithm. The causal structure learning module is connected to the multimodal feature fusion module. The game decision-making module receives the dynamic causal graph and real-time feature information from the input of the causal structure learning module. It is internally configured with multiple controlled agent units and a leader control unit, which are used to output joint operation and maintenance instructions through multi-agent reinforcement learning. The execution monitoring module is configured to receive the instructions output by the game decision-making module and apply them to physical devices, while transmitting the feedback results to the data acquisition and preprocessing module in real time.

[0020] Preferably, the causal structure learning module further includes a knowledge base verification unit, which is used to inject the prior logical relationships preset by industry experts as constraints into the search process of the dynamic causal graph, and improve the efficiency and accuracy of causal discovery by applying penalty weights to directed edges that do not conform to physical common sense.

[0021] Preferably, the controlled intelligent agent unit in the game decision-solving module is configured to have adaptive learning capabilities, which can adjust the weight coefficients of its local utility function in real time according to the aging degree of the device. For example, when the device is nearing the end of its lifespan, the proportion of the remaining lifespan term in the utility function is increased.

[0022] Preferably, the update frequency of the dynamic cause-effect graph is set to a preset time interval, or an instant update is triggered when the system senses a state change signal, so as to ensure that the decision system can respond quickly to sudden failures or environmental disturbances.

[0023] Preferably, when calculating the response strategy of the controlled agent, a prediction step size is introduced. By performing multi-step forward simulation on the dynamic causal graph, the long-term impact of the current strategy on the stability of the system at multiple future time points is evaluated, thereby enhancing the foresight of operation and maintenance decisions.

[0024] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention addresses the structural deficiency of traditional data-driven operations and maintenance (O&M) systems by deeply integrating causal inference with game theory, thus resolving the fundamental limitation of understanding the "what" but not the "why." By introducing a causal structure learning module, this invention can accurately extract the true logical evolutionary relationships from massive amounts of sensor data and alarm texts, filtering out spurious correlations caused by environmental variables or random noise. This allows the O&M system to make stable judgments based on the causal logic of the physical world even when facing unprecedented complex faults. This causal graph-based decision-making mechanism not only possesses strong robustness but also provides O&M personnel with an intuitive logical tracing path, improving the interpretability of decisions.

[0025] 2. The game theory-driven multi-agent decision-making architecture constructed in this invention resolves the sharp contradiction between global optimization and local interests in large and complex systems. By introducing the Stackelberg game model, the system can simulate the interactive feedback between operations and maintenance managers and various equipment components, enabling the global strategy formulated by the leader to fully predict and guide the behavioral tendencies of the controlled agents. Under this game framework, the agents no longer make decisions in isolation, but rather collaborate under the constraints of a causal graph. This allows the system to automatically find the Nash equilibrium state with the lowest energy consumption and least equipment wear while ensuring the core business service level, achieving optimal allocation of global resources.

[0026] 3. This invention breaks through the limitations of single time-series data-driven approaches by extracting entities and events from unstructured text data and transforming them into intervention nodes in a dynamic causal graph. This design allows the system to integrate human operation records and historical maintenance knowledge into the real-time reasoning process, using intervention calculus to simulate the expected effects of different maintenance actions, reducing the trial-and-error costs of decision-making and improving response speed in complex dynamic environments. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the overall technical solution architecture proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the present invention, which combines causal inference deep graph learning with Stackelberg game theory. Figure 3 This is a logical flowchart of the multivariate heterogeneous data preprocessing and cross-modal feature fusion in this invention; Figure 4 This is a flowchart illustrating the logical process of dynamic causal structure learning and spatiotemporal feature evolution in this invention. Figure 5 This is a schematic diagram of the hierarchical interaction relationship and data flow in the multi-agent game decision-making between leaders and followers in this invention. Detailed Implementation

[0028] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0029] Example 1: Please refer to the appendix Figure 1 To be continued Figure 5This embodiment takes the operation and maintenance of computing clusters in large-scale Internet data centers as an application background, and details the intelligent operation and maintenance decision-making method based on the fusion of diverse heterogeneous data and its specific implementation logic. The method is compatible with artificial intelligence optimized operating systems, artificial intelligence middleware and function libraries, and belongs to the application scenario of artificial intelligence application software development. In this highly complex computing environment, the system needs to process in real time the interactive data generated by tens of thousands of server nodes, cooling equipment and power supply units, as well as the operation and maintenance related data output by computer vision software and biometric recognition software. Its core objective is to minimize the total energy consumption of the system and extend the physical life of core hardware while ensuring the response speed of business requests.

