Equipment fault early warning method based on new energy station system

Abstract

The invention relates to an equipment fault early warning method based on a new energy station system, in particular to the technical field of new energy power generation monitoring, and realizes the crossing of photovoltaic inverter fault early warning from'passive response 'to'active prediction' through the architecture of edge side intelligent preprocessing and cloud deep learning cooperation. The system significantly improves the fault diagnosis accuracy through multi-source data fusion and time sequence sensitive feature analysis; generating a fault evolution path with clear physical significance and confidence evaluation by means of a knowledge graph; and finally, through a multi-mode diagnosis report, visual and reliable decision support is provided for operation and maintenance personnel, the false report and missing report rate is greatly reduced, the downtime of equipment is reduced, and the operation efficiency and the safety level of a new energy station are improved.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation monitoring technology, and more specifically, to a method for early warning of equipment faults based on a new energy power station system. Background Technology

[0002] With the rapid development of the new energy industry, the construction of large-scale wind power and photovoltaic power plants is becoming increasingly common. The number of key equipment within these plants, such as wind turbine generators and photovoltaic inverters, is continuously increasing, and the operating environment is complex and variable. These devices continuously generate massive amounts of multi-source heterogeneous monitoring data during operation, including various physical parameters such as voltage, current, temperature, and vibration frequency, with different data formats and sampling frequencies. While existing data acquisition and monitoring control systems can achieve basic data acquisition and storage functions, the assessment of equipment operating status still heavily relies on the personal experience of maintenance personnel. Maintenance personnel face a large amount of data generated by equipment from different manufacturers and models, making it difficult to quickly and accurately identify potential fault signs from this complex data. In actual operation and maintenance work, due to the lack of effective intelligent analysis tools, fault warnings are often delayed, leading to untimely equipment maintenance and seriously affecting the safety and economy of power plant operation.

[0003] Currently, some advanced new energy power station monitoring systems have begun to introduce fault early warning models based on deep learning algorithms. These models, through learning from historical data, can achieve fault early warning functions to a certain extent. However, existing technical solutions generally suffer from a significant drawback: severely insufficient interpretability of the models. Deep learning models are essentially black-box systems; when they generate early warning signals, the system can usually only output the warning level or a simple probability score, failing to clearly explain the specific reasons for the warning and the basis for decision-making. Maintenance personnel find it difficult to understand which equipment parameter changes or the degree of anomaly triggered the warning, let alone judge the evolution trend and potential impact of the fault. This lack of transparency in the decision-making process leads to a lack of trust in the warning results among maintenance personnel, often resulting in conservative strategies in actual operation. This often leads to either blindly following warning instructions, causing over-maintenance, or ignoring warning signals, resulting in missed fault detection. Existing technologies prioritize the accuracy of early warning models over their interpretability, failing to effectively correlate warning results with specific equipment parameter anomalies. This lack of effective decision support for on-site maintenance has become a key technical bottleneck restricting the widespread application of intelligent early warning technology in practical engineering. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a method for early warning of equipment failures based on a new energy power station system, thereby solving the problems mentioned in the background art.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: a method for early warning of equipment faults based on a new energy power station system, comprising the following steps: Step S1: At the edge computing node, collect multi-dimensional time-series operation data of wind turbines and photovoltaic inverters in the new energy power station in real time, perform anti-interference filtering on the multi-dimensional time-series operation data, extract time domain features and frequency domain features, and form a high-dimensional feature vector. Step S2: At the server, based on the high-dimensional feature vector generated in step S1, when the time-series deep learning early warning model generates an alarm signal, the time-sensitive SHAP algorithm is used to calculate the contribution of each feature parameter in the high-dimensional feature vector. Step S3: At the server, based on the contribution of each feature parameter obtained in step S2, the device knowledge graph is used to perform root cause reasoning of the fault, and generate fault hypotheses and corresponding confidence levels. Step S4: At the server, based on the parameter contribution obtained in step S2 and the fault reasoning result obtained in step S3, a multimodal diagnostic report containing a parameter contribution heatmap and a fault evolution path diagram is generated, and the confidence level of the diagnostic report is calculated.

