AI-powered intelligent diagnostic system for AC / DC power supplies in substations

Abstract

This application discloses an AI-powered intelligent diagnostic system for AC / DC power supplies in substations. By dynamically adjusting the data acquisition frequency and preprocessing algorithm parameters, it first achieves a balance between data acquisition quality and resource consumption. Furthermore, the adjustment process is combined with the real-time operating conditions of the equipment, making the acquired data more valuable. Through online dynamic fusion diagnosis of the first and second models, it overcomes the limitations of a single model, improves the accuracy of early fault diagnosis, and partially solves the problem of performance degradation of a single model over time. The two models use continuously generated time-series feature vector streams as the basis for diagnosis and continuously learn from diagnostic data and results as new samples, thereby continuously improving diagnostic accuracy. By combining the fusion diagnosis with the preliminary fault hypothesis set constructed for the equipment and generating classified handling suggestions through steps F1-F2, it reduces the false alarm rate of equipment and AC / DC power supply systems in substations and achieves flexible control over the urgency of maintenance.

Description

Technical Field

[0001] This application relates to the field of intelligent operation and maintenance technology for substations, specifically to an AI intelligent diagnostic system for AC / DC power supplies in substations. Background Technology

[0002] The AC / DC power supply system of a substation is an important device for ensuring reliable power supply to critical loads such as control and protection devices, automation devices, and emergency lighting. Its health status is directly related to the safe and stable operation of the substation and even the power grid.

[0003] Traditional substation power supply operation and maintenance mainly relies on periodic inspections. This method depends on personnel experience and cannot detect potential faults in a timely manner. In recent years, with the development of intelligent sensing and Internet of Things (IoT) technologies, online monitoring systems based on data acquisition and information monitoring have emerged, capable of issuing alarms for exceeding limits for basic parameters such as voltage and current. However, these systems are essentially still based on basic judgments using fixed thresholds, and their level of intelligence is not high when dealing with complex, interconnected faults in their early stages.

[0004] To enhance intelligence levels, the industry is exploring the introduction of artificial intelligence technologies. For example, algorithms such as generative adversarial networks and graph neural networks are being used to identify anomalies in AC and DC power supplies in substations. However, most of these existing anomaly identification models are static and closed, and once deployed, they remain unchanged. They cannot adapt to data distribution shifts caused by equipment aging and changes in operating modes, and the model performance degrades with operating time.

[0005] Therefore, the field expects a highly intelligent, dynamically evolving intelligent diagnostic system for AC / DC power supplies in substations to solve the aforementioned problem of poor adaptability. Summary of the Invention

[0006] This application aims to solve the technical problem that existing substation power supply anomaly identification models cannot adapt to data distribution shifts caused by equipment aging and changes in operating modes, and that model performance degrades with operating time. It provides an AI intelligent diagnostic system for AC and DC power supplies in substations.

[0007] To achieve this objective, the following technical solution is adopted in this application: A substation AC / DC power supply AI intelligent diagnostic system is provided, comprising: The data sensing and edge preprocessing unit is used to collect electrical quantities, temperature, internal resistance estimates and equipment image data of the substation power system in real time in a multimodal manner. During the acquisition process, based on the real-time operating conditions of the equipment and the preset elastic sampling strategy, the data acquisition frequency and preprocessing algorithm parameters are dynamically adjusted to generate a time-series feature vector stream. The collaborative intelligent diagnostic unit, connected to the data perception and edge preprocessing unit, is used to perform parallel computational reasoning based on the first model and anomaly detection based on the second model according to the received time-series feature vector stream, and generate a two-way fused anomaly diagnostic confidence and preliminary fault hypothesis set. The fault handling decision unit is connected to the collaborative intelligent diagnostic unit and is used to generate fault handling suggestions based on the fusion diagnostic results.

[0008] Preferably, the method for dynamically adjusting the data acquisition frequency and preprocessing algorithm parameters based on the real-time operating conditions of the equipment and a preset flexible sampling strategy includes the following steps: A1, monitor whether the conditions for switching the data sampling strategy of the device are triggered. If so, adjust the data sampling frequency and preprocessing algorithm parameters according to the triggered sampling strategy to be switched, and proceed to step A2; If not, continue monitoring; A2, monitor whether the sampling strategy rollback condition for the device is triggered. If so, the device will revert to the initial data sampling strategy before the switch. If not, continue with the sampling strategy that was successfully switched.

