Intelligent diagnosis system and method for power equipment based on multi-algorithm fusion diagnosis

Abstract

The application belongs to the technical field of power system automation, and discloses a power equipment intelligent diagnosis system and method based on multi-algorithm fusion diagnosis. Through flow calculation on multi-source monitoring data, then based on historical sliding window to generate dynamic threshold, effectively adapt to the equipment operation condition fluctuation, then based on game theory weight distribution strategy, traversal algorithm subset combination calculation performance difference, determine the dynamic weight of each diagnosis algorithm, then, use the credibility propagation logic, multiply the fault prior probability and the evidence credibility, and set it to 0 when the evidence credibility is less than 0, calculate the posterior probability of fault hypothesis to determine the fault type. Through dynamic weight distribution and credibility propagation logic, the application realizes adaptive fusion and accurate root cause analysis of multi-algorithm, and significantly improves the accuracy and robustness of power equipment fault diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of power system automation technology, and relates to an intelligent diagnostic system and method for power equipment based on multi-algorithm fusion diagnostics. Background Technology

[0002] In the daily operation and maintenance of substation equipment, real-time monitoring and fault diagnosis are core aspects of ensuring power grid safety. Existing technologies typically employ multi-algorithm fusion or fixed-threshold alarm methods to assess the condition of equipment such as transformers and GIS switchgear. For example, some existing solutions use a simple fusion of the David's triangle method and a modified three-ratio method, along with preset fixed thresholds to trigger alarms. Furthermore, traditional SCADA systems often employ a batch processing model for data acquisition, storage, and analysis.

[0003] However, the aforementioned existing technologies have significant shortcomings when facing complex operating conditions. First, the diagnostic logic of existing technologies is relatively rigid, often employing static weight fusion, meaning the proportion of each algorithm is manually set. This approach cannot adapt to feature drift caused by equipment aging, and the diagnostic accuracy drops significantly when facing mixed faults. Second, the alarm adaptability of existing technologies is poor; fixed thresholds cannot adaptively adjust to changes in ambient temperature and load current, making them prone to false alarms in specific environments such as summer. Finally, the data processing architecture of existing technologies suffers from real-time bottlenecks; data must be written to disk before being read and analyzed, resulting in high data transfer delays and failing to meet the requirements for rapid fault response.

[0004] Therefore, there is an urgent need for a diagnostic system that can dynamically adjust weights, adaptively set thresholds, and has real-time processing capabilities. Summary of the Invention

[0005] To address the problems existing in the background technology, this invention proposes an intelligent diagnostic system and method for power equipment based on multi-algorithm fusion diagnosis.

[0006] The first aspect of the present invention provides an intelligent diagnostic system for power equipment based on multi-algorithm fusion diagnostics, comprising:

[0007] Edge data sensing layer, used for real-time acquisition of multi-source monitoring data of power equipment;

[0008] The streaming computing layer is used to process the multi-source monitoring data in real time, calculate dynamic thresholds, and perform adaptive delay alarm determination.

[0009] The cognitive reasoning layer is used to construct a knowledge graph of power equipment fault evolution and perform root cause analysis of faults based on credibility propagation logic.

[0010] The cognitive reasoning layer includes a game theory weight allocation unit, which is used to evaluate the contribution value of each diagnostic algorithm based on marginal contribution and allocate dynamic weights.

[0011] The application service layer is used to receive the diagnostic results from the cognitive reasoning layer and generate operation and maintenance strategies.

[0012] Optionally, the streaming computing layer includes:

[0013] The dynamic threshold calculation subunit is used to acquire historical monitoring data within the same time period in the historical sliding window, calculate the sliding average and standard deviation of the historical monitoring data, and determine the real-time dynamic threshold by the sum of the products of the sliding average, the standard deviation and the adaptive coefficient.

[0014] The delayed alarm determination subunit is used to determine whether the current monitoring data continuously exceeds the real-time dynamic threshold and whether the duration is greater than the preset delay time. If so, an alarm event is triggered.

[0015] Optionally, the cognitive reasoning layer includes:

[0016] The knowledge graph construction module is used to construct a graph by using oil chromatography gas data, electrical quantity data and environmental data as entity nodes and fault mechanisms as relation edges.

[0017] The credibility calculation module is used to calculate the posterior probability of the failure hypothesis using credibility propagation logic. The processing procedure of the credibility propagation logic is as follows: multiply the prior probability of failure by the value of the evidence credibility. If the evidence credibility is less than 0, then set the evidence credibility to 0. The output result is the credibility of the failure hypothesis under the support of the corresponding evidence.

[0018] Optionally, the game theory weight allocation unit includes:

[0019] The marginal contribution calculation submodule is used to traverse the subset combinations of all diagnostic algorithm sets and calculate the difference in diagnostic efficiency value of each diagnostic algorithm before and after it is added to a specific subset.

[0020] The weight determination submodule is used to perform a weighted summation of all the diagnostic efficacy value differences according to the size ratio of each subset, and use the calculated value as the dynamic weight of each diagnostic algorithm.

[0021] A second aspect of the present invention provides a method for intelligent diagnosis of power equipment based on multi-algorithm fusion diagnosis, comprising:

[0022] Real-time acquisition of multi-source monitoring data from power equipment;

[0023] The multi-source monitoring data is processed by streaming computation, and a dynamic threshold is generated based on a historical sliding window. If the monitoring data continuously exceeds the dynamic threshold and the duration is greater than a preset delay time, an alarm event is generated.

[0024] Construct a knowledge graph of power equipment fault evolution, and input the alarm events as evidence into the knowledge graph;

[0025] The dynamic weights of each diagnostic algorithm are determined based on a game theory weighting strategy. The weighted diagnostic algorithms are then used to perform root cause analysis of the faults and output the fault type and confidence level.

[0026] Optionally, the determination of the dynamic weights of each diagnostic algorithm based on the game theory weight allocation strategy includes:

[0027] Obtain the historical diagnostic performance values ​​of each diagnostic algorithm;

[0028] Iterate through all subset combinations of diagnostic algorithms and calculate the difference in diagnostic performance of each subset before and after adding it to a specific subset.

[0029] The differences in all the diagnostic efficacy values ​​are weighted and summed according to the size proportion of each subset, and the calculated values ​​are used as the dynamic weights of each diagnostic algorithm.

[0030] Optionally, the step of performing root cause analysis using a weighted diagnostic algorithm includes:

[0031] The posterior probability of the failure hypothesis under the current evidence is calculated using the credibility propagation logic. The processing of the credibility propagation logic is as follows: the prior probability of failure is multiplied by the value of the evidence credibility. If the evidence credibility is less than 0, the evidence credibility is set to 0. The output result is the credibility of the failure hypothesis under the support of the corresponding evidence.

[0032] The fault type is determined based on the posterior probability.