[0006] In a preferred embodiment, the specific operation of performing anti-interference filtering on the multi-dimensional time-series running data in step S1 is as follows: Edge computing nodes use wavelet packet transform algorithm to filter multi-dimensional time-series data. The wavelet packet transform algorithm dynamically adjusts the filtering parameters according to the signal-to-noise ratio of the signal subband to suppress impulse interference and Gaussian noise.

[0007] In a preferred embodiment, the specific operations for extracting time-domain features and frequency-domain features in step S1 are as follows: Edge computing nodes extract temporal nonlinear features and frequency domain indices from filtered multi-dimensional time-series running data. The temporal nonlinear features include permutation entropy and multi-scale sample entropy, while the frequency domain indices include modulation bispectral features and frequency band energy entropy. Edge computing nodes employ the maximum correlation and minimum redundancy criterion to select features from time-domain nonlinear characteristics and frequency-domain indices, forming high-dimensional feature vectors.

[0008] In a preferred embodiment, step S2, specifically the operation of calculating the contribution of each feature parameter in the high-dimensional feature vector using the time-sensitive SHAP algorithm, is as follows: The time-sensitive SHAP algorithm adjusts the weights of historical data by introducing a time decay factor and introduces a parameter physical coupling term to characterize the synergistic effect between feature parameters, so as to calculate the time-sensitive contribution value of each feature parameter.

[0009] In a preferred embodiment, step S2, after calculating the time-sensitive contribution values ​​of each feature parameter, further includes the specific operation of performing multi-dimensional feature interaction effect analysis: The interaction effect value between any two feature parameters is calculated based on the time-sensitive SHAP algorithm to quantify the synergistic contribution or offsetting effect of the combined effect of any two feature parameters on the early warning result.

[0010] In a preferred embodiment, step S2, after obtaining the contribution of the feature parameters, further includes the following processing: A feature contribution heatmap is generated based on time-sensitive contribution values ​​for visualization; and a dynamic threshold filtering mechanism is used to select a subset of features with significant contributions from all feature parameters for subsequent root cause inference.

[0011] In a preferred embodiment, the specific operation of using a device knowledge graph for root cause reasoning in step S3 is as follows: The server maps the contribution of each feature parameter obtained in step S2 to the nodes of the device knowledge graph, forming node weights; Based on entity relationships and node weights in the device knowledge graph, a multi-hop inference algorithm is executed to generate a set of candidate fault hypotheses. The multi-hop inference algorithm is configured to prioritize traversing paths with high node weights and select paths based on the strength of graph edge relationships.

[0012] In a preferred embodiment, step S3, specifically the operation of generating the confidence level corresponding to the fault hypothesis, is as follows: For each candidate fault hypothesis, the initial confidence level is calculated based on the weighted sum of the node weights on its corresponding path; The initial confidence score is corrected by combining it with the matching degree of historical failure cases to obtain a normalized confidence score. Confidence scores are used to quantify the reliability of failure hypotheses.

[0013] In a preferred embodiment, step S4, which involves generating a multimodal diagnostic report including a parameter contribution heatmap and a fault evolution path diagram, specifically involves the following steps: Based on the contribution of each feature parameter obtained in step S2, the server uses an adaptive weighted kernel density estimation algorithm to generate a parameter contribution heatmap. Simultaneously, based on the fault reasoning results obtained in step S3, a fault evolution path diagram is constructed using a dynamic graph neural network; the heat map and the path diagram are integrated into a unified multimodal diagnostic report.

[0014] In a preferred embodiment, the specific operation for calculating the confidence level of the diagnostic report in step S4 is as follows: A multi-source evidence fusion mechanism is adopted, which integrates the consistency measure of parameter contribution and the reliability of fault reasoning results as independent evidence sources for calculation. A normalized confidence score is obtained through evidence theory algorithms to quantify the overall reliability of the diagnostic report.

