Transformer defect early warning method and device based on time sequence fragment association evidence fusion and dynamic risk map

Abstract

The invention relates to a transformer defect early warning method based on time sequence fragment association evidence fusion and a dynamic risk map. The method comprises the following steps: acquiring a standardized time sequence data set with a state label; obtaining a time sequence evidence body; calculating a dynamic fault probability and generating a dynamic risk curve; constructing and continuously iterating a dynamic risk association graph; and outputting quantitative risk early warning and generating a self-adaptive optimization operation and maintenance strategy based on the cloud edge cooperative computing framework. By introducing preprocessing based on data fluctuation characteristics and time sequence evidence body construction, deep mining and uncertainty quantification of a multi-modal long-term evolution law are realized, and the early recognition capability and evaluation reliability of latent defects are remarkably improved; through cloud edge cooperation and multi-objective optimization strategy generation, the unification of efficient utilization of computing resources and decision scientificity is realized, the output quantitative early warning and priority strategy directly supports field lean operation and maintenance, and a complete intelligent operation and maintenance closed loop is formed.

Description

Technical Field

[0001] This invention relates to the field of main transformer defect diagnosis and condition monitoring technology, and in particular to a transformer defect early warning method and device based on time segment correlation evidence fusion and dynamic risk map. Background Technology

[0002] The main transformer is a core hub device in the power grid transmission system, undertaking the critical functions of power transmission and voltage transformation. Its operational stability directly determines the reliability and quality of power supply. Main transformers are subjected to high loads and complex operating conditions for extended periods, making them prone to latent defects such as winding overheating, partial discharge, and insulation aging. These defects often exhibit multi-feature coupling and slow evolution characteristics.

[0003] Existing main transformer defect early warning technologies are insufficient to meet the needs of intelligent and lean operation and maintenance. The main problems are as follows: First, the analysis of multi-source heterogeneous data such as oil chromatography, partial discharge, and infrared spectroscopy is mostly limited to point or short-term window correlations, failing to effectively extract and correlate long-term evolution patterns across modes, and is insensitive to early defects. Second, defect judgment criteria rely heavily on fixed thresholds or static models, failing to fully consider individual equipment aging differences, operating stress, and the inherent uncertainty of monitoring data, leading to frequent false alarms and missed alarms, and poor adaptability to new equipment. Third, early warning information is primarily qualitative, lacking quantitative and forward-looking assessments of future fault probabilities and evolution trends, thus failing to support accurate operation and maintenance decisions and resource planning. Fourth, existing knowledge systems are mostly static rule bases, lacking the ability to dynamically correlate and continuously optimize real-time data, quantify uncertainties, and integrate historical cases with operation and maintenance strategies. Summary of the Invention

[0004] To address the difficulty of early warning of main transformer defects under complex operating conditions, the primary objective of this invention is to provide a transformer defect early warning method based on time-series segment correlation evidence fusion and dynamic risk map, which enables in-depth mining of the long-term evolution law of multimodal processes and quantification of uncertainty, significantly improving the early identification capability and reliability assessment of latent defects.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a transformer defect early warning method based on time-series segment correlation evidence fusion and dynamic risk map, the method comprising the following sequential steps:

[0006] (1) Obtain the multi-mode time series data of the main transformer and perform preprocessing to obtain a normalized time series dataset with state labels;

[0007] (2) Adaptive segmentation, cross-modal alignment and feature extraction are performed on the data in the normalized time series dataset, and the data is transformed into time series evidence to realize the mapping from data to evidence;

[0008] (3) Use the time-series evidence as the input sequence for fusion reasoning, calculate the dynamic failure probability and generate a dynamic risk curve;

[0009] (4) Construct, initialize, and continuously update a dynamic risk association graph;

[0010] (5) Based on the dynamic failure probability and dynamic risk correlation graph, and based on the cloud-edge collaborative computing framework, output quantitative risk warning and generate adaptive optimization operation and maintenance strategy.