[0033] Optionally, streaming computation is performed on the multi-source monitoring data to generate dynamic thresholds based on historical sliding windows, including:

[0034] Obtain historical monitoring data for the same time period within a preset time frame;

[0035] Calculate the moving average and standard deviation of the historical monitoring data;

[0036] The sum of the products of the moving average, the adaptive coefficient, and the standard deviation is determined as the real-time dynamic threshold.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] This invention discloses an intelligent diagnostic system and method for power equipment based on multi-algorithm fusion diagnosis. It achieves dynamic weight allocation of diagnostic algorithms through a game theory weighting strategy, solving the problem that static weights cannot adapt to changes in operating conditions and improving the accuracy of hybrid fault diagnosis. This invention generates dynamic thresholds through historical sliding windows and combines them with delayed alarm judgment to achieve adaptive adjustment of alarm thresholds, effectively reducing the false alarm rate. This invention adopts a streaming computing architecture, achieving millisecond-level real-time processing of monitoring data, meeting the requirements for rapid fault response. This invention utilizes credibility propagation logic for root cause analysis, effectively eliminating interfering data and enhancing diagnostic robustness by setting the credibility of negative evidence to 0. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of an intelligent diagnostic system for power equipment based on multi-algorithm fusion diagnostics in one embodiment of the present invention;

[0040] Figure 2 This is a flowchart of a power equipment intelligent diagnostic method based on multi-algorithm fusion diagnostics in one embodiment of the present invention. Detailed Implementation

[0041] 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 only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] In one embodiment, such as Figure 1 As shown, a power equipment intelligent diagnostic system based on multi-algorithm fusion diagnosis is provided. This power equipment intelligent diagnostic system based on multi-algorithm fusion diagnosis corresponds one-to-one with the power equipment intelligent diagnostic method based on multi-algorithm fusion diagnosis in the following embodiments. The power equipment intelligent diagnostic system based on multi-algorithm fusion diagnosis includes: an edge data perception layer, a streaming computing layer, a cognitive reasoning layer, and an application service layer. The functional modules are described in detail below:

[0043] The system comprises: an edge data perception layer for real-time acquisition of multi-source monitoring data from power equipment; a streaming computing layer for real-time processing of the multi-source monitoring data, calculation of dynamic thresholds, and execution of adaptive delay alarm determination; a cognitive reasoning layer for constructing a knowledge graph of power equipment fault evolution and performing root cause analysis of faults based on credibility propagation logic; wherein the cognitive reasoning layer includes a game theory weight allocation unit for evaluating the contribution value of each diagnostic algorithm based on marginal contribution and allocating dynamic weights; and an application service layer for receiving the diagnostic results from the cognitive reasoning layer and generating operation and maintenance strategies.

[0044] In one embodiment, the application service layer is deployed in an edge computing gateway or intelligent electronic device at the substation site to achieve millisecond-level real-time acquisition and preliminary processing of power equipment operating status data. The data types collected by the application service layer are not single-dimensional, but cover multi-source heterogeneous monitoring data of power equipment operating status.

[0045] Specifically, the multi-source monitoring data includes at least oil chromatography gas data, electrical quantity data, and environmental data from the power equipment. The oil chromatography gas data reflects the decomposition of the insulating oil inside the equipment. The electrical quantity data includes key operating parameters such as the equipment's load current and winding temperature. The environmental data includes the temperature information of the environment in which the equipment is located. These data serve as the foundational input for subsequent flow cytometry and cognitive reasoning, providing comprehensive data support for holographic diagnostics.

[0046] When the system of this invention is applied to internal fault diagnosis of transformers, the edge data sensing layer initiates a real-time data acquisition process. The edge gateway continuously reads gas concentration data from the transformer's online oil chromatography monitoring device. Simultaneously, the edge gateway also synchronously acquires transformer winding temperature data and current load current data. For example, in the scenario of monitoring internal overheating faults in transformers, the edge data sensing layer captures real-time changes in the ethylene gas concentration in the oil. The edge data sensing layer timestamps the acquired data and encapsulates it into a data stream using a preset communication protocol, sending it to the system's streaming computing layer for further processing.

[0047] The edge data perception layer enables comprehensive awareness of the substation equipment status. It transforms the traditional batch processing model of "collection, storage, and analysis" into a streaming processing model of "collection as analysis," significantly reducing data flow latency. The edge data perception layer can quickly acquire the operating characteristics of equipment under different conditions, laying a solid data foundation for achieving millisecond-level real-time alarms and high-accuracy fault diagnosis.

[0048] In one embodiment, the streaming computing layer is a component of the system architecture of the present invention that enables millisecond-level real-time response. The streaming computing layer overcomes the drawbacks of traditional technologies that rely on fixed thresholds and batch processing modes by deploying a lightweight streaming computing engine to receive multi-source monitoring data streams uploaded from the edge data perception layer. The streaming computing layer processes the data streams in real time, dynamically calculates alarm thresholds, and performs adaptive alarm determination in conjunction with latency logic.

[0049] Specifically, the real-time processing includes constructing a sliding window and calculating a dynamic threshold. The streaming computing layer utilizes sliding window technology to slice and analyze the real-time data stream. The algorithm logic for calculating the dynamic threshold is as follows: First, the historical moving average for the same time period over the past 24 hours is calculated to eliminate interference from diurnal temperature differences and periodic equipment operation. Then, by calculating the standard deviation of the historical data, the system generates a real-time dynamic threshold based on the above statistics. The formula for calculating the real-time dynamic threshold is the historical moving average plus an adaptive coefficient multiplied by the standard deviation of the historical data.

[0050] The adaptive coefficient is not a fixed constant but dynamically adjusted based on the real-time load rate of the equipment. When the equipment is operating under high load, the adaptive coefficient automatically tightens. This design allows the alarm threshold to adapt to the operating conditions of the equipment. When performing adaptive delay alarm judgment, the system not only determines whether the data exceeds the dynamic threshold but also introduces a duration criterion. The system will only trigger an alarm signal when the monitored data continuously exceeds the dynamic threshold and the duration is greater than a preset time threshold. This delay logic can effectively filter out instantaneous interference caused by lightning strikes or switching operations, avoiding false alarms.

[0051] For example, when the system of the present invention detects an abnormal temperature in the transformer winding, the streaming computing layer will initiate the above-mentioned judgment process. Assuming that the real-time temperature data uploaded by the edge layer is 85°C, the streaming computing layer calls historical data to calculate that the historical average temperature for the current period is 70°C with a standard deviation of 5°C. The system calculates the real-time dynamic threshold as 80°C according to the formula. Since the real-time temperature of 85°C is greater than the dynamic threshold of 80°C, the system enters the delayed judgment stage. If the high temperature state lasts for three minutes and the preset time threshold is two minutes, the system determines it as a temperature abnormality event and generates alarm information to be uploaded to the cognitive reasoning layer.