[0030] First, step 1 is executed to acquire the diverse and heterogeneous data streams of the system under monitoring, and then transform them into a unified-dimensional embedding space vector using multimodal coding technology. In the physical environment of a data center, this process begins with the synchronous acquisition of all operational data from multiple sources. Specifically, through sensor nodes pre-deployed on server motherboards and switch ports, and a log acquisition module running at the operating system kernel level, the system synchronously acquires two distinct types of raw data. The first type is structured time-series data, which mainly covers the real-time utilization percentage of the CPU, the number of bytes occupied by memory pages, the waiting time in milliseconds for disk I / O operations, and the real-time operating temperature of key components of the equipment collected by thermistors. This data flows into the data bus at a fixed millisecond-level sampling frequency.

[0031] The second category is unstructured text data, which includes strings of operating log information automatically generated by the operating system, work order texts recording historical maintenance actions, qualitative descriptions of system anomalies manually entered by computer room staff through management terminals, computer room environment image recognition annotation text output by computer audiovisual software, and computer room personnel identification event text output by biometric identification software.

[0032] After acquiring the raw data, the system normalizes the structured time-series data. This normalization process maps all continuous performance indicators to a standardized range between 0 and 1 by retrieving the maximum and minimum values ​​within the historical operating cycle, thus eliminating the influence of different physical dimensions on the accuracy of subsequent calculations. Next, using a sliding window feature extraction technique, within a preset 60-second time window, the mean, variance, skewness, and kurtosis of each indicator are calculated to generate a continuous state vector reflecting the device's performance characteristics. This vector accurately characterizes the fluctuation characteristics of the server's computing load and the evolution trend of its thermal cycle.

[0033] Concurrently, for unstructured text data, this embodiment utilizes natural language processing (NLP) techniques for in-depth data mining. First, a dedicated dictionary for the operations and maintenance domain is constructed to segment log and work order texts. Then, named entity recognition (NAME) technology is used to locate faulty component names, error codes, and operational actions. Further, event extraction logic identifies the causal semantics inherent in the text. For example, when a log describes a decrease in frequency due to temperature exceeding a threshold, the system converts this into discrete intervention signal features. This feature will serve as a proxy variable for human intervention or automatic system protection actions in subsequent logic. Finally, the continuous state vector and discrete intervention signal features are input together into a cross-modal feature fusion network constructed based on gated recurrent units. This network, through internal nonlinear transformations and weight allocation, compresses and maps the diverse original information into a fixed-dimensional embedding space vector, outputting a multi-dimensional heterogeneous fusion feature representing the current real-time operating state of the system.

[0034] In one specific embodiment, the process of generating multi-heterogeneous fusion features in step 1 specifically includes the following closed-loop data flow: First, the continuous state vector extracted by normalization and sliding window is denoted as... The discrete intervention signal features obtained through entity recognition and event extraction are mapped to... through the Embedding layer. Subsequently, the two are directly concatenated along the feature dimension to obtain the joint input vector. The concatenated vector is input into a gated recurrent unit (GRU).

[0035] Specifically, the GRU operates at time step using the following formula. Nonlinear feature compression and weight allocation:

[0036]

[0037]

[0038]

[0039] in, This represents the update gate vector, used to control the state at the previous time step. The degree of preservation; This represents the reset gate vector, used to control the degree to which historical states are ignored; This represents the hidden state of the candidate. This represents the Sigmoid activation function; Represents the Hadama product; and These represent the corresponding network weight matrix and bias vector, respectively. Through the above calculations, the GRU outputs the final hidden state. ,Should That is, the multi-dimensional heterogeneous fusion features, which represent the real-time operating state of the system, are directly passed to the causal structure learning module as input for node attributes.

[0040] Step 2 involves mining the aforementioned embedded space vectors using a causal structure learning algorithm to construct a dynamic causal graph reflecting the causal relationship between equipment operating status and environmental factors. This step is one of the core logics of this invention, aiming to eliminate spurious associations from massive statistical correlations. A constraint-based path constraint algorithm is then used to initially screen the fused feature set. The path constraint algorithm performs a high-order conditional independence test between feature pairs to determine whether there is a statistically significant direct relationship between two features given other sets of variables. If two features remain independent under all possible conditional sets, their logical connection is severed. This process quickly eliminates spurious correlations caused by common-cause interference, generating an initial undirected graph structure.