[0015] A fault early warning system for equipment based on a new energy power station system includes a preprocessing module, a feature parameter contribution calculation module, a confidence module, and an early warning module. The preprocessing module collects multi-dimensional time-series operation data of wind turbines and photovoltaic inverters in new energy power plants in real time at the edge computing nodes, performs anti-interference filtering on the multi-dimensional time-series operation data, extracts time-domain features and frequency-domain features, and forms a high-dimensional feature vector. The feature parameter contribution calculation module, located on the server, calculates the contribution of each feature parameter in the high-dimensional feature vector based on the generated high-dimensional feature vector. When the time-series deep learning early warning model generates an alarm signal, it uses the time-sensitive SHAP algorithm to calculate the contribution of each feature parameter in the high-dimensional feature vector. The confidence module, located on the server, uses a device knowledge graph to perform root cause reasoning on faults based on the contribution of each feature parameter, generating fault hypotheses and corresponding confidence levels. The early warning module, located on the server, generates a multimodal diagnostic report containing a parameter contribution heatmap and a fault evolution path diagram based on the obtained parameter contribution and fault inference results, and calculates the confidence level of the diagnostic report.

[0016] The beneficial effects of this invention are as follows: Through the collaborative architecture of edge-side intelligent preprocessing and cloud-based deep learning, the system achieves a leap from "passive response" to "active prediction" in photovoltaic inverter fault early warning. The system significantly improves the accuracy of fault diagnosis through multi-source data fusion and time-sensitive feature analysis. It generates fault evolution paths and confidence assessments with clear physical meaning using knowledge graphs. Finally, through multimodal diagnostic reports, it provides intuitive and reliable decision support for operation and maintenance personnel, greatly reducing false alarm and missed alarm rates, reducing equipment downtime, and improving the operating efficiency and safety level of new energy power plants. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0020] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0021] This embodiment provides, for example Figure 1 The method for early warning of equipment faults based on a new energy power station system, as shown, includes the following steps: Step S1: At the edge computing node, collect multi-dimensional time-series operation data of wind turbines and photovoltaic inverters in the new energy power station in real time, perform anti-interference filtering on the multi-dimensional time-series operation data, extract time domain features and frequency domain features, and form a high-dimensional feature vector. Step S2: At the server, based on the high-dimensional feature vector generated in step S1, when the time-series deep learning early warning model generates an alarm signal, the time-sensitive SHAP algorithm is used to calculate the contribution of each feature parameter in the high-dimensional feature vector. Step S3: At the server, based on the contribution of each feature parameter obtained in step S2, the device knowledge graph is used to perform root cause reasoning of the fault, and generate fault hypotheses and corresponding confidence levels. Step S4: At the server, based on the parameter contribution obtained in step S2 and the fault reasoning result obtained in step S3, a multimodal diagnostic report containing a parameter contribution heatmap and a fault evolution path diagram is generated, and the confidence level of the diagnostic report is calculated.

[0022] In this embodiment, it is specifically necessary to explain the following steps in step S1: the anti-interference filtering process for the multi-dimensional time-series running data is as follows: Edge computing nodes employ wavelet packet transform (WPT) algorithms to filter multi-dimensional time-series operational data. The WPT algorithm dynamically adjusts filtering parameters based on the signal-to-noise ratio (SNR) of the signal sub-bands to suppress impulse interference and Gaussian noise. In practice, edge computing nodes collect real-time multi-dimensional time-series operational data from wind turbines and photovoltaic inverters in renewable energy power plants via industrial bus interfaces, including current, voltage, temperature, and vibration data. Considering the complex electromagnetic environment and susceptibility of sensors to interference in renewable energy power plants, WPT algorithms are used for anti-interference filtering, expressed by the following formula:

[0023] in, This represents the filtered signal, which is used for subsequent feature extraction. These represent wavelet packet coefficients, used to capture local features of signals at different scales and locations; This represents the wavelet packet basis function, used for multi-scale decomposition of signals; θ represents the subband signal-to-noise ratio, used to evaluate the ratio of signal to noise in each subband; θ represents the adaptive threshold, which is dynamically adjusted based on historical data. This is an indicator function; it takes a value of 1 when the subband signal-to-noise ratio is greater than a threshold, and 0 otherwise. This formula effectively suppresses impulse interference and Gaussian noise through multi-scale decomposition and dynamic threshold control of the signal-to-noise ratio, while preserving the signal abrupt change characteristics. It solves the noise interference problem in data acquisition from new energy power stations, providing high-quality input for subsequent feature extraction.