[0052] In this invention, the streaming computing layer represents a leap from "static rule matching" to "dynamic cognitive reasoning." This layer solves the problem of high false alarm rates caused by fixed thresholds failing to adapt to environmental changes in existing technologies. Utilizing streaming computing technology, the layer achieves millisecond-level data flow and analysis, meeting the stringent requirements of power systems for rapid fault response. Through a dual mechanism of dynamic thresholds and delay-based determination, the streaming computing layer significantly improves the accuracy and reliability of alarms.

[0053] Furthermore, the streaming computing layer includes: a dynamic threshold calculation subunit, used to acquire historical monitoring data within the same time period in a historical sliding window, calculate the sliding average and standard deviation of the historical monitoring data, and determine the sum of the products of the sliding average, the standard deviation and the adaptive coefficient as the real-time dynamic threshold; and a delay alarm determination subunit, used to determine whether the current monitoring data continuously exceeds the real-time dynamic threshold and whether the duration is greater than a preset delay time, and if so, trigger an alarm event.

[0054] Specifically, the streaming computing layer is the hub connecting edge perception and cognitive reasoning in the system architecture of this invention. The streaming computing layer is mainly used for high-speed cleaning and feature extraction of massive real-time monitoring data. This invention achieves accurate capture of abnormal equipment states through the collaborative work of a dynamic threshold calculation subunit and a delayed alarm judgment subunit. The dynamic threshold calculation subunit is responsible for establishing an adaptive monitoring benchmark that changes with operating conditions. Traditional fixed thresholds cannot adapt to data fluctuations of power equipment under different loads. The dynamic threshold calculation subunit first acquires historical monitoring data within the same time period within a historical sliding window. The length of the historical sliding window is set according to the data periodicity. The dynamic threshold calculation subunit calculates the moving average of the historical monitoring data and simultaneously calculates the standard deviation of the historical monitoring data. The standard deviation reflects the fluctuation range of the data under normal operating conditions. The dynamic threshold calculation subunit adds the product of the moving average, the standard deviation, and the adaptive coefficient to obtain the real-time dynamic threshold. This real-time threshold can dynamically adjust with changes in the equipment's operating state, effectively avoiding false alarms caused by load fluctuations.

[0055] The delayed alarm determination subunit is used to filter transient interference signals. The power field environment is complex, and sensors are highly susceptible to transient interference such as electromagnetic pulses. If an alarm is triggered solely based on a single point data exceeding a limit, it will result in a large number of invalid alarms. The delayed alarm determination subunit determines whether the current monitored data continuously exceeds the real-time dynamic threshold, and simultaneously determines whether the duration of this exceeding state is greater than a preset delay time. Only when both conditions are met will the delayed alarm determination subunit trigger an alarm event. If the current monitored data exceeds the threshold but the duration does not meet the standard, it is considered transient interference, and the system will not trigger an alarm.

[0056] For example, when a transformer is operating under high temperature and heavy load in summer, its winding temperature reference value naturally rises. The dynamic threshold calculation subunit obtains historical temperature data for the same period over the past 7 days, calculates a high moving average with a moderate standard deviation, and calculates a high real-time dynamic threshold accordingly. If a strong electromagnetic interference causes a momentary jump in the temperature sensor reading, briefly exceeding the real-time dynamic threshold, the delayed alarm determination subunit detects that this over-limit state lasts only a very short time, less than the preset delay time, and the system determines this to be an interference signal, without triggering an alarm. Conversely, if a cooling system malfunction causes the temperature to rise continuously, and exceeds the real-time dynamic threshold for multiple consecutive sampling periods, the system immediately triggers an alarm event and sends it to the cognitive inference layer after the duration exceeds the preset delay time.

[0057] In this invention, the streaming computing layer achieves adaptability and anti-interference capability in anomaly detection. The dynamic threshold calculation subunit solves the problem of missed or false alarms under varying operating conditions caused by fixed thresholds. The delayed alarm determination subunit effectively eliminates interference from instantaneous noise. The streaming computing layer significantly reduces the number of invalid alarms, alleviates the computational burden on the subsequent cognitive inference layer, and improves the overall operating efficiency and reliability of the system.

[0058] In one embodiment, the cognitive reasoning layer is the logical unit for realizing holographic diagnosis of substation equipment in this invention, and is deployed on a cloud server or high-performance computing node. The cognitive reasoning layer is used to simulate the cognitive process of experts, specifically by constructing a knowledge graph of power equipment fault evolution and combining credibility propagation logic and game theory weight allocation mechanism to achieve in-depth root cause analysis of equipment faults.

[0059] The cognitive reasoning layer first constructs a knowledge graph of power equipment fault evolution. This knowledge graph uses oil chromatography gas data, electrical quantity data, and environmental data as entity nodes, and fault evolution mechanisms as relational edges. For example, the physical process of insulation degradation leading to partial discharge, which in turn triggers gas generation, is abstracted as a connection path in the graph. When the system receives abnormal data uploaded from the edge side, the cognitive reasoning layer initiates a graph retrieval mechanism to search the knowledge graph for fault evolution paths that match the current evidence.

[0060] The cognitive reasoning layer employs a credibility propagation logic for reasoning calculations. This logic introduces a credibility factor to assess the probability of the fault hypothesis being true. The system sets a prior probability of the fault, which is determined based on the operating years in the equipment ledger. Simultaneously, the system calculates the credibility of the evidence, which is determined based on the degree of data anomaly. Through the credibility propagation algorithm, the system dynamically calculates the posterior probability of the fault hypothesis being true, supported by the current evidence. This represents a leap from static rule matching to dynamic probabilistic reasoning, significantly reducing the misjudgment rate of mixed faults.

[0061] The cognitive reasoning layer includes a game theory weight allocation unit. This unit addresses the weight allocation problem in multi-algorithm fusion. Traditional methods use manually set static weights, which cannot adapt to changes in device state. This invention uses the Shapley value method to calculate the dynamic weights of each diagnostic algorithm. The diagnostic algorithms include the David's triangle method, the modified three-ratio method, and the neural network method. The game theory weight allocation unit evaluates the marginal contribution of each algorithm to the diagnostic results in real time. When the device is in a low-energy discharge stage, the system automatically identifies the David's triangle method as having a higher contribution value and increases its weight accordingly. When the device is in an overheating stage, the system automatically identifies the three-ratio method as having a higher contribution value and increases its weight accordingly.

[0062] For example, when a transformer experiences a winding hotspot overheating fault, the cognitive reasoning layer receives an abnormal event from the streaming computing layer. This event includes a sudden increase in ethylene gas concentration, a rise in winding temperature, and high load current data. The game theory weight allocation unit calculates that the three-ratio method has the largest marginal contribution under this operating condition and dynamically adjusts its weight to the highest. Simultaneously, the cognitive reasoning layer searches for causal chains in the fault evolution knowledge graph. The system finds that the increased load leads to increased copper losses, which in turn causes an increase in oil temperature and ethylene precipitation—a path that highly matches the current evidence. Combined with highly credible overheating evidence, the system ultimately determines that the root cause of the fault is winding hotspot overheating.