[0011] Step (1) specifically includes the following steps:

[0012] (1a) Install oil chromatography online monitoring devices, ultra-high frequency partial discharge sensors, ultrasonic partial discharge sensors, infrared thermal imagers, and acoustic fingerprint sensors on key components of the main transformer to ensure coverage of key monitoring areas;

[0013] (1b) Collect multimodal time series data and first remove outliers: use the three-standard-deviation method to identify and remove obvious outliers; for negative values ​​or abnormally large values ​​in oil chromatography data caused by sampling or transmission, mark them as invalid; then perform noise filtering: for high-frequency signals such as partial discharge and acoustic signature, use wavelet thresholding to denoise, select the db4 wavelet basis for 3-level decomposition, and use soft thresholding for detail coefficients; for low-frequency slowly varying signals such as oil chromatography and temperature, use moving average filtering.

[0014] (1c) The variance and first difference of the data are calculated using a sliding time window, and the fluctuation state of each data point is automatically identified and marked. The fluctuation state is divided into steady state, trend state and sudden jump state.

[0015] (1c) For the time series data of each modality, the sliding time window method is used for analysis, and the variance is calculated within each window W. and the first difference mean It automatically identifies and marks the fluctuation state of each data point;

[0016] ;

[0017] in, All are threshold values;

[0018] (1d) For different state segments, a data completion algorithm is used to process missing values ​​and all data are normalized to form a normalized time series dataset with state labels. The state labels include three types of states: steady state, trend, and sudden jump.

[0019] Step (2) specifically includes the following steps:

[0020] (2a) Based on the status labels and combined with the change point detection algorithm, the continuous time series data of each modality are automatically divided into time series segments with consistent trend characteristics;

[0021] (2b) Using a unified time axis as a reference, align the temporal segment boundaries of all modes to form cross-modal aligned temporal analysis units;

[0022] (2c) For each aligned time series analysis unit, the statistical features and trend features of each modality within it, as well as the correlation features between cross-modal data sequences within the unit, are extracted through the lightweight feature extraction module. All statistical features, trend features, and correlation features are concatenated to form a high-dimensional fusion feature vector, which is used to characterize the multimodal collaborative state of the equipment during that time period.

[0023] (2d) Input the fused feature vector into an evidence generation module. The evidence generation module adopts a multilayer perceptron network. The output layer of the multilayer perceptron network has 5 neurons, which correspond to four health states and an uncertainty measure. The four health states are healthy, attentive, abnormal, and severe. The multilayer perceptron network is trained using evidence deep learning methods. The output basic probability assignment value and the uncertainty measure together constitute a temporal evidence body.

[0024] Step (3) specifically includes the following steps:

[0025] (3a) The time-series evidence generated in chronological order is used as the input sequence and fed into the time-series evidence fusion model. According to the rules of evidence theory, the time-series evidence fusion model regards each piece of evidence in the input sequence as a basic reliability assignment and considers the temporal dependency between evidences. It models the reliability transmission and evolution of evidence in the time dimension through a long short-term memory network (LSTM). Finally, the time-series evidence fusion model outputs the reliability distribution that represents the current comprehensive health status of the device.

[0026] (3b) Constructing a dynamic proportional risk model:

[0027] ;

[0028] in, This is the baseline risk function fitted based on historical failure data; It is the sum of the reliability scores for abnormal and severe states in the overall reliability distribution of health status; An aging factor calculated based on the equipment's years of operation; This is the real-time operating condition stress factor calculated based on the current load rate and ambient temperature. All parameters are adjustable;

[0029] The conditional probability of equipment failure in the future, i.e., the dynamic failure probability, is calculated using a dynamic proportional risk model.

[0030] ;

[0031] In the formula, The dynamic failure probability refers to the conditional probability that the equipment will fail within a future time period ∆t at the current time t. Dynamic proportional risk model;

[0032] By calculating and mapping different futures corresponding The value generates dynamic failure probability curves for multiple time scales in the future, intuitively showing the development trend of risk over time.