[0041] As a preferred implementation, step 2, which uses a constraint-based path constraint algorithm to generate the initial undirected graph, includes the following mathematical verification logic: For any two node features in the multi-heterogeneous fusion feature and The system calculates its condition set. Partial correlation coefficient The calculation formula is as follows:

[0042] in, Let be the feature variable of the k-th condition node in the condition set Z. This represents the set of remaining conditions after removing the feature variable of the k-th condition node from the condition set Z; if (in If the preset significance level threshold is used (in this embodiment, the value is 0.05), then it is determined that... and If the conditions are independent, cut the connecting edge between the two in the undirected graph.

[0043] Furthermore, when applying a score-based greedy equivalence search algorithm to optimize the initial undirected graph, the following Bayesian information criterion scoring function is used:

[0044] in, Represents the candidate cause-effect graph structure. This represents the observed fused feature dataset. Represents a node In the figure The set of parent nodes in the middle, This represents the local parameters obtained from the maximum likelihood estimation. Representation diagram The total number of directed edges in the region. This represents the total number of samples. The system continuously performs edge addition, deletion, or reversal operations until... The value of no longer increases, thus the initial causal structure is output and passed to the graph neural network module.

[0045] A score-based greedy equivalence search algorithm is applied to optimize the initial undirected graph. In this process, the system starts from an empty graph with no connections and attempts to add a directed edge in each iteration that maximizes the model score. The model score is based on the maximum Bayesian information criterion, which, while evaluating the model's interpretability of the observed data, penalizes the number of edges in the graph to prevent overly complex overfitting structures. By continuously trying to add edges, remove redundant edges, or reverse the direction of edges, the algorithm eventually converges to a locally optimal directed acyclic graph structure, thus forming the initial causal structure. The initial causal structure clearly defines the causal chain, such as a surge in request volume leading to increased CPU load, which in turn causes an increase in device temperature.

[0046] To address the dynamically changing operating environment of data centers, this embodiment further introduces a graph neural network to dynamically update the initial causal structure. Multi-dimensional heterogeneous fusion features are input as node attributes into a spatiotemporal graph convolutional network. This network includes temporal convolutional layers and spatial graph convolutional layers. The temporal convolutional layer captures the evolution pattern of individual monitoring indicators along a preset timeline, such as identifying the periodic increase in CPU load. The spatial graph convolutional layer aggregates the feature information of neighboring nodes based on the connection strength of causal edges. For example, when the temperature of a computing node rises, the cooling fan speed feature causally related to it automatically receives a higher aggregation weight. By introducing an attention mechanism, the system can automatically identify which causal paths have the greatest impact on system stability in the current state. Through residual connections and normalization layer processing, a dynamic node feature representation containing predictive performance is output, ultimately generating a dynamic causal graph that reflects the system's causal logic in real time. In the dynamic causal graph, nodes represent various monitoring indicators, directed edges represent causal driving directions, and edge weights quantify the strength of this driving relationship.

[0047] When constructing a dynamic causal graph that reflects the physical logic of the system in real time, the graph neural network performs the following graph convolution operation: First, construct the adjacency matrix based on the initial causal structure. (in To monitor the total number of indicator nodes, if the number of nodes Pointing to node but (otherwise it is 0), and a degree matrix is ​​introduced. Symmetric normalization is performed to obtain the renormalized spectral Laplacian matrix. :

[0048] In the spatial graph convolutional layer, the node feature matrix of the previous layer is... (in Using the feature dimension as input, the causal feature information of neighboring nodes is aggregated using the following formula:

[0049] in, The spatial convolution weight matrix, This is the ReLU activation function.

[0050] In the temporal convolutional layer, a one-dimensional extended convolution is used to pair the spatial feature matrix. In the time dimension

[0051] The evolution trend of the above extraction:

[0052] in, The preset time window size is The convolutional kernels are then processed through residual connections and normalization layers. ,Should That is, as a dynamic node feature representation containing predictive performance, it is passed to the game decision-solving module as input to the state space.

[0053] Next, step 3 is executed, constructing a multi-agent game decision-making model based on a dynamic causal graph. In the data center operation and maintenance scenario, this embodiment defines each core hardware component in the system, such as a single server, an independent cooling control unit, and an uninterruptible power supply module, as a controlled agent. Each controlled agent is assigned a specific local utility function, which comprehensively considers the instantaneous energy consumption of the target core hardware component, the predicted remaining lifespan based on operating time and workload, and the throughput efficiency of the current task execution. Simultaneously, the global operation and maintenance optimization goal of the data center is defined as the leader agent. The global goal of the leader agent is set to minimize the total operating cost of the entire data center by formulating a global resource allocation scheme or incentive strategy, while strictly meeting the preset service level agreement.