[0063] In this invention, the cognitive reasoning layer enables in-depth analysis of the fault mechanism. This layer can not only identify the fault symptoms but also trace the root cause of the fault. The game theory weight allocation unit enables the diagnostic strategy to adaptively adjust according to the equipment operating conditions, effectively solving the feature drift problem caused by equipment aging or changes in operating conditions, and significantly improving the accuracy and robustness of the diagnostic results.

[0064] Furthermore, the cognitive reasoning layer includes: a knowledge graph construction module, used to construct a graph by using oil chromatography gas data, electrical quantity data, and environmental data as entity nodes and fault mechanisms as relation edges; and a credibility calculation module, used to calculate the posterior probability of the fault hypothesis using credibility propagation logic, wherein the processing procedure of the credibility propagation logic is as follows: multiply the prior probability of the fault by the value of the evidence credibility; if the evidence credibility is less than 0, then set the evidence credibility to 0, and the output result is the credibility of the fault hypothesis under the support of the corresponding evidence.

[0065] Specifically, the cognitive reasoning layer is the core analysis unit of the system of this invention. Deployed on a cloud server, this layer integrates a knowledge graph construction module and a credibility calculation module. These two modules work together to achieve in-depth understanding and quantitative assessment of power equipment faults. The knowledge graph construction module is responsible for building a structured fault knowledge base. This module defines oil chromatography gas data, electrical quantity data, and environmental data as entity nodes, and fault mechanisms as relational edges. For example, the module establishes connections between gas concentration nodes (such as hydrogen and acetylene) and fault type nodes (such as partial discharge and high-temperature overheating). These connections represent the physicochemical mechanisms between gas evolution and fault types. In this way, a fault evolution network covering multi-source data is constructed, providing a topological foundation for subsequent reasoning. The credibility calculation module uses credibility propagation logic to process uncertain information. In power equipment diagnosis, single evidence is often insufficient to completely determine a fault; it is necessary to synthesize the credibility of multiple pieces of evidence. The credibility calculation module first obtains the prior probability of the fault hypothesis, which is determined based on the historical operating data of the equipment. Subsequently, the credibility calculation module obtains the current monitoring data as evidence credibility. The module executes credibility propagation logic, multiplying the prior probability of the fault by the value of the evidence credibility to calculate the credibility of the fault hypothesis supported by the corresponding evidence. This credibility propagation logic includes a specific numerical correction mechanism. During the calculation process, the credibility calculation module continuously assesses the evidence credibility value. If the evidence credibility is less than 0, the module sets the evidence credibility to 0. This means that when a piece of evidence does not support the fault hypothesis, or when the evidence has a reverse indicative effect, the system cuts off the negative contribution of that evidence to the fault hypothesis, preventing it from unreasonably lowering the overall diagnostic credibility. The final output is the corrected fault hypothesis credibility.

[0066] For example, when the system detects a significant increase in acetylene gas concentration in transformer oil, the knowledge graph construction module locates the connection between the acetylene node and the arc discharge fault node in the graph. The credibility calculation module retrieves the prior probability of the arc discharge fault and simultaneously calculates the credibility of the evidence of abnormal acetylene concentration. Assuming the calculated credibility is negative, it indicates that the concentration change may be caused by other non-fault factors. The credibility propagation logic sets this negative value to 0, and the system outputs that the support of this evidence for the arc discharge fault is 0, avoiding misdiagnosis. If the credibility is positive, the system multiplies it by the prior probability, outputting a higher fault credibility. In this invention, the cognitive reasoning layer implements intelligent reasoning based on the knowledge graph. The knowledge graph construction module semantically associates multi-source heterogeneous data, and the credibility calculation module handles the uncertainty in diagnosis through specific propagation logic. The method of setting negative evidence to 0 effectively avoids the misleading influence of interfering data on the diagnostic results, significantly improving the robustness and accuracy of fault diagnosis under complex operating conditions.

[0067] Furthermore, the game theory weight allocation unit includes: a marginal contribution calculation submodule, used to traverse all subset combinations of the diagnostic algorithm set and calculate the difference in diagnostic efficacy value of each diagnostic algorithm before and after it is added to a specific subset; and a weight determination submodule, used to perform a weighted summation of all the diagnostic efficacy value differences according to the size ratio of each subset, and use the calculated value as the dynamic weight of each diagnostic algorithm.

[0068] Specifically, the game theory weight allocation unit is a core component of the cognitive reasoning layer for achieving adaptive fusion of multiple algorithms. This unit aims to solve the problem that traditional fixed weights cannot adapt to complex working conditions. Internally, it includes a marginal contribution calculation submodule and a weight determination submodule. These two submodules work together to achieve accurate quantification of the contribution of each diagnostic algorithm.

[0069] The marginal contribution calculation submodule is responsible for evaluating the incremental value of a single algorithm in a specific combination. This submodule first iterates through subset combinations of all diagnostic algorithms. For each specific subset combination, the submodule calculates the overall diagnostic efficiency value when the target diagnostic algorithm is included, and the overall diagnostic efficiency value when it is not included. Then, the submodule subtracts these two diagnostic efficiency values; the difference is the marginal contribution of the target diagnostic algorithm before and after adding it to that specific subset. This process objectively reflects the actual improvement in diagnostic results of the algorithm under the current combination.

[0070] The weight determination submodule is used to comprehensively evaluate and output the final dynamic weights. This submodule obtains the differences in diagnostic efficacy values ​​output by the marginal contribution calculation submodule. To reflect the differences in the probability of occurrence of subsets of different sizes, this submodule performs a weighted summation of these differences according to the size proportion of each subset. Specifically, smaller or larger subsets have different weighting factors than subsets of intermediate size. After weighted summation, this submodule uses the final calculated value as the dynamic weight of the corresponding diagnostic algorithm. This Shapley value-based calculation method ensures the fairness and rationality of the weight allocation.

[0071] For example, when the system needs to use the David's Triangle method, the modified three-ratio method, and the neural network method simultaneously for fault diagnosis, the game theory weight allocation unit initiates the calculation process. The marginal contribution calculation submodule iterates through all possible combinations of these three algorithms. For instance, in a subset containing only the David's Triangle method, it calculates the difference in diagnostic efficacy after introducing the modified three-ratio method. This difference reflects the marginal contribution of the modified three-ratio method on this basis. Subsequently, the weight determination submodule collects all efficacy difference values. This submodule performs a weighted sum based on the proportion of subset sizes containing one, two, and three algorithms, ultimately deriving the dynamic weight of each algorithm under the current data characteristics. If the current device is in an overheated state, the system automatically assigns a higher dynamic weight to the three-ratio method.

[0072] In this invention, the game theory weight allocation unit enables intelligent adjustment of the diagnostic strategy. The marginal contribution calculation submodule accurately identifies the true utility of each algorithm. The weight determination submodule eliminates subjective bias through a scientific weighting mechanism. This unit allows the system to automatically optimize the contribution of algorithm combinations based on the characteristics of real-time monitoring data, significantly improving the diagnostic accuracy and robustness of the system under varying operating conditions.

