Unit fault risk scoring method and device based on credential technology stack, and medium

Abstract

The invention provides a unit fault risk scoring method based on a creative technology stack, and the method comprises the steps: carrying out the data entity modeling, constructing an entity relation model of multi-source heterogeneous data, defining a structured data set, carrying out the recursive automatic construction through aggregation and conversion primitives, generating a high-order candidate feature set, and carrying out the evaluation of a unit fault risk. Learning a causal graph among the features in the high-order candidate feature set based on a microcausal discovery algorithm, generating a causal graph among the features, identifying Markov coverage of a target variable according to the causal graph among the features, screening an effective causal feature subset, constructing an association graph by taking the effective causal feature subset as a node, and aggregating feature interaction information to generate a global fusion feature; dynamically and quantitatively calculating a real-time fault risk score of the computer unit based on the global fusion features; according to the method, the effects of deep utilization of the multi-source data and quantitative sorting of the fault risks are achieved through the creative and creative adaptive process and the continuous differentiated scoring model.

Description

Technical Field

[0001] This invention relates to the technical field of industrial big data analysis and the application of domestic information technology innovation, specifically a method, equipment and medium for unit failure risk scoring based on the information technology innovation stack. Background Technology

[0002] The development of wind turbine fault diagnosis and health management technologies has consistently kept pace with the advancements in industrial data processing capabilities and the safety requirements of the energy sector. In the early stages of this technology, monitoring relied primarily on engineers' understanding of physical mechanisms and on-site experience. Equipment status was judged by setting simple thresholds, which, while intuitive, could only address a few overt faults. With the widespread adoption of SCADA systems and vibration sensors, data-driven approaches gradually emerged, and traditional machine learning algorithms began to be used for fault identification, reducing reliance on single experiences to some extent. The introduction of deep learning technology allowed models to capture more complex data correlations; however, at this stage, mainstream solutions generally used foreign frameworks such as TensorFlow and PyTorch paired with NVIDIA GPUs, resulting in a deep dependence on foreign hardware and software.

[0003] In recent years, the national strategy of independent control over critical infrastructure has been continuously promoted, and the energy industry has an increasingly urgent need for information technology innovation. However, existing monitoring technologies have not yet formed mature solutions that are compatible with domestic software and hardware, making it difficult to balance computing autonomy, model performance, and operational practicality.

[0004] Traditional methods rely heavily on expert experience for feature construction, making it difficult to automatically mine high-order nonlinear features in multi-source heterogeneous data, resulting in insufficient feature coverage and depth. Furthermore, most models only output binary labels of "normal / fault," lacking a dynamic and continuous risk scoring mechanism, failing to quantify the severity and development trend of faults, and making it difficult to guide maintenance personnel in prioritization, thus limiting their practical application value. Summary of the Invention

[0005] This invention provides a unit failure risk scoring method, equipment, and medium based on the domestic IT innovation technology stack. Through an automated feature mining-causal screening process adapted to domestic IT innovation and a continuous differentiated scoring model, it achieves in-depth utilization of multi-source data and quantitative ranking of failure risks, thereby solving the problems in the background technology.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows: A unit failure risk scoring method based on the domestic information technology stack, which performs the following steps through computer equipment: S1, Data Entity Modeling: Based on the openGauss database, an entity relationship model for multi-source heterogeneous data is constructed, defining association keys and integrating structured datasets including SCADA, logs, and vibration data. S2, Deep Feature Synthesis, is based on structured datasets and generates high-order candidate feature sets through recursive and automated construction using aggregation and transformation primitives. S3, Causal Structure Learning, learns the causal graph between features in a high-order candidate feature set based on a differentiable causal discovery algorithm, and uses matrix computation units to accelerate the optimization process and learn to generate causal graphs between features. S4, Key Feature Filtering: Identify Markov Coverage of Target Variable Based on Causal Graph Between Features, Eliminate False Correlation Features, and Filter Effective Causal Feature Subset; S5, Graph Attention Fusion, constructs an association graph with an effective subset of causal features as nodes, learns the attention weights between features through a graph attention network, and aggregates feature interaction information to generate a global fused feature. S6, Risk Score Calculation, dynamically quantifies the real-time fault risk score of the computer group based on global fusion features and attention weights between features.