[0024] In step S1, the specific operations for extracting time-domain and frequency-domain features are as follows: Edge computing nodes extract temporal nonlinear features and frequency domain indices from filtered multi-dimensional time-series data. Temporal nonlinear features include permutation entropy and multi-scale sample entropy, while frequency domain indices include modulation bispectral features and band energy entropy. Edge computing nodes use the maximum correlation and minimum redundancy criterion to select features from these temporal nonlinear features and frequency domain indices, forming a high-dimensional feature vector. In specific implementation, edge computing nodes extract temporal nonlinear features and frequency domain indices from filtered multi-dimensional time-series data: temporal nonlinear features include permutation entropy, used to quantify the complexity of the signal sequence; multi-scale sample entropy, used to assess the irregularity of the signal at different time scales; frequency domain indices include modulation bispectral features, used to capture nonlinear coupling effects in the signal; and band energy entropy, used to measure the uniformity of band energy distribution. Subsequently, edge computing nodes use the maximum correlation and minimum redundancy criterion for feature selection, the formula of which is expressed as:

[0025] in, The score represents the feature subset, which is used to optimize feature selection; This indicates the correlation between the feature subset and the fault category; a higher value indicates a stronger discriminative ability of the feature. This indicates the redundancy within a feature subset; a lower value indicates better feature independence. This represents the trade-off parameter used to balance correlation and redundancy. This formula maximizes the correlation between features and fault categories while minimizing redundancy between features, selecting the most discriminative feature subset to form a high-dimensional feature vector. This solves the problem of excessive redundant information in high-dimensional features, improving the efficiency and accuracy of subsequent fault warning models.

[0026] In this embodiment, it is specifically necessary to explain the specific operation in step S2, which uses the time-sensitive SHAP algorithm to calculate the contribution of each feature parameter in the high-dimensional feature vector as follows: The time-sensitive SHAP algorithm adjusts the weights of historical data by introducing a time decay factor and a parameter physical coupling term to characterize the synergistic effect between feature parameters, thereby calculating the time-sensitive contribution value of each feature parameter. In practical implementation, after the time-series deep learning early warning model generates an alarm signal, the server starts the time-sensitive SHAP algorithm to analyze the high-dimensional feature vector received from step S1. The core innovation of this algorithm is reflected in the following formula:

[0027] in The time-sensitive SHAP value of feature i is used to quantify the contribution of this feature to the early warning result; Represents a subset of features; Indicates the total number of features; This represents the model output when using a feature subset S; Let be the time decay function, where It is the attenuation coefficient. It is the current timestamp. This is a reference time point; this function is used to assign higher weight to recent data. The physical coupling coefficient between features i and j is calculated using the device mechanism model and is used to characterize the synergistic effect among multiple parameters. This formula, by introducing time decay and physical coupling terms, overcomes the limitations of the traditional SHAP algorithm in processing dynamic time-series data, and can more accurately reflect the true contribution of features during device operation.

[0028] In step S2, after calculating the time-sensitive contribution values ​​of each feature parameter, the specific operation of performing multi-dimensional feature interaction effect analysis is as follows: The interaction effect value between any two feature parameters is calculated based on the time-sensitive SHAP algorithm to quantify the synergistic contribution or offsetting effect of the combined effect of any two feature parameters on the early warning result. In specific implementation, after obtaining the time-sensitive SHAP values ​​of each feature, the server further employs an improved interaction effect calculation model to analyze the synergistic relationship between features:

[0029] in, This represents the interaction effect value between features i and j, used to quantify the additional effect produced by the combined action of the two features; when A value greater than 0 indicates a synergistic enhancement effect. A value less than 0 indicates a canceling effect. This formula calculates the interaction effect value between two features, i.e., whether their combined contribution to the model prediction is greater than the sum of their individual contributions. Based on the definition of SHAP interaction value, the formula quantifies the interaction effect by comparing the changes in model output when feature subsets i and j are included. This allows for the identification of synergistic (positive interaction) or canceling (negative interaction) effects between features, helping to understand the mechanisms by which multiple features jointly influence early warning results. This provides deeper insights into root cause analysis, such as discovering the synergistic effect of "increased temperature and increased vibration."