[0073] In one embodiment, the application service layer is the top-level interaction unit of the system architecture of the present invention. This layer is responsible for receiving the deep diagnostic conclusions output by the cognitive reasoning layer and transforming the abstract diagnostic data into specific operation and maintenance execution instructions. This layer realizes closed-loop control from intelligent diagnosis to actual production operation and maintenance, and is used to generate operable equipment maintenance strategies.

[0074] The application service layer first receives a root cause analysis report from the cognitive reasoning layer. This report includes the specific location of the fault, its type and attributes, and a confidence assessment of the diagnostic results. The application service layer incorporates an operation and maintenance strategy mapping engine, which associates fault types with standard operating procedures based on preset power equipment maintenance protocols. The system automatically classifies operation and maintenance response levels according to the severity of the fault. For example, for critical defects, the system generates an emergency strategy for immediate power outage maintenance; for general defects, the system generates a routine strategy for tracking and monitoring or scheduling maintenance in conjunction with a power outage plan.

[0075] The operation and maintenance strategy generated by the application service layer includes specific handling suggestions based on the analysis results of the fault mechanism. The system not only identifies where the equipment malfunctions but also clearly indicates how to handle it. The operation and maintenance strategy is presented in the form of a structured report. This report can be directly pushed to mobile terminals at the substation site or integrated into the power production management system, enabling operation and maintenance personnel to quickly understand the equipment status and accurately perform corresponding maintenance operations.

[0076] For example, when the cognitive reasoning layer determines that the transformer has a winding hotspot overheating fault, the application service layer receives the diagnostic result. The system identifies the fault as an overheating fault, and the location points to the winding area. The confidence assessment shows that the diagnostic result is reliable. The application service layer calls the operation and maintenance strategy mapping engine to retrieve the corresponding standard handling procedure. The system generates a diagnostic report containing the fault location, fault type, and specific handling suggestions. The handling suggestions clearly indicate that the tightness of the tap changer should be checked, and the strategy is pushed to the handheld device of the on-site operation and maintenance personnel. The operation and maintenance personnel conduct targeted troubleshooting according to the strategy, quickly eliminating potential equipment hazards.

[0077] In this invention, the application service layer realizes the value transformation of diagnostic results. This layer transforms complex algorithm reasoning results into intuitive instructions that front-line operation and maintenance personnel can understand, effectively shortening the time interval between fault assessment and maintenance decision-making. This layer improves the intelligence level of substation equipment operation and maintenance and ensures the safe and stable operation of the power system.

[0078] Specific limitations regarding the intelligent diagnostic system for power equipment based on multi-algorithm fusion diagnosis can be found in the limitations of the intelligent diagnostic method for power equipment based on multi-algorithm fusion diagnosis below, and will not be repeated here. Each module in the aforementioned intelligent diagnostic system for power equipment based on multi-algorithm fusion diagnosis can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0079] In one embodiment, such as Figure 2 As shown, a method for intelligent diagnosis of power equipment based on multi-algorithm fusion diagnosis is provided, and this method is applied to... Figure 2 Taking China as an example, the following specific steps will be used:

[0080] S10: Real-time acquisition of multi-source monitoring data of power equipment.

[0081] S20: Perform streaming computation on the multi-source monitoring data, generate a dynamic threshold based on the historical sliding window, and generate an alarm event if the monitoring data continuously exceeds the dynamic threshold and the duration is greater than the preset delay time.

[0082] Specifically, this invention first collects multi-source monitoring data from power equipment in real time. This multi-source monitoring data covers multi-dimensional information reflecting the equipment's operating status, including physical quantities such as transformer oil chromatographic gas concentration, bushing dielectric loss value, core grounding current, and ambient temperature. The data acquisition process is achieved through intelligent sensing terminals deployed on the equipment itself. The sensing terminals synchronously capture the aforementioned physical quantities at millisecond-level frequencies and convert the collected raw analog signals into digital signals, forming a continuous data stream. Collecting multi-source monitoring data from power equipment ensures a highly timely data foundation for subsequent analysis.

[0083] Subsequently, the system performs streaming computation on the multi-source monitoring data. Streaming computation refers to processing data in real time as it is generated. The system does not need to wait for all data to be stored; instead, it directly cleans and extracts features from the data stream in memory. During streaming computation, the system generates a dynamic threshold based on a historical sliding window. This historical sliding window refers to a data interval that traces back a preset time length from the current moment. The system reads the historical monitoring data within this window and calculates the moving average and standard deviation of the historical data. The moving average reflects the baseline level of the equipment under the current operating conditions, and the standard deviation reflects the fluctuation range of the data. The system determines the real-time dynamic threshold by summing the products of the moving average, standard deviation, and adaptive coefficient. This threshold can adaptively adjust according to changes in equipment load and environment.

[0084] After generating the dynamic threshold, the system executes anomaly detection logic to determine whether the current monitoring data continuously exceeds the dynamic threshold. If the current monitoring data only momentarily exceeds the threshold, the system considers it noise interference. If the current monitoring data exceeds the threshold, the system starts a timer to determine whether the duration of the over-limit state is greater than a preset delay time. Only when the monitoring data continuously exceeds the dynamic threshold and the duration is greater than the preset delay time, the system determines that the device has a substantial anomaly. At this time, the system generates an alarm event, which includes the abnormal device ID, the abnormal data type, and the time when the anomaly occurred.

[0085] For example, the system collects real-time top-layer oil temperature data of a main transformer. Assume the current ambient temperature is 30℃ and the transformer is operating under high load. Based on a historical sliding window of the past 24 hours, the system calculates the dynamic threshold for the current top-layer oil temperature to be 85℃. At this moment, due to external electromagnetic interference, the sensor collects an instantaneous jump value of 90℃. Although this value exceeds the dynamic threshold, it only lasts for one sampling cycle, less than the preset delay time. The system determines this as interference and does not generate an alarm event. Subsequently, due to a cooler malfunction, the top-layer oil temperature continues to rise, with data collected for 10 consecutive minutes exceeding 85℃, and the duration exceeding the preset delay time. The system determines that the transformer has a potential overheating hazard, immediately generates an alarm event, and uploads it to the monitoring center.

[0086] This invention achieves precise sensing of power equipment status. Based on a dynamic threshold generation mechanism using a historical sliding window, it solves the problem that fixed thresholds cannot adapt to varying operating conditions. Combined with streaming calculation logic based on duration determination, it effectively filters out transient pulse interference, reducing the false alarm rate. This invention can promptly detect latent equipment faults, ensuring the safe and stable operation of the power system.

[0087] S30: Construct a knowledge graph of power equipment fault evolution, and input the alarm event as evidence into the knowledge graph.