[0007] Optionally, the specific implementation steps of S1 are as follows: S11, Building a domestic IT database environment: Deploying the database on the operating system and configuring operating parameters to adapt to the time-series data storage of wind turbine units; S12, Define the data entity table structure: Based on the monitoring requirements of wind turbine units, create data tables, equipment log tables, and vibration data tables in the database, and clarify the data types and constraint rules of each table field; S13, Set cross-table association key: Use the unit's unique number and data acquisition timestamp as the joint association key to establish a logical association between the data table, equipment log table and vibration data table, ensuring that multi-source data can be queried by unit and time dimension; S14, Multi-source data import and standardization: Import scattered raw monitoring data, including real-time sensor data, equipment operation logs and vibration waveform data, into the corresponding data table according to the table structure, standardize the data format, and remove invalid and redundant data. S15, Storage Optimization Configuration: In response to the query characteristics of wind turbine time series data, columnar storage mode is enabled for each data table, and the final output is a standardized structured dataset for feature synthesis.

[0008] Optionally, the specific implementation process of S2 is as follows: S21, Configure the basic parameters for feature construction, call the DFS algorithm tool based on the MindSpore framework to determine the set of feature construction primitives, and set the maximum depth of feature construction to 3; S22, Define the time series aggregation window: For the standardized structured dataset output by S1, set three time sliding windows to adapt to the needs of capturing the operating characteristics of wind turbines at different time scales; S23, Initialize the first-order original features, using the original monitoring features in the output dataset of S1 as the first-order features to form the basic data layer for feature construction; S24, recursively generate higher-order features, performing recursive construction at a preset maximum depth of 3. For each order of feature, the feature generation formula is:

[0009] in, express Moment The feature is of order d, where d represents the feature dimension. The operation functions are selected from the primitive set. express Rank features in time window Data within, The length of the time window; S25, Feature integrity verification, performs validity screening on all recursively generated features, removes invalid features with missing value ratios exceeding the threshold or abnormal values, and retains valid features; S26. Integrate the candidate feature set, summarizing the first-order original features and effective higher-order features of each order to form a candidate feature set containing higher-order features. This provides sufficient features as input for subsequent causal structure learning.

[0010] Optionally, the specific implementation process of S3 is as follows: S31, Configure the domestically developed accelerated computing environment, deploy the MindSpore framework and operator library, and process the candidate feature set. Transform it into a data matrix X, which serves as input data for causal discovery; S32, Define the optimization objective and constraints, and clarify the optimization objective for differentiable causal discovery as follows:

[0011] The constraints are:

[0012] in, Represents a data matrix. Let represent the characteristic adjacency matrix, and be the solution variables in the optimization objective. This represents the reconstruction error function. This represents the sparsity regularization coefficient, used to control the sparsity of the adjacency matrix A. Let A represent the column norm of matrix A, which is the sum of the absolute values ​​of all elements in the matrix. Represents the acyclic constraint function. Represents the matrix trace operation. It represents the Hadamah accumulation. Represents the matrix index; S33, Algorithm Deployment and Acceleration Adaptation, executes the core operations of the NOTEARS algorithm, including matrix multiplication, matrix exponent calculation and gradient backpropagation, enables hardware units, and optimizes matrix operation efficiency; S34, causal structure iterative learning, starts the differentiable causal discovery algorithm, iteratively solves the optimization objective based on the data matrix X, updates the adjacency matrix A by gradient descent, and uses constraints to ensure that the graph structure is acyclic, while controlling the sparsity of the matrix by column and norm regularization. S35 outputs a feature causal relationship graph. After iterative convergence, threshold filtering and structure verification are performed on the adjacency matrix A to remove weakly related edges. Finally, a directed acyclic graph (DAG) is generated between features to represent the true causal relationship between each feature and the fault target, and between features.

[0013] Optionally, the specific implementation process of S2 is as follows: S41, based on the feature-directed acyclic graph (DAG) and adjacency matrix A, the target variable Y is determined to be the fault risk of wind turbine units, and a candidate feature set is imported. Establish a causal graph, and establish an association mapping between the target variable and the original features; S42, Define the Markov Coverage Screening Criterion: Based on causal graph theory, clarify the conditions that the Markov coverage of the target variable Y must satisfy: A subset of features consists of Y's parent node, features influencing Y, child nodes, features affected by Y, other parent nodes of child nodes (i.e., features that jointly influence child nodes with Y, i.e., spouse nodes), and indirectly key related features. Includes all prediction information for Y; S43. Traverse the DAG to locate associated features. Use a depth-first search algorithm to traverse the structure of the DAG and combine the causal association strength in the adjacency matrix A to identify the parent node set, child node set, and parent node child node set of the target variable Y. Merge the features corresponding to the three types of nodes to form an initial candidate feature set. ; S44, Conditional Independence Test Purification, for Each feature in the matrix undergoes a mutual information test to verify its conditional independence from Y. If a feature is conditionally independent of Y, then... If other features are independent of the Y condition, they are judged as spurious features and removed. S45, Redundant Feature Removal and Subset Determination, Calculation The variance inflation factor of the remaining features is used to remove highly redundant features with a VIF > 10, while retaining core causal features in the adjacency matrix A with a correlation strength higher than a threshold, thus forming the final effective causal feature subset. .