[0033] Step (4) specifically includes the following steps:

[0034] (4a) Design and construct graph patterns in the Neo4j graph database: construct multiple types of nodes including device nodes, time-series segment pattern nodes, evidence nodes, risk probability nodes, historical case nodes and operation and maintenance strategy nodes;

[0035] (4b) Define and create directed relation edges that connect nodes of multiple types, and assign quantified weight and confidence attributes to each directed relation edge to characterize the strength and reliability of the association;

[0036] (4c) The graph framework is composed of the nodes constructed in step (4a) and the directed relation edges created in step (4b). Graph initialization is performed: the normalized time series dataset, time series evidence body and dynamic fault probability are injected into the graph framework and assigned initial weights and confidence to obtain an initialized graph.

[0037] (4d) Establish a dynamic graph update mechanism: The system records each process from fault warning, on-site verification and fault handling as a learning sample, driving the graph learning algorithm to automatically adjust the weight and confidence of the relevant directed edges, so as to realize the continuous evolution of graph knowledge.

[0038] Step (5) specifically includes the following steps:

[0039] (5a) Adopting a cloud-edge collaborative architecture, the time-series evidence fusion model, dynamic proportional risk model and dynamic risk association map are deployed in the cloud, and a lightweight feature extraction and evidence generation module is deployed at the edge;

[0040] (5b) Generate a quantitative early warning report based on the dynamic failure probability, which includes specific probability values, dominant defect modes, risk evolution rate and expected development time.

[0041] (5c) Based on the risk characteristics in the quantitative early warning report, retrieve matching historical cases and related strategies from the dynamic risk association map, and comprehensively consider the power grid operation constraints and on-site resource conditions, and generate an operation and maintenance decision scheme with priority ranking and recommended time window through a multi-objective optimization algorithm.

[0042] (5d) Push the quantitative early warning report and operation and maintenance decision-making plan to the operation and maintenance personnel, and collect the handling feedback to drive the iterative update of the dynamic risk correlation map.

[0043] Another object of the present invention is to provide an electronic device comprising:

[0044] Processor; and

[0045] The memory stores computer program instructions that, when executed by the processor, cause the processor to perform the transformer defect early warning method based on time-series segment association evidence fusion and dynamic risk map as described above.

[0046] The present invention also provides a computer-readable storage medium having stored thereon computer program instructions, which, when executed by a processor, cause the processor to perform the transformer defect early warning method based on time-series segment association evidence fusion and dynamic risk map as described above.

[0047] As can be seen from the above technical solution, the beneficial effects of the present invention are as follows: First, by introducing preprocessing based on data fluctuation characteristics and constructing time-series evidence, the deep mining of long-term evolution laws of multimodal systems and the quantification of uncertainty are realized, significantly improving the early identification capability and assessment reliability of latent defects; Second, by using time-series evidence fusion and dynamic risk probability calculation, multi-dimensional state information is transformed into forward-looking quantitative risk indicators, realizing the leap from qualitative to quantitative defect assessment and providing a core basis for accurate decision-making; Third, by constructing and iterating dynamic risk association graphs, not only is the digitization and visualization of operation and maintenance knowledge realized, but the system is also endowed with the ability to autonomously optimize association rules through closed-loop feedback, solving the rigidity problem of static knowledge bases; Fourth, through cloud-edge collaboration and multi-objective optimization strategy generation, the unity of efficient utilization of computing resources and scientific decision-making is realized, and the output quantitative early warning and priority strategy directly support lean operation and maintenance on site, forming a complete intelligent operation and maintenance closed loop. Attached Figure Description

[0048] Figure 1 This is a flowchart of the method of the present invention;

[0049] Figure 2 This is a system architecture diagram of the present invention. Detailed Implementation

[0050] like Figure 1 As shown, a transformer defect early warning method based on temporal segment correlation evidence fusion and dynamic risk map is proposed. The method includes the following steps in sequence:

[0051] (1) Obtain the multi-mode time series data of the main transformer and perform preprocessing to obtain a normalized time series dataset with state labels;

[0052] (2) Adaptive segmentation, cross-modal alignment and feature extraction are performed on the data in the normalized time series dataset, and the data is transformed into time series evidence to realize the mapping from data to evidence;

[0053] (3) Use the time-series evidence as the input sequence for fusion reasoning, calculate the dynamic failure probability and generate a dynamic risk curve;

[0054] (4) Construct, initialize, and continuously update a dynamic risk association graph;

[0055] (5) Based on the dynamic failure probability and dynamic risk correlation graph, and based on the cloud-edge collaborative computing framework, output quantitative risk warning and generate adaptive optimization operation and maintenance strategy.