[0054] In this game theory model, a Stackelberg game relationship is formed between the leader and the controlled agents. The leader first issues a global policy vector, such as power consumption limits for different regions. Subsequently, each controlled agent, acting as a follower, receives this signal and makes a decision based on the physical evolution revealed by a dynamic causal graph. For example, upon receiving a power reduction instruction, a server agent will predict the impact of reducing the clock speed on task latency using a dynamic causal graph and, combined with its own lifetime decay model, select the optimal operating frequency. Through this hierarchical game architecture, the system can simulate the interaction and feedback between global goals and local interests, ensuring that the behavior of each component tends towards the global optimum while adhering to physical causal logic.

[0055] In step 3, the operational physical logic must be transformed into a rigorous mathematical programming problem. For the first... A controlled intelligent agent (follower) has a local utility function It is defined as the following weighted sum:

[0056] in, Controlled intelligent agent Action vectors (such as adjusting frequency and allocating computing power). For the set of actions of other intelligent agents, The global policy vector issued to the leader agent; Instantaneous operating energy consumption This is the rated maximum power consumption; This is the predicted remaining useful life. This is a reference value for the design life. For task execution throughput; The weighting coefficients and For the sake of obedience to the leader's strategy The incentive compensation received.

[0057] For the leader agent, its global optimization objective function is... Set as:

[0058] in, For the first The operating cost of each component In a given leader strategy The optimal response action of a controlled agent when it reaches Nash equilibrium. To ensure the overall service level agreement compliance rate of the system, This represents the preset minimum business assurance threshold. The above function establishes a strict hierarchical constraint relationship, which directly serves as the basis for generating reward signals in reinforcement learning.

[0059] Finally, step 4 is executed, using a multi-agent reinforcement learning algorithm to solve for the Nash equilibrium state of the game model and outputting the optimal joint operation and maintenance decision instruction. This embodiment employs a deep deterministic policy gradient algorithm with a centralized training and distributed execution architecture. During the training phase, the system initializes the policy network and value network parameters of each agent. The leader agent outputs a global policy vector of a preset dimension based on the current dynamic causal graph state. Upon receiving the signal, each controlled agent outputs its own action selection, such as adjusting fan speed or migrating computing tasks. The system then feeds back a comprehensive reward signal based on the actual impact of the joint actions of all agents on the global environment.

[0060] To overcome the non-stationarity problem in multi-agent environments, this embodiment utilizes a global critic network to aggregate the observation and action information of all agents for unified evaluation. In each iteration, the gradient descent algorithm is used to calculate the rate of change of the loss function with respect to the policy network parameters, and the network weights are updated accordingly. This process is repeated until the policies of all agents reach Nash equilibrium. In this state, no controlled agent can gain higher local utility by changing its own actions without altering the policies of other agents, and the global cost of the leader is also minimized. In the actual execution phase, each agent only needs to use its own perceived local observation data and the received synchronization signal from the leader to output the optimal operation and maintenance instructions in real time within milliseconds, achieving intelligent and precise control of the computing power cluster.

[0061] In step 4, when the deep deterministic policy gradient algorithm with a centralized training and distributed execution architecture is used to solve the problem, the core lies in the collaborative iteration of the commentator network and the actor network.

[0062] During the intensive training phase, the global commentator network targets the first... Loss function of each agent The timing difference error is defined as follows:

[0063] Among them, the target value The calculation formula is:

[0064] in, The global state that the system transitions to at the next moment after executing the current joint action; For the next time step, all agents will base their strategies on the target actor's policy. Output target joint action vector; This represents the current global state (composed of node features from a dynamic cause-effect graph). This is the joint action vector for all agents. For the first The local reward received by each agent (as defined by the aforementioned utility function) (Calculated) This is the discount factor, with a value range of [0.9, 0.99]. For the current commentator network parameters, For target commentator network parameters; This indicates the target actor's strategy.

[0065] For a controlled agent (follower) policy network, its gradient update formula strictly follows the Stackelberg follower response logic:

[0066] in, For intelligent agents The local observation status, These are the network parameters for the actors. Specifically, during action selection, the agent... The global strategy vector issued by the leader With its own local observation The concatenated data serves as the input to the actor network. Through backpropagation and gradient descent using the aforementioned loss function, the process continues until the policy gradients of all agents approach zero. and Once the Nash equilibrium state is reached, the actor network directly outputs the optimal joint operation and maintenance decision instruction to the execution monitoring module.