[0088] Specifically, the construction of the power equipment fault evolution knowledge graph aims to transform the scattered fault data of power equipment into a structured knowledge network. This invention first designs the ontology layer, which defines the skeleton of the knowledge graph, including entity types, relation types, and attribute definitions. Entity types cover power equipment objects such as transformers, circuit breakers, and bushings, as well as fault modes such as partial discharge, overheating, and insulation aging. Relationship types describe the logical connections between entities, such as the inclusion relationship between equipment and fault phenomena, and the causal relationship between fault causes and fault results.

[0089] After completing the ontology layer design, the data layer is instantiated. The system extracts data from historical maintenance records, fault reports, and expert knowledge bases. Natural language processing technology is used to identify entities and relationships in the text, transforming unstructured text into triple structures. Each triple consists of a head entity, a relation, and a tail entity. For example, "transformer," "occurrence," and "partial discharge" constitute a triple. The system stores all triples in a graph database, forming a knowledge graph of power equipment fault evolution. This graph fully records the complete evolution chain of a fault, from its cause to its phenomenon and then to its consequence.

[0090] The step of inputting the alarm event as evidence into the knowledge graph is the initiation stage of the reasoning process. The alarm event is generated by the aforementioned streaming computing layer and carries specific fault characteristic information. The system maps the key information in the alarm event to entity nodes in the knowledge graph. These entity nodes constitute the evidence set for reasoning. For example, if the alarm event indicates that acetylene gas exceeds the standard, the system will locate the acetylene node in the knowledge graph and activate it as an evidence node.

[0091] For example, when the system receives an alarm event regarding excessively high transformer oil temperature and abnormal acetylene concentration, the system first parses the alarm content, extracting two features: excessively high oil temperature and abnormal acetylene concentration. Then, it searches for corresponding nodes in the constructed power equipment fault evolution knowledge graph. The system finds a correlation between the excessively high oil temperature node and the high-temperature overheating fault node, and a correlation between the abnormal acetylene concentration node and the arc discharge fault node. The system uses these alarm features as conclusive evidence and inputs them into the inference engine. The inference engine backtracks based on the paths in the graph to analyze the upstream causes leading to these phenomena.

[0092] The construction of a knowledge graph for the evolution of power equipment faults enables deep association of fault knowledge. By inputting alarm events as evidence into the knowledge graph, the system can perform logical deduction based on factual data, breaking through the limitations of rigid rules in traditional expert systems. It can effectively handle complex and ever-changing composite faults and significantly improve the accuracy and interpretability of fault diagnosis.

[0093] S40: Determine the dynamic weights of each diagnostic algorithm based on a game theory weight allocation strategy, use the weighted diagnostic algorithm to perform root cause analysis of the fault, and output the fault type and confidence level.

[0094] Specifically, determining the dynamic weights of each diagnostic algorithm based on a game theory weight allocation strategy is key to improving the accuracy of fault diagnosis in this invention. In the process of power equipment fault diagnosis, a single diagnostic algorithm often has limitations, and different algorithms perform significantly differently under different fault types. To fully leverage the advantages of each algorithm, this invention introduces the Shapley value concept from game theory to calculate the weights. The system first constructs a set containing multiple diagnostic algorithms, including the three-ratio method based on dissolved gas analysis in oil, the Duval triangular method, and a neural network diagnostic model based on machine learning. The system traverses all subset combinations of this set and calculates the marginal contribution of each diagnostic algorithm to the overall diagnostic result under different subset combinations.

[0095] The marginal contribution refers to the improvement in the overall accuracy of the diagnostic system after adding a specific diagnostic algorithm. The system performs a weighted sum of the marginal contributions for all subset combinations. The weight coefficient is related to the number of algorithms included in the subset combination. The Shapley value of each diagnostic algorithm is calculated, and this Shapley value is determined as the dynamic weight of that algorithm under the current operating conditions. For example, when monitoring data shows that the equipment is in the overheating fault characteristic range, the Shapley value of the three-ratio method is higher, and the system automatically assigns a larger dynamic weight to the three-ratio method. When the monitoring data exhibits complex nonlinear characteristics, the Shapley value of the neural network diagnostic model increases, and the system correspondingly increases the dynamic weight of that model. This dynamic adjustment mechanism ensures the scientific and adaptive nature of the weight allocation.

[0096] The system's reasoning process involves using weighted diagnostic algorithms to perform root cause analysis. The system receives pre-processed multi-source monitoring data from power equipment and simultaneously inputs this data into multiple diagnostic algorithms. Each algorithm outputs a preliminary diagnostic result and its corresponding initial confidence level based on its own logic. Subsequently, the system calls upon dynamic weights calculated using a game-theoretic weighting strategy, multiplying each algorithm's initial confidence level by its corresponding dynamic weight to obtain a weighted confidence score. The system then merges the weighted results from all algorithms and determines the final root cause of the fault through a voting mechanism or a probabilistic fusion algorithm.

[0097] For example, suppose the system detects an anomaly in a transformer. The three-ratio method identifies it as a high-energy discharge with an initial confidence level of 0.8. The Duval trigonometry method identifies it as a low-energy discharge with an initial confidence level of 0.6. The neural network model identifies it as a high-energy discharge with an initial confidence level of 0.9. At this point, the game theory weighting strategy, based on the current data characteristics, calculates the dynamic weights for the three-ratio method as 0.4, the Duval trigonometry method as 0.2, and the neural network model as 0.4. The system calculates the weighted scores for each algorithm: the three-ratio method has a weighted score of 0.32, the Duval trigonometry method has a weighted score of 0.12, and the neural network model has a weighted score of 0.36. After comprehensive comparison, the system determines the fault type as a high-energy discharge.

[0098] Outputting the fault type and confidence level is the final step in the diagnostic process. The system outputs the fused analysis results in a formatted format. The fault type clearly indicates the specific nature of the equipment fault, such as partial discharge, overheating, or arcing. The confidence level is a value between 0 and 1, which quantifies the reliability of the diagnostic results. Maintenance personnel can determine the maintenance strategy based on the confidence level. If the confidence level is high, an immediate power outage maintenance is arranged; if the confidence level is low, the monitoring frequency is increased.

[0099] This invention achieves intelligent and precise fault diagnosis. Based on a game theory-based weight allocation strategy, it effectively solves the weight allocation problem in multi-algorithm fusion, avoiding the subjectivity and bias of manually setting weights. By utilizing the weighted diagnostic algorithm for root cause analysis, it significantly improves the system's diagnostic robustness under complex operating conditions. The output fault types and confidence levels provide strong decision support for condition-based maintenance of power equipment.

[0100] In one embodiment, determining the dynamic weights of each diagnostic algorithm based on a game theory weight allocation strategy includes: obtaining the historical diagnostic performance values ​​of each diagnostic algorithm; traversing all subset combinations of the diagnostic algorithm set and calculating the difference in diagnostic performance values ​​of each subset before and after each diagnostic algorithm is added to a specific subset; weighting and summing all the differences in diagnostic performance values ​​according to the size ratio of each subset, and using the calculated values ​​as the dynamic weights of each diagnostic algorithm.