[0056] Step (1) specifically includes the following steps:

[0057] (1a) Install oil chromatography online monitoring devices, ultra-high frequency partial discharge sensors, ultrasonic partial discharge sensors, infrared thermal imagers, and acoustic fingerprint sensors on key components of the main transformer to ensure coverage of key monitoring areas;

[0058] (1b) Collect multimodal time series data and first remove outliers: use the three-standard-deviation method to identify and remove obvious outliers; for negative values ​​or abnormally large values ​​in oil chromatography data caused by sampling or transmission, mark them as invalid; then perform noise filtering: for high-frequency signals such as partial discharge and acoustic signature, use wavelet thresholding to denoise, select the db4 wavelet basis for 3-level decomposition, and use soft thresholding for detail coefficients; for low-frequency slowly varying signals such as oil chromatography and temperature, use moving average filtering.

[0059] (1c) The variance and first difference of the data are calculated using a sliding time window, and the fluctuation state of each data point is automatically identified and marked. The fluctuation state is divided into steady state, trend state and sudden jump state.

[0060] (1c) For the time series data of each modality, the sliding time window method is used for analysis, and the variance is calculated within each window W. and the first difference mean It automatically identifies and marks the fluctuation state of each data point;

[0061] ;

[0062] in, All are threshold values;

[0063] (1d) For different state segments, a data completion algorithm is used to process missing values ​​and all data are normalized to form a normalized time series dataset with state labels. The state labels include three types of states: steady state, trend, and sudden jump.

[0064] For missing points marked as being within steady-state segments, linear interpolation is used for completion. For missing points marked as being within trend segments, cubic spline interpolation is used to maintain trend continuity. For missing points marked as being within abrupt jump segments, which may contain important fault information, interpolation is not performed to avoid introducing spurious smoothing; instead, they are retained as missing markers and handled specially during subsequent segmentation. For multimodal data, all modalities are ensured to be aligned on timestamps, and missing time periods that cannot be aligned are removed entirely. Max-min normalization is performed on all completed modal data, mapping the values ​​to the [0,1] interval. Finally, a structured, normalized time-series dataset with temporal state labels is generated.

[0065] Step (2) specifically includes the following steps:

[0066] (2a) Based on the status labels and combined with the PELT (Pruned Exact Linear Time) change point detection algorithm, the continuous time series data of each modality is automatically divided into time series segments with consistent trend characteristics; a larger penalty coefficient is set in the steady-state segment to reduce missegmentation, and a smaller penalty coefficient is set near the trend segment or abrupt segment to capture changes. The optimal segmentation point set for each modality data is output, thereby dividing the continuous time series into multiple time series segments with consistent internal trends;

[0067] (2b) Using a unified time axis as a reference, align the temporal segment boundaries of all modes to form cross-modal aligned temporal analysis units;

[0068] (2c) For each aligned time series analysis unit, the statistical features and trend features of each modality within it, as well as the correlation features between cross-modal data sequences within the unit, are extracted through the lightweight feature extraction module. All statistical features, trend features, and correlation features are concatenated to form a high-dimensional fusion feature vector, which is used to characterize the multimodal collaborative state of the equipment during that time period.

[0069] (2d) Input the fused feature vector into an evidence generation module. The evidence generation module adopts a multilayer perceptron network. The output layer of the multilayer perceptron network has 5 neurons, which correspond to four health states and an uncertainty measure. The four health states are healthy, attentive, abnormal, and severe. The multilayer perceptron network is trained using evidence deep learning methods. The output basic probability assignment value and the uncertainty measure together constitute a temporal evidence body.