[0067] Example 2: This example applies the method of the present invention to the operation and maintenance scenario of a smart factory in the field of precision electronics manufacturing. The method is compatible with artificial intelligence optimized operating systems, artificial intelligence middleware, and function libraries, representing a practical application scenario of artificial intelligence application software development in the industrial manufacturing field. Unlike data centers, the core of smart factory operation and maintenance lies in predictive maintenance of robotic arms, conveyor belts, and precision machine tools on the production line. Its diverse and heterogeneous data encompasses the entire process of operation and maintenance related data output by computer vision and audiovisual software and biometric recognition software. The physical meaning and causal logic of the data have stronger spatiotemporal coupling.

[0068] In the data acquisition phase of step 1, this embodiment acquires structured time-series data through a high-speed acquisition gateway deployed in an industrial Ethernet network. This data specifically includes real-time current waveforms of the motors at each joint of the robotic arm, the vibration frequency spectrum of the spindle rotation, pressure fluctuations in the hydraulic system, and environmental humidity parameters. Unstructured text data primarily originates from qualitative fault descriptions such as lead screw wear and insufficient lubrication in the equipment maintenance log, as well as defect classification and identification results from the machine vision inspection system.

[0069] When preprocessing these heterogeneous data, for high-frequency signals such as vibration frequencies, the system uses wavelet transform for sliding window feature extraction to generate an energy distribution vector that reflects the health status of the mechanical structure. For discrete labels from visual inspection outputs and text in maintenance logs, a pre-trained industrial semantic encoder is used to map them into intervention signal features. For example, adding lubricating oil is identified as an intervention action, which has a specific directionality in the feature space and is used to simulate the improvement effect of maintenance actions on equipment status. The cross-modal feature fusion network deeply fuses continuous values ​​from physical sensors with qualitative maintenance action features to form a fused feature vector that can characterize the health evolution trend of the production line.

[0070] In the causal structure learning step 2, considering the strong interference characteristics in the industrial environment, this embodiment introduces a higher level threshold in the path constraint algorithm to ensure that only statistically extreme physical connections are retained as edges in the graph. For example, analysis reveals a clear causal path between small fluctuations in ambient humidity and circuit board mounting defect rates, rather than a simple statistical correlation. When applying the greedy equivalence search algorithm for structure optimization, the system introduces a prior knowledge base based on industrial mechanisms. This knowledge base, acting as a verification unit, imposes a significant penalty weight on causal edges that do not conform to the principles of mechanical transmission; for example, it strictly prohibits causal directions that violate the physical time sequence, such as yield affecting motor current. The resulting dynamic causal graph can reflect the deep logic between production process parameters and equipment losses in real time.

[0071] In step 3, when constructing the game model, each machine tool is defined as a controlled agent. Its local utility function is configured to focus on minimizing the weighted sum of the average processing cost per unit and the risk of unplanned equipment downtime. The leader agent is the factory's production scheduling center, whose global objective is to optimize the energy efficiency of the entire production line while meeting delivery deadlines. The dynamic cause-effect graph, as a state transition constraint, specifies physical limitations, such as the inevitable increase in spindle speed leading to increased tool wear, ensuring that the decisions of each agent are made within physically feasible limits.

[0072] In the reinforcement learning solution process of step 4, to enhance the foresight of the decision-making, this embodiment introduces a prediction step size of 100 time steps. When each agent calculates the response strategy, it first uses a dynamic causal graph to perform multi-step forward simulation in a virtual environment to evaluate the impact of the current maintenance instruction on the evolution of the equipment's health status over the next week. Through this policy iteration based on causal simulation, the Nash equilibrium instruction finally output by the system can not only address the current failure risk, but also automatically compensate for potential mechanical fatigue by fine-tuning the current production parameters.

[0073] Example 3: This example further illustrates an intelligent operation and maintenance decision-making system based on multi-source heterogeneous data fusion involved in this invention. This system is an engineering implementation platform for artificial intelligence application software development, adapted for deployment and operation of AI-optimized operating systems, AI middleware, and function libraries. Through a specific hardware architecture and the collaborative efforts of software modules, this system achieves the engineering implementation of the aforementioned methods.