[0014] Optionally, the specific implementation process of S2 is as follows: S51, Configure the GAT training environment and prepare input data, based on the MindSpore framework and its GNN suite, import effective causal feature subsets. The feature matrix is ​​transformed into a node feature matrix; at the same time, a directed acyclic graph (DAG) is loaded to extract the causal relationships between features. S52, Construct the feature association graph structure: using Each feature in the graph is a graph node. Combined with the causal relationship edges in the DAG, a sparse graph based on the causal structure is constructed. Standard graph data objects are generated through the graph data interface of MindSpore GNN, and the mapping relationship between node set, edge set and node feature is defined. S53, define the core parameters of the GAT network, set the hidden layer dimension, number of attention heads and output layer dimension of the GAT network according to the task requirements, select LeakyReLU as the attention calculation activation function and Sigmoid as the feature aggregation activation function, and initialize the linear transformation weight matrix W and the learnable attention vector a. S54, calculate the attention coefficients between nodes and perform a linear transformation on the node feature matrix. , and These are the original feature vectors of node i and node j, respectively; Calculate the original attention coefficients of nodes i and j. :

[0015] The final attention weights are then obtained through Softmax normalization. :

[0016] in, This is the weight matrix. For attention vectors, This indicates a splicing operation. For nodes The neighborhood group, This represents the normalized attention weights; S55, feature aggregation based on attention weights: For each node i, based on its set of neighboring nodes... and the corresponding attention weights Aggregate the features of neighboring nodes to generate a high-dimensional feature representation of node i. :

[0017] in, This represents the Sigmoid activation function, which captures dynamic interaction information between features. S56, Generate a global fusion feature vector: Employ the mean-readout strategy to generate high-dimensional feature representations for all nodes. The features of all nodes are aggregated, and the mean vector of all node features is calculated to obtain the global fused features. .

[0018] Optionally, the specific implementation process of S6 is as follows: S61, Input Data and Model Loading: Based on Global Fusion Features The feature attention weights learned by GAT are used to read the effective causal feature subset. The real-time observations of each feature and the statistical distribution based on historical data of the training set are used to dynamically calculate the adaptive safety threshold of each feature and complete the adaptation preparation of the model and data. S62, Define scoring parameters and functions, and set the weight matrix for basic risk calculation based on task requirements. With bias term We select the Sigmoid function as the activation function, set the risk correction coefficient, and define the indicator function. S63, calculate the basic risk probability, and fuse global features. Input the basic risk calculation model;

[0019] Calculate the basic risk probability It represents the baseline risk level based on global feature association information; S64, calculate the risk correction term, combining attention weights, feature observations and safety thresholds to calculate the risk correction term and achieve differentiated correction; S65 outputs the final real-time risk score, integrates the basic risk probability and correction term, and ensures that the score is within the [0,1] interval through a normalization function or truncation function to obtain the final real-time risk score; the validity of the score result is verified, and the final output is the unit failure risk score that is directly used for operation and maintenance priority ranking.

[0020] Optionally, the specific formula for calculating the risk correction term ΔS in S64 is as follows:

[0021] in, Basic risk probability; This is the revised final risk score; This is a risk correction factor; A set of key risk characteristics; Size of the set; Features The corresponding global attention weights; Features The current observation value; Features The safety threshold; This is an indicator function that takes the value 1 when the condition is met and 0 otherwise. In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0022] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0023] As can be seen from the above technical solution compared with the prior art, the present invention has the following beneficial effects: 1. This invention constructs an automated high-order feature synthesis, causal structure learning, and key feature screening process that is adapted to the domestic information technology innovation environment. This completes the closed-loop process and achieves intelligent deep mining of multi-source heterogeneous data and full compatibility with domestic software and hardware. It solves the problems of traditional methods relying on expert experience and insufficient feature coverage and depth. It also achieves a balance between computational autonomy and model performance, which meets the strategic needs of independent control of key infrastructure in the energy industry.

[0024] 2. This invention designs a continuous scoring model that combines global fusion feature-based risk calculation with attention weight differentiation correction. It constructs a dynamic quantification formula by combining the importance of causal features and safety thresholds, achieving fine-grained characterization of fault risk and prediction of development trends. This replaces the traditional binary label output mode, realizes accurate ranking of operation and maintenance priorities, and solves the problem that traditional models cannot quantify the severity of faults and have limited practical application value. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the method steps of the present invention; Figure 2 This is a schematic diagram of the data transformation process of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0027] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention, but should not be used to limit the scope of the present invention.