[0070] Step (3) specifically includes the following steps:

[0071] (3a) The time-series evidence generated in chronological order is used as the input sequence and fed into the time-series evidence fusion model. According to the rules of evidence theory, the time-series evidence fusion model regards each piece of evidence in the input sequence as a basic reliability assignment and considers the temporal dependency between evidences. It models the reliability transmission and evolution of evidence in the time dimension through a long short-term memory network (LSTM). Finally, the time-series evidence fusion model outputs the reliability distribution that represents the current comprehensive health status of the device.

[0072] (3b) Constructing a dynamic proportional risk model:

[0073] ;

[0074] in, This is the baseline risk function fitted based on historical failure data; It is the sum of the reliability scores for abnormal and severe states in the overall reliability distribution of health status; An aging factor calculated based on the equipment's years of operation; This is the real-time operating condition stress factor calculated based on the current load rate and ambient temperature. All parameters are adjustable;

[0075] The conditional probability of equipment failure in the future, i.e., the dynamic failure probability, is calculated using a dynamic proportional risk model.

[0076] ;

[0077] In the formula, The dynamic failure probability refers to the conditional probability that the equipment will fail within a future time period ∆t at the current time t. Dynamic proportional risk model;

[0078] By calculating and mapping different futures corresponding The value generates dynamic failure probability curves for multiple time scales in the future, intuitively showing the development trend of risk over time.

[0079] Step (4) specifically includes the following steps:

[0080] (4a) Design and construct graph patterns in the Neo4j graph database: construct multiple types of nodes including device nodes, time-series segment pattern nodes, evidence nodes, risk probability nodes, historical case nodes and operation and maintenance strategy nodes;

[0081] (4b) Define and create directed relation edges that connect nodes of multiple types, and assign quantified weight and confidence attributes to each directed relation edge to characterize the strength and reliability of the association;

[0082] (4c) The graph framework is composed of the nodes constructed in step (4a) and the directed relation edges created in step (4b). Graph initialization is performed: the normalized time series dataset, time series evidence body and dynamic fault probability are injected into the graph framework and assigned initial weights and confidence to obtain an initialized graph.

[0083] (4d) Establish a dynamic graph update mechanism: The system records each process from fault warning, on-site verification and fault handling as a learning sample, driving the graph learning algorithm to automatically adjust the weight and confidence of the relevant directed edges, so as to realize the continuous evolution of graph knowledge.

[0084] like Figure 2 As shown, step (5) specifically includes the following steps:

[0085] (5a) Adopting a cloud-edge collaborative architecture, the time-series evidence fusion model, dynamic proportional risk model and dynamic risk association map are deployed in the cloud, and a lightweight feature extraction and evidence generation module is deployed at the edge;

[0086] (5b) Generate a quantitative early warning report based on the dynamic failure probability, which includes specific probability values, dominant defect modes, risk evolution rate and expected development time.

[0087] (5c) Based on the risk characteristics in the quantitative early warning report, retrieve matching historical cases and related strategies from the dynamic risk association map, and comprehensively consider the power grid operation constraints and on-site resource conditions, and generate an operation and maintenance decision scheme with priority ranking and recommended time window through a multi-objective optimization algorithm.

[0088] (5d) Push the quantitative early warning report and operation and maintenance decision-making plan to the operation and maintenance personnel, and collect the handling feedback to drive the iterative update of the dynamic risk correlation map.