[0074] The core component of the system is the data acquisition and preprocessing module. Physically, this module consists of embedded acquisition daughter cards distributed throughout the monitored system. These daughter cards possess high-precision analog-to-digital signal conversion capabilities, enabling sampling of physical quantities such as voltage and current at frequencies no lower than 10 kHz. Simultaneously, the data acquisition and preprocessing module also includes a high-performance log parsing engine, specifically responsible for extracting asynchronously generated text information from the system buffer in real time. The preprocessing unit, through a built-in field-programmable gate array (FPGA) chip, achieves real-time normalization and feature extraction of large-scale data, ensuring extremely low latency in data processing.

[0075] The multimodal feature fusion module is connected to the output of the data acquisition module. Internally, the multimodal feature fusion module integrates a deep neural network accelerator, specifically designed to run cross-modal coding algorithms adapted to AI-optimized operating systems, AI middleware, and function libraries. It is compatible with processing multi-dimensional operational data output from computer vision and biometric recognition software. It receives continuous and discrete features from the preprocessing unit via a high-speed bus and projects them into a high-dimensional manifold space. The output of this module is a highly cohesive, low-redundancy fused feature vector, providing a standardized data foundation for subsequent inference.

[0076] The causal structure learning module is the system's logical brain. It consists of a causal inference computation cluster, pre-installed with path constraint algorithm plugins and a greedy search optimization engine. A key feature of this module is its knowledge base verification unit. This unit stores a wealth of industry expert experience and physical laws. During the construction of the dynamic causal graph, the knowledge base verification unit scans the generated directed edges in real time and compares them with the established physical logic. If the logic automatically generated by the algorithm conflicts with known physical common sense, the knowledge base verification unit will force the algorithm to re-search for a more reasonable structure by adjusting the penalty items in the scoring criteria. The causal structure learning module also includes dynamic update triggers. When the system senses a sudden state change signal, such as a voltage surge or a critical alarm trigger, it immediately initiates a real-time update process to recalibrate the edge weights of the causal graph, ensuring the timeliness of decision-making.

[0077] The game theory decision-making module is the core of instruction generation and execution. Internally, it comprises multiple virtualized controlled agent units and a leader control unit. Each controlled agent unit has independent storage space to store its own policy network parameters and weight coefficients of the local utility function. These agents possess adaptive learning capabilities, allowing them to adjust the weights of various indicators in the utility function in real time based on equipment aging data fed back from the execution monitoring module. For example, when the bearing vibration value is detected to be approaching a critical point, the controlled agent unit will automatically increase the priority of the remaining lifespan term, tending to choose a more conservative and protective action plan in the game.

[0078] The execution monitoring module, serving as the closed-loop feedback link of the system, is responsible for issuing the optimal joint operation and maintenance commands output by the game decision-making module to the physical device actuators. Simultaneously, the execution monitoring module uses sensors distributed on the actuators to monitor the physical feedback effects after command execution in real time, and transmits these results back to the data acquisition and preprocessing module in real time, forming a complete closed-loop control chain.

[0079] In this embodiment, by combining the real physical logic identified through causal inference with a global collaborative strategy optimized by game theory during system operation, transparent management of complex systems is achieved. Maintenance personnel no longer face inexplicable black-box predictions, but instead, through the dynamic causal graph displayed in the system, intuitively see the logical chain leading to the failure and the cost-benefit balance point behind each decision instruction. This high degree of interpretability, combined with the global optimization capability under Nash equilibrium, enables the system to exhibit high stability and decision-making accuracy when handling large-scale computing clusters or automated production lines.

[0080] In this invention, the dynamic causal graph update frequency is set to a preset time interval, such as a full structure check every 10 minutes. Between two intervals, the weights of the edges are fine-tuned at the millisecond level only through the attention mechanism of the graph neural network. This design ensures that the system can keenly capture subtle environmental disturbances while avoiding the huge computational overhead caused by frequent graph structure reconstruction. When calculating the agent's response strategy, the introduced prediction step size allows the system to anticipate certain causal effects with lag, such as the long-term impact of heat accumulation on the failure rate of electronic components. This forward-looking decision-making mechanism enables the system to transform from a traditional passive fault response mode to a truly proactive predictive intelligent operation and maintenance.

[0081] Through the detailed description of the above embodiments, it can be seen that the present invention, by deeply fusing diverse heterogeneous data and mining causal logic, combined with multi-objective optimization using Stackelberg games, successfully solves the problems of poor robustness, weak interpretability, and difficulty in achieving global optimum when dealing with complex and compound faults in existing operation and maintenance systems. All logical operations, data transformations, and parameter optimization processes are described in detail using rigorous engineering natural language, ensuring that those skilled in the art can fully understand and implement the various technical solutions of the present invention.