[0030] In step S2, after obtaining the contribution of the feature parameters, the following processing is also included: A feature contribution heatmap is generated based on time-sensitive contribution values ​​for visualization. A dynamic threshold filtering mechanism is then employed to select a subset of features with significant contributions from all feature parameters for subsequent root cause inference. The implementation involves two key steps: First, the server generates a feature contribution heatmap based on time-sensitive SHAP values ​​to visually display the importance ranking of each feature. Second, a dynamic threshold filtering mechanism is used to select a significant feature subset (the selection criteria are as follows: ...).

[0031] in, Represents a subset of salient features; and These represent the mean and standard deviation of the feature contribution of the current batch, respectively. This is an adjustable sensitivity coefficient, adjusted between 2 and 3 depending on the actual application scenario. This formula is used to filter feature subsets with significant contributions, setting a dynamic threshold based on the mean and standard deviation of the contribution. Features with contribution values ​​higher than the threshold (…) Features that are considered salient are filtered out using a dynamic threshold, removing noisy or irrelevant features and retaining only those that have a critical impact on fault diagnosis, thus improving the efficiency and accuracy of subsequent root cause analysis. The threshold is dynamically adjusted according to the data distribution to adapt to different operating conditions.

[0032] In this embodiment, it is specifically necessary to explain the specific operation of using a device knowledge graph (KG) for root cause reasoning in step S3 as follows: The server (SVR) will calculate the contribution of each feature parameter obtained in step S2. Mapped to nodes in the device knowledge graph (KG), forming node weights. In practice, an adaptive weighted algorithm based on semantic similarity is used to complete the mapping process, and its calculation formula is as follows:

[0033] This formula is used to calculate the weight of each node in a knowledge graph. The weight is based on the feature contribution. Semantic similarity between features and entities And an indicator function (when similarity exceeds a threshold) The formula is a weighted sum (effective at the time of application). It reflects the fusion of feature contribution and semantic similarity, ensuring that high-contribution features receive higher weights in the knowledge graph. These weights are used to map the feature contributions obtained in step S2 onto nodes in the knowledge graph, forming node weights and providing a focus for subsequent multi-hop inference. Node weights reflect the importance of nodes in fault diagnosis, guiding the inference algorithm to prioritize high-weight nodes.

[0034] in Representing nodes in a knowledge graph (KG) The weight value; Representation of features Contribution value; Features With entity The semantic similarity, with a value range of [0,1], where 1 indicates complete similarity; This is the similarity threshold, with a typical value of 0.7 and a range of [0.5, 0.9]. For indicator functions; The summation range is determined by the total number of features; there is no fixed value, but the typical range is 10-100. This mapping process ensures that high-contribution features receive higher weights in the knowledge graph (KG), providing a focus for subsequent reasoning.

[0035] Based on entity relationships and node weights in a device knowledge graph (KG) The Multi-Hop Inference Algorithm (MHR) is executed to generate a candidate fault hypothesis set (FHS). Specifically, a random walk-based MHR algorithm is used, with the path evaluation formula as follows:

[0036] This formula calculates the evaluation score of a path by assigning weights to each node on the path. The strength of the relationship between nodes The formula, obtained through multiplication, reflects the combined influence of node weights and relationship strength along a path. It's used to assess the importance of a path and, in multi-hop inference, to evaluate the scores of different paths, prioritizing paths with high node weights and strong relationships to generate a high-quality set of candidate fault hypotheses. A higher path score indicates a greater likelihood that the path points to the root cause of the fault.

[0037] in, Representing a path The assessment score For path length, The strength of the relationship between entities is represented by the definition used during knowledge graph construction, based on historical data or domain knowledge, with a value range of [0,1], where 1 indicates a strong relationship; the multi-hop reasoning algorithm (MHR) is configured to prioritize traversing node weights. The path with high probability is selected, and the path is filtered by combining the strength of graph edge relationships to generate the most probable set of fault hypotheses (FHS).