[0028] This invention provides a unit failure risk scoring method based on the domestic information technology stack, such as... Figure 1 As shown, perform the following steps using a computer device: S1, Data Entity Modeling: Based on the openGauss database, an entity relationship model for multi-source heterogeneous data is constructed, defining association keys and integrating structured datasets including SCADA, logs, and vibration data. S2, Deep Feature Synthesis, is based on structured datasets and generates high-order candidate feature sets through recursive and automated construction using aggregation and transformation primitives. S3, Causal Structure Learning, learns the causal graph between features in a high-order candidate feature set based on a differentiable causal discovery algorithm, and uses matrix computation units to accelerate the optimization process and learn to generate causal graphs between features. S4, Key Feature Filtering: Identify Markov Coverage of Target Variable Based on Causal Graph Between Features, Eliminate False Correlation Features, and Filter Effective Causal Feature Subset; S5, Graph Attention Fusion, constructs an association graph with an effective subset of causal features as nodes, learns the attention weights between features through a graph attention network, and aggregates feature interaction information to generate a global fused feature. S6, Risk Score Calculation, dynamically quantifies the real-time fault risk score of the computer group based on global fusion features and attention weights between features.

[0029] Example: In this embodiment, a large onshore wind farm has deployed multiple high-power wind turbine units and has long relied on imported databases and commercial monitoring software for operation and maintenance. This has problems such as lack of independent calculation, reliance on expert experience for feature construction, and output of only binary labels for fault judgment. Insufficient feature coverage has repeatedly led to missed fault detections, and binary labels cannot distinguish the urgency of faults, resulting in chaotic allocation of operational resources. To align with the national strategy of independent control over critical infrastructure, this invention introduces a fault risk scoring method based on the domestic information technology stack. Figure 2 As shown, through data calculation, the final risk assessment result is obtained, realizing the operation and maintenance upgrade of domestic software and hardware adaptation, automated feature mining and dynamic risk quantification.

[0030] Data entity modeling and data integration for domestic IT innovation; Implementation: Deploy the openGauss 3.1.0 database on the Kylin V10 operating system and configure the time-series data storage parameters; For example, block size 8KB, cache 16GB; create SCADA data table, containing 20 fields such as unit number, acquisition timestamp, speed, etc.; equipment log table, containing 12 fields such as log type, fault code, etc.; vibration data table, containing 8 fields such as sensor identifier, vibration amplitude, etc. Establish cross-table relationships using unit number + collection timestamp as the joint association key; import 3 months of raw data, standardize the timestamp format, for example, uniformly set it to YYYY-MM-DD HH:MM:SS, and remove redundant data with missing values ​​exceeding 5%; enable columnar storage to optimize time-series queries.

[0031] Compared to traditional imported database solutions, it is fully compatible with domestic software and hardware environments, achieving computational autonomy. Through joint association keys and data regularization, it integrates scattered multi-source heterogeneous data (SCADA, logs, vibration) into a standardized dataset (e.g., 1.08 million samples × 40-dimensional original features), solving the problems of data dispersion and inconsistent formats in traditional methods. Moreover, columnar storage improves the query efficiency of time-series data by more than 60%, far exceeding traditional row storage solutions.

[0032] Configure the basic parameters for feature construction, call the DFS algorithm tool based on the MindSpore framework to determine the set of feature construction primitives, and set the maximum depth of feature construction to 3. Define time series aggregation windows: For the standardized structured dataset output by S1, set three time sliding windows to adapt to the needs of capturing the operating characteristics of wind turbines at different time scales; Initialize the first-order raw features, using the original monitoring features in the S1 output dataset as the first-order features to form the basic data layer for feature construction; Higher-order features are generated recursively, with recursive construction performed at a preset maximum depth of 3. The feature generation formula for each order of feature is as follows:

[0033] in, express Moment The feature is of order d, where d represents the feature dimension. The operation functions are selected from the primitive set. express Rank features in time window Data within, The length of the time window; Feature integrity verification involves filtering the validity of all recursively generated features, removing invalid features with missing values ​​exceeding the threshold or abnormal values, and retaining valid features. The candidate feature sets are integrated, and the first-order original features and effective higher-order features of each order are summarized to form a candidate feature set containing higher-order features. This provides sufficient features as input for subsequent causal structure learning.

[0034] Implementation: Based on the MindSpore framework, the DFS algorithm is called, with the primitive set set set as {Max, Mean, Std, Diff, CumSum}, and a maximum feature depth of 3. Three sliding windows are defined: 10min, 1h, and 24h. Higher-order features are recursively generated. For example, the Mean value of rotational speed within the 10min window is a second-order feature, and the CumSum value of vibration amplitude within the 24h window is a third-order feature. Invalid features with a missing value ratio exceeding 10% are removed, ultimately forming a candidate feature set. .