[0082] This embodiment also focuses on how to transform human intervention signals into intervention nodes in a dynamic causal graph. In actual operation and maintenance processes, every manual optimization or maintenance replacement by the operator is captured by the system in the form of logs. The system calculates the changes in the conditional probability distribution of each node in the graph after the intervention to quantitatively assess the contribution or potential risk of the human action to the overall stability of the system. For example, through simulation calculations, it can predict whether adding a backup server now can alleviate the input / output waiting pressure of the current network switch. This simulation function based on intervention calculations provides a reliable digital twin environment for operation and maintenance decisions, greatly reducing the risks and costs of conducting operation and maintenance experiments on real physical systems.

[0083] In the execution of multi-agent reinforcement learning, global information acquired through a centralized commentator network can eliminate decision-making oscillations caused by the limited perspective of individual devices. For example, when a cooling system experiences a significant increase in power consumption due to localized high temperatures, the leader agent can coordinate computational agents to reduce the load on this high-temperature area from a global energy efficiency perspective, achieving seamless collaboration between subsystems. This global collaborative capability is a key guarantee for the sustainable and efficient operation of smart factories and large data centers.

[0084] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention; however, any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An intelligent operation and maintenance decision-making method based on the fusion of diverse heterogeneous data, characterized in that, The method is adapted to AI-optimized operating systems, AI middleware, and function libraries, and includes the following steps: S1. Acquire the multi-dimensional heterogeneous data stream of the system to be monitored; S2. Use a constraint-based path constraint algorithm to perform preliminary screening of multi-dimensional heterogeneous fusion features and generate a dynamic causal graph that reflects the physical logic of the system in real time; S3. Construct a multi-agent game decision-making model based on the dynamic causal graph, define the hardware components or functional modules in the system to be monitored as controlled agents, assign local utility functions to the controlled agents, define the global optimization objective of operation and maintenance as the leader agent, guide the decision-making behavior of the controlled agents by formulating a global policy vector, establish state transition constraints between the controlled agents based on the dynamic causal graph, and form a Stackelberg game relationship between the leader agent and the controlled agents; S4. Initialize the policy network and value network parameters of each agent, use the multi-agent reinforcement learning algorithm to solve the Nash equilibrium state of the multi-agent game decision model, and output the optimal joint operation and maintenance decision instruction.

2. The intelligent operation and maintenance decision-making method based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The multi-dimensional heterogeneous data stream includes structured time-series data and unstructured text data. The structured time-series data covers structured detection and recognition results output by computer vision and audiovisual software and biometric recognition software. Normalization processing and sliding window feature extraction are performed on the structured time-series data to generate a continuous state vector reflecting the performance characteristics of the device. Entity recognition and event extraction are performed on the unstructured text data to transform it into discrete intervention signal features. The continuous state vector and the intervention signal features are input into a cross-modal feature fusion network adapted to an AI-optimized operating system, AI middleware, and function library. Through nonlinear transformations within the cross-modal feature fusion network, information from different dimensions is mapped to a unified-dimensional embedding space, outputting multi-dimensional heterogeneous fusion features that characterize the real-time operating status of the system under monitoring.

3. The intelligent operation and maintenance decision-making method based on multi-source heterogeneous data fusion according to claim 1, characterized in that, By performing high-order conditional independence tests between feature pairs, the conditional independence relationships between nodes are determined to generate an initial undirected graph. A score-based greedy equivalence search algorithm is applied to score and optimize the initial undirected graph. The directed edge directions between nodes are determined by the maximization Bayesian information criterion to form an initial causal structure. A graph neural network is introduced to dynamically update the initial causal structure. An attention mechanism is used to capture the nonlinear evolution of node features over time, generating a dynamic causal graph that reflects the physical logic of the system in real time.

4. The intelligent operation and maintenance decision-making method based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The leader agent outputs a global policy vector of a preset dimension based on the current state of the dynamic causal graph. Each controlled agent, as a follower, outputs an action selection based on the local utility function after receiving the global policy vector. The system outputs a reward signal based on the joint action and iteratively updates the parameters using the gradient descent algorithm until the policies of all agents reach an equilibrium point where they cannot obtain higher returns by changing their own policies without changing the policies of other agents.