[0038] In step S3, the specific operation for generating the confidence level corresponding to the fault hypothesis is as follows: For each candidate failure hypothesis (FH), based on the node weights on its corresponding path... The initial confidence level is calculated by weighted summation. In practice, the initial confidence level is calculated using the following formula. :

[0039] This formula calculates the initial confidence level of the failure hypothesis by summing the weights of all nodes on the hypothetical path and dividing by the path length to obtain the average weight. The formula provides a preliminary quantification of the reliability of the failure hypothesis based on the average importance of the path nodes. It is used to generate the initial confidence level of the failure hypothesis, serving as the basis for subsequent revisions. The initial confidence level reflects the average weight of the nodes on the path, avoiding the influence of path length and making paths of different lengths comparable.

[0040] in, Indicating fault assumptions The initial confidence level; This indicates the path corresponding to the hypothesis; The path length is a positive integer, typically ranging from 2 to 10. This calculation method ensures a balance between path weight and path length.

[0041] Initial confidence level Matching degree with historical failure cases By combining and correcting these factors, a normalized confidence score is obtained. In practice, the Bayesian correction method is used:

[0042] The formula uses Bayesian correction to adjust the initial confidence level. Matching degree with historical cases The scores are combined and normalized to obtain a normalized confidence score. The formula ensures that the sum of the confidence scores for all failure hypotheses is 1, providing a probabilistic interpretation. This score is used to generate the final confidence score, comprehensively considering current evidence and historical data to provide a quantitative basis for operational decisions. The normalization process makes the scores of different hypotheses comparable, facilitating the prioritization of high-confidence failures.

[0043] in, Represents the normalized confidence score; Indicates a hypothesis Matching degree with historical failure cases, value range [0,1], 1 indicates a high match. Confidence score. It is used to quantify the reliability of failure assumptions and provide a reliable basis for subsequent decision-making.

[0044] In this embodiment, the specific operation of generating a multimodal diagnostic report containing a parameter contribution heatmap and a fault evolution path diagram in step S4 is as follows: The server (SVR) contributes to each feature parameter obtained in step S2. An adaptive weighted kernel density estimation algorithm is used to generate a parameter contribution heatmap (HMP). Specifically, the parameter contributions are first spatiotemporally aligned, and then density estimation is performed using a radial basis function kernel.

[0045] This formula generates a heatmap of parameter contributions, visualizing the contribution of feature parameters in the spatiotemporal dimensions through a weighted kernel density estimation method. The formula combines spatial proximity (through the kernel function) and temporal decay effects (through the exponential term) to achieve dynamic visualization of parameter importance. This transforms abstract feature parameter contribution data into an intuitive heatmap, enabling operations and maintenance personnel to quickly identify the spatial distribution and temporal trends of key parameters, supporting rapid location of abnormal areas.

[0046] in, Indicates position and time The heat value at that point The contribution value of feature i. For Gaussian kernel function, For bandwidth parameters, This represents the time decay coefficient. This heatmap generation method can simultaneously reflect the spatial distribution and temporal trend of parameter importance.

[0047] Simultaneously, based on the Fault Reasoning Result (FRR) obtained in step S3, a Fault Evolution Path Graph (FEP) is constructed using a Dynamic Graph Neural Network (DGNN). Specifically, a spatiotemporal graph structure is constructed:

[0048] This formula defines the mathematical structure of a fault evolution path graph, including a node set (fault assumptions), an edge set (propagation paths), and a weight matrix (relationship strength). The formula uses a spatiotemporal graph to capture the dynamic characteristics of fault propagation. This allows for the construction of a graph model of fault propagation, providing a data structure foundation for subsequent path analysis and visualization, and accurately describing the evolution of faults in both time and space dimensions.

[0049] Among them, node set Denotes the Fault Hypothesis Set (FHS) and the Edge Set. Represents the fault propagation path, weight matrix It is determined by both the confidence score (CS) and the strength of the temporal association.

[0050] The heatmap (HMP) and pathmap (FEP) are integrated into a unified multimodal diagnostic report (MDR) through spatiotemporal alignment and scale normalization. Specifically, a spatiotemporal alignment algorithm based on feature point matching is used.