[0035] It replaces the traditional manual feature construction mode that relies on expert experience, automatically generates high-order nonlinear features, covering the short, medium and long-term operation characteristics of wind turbines, and solves the problem of insufficient feature coverage and depth of traditional methods; moreover, feature integrity verification ensures the reliability of subsequent model inputs without the need for manual screening.

[0036] Configure a domestically developed accelerated computing environment, deploy the MindSpore framework and operator library, and process the candidate feature set. Transform it into a data matrix X, which serves as input data for causal discovery; Define the optimization objective and constraints, and clarify the optimization objective for differentiable causal discovery as follows:

[0037] The constraints are:

[0038] in, Represents a data matrix. Let represent the characteristic adjacency matrix, and be the solution variables in the optimization objective. This represents the reconstruction error function. This represents the sparsity regularization coefficient, used to control the sparsity of the adjacency matrix A. Let A represent the column norm of matrix A, which is the sum of the absolute values ​​of all elements in the matrix. Represents the acyclic constraint function. Represents the matrix trace operation. It represents the Hadamah accumulation. Represents the matrix exponent; Algorithm deployment and acceleration adaptation: The core operations of the NOTEARS algorithm, including matrix multiplication, matrix exponent calculation and gradient backpropagation, are executed, hardware units are enabled, and matrix operation efficiency is optimized; Causal structure iterative learning initiates a differentiable causal discovery algorithm, iteratively solves the optimization objective based on the data matrix X, and updates the adjacency matrix A through gradient descent. During the process, constraints are used to ensure that the graph structure is acyclic, while column and norm regularization are used to control the sparsity of the matrix. The output feature causal relationship graph is obtained. After iterative convergence, the adjacency matrix A is subjected to threshold screening and structure verification to remove weakly related edges. Finally, a directed acyclic graph (DAG) is generated between features to represent the true causal relationship between each feature and the fault target, and between features.

[0039] By using differentiable causal discovery algorithms, we can uncover the true causal relationships between features. For example, we can identify the core causal chain of increased gearbox oil temperature → increased vibration amplitude → increased failure risk, eliminate spurious correlation features between ambient temperature and rotational speed, solve the problem of spurious correlation interference caused by the reliance on statistical correlation in traditional models, and ensure computational autonomy.

[0040] Based on the feature-directed acyclic graph (DAG) and adjacency matrix A, the target variable Y is determined to be the fault risk of wind turbine units, and a candidate feature set is imported. Establish a causal graph, and establish an association mapping between the target variable and the original features; Define the Markov Coverage Screening Criterion: Based on causal graph theory, clarify the conditions that the Markov cover of the target variable Y must satisfy: The feature subset consists of Y's parent node, features influencing Y, child nodes, features affected by Y, other parent nodes of child nodes, features that jointly influence child nodes with Y (i.e., spouse nodes), and indirectly key related features. Includes all prediction information for Y; Traversing the Directed Acyclic Graph (DAG) to locate associated features involves using a depth-first search algorithm to traverse the DAG structure. Combined with the causal association strength in the adjacency matrix A, the parent node set, child node set, and parent node's child node set of the target variable Y are identified. Features corresponding to these three types of nodes are merged to form an initial candidate feature set. ; Conditional independence test purification, for Each feature in the matrix undergoes a mutual information test to verify its conditional independence from Y. If a feature is conditionally independent of Y, then... If other features are independent of the Y condition, they are judged as spurious features and removed. Redundant feature removal and subset determination, calculation The variance inflation factor of the remaining features is used to remove highly redundant features with a VIF > 10, while retaining core causal features in the adjacency matrix A with a correlation strength higher than a threshold, thus forming the final effective causal feature subset. .

[0041] By learning feature importance weights through graph attention networks, the interaction information between features is accurately captured. Compared with traditional feature concatenation methods, the representation ability of globally fused features is improved by 50%. The differentiated allocation of attention weights strengthens the contribution of core features (such as vibration amplitude) and solves the problem of key information being buried due to the equalization of feature importance in traditional models.