5. The intelligent operation and maintenance decision-making method based on multi-source heterogeneous data fusion according to claim 2, characterized in that, The acquired structured time-series data includes CPU utilization, memory usage, input / output latency, real-time operating temperature of the device, structured data of device defect detection output by computer audiovisual software, and structured data of operation and maintenance permission verification output by biometric recognition software. The unstructured text data includes equipment operation log information, historical maintenance work order records, system status descriptions entered into the management terminal, image recognition annotation text output by computer audiovisual software, and identity recognition event text output by biometric recognition software.

6. The intelligent operation and maintenance decision-making method based on multi-source heterogeneous data fusion according to claim 2, characterized in that, The process of generating the multi-heterogeneous fusion features includes: the cross-modal feature fusion network is constructed based on a gated recurrent unit, adapted to the operation of an AI-optimized operating system, AI middleware, and function library; the gated recurrent unit performs feature compression and weight allocation on the received continuous state vector and the intervention signal features; and projects the heterogeneous original information into the embedding space with a fixed dimension to achieve an integrated representation of the real-time operating state.

7. The intelligent operation and maintenance decision-making method based on multi-source heterogeneous data fusion according to claim 3, characterized in that, The process of applying the score-based greedy equivalence search algorithm includes: starting from an empty graph structure, attempting to add directed edges that can improve the model score in each iteration, entering the pruning phase when the score stops growing and attempting to remove redundant directed edges, and performing operations of adding, deleting, or reversing edges in the graph space until convergence to a locally optimal equivalence class graph structure. The model score is determined by applying a penalty term to the number of edges in the graph, based on an evaluation of the model's ability to interpret the observed data.

8. The intelligent operation and maintenance decision-making method based on multi-source heterogeneous data fusion according to claim 3, characterized in that, The process of dynamically updating the initial causal structure using a graph neural network includes: inputting the multivariate heterogeneous fusion features as node attributes into a spatiotemporal graph convolutional network; using a temporal convolutional layer to capture the evolution trend of a single performance indicator within a preset time window; using a spatial graph convolutional layer to aggregate the feature information of neighboring nodes based on the connection strength of causal edges; and processing the output of the spatiotemporal graph convolutional network through residual connections and normalization layers to generate a dynamic node feature representation containing the predicted performance.

9. The intelligent operation and maintenance decision-making method based on multi-source heterogeneous data fusion according to claim 4, characterized in that, The multi-agent reinforcement learning algorithm employs a deep deterministic policy gradient algorithm with a centralized training and distributed execution architecture. During the training phase, a global critic network is used to aggregate the observation and action information of all agents for unified value assessment. During the execution phase, each agent outputs the optimal joint operation and maintenance decision instruction in real time based on local observation data and the received global policy vector.

10. An intelligent operation and maintenance decision-making system based on the fusion of diverse heterogeneous data, characterized in that, To implement the method of any one of claims 1 to 9, adapted for deployment and operation of AI-optimized operating systems, AI middleware, and function libraries, and belonging to the carrier of AI application software development technology, it includes: a data acquisition and preprocessing module, configured to acquire and synchronize multi-dimensional heterogeneous data streams of the system to be monitored, wherein the multi-dimensional heterogeneous data streams include structured time-series data and unstructured text data, and perform normalization and feature extraction on the structured time-series data, and perform entity recognition and event extraction on the unstructured text data; a multi-modal feature fusion module, the input end of which is connected to the output end of the data acquisition and preprocessing module, and internally integrates a deep neural network accelerator, configured to map data of different formats to a unified-dimensional embedding space, and output multi-dimensional heterogeneous fusion features representing the real-time operating status of the system; The causal structure learning module is configured to construct a dynamic causal graph using a causal inference algorithm. This module is connected to the multimodal feature fusion module. The causal structure learning module includes a knowledge base verification unit, which injects preset prior logical relationships as constraints into the causal structure search process and verifies the causal discovery process by applying penalty weights to directed edges that do not conform to physical common sense. The game theory decision-making module receives the dynamic causal graph and the multimodal heterogeneous fusion features at its input end. It internally contains multiple controlled agent units and a leader control unit, and is configured to establish a game model based on the dynamic causal graph and use a multi-agent reinforcement learning algorithm to solve for the Nash equilibrium state to output the optimal joint operation and maintenance decision command. The execution monitoring module receives the commands output by the game theory decision-making module and applies them to physical devices. Simultaneously, it collects the feedback results after execution and transmits them in real time to the data acquisition and preprocessing module, forming a closed-loop control loop.

Images