[0009] Preferably, the method for generating the time-series feature vector stream from the real-time acquired raw data includes the following steps: B1, divide the sliding time window, and calculate the corresponding standardized features for each type of preprocessed data within the same time window; B2, the standardized features calculated within the same time window are concatenated into a one-dimensional vector, and the one-dimensional vectors corresponding to each time window are sorted in chronological order to form the time-series feature vector stream; For the preprocessed voltage and current data, the standardization features include one or more of the following: mean, standard deviation, peak-to-peak value, and RMS ripple value of a specified frequency band. For temperature data, the standardized features include temperature gradient and / or spatial temperature difference; For equipment image data, the standardized features include one or more of the following: the coordinates and temperature value of the highest temperature point in the image, the area ratio of the high-temperature region, and the percentage of rusted pixels on the connecting strip.

[0010] Preferably, the first model performs computational inference based on the input temporal feature vector stream according to the following steps: C1 identifies the device type corresponding to the state represented by the current input time-series feature vector stream, and then calls the computational inference model associated with that type of device; C2, based on the invoked computational reasoning model, performs computational reasoning according to a preset strategy to generate a first diagnostic vector.

[0011] Preferably, the second model performs anomaly detection based on the input temporal feature vector stream according to the following steps: D1, based on the pre-learned feature associations, extract the feature array with associations from the temporal feature vector stream; D2, calculate the reconstruction error of the feature array; D3 generates the second diagnostic vector based on the reconstruction error calculation results.

[0012] Preferably, the first feature and the second feature in the feature array of each time window in the continuous time window are used as the independent and dependent variables of the higher-order equation, respectively, and the term coefficients of the higher-order equation are solved by reverse deduction. Then, the mean value of the deviation from the corresponding term coefficients of the higher-order equation learned in advance is calculated as the reconstruction error of the feature array.

[0013] Preferably, the feature elements in the second diagnostic vector include the anomaly probability calculated for the device, wherein the anomaly probability is the mean of the reconstruction error corresponding to each of the feature arrays extracted in step D1.

[0014] Preferably, the method for fusing the confidence scores of the abnormal diagnosis is as follows: The first anomaly confidence and the first anomaly probability are extracted from the first diagnostic vector obtained from the first model inference, and the second anomaly confidence and the second anomaly probability are extracted from the second diagnostic vector obtained from the second model detection. Obtain the weights of the first anomaly confidence level and the second anomaly confidence level; the dynamic weight of the first anomaly confidence level is: the historical accuracy of judging similar faults of the device at a preset time interval from the current time; the weight of the second anomaly confidence level is set to 1; The dynamic weight of the anomaly confidence level is calculated as: the product of the first anomaly confidence level or the second confidence level and the corresponding weight obtained; The dynamic weights of the outlier confidence scores are normalized by calculating the ratio of the dynamic weight of the first or second outlier confidence score to the sum of the dynamic weights of the two outlier confidence scores as the normalized dynamic weights. The product of the first normalized dynamic weight and the first anomaly probability of the first anomaly confidence level is calculated, and the sum of the product of the second normalized dynamic weight and the second anomaly probability of the second anomaly confidence level is used as the fused anomaly diagnosis confidence level.

[0015] Preferably, the method for generating the preliminary set of fault hypotheses includes the following steps: E1, extract abnormal correlation features from the second diagnostic vector generated by the second model, and then directly match the main hypothesis corresponding to the abnormal correlation features from the pre-built knowledge graph; E2, based on the first diagnostic vector generated by the first model and the second diagnostic vector generated by the second model, identify the associated hypotheses and / or derived hypotheses related to the main hypothesis, and calculate the confidence level of the associated hypotheses and / or the derived hypotheses; E3, after arranging the main hypothesis, the related hypothesis, and the derived hypothesis in descending order of confidence level, forms the preliminary fault hypothesis set.

[0016] Preferably, the method by which the fault handling decision unit generates the fault handling recommendation includes the following steps: F1 determines whether the confidence levels of all hypotheses in the preliminary fault hypothesis set are higher than the corresponding set confidence thresholds, and whether the primary hypothesis is ranked first in the set. If so, mark the currently diagnosed device as "urgent repair" and push the marking information and the device's location information to the repair personnel; If not, proceed to step F2; F2 determines whether the primary hypothesis is ranked first in the set and whether the confidence levels of all the associated hypotheses and / or the derived hypotheses are lower than the corresponding set confidence thresholds. If so, then based on the main assumptions, a "maintenance suggestion" is generated for the equipment; If not, the equipment is marked as "planned maintenance," and the marking information and the equipment's location information are pushed to the maintenance personnel.