[0051] This formula uses an optimization method to find the optimal alignment parameters between the heatmap and the path map in the spatiotemporal dimensions, including time offset and scale factor. The formula employs the least squares method to minimize the differences between the two data sources. This ensures the consistency of the heatmap and path map in the spatiotemporal dimensions, providing accurate coordinate correspondences for multimodal data fusion and avoiding misjudgments caused by spatiotemporal inconsistencies.

[0052] in, This represents the optimal alignment parameter. This is the time offset. is the scale factor.

[0053] In step S4, the specific operation for calculating the confidence level of the diagnostic report is as follows: A multi-source evidence fusion (MEF) mechanism is employed, fusing parameter contribution consistency measure (PCC) and fault reasoning result reliability (FRR) as independent evidence sources. In practice, basic probability allocation is performed based on DS evidence theory.

[0054]

[0055] These two formulas calculate the basic probability allocation function based on parameter contribution consistency and fault reasoning reliability, respectively. The first formula reflects data consistency based on contribution variance, while the second formula reflects reasoning reliability based on confidence scores. These two formulas provide basic probabilistic inputs for DS evidence theory, transforming evidence from different sources into a unified probabilistic representation, supporting subsequent evidence fusion calculations.

[0056] in, This represents the basic probability allocation function based on the consistency of parameter contribution. This represents the basic probability assignment function for reliability-based fault reasoning. Indicates the variance of contribution. To adjust the parameters.

[0057] The normalized confidence score (CS) is obtained through evidence theory algorithms. In practice, Dempster's combination rule is used for evidence synthesis.

[0058] in, This represents the overall confidence score of the diagnostic report, used to quantify the overall reliability of the diagnostic report.

[0059] This formula, the core of the DS evidence theory, is used to fuse multiple basic probability assignment functions into a unified confidence score. By considering the intersection of evidence and the non-empty set condition, the formula achieves a reasonable combination of probabilities. This generates a final comprehensive confidence score, providing a quantitative basis for the reliability of diagnostic reports and supporting confidence-based decision-making.

[0060] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0061] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0062] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0063] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0064] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for early warning of equipment faults based on a new energy power station system, characterized in that, Includes the following steps: Step S1: At the edge computing node, collect multi-dimensional time-series operation data of wind turbines and photovoltaic inverters in the new energy power station in real time, perform anti-interference filtering on the multi-dimensional time-series operation data, extract time domain features and frequency domain features, and form a high-dimensional feature vector. Step S2: At the server, based on the high-dimensional feature vector generated in step S1, when the time-series deep learning early warning model generates an alarm signal, the time-sensitive SHAP algorithm is used to calculate the contribution of each feature parameter in the high-dimensional feature vector. Step S3: At the server, based on the contribution of each feature parameter obtained in step S2, the device knowledge graph is used to perform root cause reasoning of the fault, and generate fault hypotheses and corresponding confidence levels. Step S4: At the server, based on the parameter contribution obtained in step S2 and the fault reasoning result obtained in step S3, a multimodal diagnostic report containing a parameter contribution heatmap and a fault evolution path diagram is generated, and the confidence level of the diagnostic report is calculated.

2. The equipment fault early warning method based on a new energy power station system according to claim 1, characterized in that: In step S1, the specific operation of performing anti-interference filtering on the multi-dimensional time-series running data is as follows: Edge computing nodes use wavelet packet transform algorithm to filter multi-dimensional time-series data. The wavelet packet transform algorithm dynamically adjusts the filtering parameters according to the signal-to-noise ratio of the signal subband to suppress impulse interference and Gaussian noise.

3. The equipment fault early warning method based on a new energy power station system according to claim 2, characterized in that: In step S1, the specific operations for extracting time-domain and frequency-domain features are as follows: Edge computing nodes extract temporal nonlinear features and frequency domain indices from filtered multi-dimensional time-series running data. The temporal nonlinear features include permutation entropy and multi-scale sample entropy, while the frequency domain indices include modulation bispectral features and frequency band energy entropy. Edge computing nodes employ the maximum correlation and minimum redundancy criterion to select features from time-domain nonlinear characteristics and frequency-domain indices, forming high-dimensional feature vectors.