[0089] In summary, this invention, by introducing preprocessing based on data fluctuation characteristics and constructing time-series evidence, achieves in-depth mining of long-term multimodal evolution patterns and quantification of uncertainty, significantly improving the early identification capability and assessment reliability of latent defects. Through time-series evidence fusion and dynamic risk probability calculation, multidimensional state information is transformed into forward-looking quantitative risk indicators, realizing a leap from qualitative to quantitative defect assessment and providing a core basis for accurate decision-making. By constructing and iterating a dynamic risk association graph, not only is the digitization and visualization of operation and maintenance knowledge realized, but the system is also endowed with the ability to autonomously optimize association rules through closed-loop feedback, solving the rigidity problem of static knowledge bases. Through cloud-edge collaboration and multi-objective optimization strategy generation, the efficient utilization of computing resources and the scientific nature of decision-making are unified, and the output quantitative early warning and priority strategies directly support lean operation and maintenance on site, forming a complete intelligent operation and maintenance closed loop.

[0090] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A transformer defect early warning method based on temporal segment association evidence fusion and dynamic risk map, characterized in that: The method includes the following steps in sequence: (1) Obtain the multi-mode time series data of the main transformer and perform preprocessing to obtain a normalized time series dataset with state labels; (2) Adaptive segmentation, cross-modal alignment and feature extraction are performed on the data in the normalized time series dataset, and the data is transformed into time series evidence to realize the mapping from data to evidence; (3) Use the time-series evidence as the input sequence for fusion reasoning, calculate the dynamic failure probability and generate a dynamic risk curve; (4) Construct, initialize, and continuously update a dynamic risk association graph; (5) Based on the dynamic failure probability and dynamic risk correlation graph, and based on the cloud-edge collaborative computing framework, output quantitative risk warning and generate adaptive optimization operation and maintenance strategy.

2. The transformer defect early warning method based on time-series segment association evidence fusion and dynamic risk map according to claim 1, characterized in that: Step (1) specifically includes the following steps: (1a) Install oil chromatography online monitoring devices, ultra-high frequency partial discharge sensors, ultrasonic partial discharge sensors, infrared thermal imagers, and acoustic fingerprint sensors on key components of the main transformer to ensure coverage of key monitoring areas; (1b) Collect multimodal time series data and first remove outliers: use the three-standard-deviation method to identify and remove obvious outliers; for negative values ​​or abnormally large values ​​in oil chromatography data caused by sampling or transmission, mark them as invalid; then perform noise filtering: for high-frequency signals such as partial discharge and acoustic signature, use wavelet thresholding to denoise, select the db4 wavelet basis for 3-level decomposition, and use soft thresholding for detail coefficients; for low-frequency slowly varying signals such as oil chromatography and temperature, use moving average filtering. (1c) The variance and first difference of the data are calculated using a sliding time window, and the fluctuation state of each data point is automatically identified and marked. The fluctuation state is divided into steady state, trend state and sudden jump state. (1c) For the time series data of each modality, the sliding time window method is used for analysis, and the variance is calculated within each window W. and the first difference mean It automatically identifies and marks the fluctuation state of each data point; ； in, All are threshold values; (1d) For different state segments, a data completion algorithm is used to process missing values ​​and all data are normalized to form a normalized time series dataset with state labels. The state labels include three types of states: steady state, trend, and sudden jump.

3. The transformer defect early warning method based on time-series segment association evidence fusion and dynamic risk map according to claim 1, characterized in that: Step (2) specifically includes the following steps: (2a) Based on the status labels and combined with the change point detection algorithm, the continuous time series data of each modality are automatically divided into time series segments with consistent trend characteristics; (2b) Using a unified time axis as a reference, align the temporal segment boundaries of all modes to form cross-modal aligned temporal analysis units; (2c) For each aligned time series analysis unit, the statistical features and trend features of each modality within it, as well as the correlation features between cross-modal data sequences within the unit, are extracted through the lightweight feature extraction module. All statistical features, trend features, and correlation features are concatenated to form a high-dimensional fusion feature vector, which is used to characterize the multimodal collaborative state of the equipment during that time period. (2d) Input the fused feature vector into an evidence generation module. The evidence generation module adopts a multilayer perceptron network. The output layer of the multilayer perceptron network has 5 neurons, which correspond to four health states and an uncertainty measure. The four health states are healthy, attentive, abnormal, and severe. The multilayer perceptron network is trained using evidence deep learning methods. The output basic probability assignment value and the uncertainty measure together constitute a temporal evidence body.