[0101] Specifically, the process of determining the dynamic weights of each diagnostic algorithm based on a game theory weight allocation strategy aims to objectively quantify the true contribution of each algorithm in the current data environment. This process first obtains the historical diagnostic performance values ​​of each algorithm. The system evaluates its basic performance by statistically analyzing the diagnostic accuracy and recall of each algorithm in past test sets or validation sets. These historical diagnostic performance values ​​serve as benchmark parameters for subsequent calculations.

[0102] Subsequently, the system iterates through subset combinations of all diagnostic algorithm sets. Assuming the system contains three diagnostic algorithms, subset combinations include combinations containing only a single algorithm, combinations containing two algorithms, and the complete set containing all three algorithms. For each specific subset combination, the system calculates the overall diagnostic efficiency value when the target diagnostic algorithm is included, and the overall diagnostic efficiency value when the target diagnostic algorithm is not included. These two efficiency values ​​are subtracted to obtain the difference in diagnostic efficiency value of each subset before and after the addition of each diagnostic algorithm. This difference is the marginal contribution, reflecting the actual improvement effect of the algorithm on the current combination.

[0103] After obtaining all diagnostic efficacy value differences, the system performs a weighted summation of all differences according to the size of each subset. Subsets of different sizes have different weights in the calculation. Specifically, small subsets containing fewer algorithms and large subsets containing more algorithms have different weighting factors. This weighting mechanism follows the fairness principle of Shapley values, ensuring that the contribution of each algorithm is reasonably distributed. After the weighted summation, the system uses the final calculated value as the dynamic weight of each diagnostic algorithm.

[0104] For example, when a transformer is operating under high load, the gas characteristics of the oil chromatography change significantly. The system obtains the difference in diagnostic effectiveness values ​​between the three-ratio method, the David's triangle method, and the neural network method. Calculations show that under this operating condition, the three-ratio method has the largest difference in effectiveness value after adding any subset. Furthermore, since the subset containing the two algorithms is the largest, it occupies a higher proportion in the weighted summation. After comprehensive weighted summation, the dynamic weight of the three-ratio method is significantly higher than that of other algorithms. Based on this, the system automatically increases the influence of the three-ratio method in the fusion decision-making process.

[0105] This invention achieves adaptive optimization through multi-algorithm fusion. Obtaining historical diagnostic performance values ​​provides reliable foundational data for weight calculation. Traversing subsets and calculating performance value differences accurately reveals the true utility of each algorithm. Weighted summation based on subset size eliminates biases caused by subjectively set weights. This strategy enables the diagnostic system to automatically adjust algorithm configurations based on the characteristics of real-time monitoring data, significantly improving diagnostic accuracy and robustness in complex fault scenarios.

[0106] In one embodiment, the step of performing root cause analysis of a fault using a weighted diagnostic algorithm includes: calculating the posterior probability of a fault hypothesis under the current evidence using a credibility propagation logic, wherein the processing procedure of the credibility propagation logic is as follows: multiplying the prior probability of the fault by the value of the evidence credibility; if the evidence credibility is less than 0, setting the evidence credibility to 0; and outputting the credibility of the fault hypothesis under the support of the corresponding evidence; and determining the fault type based on the posterior probability.

[0107] Specifically, the core step of this invention—using weighted diagnostic algorithms for root cause analysis—is to achieve accurate fault location. This step, based on probabilistic reasoning theory, transforms the outputs of each diagnostic algorithm into the degree of support for the fault hypothesis. In practice, the system first obtains the prior probability of the fault, which refers to the initial probability of a certain fault type occurring based on historical operating data of power equipment and fault statistics, before obtaining current real-time monitoring evidence. Simultaneously, the system obtains the credibility of the evidence outputs of each weighted diagnostic algorithm at the current moment. This credibility reflects the strength of support of the current monitoring data features for the specific fault.

[0108] Subsequently, the system executes the credibility propagation logic, which aims to calculate the posterior probability of the failure hypothesis under the current evidence. The system reads the prior probability of the failure and the credibility of the evidence, and multiplies the two. During the multiplication, the system introduces a non-negativity constraint mechanism. If the credibility of the evidence is less than 0, it indicates that the current evidence does not support the failure hypothesis or that there is interference. In this case, the system sets the credibility of the evidence to 0. By setting negative values ​​to 0, the system effectively blocks the negative impact of erroneous evidence on failure reasoning. After the multiplication, the output result is the credibility of the failure hypothesis supported by the corresponding evidence. This result is the posterior probability, which integrates historical experience and real-time evidence.

[0109] Determining the fault type based on the posterior probability is the final step in root cause analysis. The system compares the posterior probabilities of all calculated fault hypotheses and selects the fault hypothesis with the highest posterior probability as the final diagnosis result. If the maximum posterior probability exceeds a preset confidence threshold, the system determines that the fault type is valid. If the maximum posterior probability is below the threshold, the system determines that the equipment is in normal condition or is in an early latent fault stage, requiring continued monitoring.

[0110] For example, assume a partial discharge fault exists in the power equipment. The system obtains a prior probability of 0.2 for this fault. At this time, the diagnostic algorithm based on ultrasonic monitoring outputs an evidence confidence score of 0.8, indicating that a significant ultrasonic signal was detected. The diagnostic algorithm based on UHF monitoring outputs an evidence confidence score of -0.1, indicating that no UHF signal was detected. The system executes the confidence score propagation logic. For ultrasonic evidence, 0.2 is multiplied by 0.8 to obtain a posterior probability of 0.16. For UHF evidence, since its evidence confidence score is less than 0, the system sets it to 0. 0.2 is multiplied by 0, resulting in 0. By integrating evidence from multiple sources, the system ultimately determines that the posterior probability of the partial discharge fault is significantly increased, thus locking the fault type as partial discharge.

[0111] This invention enables quantitative analysis of fault root causes by organically combining prior knowledge with real-time monitoring data using credibility propagation logic. A non-negativity constraint mechanism is introduced during the calculation process, setting the credibility of evidence with a value less than 0 to 0. This effectively eliminates interference from invalid or contradictory evidence, avoiding misjudgments. Determining the fault type based on posterior probability gives the diagnostic results clear physical meaning and statistical basis, significantly improving the reliability and intelligence level of power system fault diagnosis.

[0112] In one embodiment, the step of performing streaming computation on the multi-source monitoring data and generating a dynamic threshold based on a historical sliding window includes: acquiring historical monitoring data for the same time period within a preset time period; calculating the sliding average and standard deviation of the historical monitoring data; and determining the sum of the products of the sliding average, the adaptive coefficient, and the standard deviation as the real-time dynamic threshold.

[0113] Specifically, the process of performing streaming computation on the multi-source monitoring data and generating dynamic thresholds is a key step in achieving adaptive anomaly detection in this invention. This process first acquires historical monitoring data from the same time period within a preset time frame. The operating status of power equipment often exhibits obvious periodicity and seasonality. By extracting data from the same historical period, the system can accurately reflect the baseline operating level of the equipment under similar current operating conditions.