4. The equipment fault early warning method based on a new energy power station system according to claim 3, characterized in that: In step S2, the specific operation of calculating the contribution of each feature parameter in the high-dimensional feature vector using the time-sensitive SHAP algorithm is as follows: The time-sensitive SHAP algorithm adjusts the weights of historical data by introducing a time decay factor and introduces a parameter physical coupling term to characterize the synergistic effect between feature parameters, so as to calculate the time-sensitive contribution value of each feature parameter.

5. The equipment fault early warning method based on a new energy power station system according to claim 4, characterized in that: In step S2, after calculating the time-sensitive contribution values ​​of each feature parameter, the specific operation of performing multi-dimensional feature interaction effect analysis is as follows: The interaction effect value between any two feature parameters is calculated based on the time-sensitive SHAP algorithm to quantify the synergistic contribution or offsetting effect of the combined effect of any two feature parameters on the early warning result.

6. The equipment fault early warning method based on a new energy power station system according to claim 5, characterized in that: In step S2, after obtaining the contribution of the feature parameters, the following processing is also included: A feature contribution heatmap is generated based on time-sensitive contribution values ​​for visualization; and a dynamic threshold filtering mechanism is used to select a subset of features with significant contributions from all feature parameters for subsequent root cause inference.

7. The equipment fault early warning method based on a new energy power station system according to claim 6, characterized in that: In step S3, the specific operation of using the device knowledge graph for root cause reasoning of the fault is as follows: The server maps the contribution of each feature parameter obtained in step S2 to the nodes of the device knowledge graph, forming node weights; Based on entity relationships and node weights in the device knowledge graph, a multi-hop inference algorithm is executed to generate a set of candidate fault hypotheses. The multi-hop inference algorithm is configured to prioritize traversing paths with high node weights and select paths based on the strength of graph edge relationships.

8. The equipment fault early warning method based on a new energy power station system according to claim 7, characterized in that: In step S3, the specific operation for generating the confidence level corresponding to the fault hypothesis is as follows: For each candidate fault hypothesis, the initial confidence level is calculated based on the weighted sum of the node weights on its corresponding path; The initial confidence score is corrected by combining it with the matching degree of historical failure cases to obtain a normalized confidence score. Confidence scores are used to quantify the reliability of failure hypotheses.

9. A method for early warning of equipment faults based on a new energy power station system according to claim 8, characterized in that: In step S4, the specific steps for generating a multimodal diagnostic report that includes a heatmap of parameter contribution and a fault evolution path diagram are as follows: Based on the contribution of each feature parameter obtained in step S2, the server uses an adaptive weighted kernel density estimation algorithm to generate a parameter contribution heatmap. Simultaneously, based on the fault reasoning results obtained in step S3, a fault evolution path diagram is constructed using a dynamic graph neural network; the heat map and the path diagram are integrated into a unified multimodal diagnostic report.

10. An equipment fault early warning system for a new energy power station system using the method of claim 1, characterized in that, It includes a preprocessing module, a feature parameter contribution calculation module, a confidence module, and an early warning module: The preprocessing module collects multi-dimensional time-series operation data of wind turbines and photovoltaic inverters in new energy power plants in real time at the edge computing nodes, performs anti-interference filtering on the multi-dimensional time-series operation data, extracts time-domain features and frequency-domain features, and forms a high-dimensional feature vector. The feature parameter contribution calculation module, located on the server, calculates the contribution of each feature parameter in the high-dimensional feature vector based on the generated high-dimensional feature vector. When the time-series deep learning early warning model generates an alarm signal, it uses the time-sensitive SHAP algorithm to calculate the contribution of each feature parameter in the high-dimensional feature vector. The confidence module, located on the server, uses a device knowledge graph to perform root cause reasoning on faults based on the contribution of each feature parameter, generating fault hypotheses and corresponding confidence levels. The early warning module, located on the server, generates a multimodal diagnostic report containing a parameter contribution heatmap and a fault evolution path diagram based on the obtained parameter contribution and fault inference results, and calculates the confidence level of the diagnostic report.

Images