[0017] This application achieves a balance between data acquisition quality and resource consumption by dynamically adjusting the data acquisition frequency and preprocessing algorithm parameters. Furthermore, the adjustment process incorporates real-time equipment operating conditions, making the acquired data more valuable. Through online dynamic fusion diagnosis of the first and second models, the limitations of a single model are overcome, improving the accuracy of early fault diagnosis and partially addressing the performance degradation of a single model over time. Both models use continuously generated time-series feature vector streams as diagnostic basis and continuously learn from diagnostic data and results as new samples, resulting in continuously improved diagnostic accuracy. By combining fusion diagnosis with a preliminary fault hypothesis set constructed for the equipment and generating categorized handling suggestions through steps F1-F2, the false alarm rate for equipment and substation AC / DC power supply systems is reduced, and flexible control over the urgency of maintenance is achieved. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments of this application will be briefly described below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the substation AC / DC power supply AI intelligent diagnostic system provided in this embodiment. Detailed Implementation

[0020] The technical solution of this application will be further described below with reference to the accompanying drawings and specific embodiments.

[0021] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual images. They should not be construed as limiting the scope of this application. To better illustrate the embodiments of this application, some components in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0022] In the accompanying drawings of the embodiments of this application, the same or similar reference numerals correspond to the same or similar components. In the description of this application, it should be understood that if terms such as "upper," "lower," "left," "right," "inner," and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting this application. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0023] In the description of this application, unless otherwise expressly specified and limited, the term "connection" or similar designation indicating a connection between components should be interpreted broadly. For example, it can refer to a fixed connection, a detachable connection, or an integral part; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can refer to the internal communication between two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0024] The substation AC / DC power supply AI intelligent diagnostic system provided in this application embodiment, such as Figure 1 As shown, it includes: The data sensing and edge preprocessing unit is used to collect electrical quantities, temperature, internal resistance estimates and equipment image data of the substation power system in real time in a multimodal manner. During the acquisition process, based on the real-time operating conditions of the equipment and the preset elastic sampling strategy, the data acquisition frequency and preprocessing algorithm parameters are dynamically adjusted to generate a time-series feature vector stream. Taking a 220V DC system within a substation that supplies power to relay protection devices as an example, which includes a charging panel and battery banks, the data sensing and edge preprocessing unit collects electrical quantity data through sensors installed in the DC system circuit. This includes collecting three-phase voltage and current on the AC input side of the charging panel by deploying voltage and current sensors; collecting bus voltage and total load current on the DC output bus; collecting load current for each DC feeder; and collecting the voltage of each individual battery, the charging current of the entire battery bank, and the discharging current of the battery bank.

[0025] Temperature data is collected through temperature sensors, including monitoring the temperature of the rectifier module IGBT heat sink inside the charging screen, and monitoring the ambient temperature and the surface temperature of the individual battery casing inside the battery cabinet.

[0026] The estimated internal resistance is obtained by the battery monitoring unit using either the DC discharge method or the AC injection method. For example, under light system load conditions, the battery monitoring unit periodically applies a transient small current load to a single cell and estimates the internal resistance of that cell by analyzing its voltage transient response curve and using a pre-built model.

[0027] Equipment image data is collected via cameras or infrared imagers deployed inside the equipment compartment. These can be set to trigger periodically, such as weekly, or upon detecting an abnormal event, such as exceeding the equipment compartment temperature limit, to capture visible light images of the internal components of the charging screen, battery connections, and circuit breaker status. The infrared thermal images captured by the infrared imager can generate temperature distribution cloud maps, enabling the detection of potential hazards such as overheating at various connection points within the equipment.

[0028] In this embodiment, the method by which the data perception and edge preprocessing unit dynamically adjusts the data acquisition frequency and preprocessing algorithm parameters based on the real-time operating conditions of the device and a preset flexible sampling strategy includes the following steps: A1, monitor whether the conditions for switching the data sampling strategy of the device are triggered. If so, adjust the data sampling frequency and preprocessing algorithm parameters according to the triggered sampling strategy to be switched, and proceed to step A2; If not, continue monitoring; For example, the predefined data sampling strategies for this 220V DC system include two types: steady-state sampling strategy and transient sampling strategy. The steady-state sampling strategy is defined, for instance, as follows: when the rate of change of the system's total DC load current remains within a preset threshold range, system operating data is collected using a basic sampling frequency and a basic preprocessing algorithm. For example, electrical quantity data is collected at a frequency of 1Hz, temperature data at a frequency of 0.1Hz, and the collected data is preprocessed using a recursive averaging filter to remove random noise.