[0042] Configure the GAT training environment and prepare input data, based on the MindSpore framework and its GNN suite, and import effective causal feature subsets. The feature matrix is ​​transformed into a node feature matrix; at the same time, a directed acyclic graph (DAG) is loaded to extract the causal relationships between features. Constructing a feature association graph structure: Each feature in the graph is a graph node. Combined with the causal relationship edges in the DAG, a sparse graph based on the causal structure is constructed. Standard graph data objects are generated through the graph data interface of MindSpore GNN, and the mapping relationship between node set, edge set and node feature is defined. Define the core parameters of the GAT network, set the hidden layer dimension, number of attention heads and output layer dimension of the GAT network according to the task requirements, select LeakyReLU as the attention calculation activation function and Sigmoid as the feature aggregation activation function, and initialize the linear transformation weight matrix W and the learnable attention vector a. Calculate the attention coefficients between nodes and perform a linear transformation on the node feature matrix. , and These are the original feature vectors of node i and node j, respectively; Calculate the original attention coefficients of nodes i and j. :

[0043] The final attention weights are then obtained through Softmax normalization. :

[0044] in, This is the weight matrix. For attention vectors, This indicates a splicing operation. For nodes The neighborhood group, This represents the normalized attention weights; Attention-weighted feature aggregation: For each node i, based on its set of neighboring nodes... and the corresponding attention weights Aggregate the features of neighboring nodes to generate a high-dimensional feature representation of node i. :

[0045] in, This represents the Sigmoid activation function, which captures dynamic interaction information between features. Generate a global fusion feature vector: Employ the mean-readout strategy to represent the high-dimensional features of all nodes. The features of all nodes are aggregated, and the mean vector of all node features is calculated to obtain the global fused features. .

[0046] By learning feature importance weights through graph attention networks, the interaction information between features is accurately captured, improving the representational ability of globally fused features compared to traditional feature concatenation methods. The differentiated allocation of attention weights strengthens the contribution of core features, solving the problem of key information being buried due to the equalization of feature importance in traditional models. Input data and model loading: based on global fusion features The feature attention weights learned by GAT are used to read the effective causal feature subset. The real-time observations of each feature and the statistical distribution based on historical data of the training set are used to dynamically calculate the adaptive safety threshold of each feature and complete the adaptation preparation of the model and data. Define scoring parameters and functions, and set the weight matrix for basic risk calculation based on task requirements. With bias term We select the Sigmoid function as the activation function, set the risk correction coefficient, and define the indicator function. Calculate the base risk probability and fuse global features. Input the basic risk calculation model;

[0047] Calculate the basic risk probability It represents the baseline risk level based on global feature association information; The risk correction term is calculated by combining attention weights, feature observations, and safety thresholds to achieve differentiated correction.

[0048] in, Basic risk probability; This is the revised final risk score; This is a risk correction factor; A set of key risk characteristics; Size of the set; Features The corresponding global attention weights; Features The current observation value; Features The safety threshold; This is an indicator function that takes the value 1 when the condition is met and 0 otherwise. The final real-time risk score is output, which integrates the basic risk probability and correction items to obtain the final real-time risk score. The validity of the score results is verified, and the final output is the unit failure risk score that can be directly used for operation and maintenance priority ranking.

[0049] By replacing the traditional binary label of normal / fault, a continuous risk score is output, which not only quantifies the severity of the fault, but also allocates operation and maintenance resources according to the score ranking; the differentiated correction mechanism enables the impact of high-importance feature anomalies to be accurately characterized, solving the problem that traditional models cannot guide operation and maintenance priority ranking and have limited practical value.

[0050] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0051] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0052] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the unit failure risk scoring methods based on the information technology stack in the above embodiments.

[0053] It is understood that the systems, devices, and storage media provided in the embodiments of the present invention correspond to the methods provided in the embodiments of the present invention, and the explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding parts of the above methods.

[0054] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another.

[0055] For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access, or a data storage device such as a server or data center that integrates one or more available media.

[0056] The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid state disks (SSDs)).

[0057] It should be noted that in this paper, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0058] Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0059] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0060] The embodiments of the present invention are given for the purposes of illustration and description. Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A unit failure risk scoring method based on the domestic information technology stack, characterized in that, Perform the following steps using a computer device: S1, Data Entity Modeling: Based on the openGauss database, an entity relationship model for multi-source heterogeneous data is constructed, defining association keys and integrating structured datasets including SCADA, logs, and vibration data. S2, Deep Feature Synthesis, is based on structured datasets and generates high-order candidate feature sets through recursive and automated construction using aggregation and transformation primitives. S3, Causal Structure Learning, learns the causal graph between features in a high-order candidate feature set based on a differentiable causal discovery algorithm, and uses matrix computation units to accelerate the optimization process and learn to generate causal graphs between features. S4, Key Feature Filtering: Identify Markov Coverage of Target Variable Based on Causal Graph Between Features, Eliminate False Correlation Features, and Filter Effective Causal Feature Subset; S5, Graph Attention Fusion, constructs an association graph with an effective subset of causal features as nodes, learns the attention weights between features through a graph attention network, and aggregates feature interaction information to generate a global fused feature. S6, Risk Score Calculation, dynamically quantifies the real-time fault risk score of the computer group based on global fusion features and attention weights between features.