[0114] Subsequently, the system performs statistical characteristic calculations based on the acquired historical monitoring data, specifically including calculating the moving average and standard deviation of the historical monitoring data. The moving average reflects the central tendency of the equipment under normal conditions. The standard deviation quantifies the dispersion and fluctuation range of the data under normal conditions. These two statistics together characterize the normal operating boundary of the equipment under its current operating conditions.

[0115] After obtaining the statistical characteristics, the system determines the real-time dynamic threshold. The system determines the real-time dynamic threshold by summing the products of the moving average, the adaptive coefficient, and the standard deviation. The adaptive coefficient is used to adjust the sensitivity of the threshold. When the equipment is under high load or in a harsh environment with large fluctuations, the standard deviation naturally increases, and the real-time dynamic threshold automatically generated by the system will also be relaxed accordingly, effectively avoiding false alarms caused by fluctuations in normal operating conditions.

[0116] For example, the system sets the preset time period to the past 7 days. When it is necessary to assess whether the top oil temperature of a transformer is abnormal at the current moment, the system obtains historical oil temperature data at the same moment in the past 7 days. The system calculates the moving average of these historical data to be 65℃ and the standard deviation to be 3℃. If the adaptive coefficient is set to 2, the real-time dynamic threshold is 65 + 2 × 3, that is, 71℃. At this time, if the current oil temperature is 70℃, although it is higher than the historical average, it does not exceed the real-time dynamic threshold of 71℃, and the system determines that the equipment is operating normally. If the current oil temperature continues to rise to 75℃, the subsequent alarm logic is triggered.

[0117] This invention achieves dynamic adaptation of anomaly detection standards. By acquiring historical monitoring data over the same time period, it fully considers the periodic operating patterns of power equipment. The real-time dynamic threshold constructed using the moving average and standard deviation completely overcomes the limitations of traditional fixed thresholds. This method significantly improves the accuracy and robustness of anomaly detection under complex and variable operating conditions.

[0118] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above.

[0119] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A power equipment intelligent diagnostic system based on multi-algorithm fusion diagnosis, characterized in that, include: Edge data sensing layer, used for real-time acquisition of multi-source monitoring data of power equipment; The streaming computing layer is used to process the multi-source monitoring data in real time, calculate dynamic thresholds, and perform adaptive delay alarm determination. The streaming computing layer includes: a dynamic threshold calculation subunit, used to acquire historical monitoring data within the same time period in a historical sliding window, calculate the sliding average and standard deviation of the historical monitoring data, and determine the sum of the products of the sliding average, the standard deviation and the adaptive coefficient as the real-time dynamic threshold; and a delay alarm determination subunit, used to determine whether the current monitoring data continuously exceeds the real-time dynamic threshold and whether the duration is greater than a preset delay time, and if so, trigger an alarm event. The cognitive reasoning layer is used to construct a knowledge graph of power equipment fault evolution and perform root cause analysis of faults based on credibility propagation logic. The cognitive reasoning layer includes a game theory weight allocation unit, which is used to evaluate the contribution value of each diagnostic algorithm based on marginal contribution and allocate dynamic weights. The application service layer is used to receive the diagnostic results from the cognitive reasoning layer and generate operation and maintenance strategies.

2. The intelligent diagnostic system for power equipment based on multi-algorithm fusion diagnostics according to claim 1, characterized in that, The cognitive reasoning layer includes: The knowledge graph construction module is used to construct a graph by using oil chromatography gas data, electrical quantity data and environmental data as entity nodes and fault mechanisms as relation edges. The credibility calculation module is used to calculate the posterior probability of the failure hypothesis using credibility propagation logic. The processing procedure of the credibility propagation logic is as follows: multiply the prior probability of failure by the value of the evidence credibility. If the evidence credibility is less than 0, then set the evidence credibility to 0. The output result is the credibility of the failure hypothesis under the support of the corresponding evidence.

3. The intelligent diagnostic system for power equipment based on multi-algorithm fusion diagnostics according to claim 1, characterized in that, The game theory weight allocation unit includes: The marginal contribution calculation submodule is used to traverse the subset combinations of all diagnostic algorithm sets and calculate the difference in diagnostic efficiency value of each diagnostic algorithm before and after it is added to a specific subset. The weight determination submodule is used to perform a weighted summation of all the diagnostic efficacy value differences according to the size ratio of each subset, and use the calculated value as the dynamic weight of each diagnostic algorithm.

4. A method for intelligent diagnosis of power equipment based on multi-algorithm fusion diagnosis, used to implement the system described in any one of claims 1 to 3, characterized in that, include: Real-time acquisition of multi-source monitoring data from power equipment; The multi-source monitoring data is processed by streaming computation, and a dynamic threshold is generated based on a historical sliding window. If the monitoring data continuously exceeds the dynamic threshold and the duration is greater than a preset delay time, an alarm event is generated. Construct a knowledge graph of power equipment fault evolution, and input the alarm events as evidence into the knowledge graph; The dynamic weights of each diagnostic algorithm are determined based on a game theory weighting strategy. The weighted diagnostic algorithms are then used to perform root cause analysis of the faults and output the fault type and confidence level.

5. The intelligent diagnostic method for power equipment based on multi-algorithm fusion diagnostics according to claim 4, characterized in that, The dynamic weights of each diagnostic algorithm are determined using a game-theoretic weight allocation strategy, including: Obtain the historical diagnostic performance values ​​of each diagnostic algorithm; Iterate through all subset combinations of diagnostic algorithms and calculate the difference in diagnostic performance of each subset before and after adding it to a specific subset. The differences in all the diagnostic efficacy values ​​are weighted and summed according to the size proportion of each subset, and the calculated values ​​are used as the dynamic weights of each diagnostic algorithm.

6. The intelligent diagnostic method for power equipment based on multi-algorithm fusion diagnostics according to claim 4, characterized in that, The root cause analysis of the fault using the weighted diagnostic algorithm includes: The posterior probability of the failure hypothesis under the current evidence is calculated using the credibility propagation logic. The processing of the credibility propagation logic is as follows: the prior probability of failure is multiplied by the value of the evidence credibility. If the evidence credibility is less than 0, the evidence credibility is set to 0. The output result is the credibility of the failure hypothesis under the support of the corresponding evidence. The fault type is determined based on the posterior probability.

7. The intelligent diagnostic method for power equipment based on multi-algorithm fusion diagnostics according to claim 4, characterized in that, Streaming computation is performed on the multi-source monitoring data, and dynamic thresholds are generated based on historical sliding windows, including: Obtain historical monitoring data for the same time period within a preset time frame; Calculate the moving average and standard deviation of the historical monitoring data; The sum of the products of the moving average, the adaptive coefficient, and the standard deviation is determined as the real-time dynamic threshold.

Images