4. The transformer defect early warning method based on time-series segment association evidence fusion and dynamic risk map according to claim 1, characterized in that: Step (3) specifically includes the following steps: (3a) The time-series evidence generated in chronological order is used as the input sequence and fed into the time-series evidence fusion model. According to the rules of evidence theory, the time-series evidence fusion model regards each piece of evidence in the input sequence as a basic reliability assignment and considers the temporal dependency between evidences. It models the reliability transmission and evolution of evidence in the time dimension through a long short-term memory network (LSTM). Finally, the time-series evidence fusion model outputs the reliability distribution that represents the current comprehensive health status of the device. (3b) Constructing a dynamic proportional risk model: ； in, This is the baseline risk function fitted based on historical failure data; It is the sum of the reliability scores for abnormal and severe states in the overall reliability distribution of health status; An aging factor calculated based on the equipment's years of operation; This is the real-time operating condition stress factor calculated based on the current load rate and ambient temperature. All parameters are adjustable; The conditional probability of equipment failure in the future, i.e., the dynamic failure probability, is calculated using a dynamic proportional risk model. ； In the formula, The dynamic failure probability refers to the conditional probability that the equipment will fail within a future time period ∆t at the current time t. Dynamic proportional risk model; By calculating and mapping different futures corresponding The value generates dynamic failure probability curves for multiple time scales in the future, intuitively showing the development trend of risk over time.

5. The transformer defect early warning method based on time-series segment association evidence fusion and dynamic risk map according to claim 1, characterized in that: Step (4) specifically includes the following steps: (4a) Design and construct graph patterns in the Neo4j graph database: construct multiple types of nodes including device nodes, time-series segment pattern nodes, evidence nodes, risk probability nodes, historical case nodes and operation and maintenance strategy nodes; (4b) Define and create directed relation edges that connect nodes of multiple types, and assign quantified weight and confidence attributes to each directed relation edge to characterize the strength and reliability of the association; (4c) The graph framework is composed of the nodes constructed in step (4a) and the directed relation edges created in step (4b). Graph initialization is performed: the normalized time series dataset, time series evidence body and dynamic fault probability are injected into the graph framework and assigned initial weights and confidence to obtain an initialized graph. (4d) Establish a dynamic graph update mechanism: The system records each process from fault warning, on-site verification and fault handling as a learning sample, driving the graph learning algorithm to automatically adjust the weight and confidence of the relevant directed edges, so as to realize the continuous evolution of graph knowledge.

6. The transformer defect early warning method based on time-series segment association evidence fusion and dynamic risk map according to claim 1, characterized in that: Step (5) specifically includes the following steps: (5a) Adopting a cloud-edge collaborative architecture, the time-series evidence fusion model, dynamic proportional risk model and dynamic risk association map are deployed in the cloud, and a lightweight feature extraction and evidence generation module is deployed at the edge; (5b) Generate a quantitative early warning report based on the dynamic failure probability, which includes specific probability values, dominant defect modes, risk evolution rate and expected development time. (5c) Based on the risk characteristics in the quantitative early warning report, retrieve matching historical cases and related strategies from the dynamic risk association map, and comprehensively consider the power grid operation constraints and on-site resource conditions, and generate an operation and maintenance decision scheme with priority ranking and recommended time window through a multi-objective optimization algorithm. (5d) Push the quantitative early warning report and operation and maintenance decision-making plan to the operation and maintenance personnel, and collect the handling feedback to drive the iterative update of the dynamic risk correlation map.

7. An electronic device, comprising: processor; as well as A memory storing computer program instructions, which, when executed by the processor, cause the processor to perform the transformer defect early warning method based on time-series segment association evidence fusion and dynamic risk map as described in any one of claims 1-6.

8. A computer-readable storage medium having stored thereon computer program instructions, which, when executed by a processor, cause the processor to perform the transformer defect early warning method based on time-series segment association evidence fusion and dynamic risk map as described in any one of claims 1-6.

Images