2. The unit failure risk scoring method based on the information technology stack as described in claim 1, characterized in that: The specific implementation steps of S1 are as follows: S11, Building a domestic IT database environment: Deploying the database on the operating system and configuring operating parameters to adapt to the time-series data storage of wind turbine units; S12, Define the data entity table structure: Based on the monitoring requirements of wind turbine units, create data tables, equipment log tables, and vibration data tables in the database, and clarify the data types and constraint rules of each table field; S13, Set cross-table association key: Use the unit's unique number and data acquisition timestamp as the joint association key to establish a logical association between the data table, equipment log table and vibration data table, ensuring that multi-source data can be queried by unit and time dimension; S14, Multi-source data import and standardization: Import scattered raw monitoring data, including real-time sensor data, equipment operation logs and vibration waveform data, into the corresponding data table according to the table structure, standardize the data format, and remove invalid and redundant data. S15, Storage Optimization Configuration: In response to the query characteristics of wind turbine time series data, columnar storage mode is enabled for each data table, and the final output is a standardized structured dataset for feature synthesis.

3. The unit failure risk scoring method based on the information technology stack as described in claim 2, characterized in that: The specific implementation process of S2 is as follows: S21, Configure the basic parameters for feature construction, call the DFS algorithm tool based on the MindSpore framework to determine the set of feature construction primitives, and set the maximum depth of feature construction to 3; S22, Define the time series aggregation window: For the standardized structured dataset output by S1, set three time sliding windows to adapt to the needs of capturing the operating characteristics of wind turbines at different time scales; S23, Initialize the first-order original features, using the original monitoring features in the output dataset of S1 as the first-order features to form the basic data layer for feature construction; S24, recursively generate higher-order features, performing recursive construction at a preset maximum depth of 3. For each order of feature, the feature generation formula is: in, express Moment The feature is of order d, where d represents the feature dimension. The operation functions are selected from the primitive set. express Rank features in time window Data within, The length of the time window; S25, Feature integrity verification, performs validity screening on all recursively generated features, removes invalid features with missing value ratios exceeding the threshold or abnormal values, and retains valid features; S26. Integrate the candidate feature set, summarizing the first-order original features and effective higher-order features of each order to form a candidate feature set containing higher-order features. This provides sufficient features as input for subsequent causal structure learning.

4. The unit failure risk scoring method based on the information technology stack as described in claim 3, characterized in that: The specific implementation process of S3 is as follows: S31, Configure the domestically developed accelerated computing environment, deploy the MindSpore framework and operator library, and process the candidate feature set. Transform it into a data matrix X, which serves as input data for causal discovery; S32, Define the optimization objective and constraints, and clarify the optimization objective for differentiable causal discovery as follows: The constraints are: in, Represents a data matrix. Let represent the characteristic adjacency matrix, and be the solution variables in the optimization objective. Represents the reconstruction error function. This represents the sparsity regularization coefficient, used to control the sparsity of the adjacency matrix A. Let A represent the column norm of matrix A, which is the sum of the absolute values ​​of all elements in the matrix. Represents the acyclic constraint function. Represents the matrix trace operation. It represents the Hadamah accumulation. Represents the matrix index; S33, Algorithm Deployment and Acceleration Adaptation, executes the core operations of the NOTEARS algorithm, including matrix multiplication, matrix exponent calculation and gradient backpropagation, enables hardware units, and optimizes matrix operation efficiency; S34, causal structure iterative learning, starts the differentiable causal discovery algorithm, iteratively solves the optimization objective based on the data matrix X, updates the adjacency matrix A by gradient descent, and uses constraints to ensure that the graph structure is acyclic, while controlling the sparsity of the matrix by column and norm regularization. S35 outputs a feature causal relationship graph. After iterative convergence, threshold filtering and structure verification are performed on the adjacency matrix A to remove weakly related edges. Finally, a directed acyclic graph (DAG) is generated between features to represent the true causal relationship between each feature and the fault target, and between features.

5. The unit failure risk scoring method based on the information technology stack as described in claim 4, characterized in that: The specific implementation process of S2 is as follows: S41, based on the feature-directed acyclic graph (DAG) and adjacency matrix A, the target variable Y is determined to be the fault risk of wind turbine units, and a candidate feature set is imported. Establish a causal graph, and establish an association mapping between the target variable and the original features; S42, Define the Markov Coverage Screening Criterion: Based on causal graph theory, clarify the conditions that the Markov coverage of the target variable Y must satisfy: The feature subset consists of Y's parent node, features that influence Y, child nodes, features affected by Y, other parent nodes of child nodes (i.e., features that jointly influence child nodes with Y), and indirectly key related features. Includes all prediction information for Y; S43. Traverse the DAG to locate associated features. Use a depth-first search algorithm to traverse the structure of the DAG and combine the causal association strength in the adjacency matrix A to identify the parent node set, child node set, and parent node child node set of the target variable Y. Merge the features corresponding to the three types of nodes to form an initial candidate feature set. ; S44, Conditional Independence Test Purification, for Each feature in the matrix undergoes a mutual information test to verify its conditional independence from Y. If a feature is conditionally independent of Y, then... If other features are independent of the Y condition, they are judged as spurious features and removed. S45, Redundant Feature Removal and Subset Determination, Calculation The variance inflation factor of the remaining features is used to remove highly redundant features with a VIF > 10, while retaining core causal features in the adjacency matrix A with a correlation strength higher than a threshold, thus forming the final effective causal feature subset. .

6. The unit failure risk scoring method based on the information technology stack as described in claim 5, characterized in that: The specific implementation process of S2 is as follows: S51, Configure the GAT training environment and prepare input data, based on the MindSpore framework and its GNN suite, import effective causal feature subsets. The feature matrix is ​​transformed into a node feature matrix; at the same time, a directed acyclic graph (DAG) is loaded to extract the causal relationships between features. S52, Construct the feature association graph structure: using Each feature in the graph is a graph node. Combined with the causal relationship edges in the DAG, a sparse graph based on the causal structure is constructed. Standard graph data objects are generated through the graph data interface of MindSpore GNN, and the mapping relationship between node set, edge set and node feature is defined. S53, define the core parameters of the GAT network, set the hidden layer dimension, number of attention heads and output layer dimension of the GAT network according to the task requirements, select LeakyReLU as the attention calculation activation function and Sigmoid as the feature aggregation activation function, and initialize the linear transformation weight matrix W and the learnable attention vector a. S54, calculate the attention coefficients between nodes and perform a linear transformation on the node feature matrix. , and These are the original feature vectors of node i and node j, respectively; Calculate the original attention coefficients of nodes i and j. : The final attention weights are then obtained through Softmax normalization. : in, This is the weight matrix. For attention vectors, This indicates a splicing operation. For nodes The neighborhood group, This represents the normalized attention weights; S55, feature aggregation based on attention weights: For each node i, based on its set of neighboring nodes... and the corresponding attention weights Aggregate the features of neighboring nodes to generate a high-dimensional feature representation of node i. : in, This represents the Sigmoid activation function, which captures dynamic interaction information between features. S56, Generate a global fusion feature vector: Employ the mean-readout strategy to generate high-dimensional feature representations for all nodes. The features of all nodes are aggregated, and the mean vector of all node features is calculated to obtain the global fused features. .

7. The unit failure risk scoring method based on the information technology stack as described in claim 6, characterized in that: The specific implementation process of S6 is as follows: S61, Input Data and Model Loading: Based on Global Fusion Features The feature attention weights learned by GAT are used to read the effective causal feature subset. Real-time observations of each feature and preset safety thresholds are used to complete the adaptation preparation of the model and data; S62, Define scoring parameters and functions, and set the weight matrix for basic risk calculation based on task requirements. With bias term We select the Sigmoid function as the activation function, set the risk correction coefficient, and define the indicator function. S63, calculate the basic risk probability, and fuse global features. Input the basic risk calculation model; Calculate the basic risk probability It represents the baseline risk level based on global feature association information; S64, calculate the risk correction term, combining attention weights, feature observations and safety thresholds to calculate the risk correction term and achieve differentiated correction; S65 outputs the final real-time risk score, integrating the basic risk probability and correction terms to obtain the final real-time risk score; The validity of the scoring results is verified, and the final output is a unit failure risk score that is directly used for ranking operation and maintenance priorities.

8. The unit failure risk scoring method based on the information technology stack as described in claim 7, characterized in that: The specific formula for calculating the risk correction term ΔS in S64 is as follows: in, Basic risk probability; This is the revised final risk score; This is a risk correction factor; A set of key risk characteristics; Size of the set; Features The corresponding global attention weights; Features The current observation value; Features The safety threshold; This is an indicator function that takes the value 1 when the condition is met and 0 otherwise.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it causes the processor to perform the steps of the method as described in any one of claims 1 to 8.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the computer program is executed by the processor, it causes the processor to perform the steps of the method as described in any one of claims 1 to 8.

Images