[0029] Transient sampling strategies can be defined as follows: when the rate of change of the total DC load current of the system is outside the preset threshold range, it indicates that the system state has changed drastically. At this time, high-frequency sampling and enhanced preprocessing are used to collect system operating data. For example, the DC bus voltage is collected by adjusting from 1Hz to 100Hz, and the data preprocessing algorithm is switched to low-pass filtering or Kalman filtering.

[0030] For example, the system continuously monitors the rate of change of the total DC load current. If the monitoring detects that the load current suddenly increases by 50A within 100ms due to the transmission test of the relay protection device, exceeding the preset threshold range, the system is instructed to switch from a steady-state sampling strategy to a transient sampling strategy, adjusting the sampling frequency of the bus voltage and related feeder current, for example, from 1Hz to 100Hz. Simultaneously, the data preprocessing algorithm is changed from recursive averaging filtering to remove random noise to Kalman filtering.

[0031] A2, monitor whether the sampling strategy rollback condition for the device is triggered. If so, the device will revert to the initial data sampling strategy before the switchover. If not, continue with the sampling strategy that was successfully switched.

[0032] Continuing with the above example, for instance, when the load is detected to have stabilized, such as when the total DC load current returns to the stable current range before the surge, the sampling strategy rollback condition for the DC system is triggered, and the data sampling strategy for the DC system is switched from the transient sampling strategy back to the steady-state sampling strategy.

[0033] In this embodiment, the method for the data perception and edge preprocessing unit to generate a time-series feature vector stream based on the raw data acquired through multimodal acquisition includes the following steps: B1, divide the sliding time window, and calculate the corresponding standardized features for each type of preprocessed data within the same time window; For example, the system uses a 10-second analysis window with a sliding step of 5 seconds. For the preprocessed voltage and / or current data, it calculates the mean, standard deviation, peak-to-peak value, and RMS ripple value within the specified frequency band, such as the RMS ripple value in the 100-150Hz range. Ripple can reflect the fault condition of the rectifier module in the DC system and serves as a standardized feature of the voltage and current data.

[0034] Peak-to-peak value is the difference between the maximum and minimum values ​​of a signal within a time window, such as the difference between the maximum and minimum voltage. Ripple is the AC component superimposed on a stable DC voltage or current. The RMS value of ripple is the root mean square value of this AC component calculated within a time window to quantify the intensity or magnitude of this AC noise.

[0035] For temperature data, standardized characteristics are represented by calculating temperature gradient and spatial temperature difference. Temperature gradient is the temperature change per unit time within a time window. Spatial temperature difference refers to the difference between the highest and lowest individual cell temperatures within a device, such as a battery pack.

[0036] For equipment image data, standardized features include the coordinates and temperature value of the highest temperature point in the infrared image, and the area ratio of the high-temperature region; and the percentage of rusted pixels in the connecting strip in the visible light image, where rust is usually represented by green.

[0037] After calculating the standardized features of the preprocessed data in step B1, the method for generating the time-series feature vector stream proceeds to the next step: B2 concatenates the standardized features calculated within the same time window into a one-dimensional vector. The one-dimensional vectors corresponding to each time window are sorted in chronological order to form a time-series feature vector stream.

[0038] For example, for time windows , Representing the start and end times of the time window respectively, the resulting time-series feature vector stream can be expressed as follows: [Average bus voltage, bus voltage ripple, average load current, load current change rate, average battery pack voltage, battery pack voltage range, highest single cell temperature, single cell temperature range, percentage of high-temperature areas in infrared image, percentage of corroded pixels in connecting strips, ...].

[0039] like Figure 1 As shown, the substation AC / DC power supply AI intelligent diagnostic system provided in this embodiment also includes: The collaborative intelligent diagnostic unit connects the data perception and edge preprocessing unit. It is used to perform computational reasoning based on the first model and anomaly detection based on the second model in parallel according to the received time-series feature vector stream, and generate a fused anomaly diagnosis confidence and preliminary fault hypothesis set. The following uses the received timing feature vector stream of the battery pack as an example to illustrate the working process of the collaborative intelligent diagnostic unit: Suppose that the received temporal feature vector stream at a certain moment includes: [Timestamp, total battery pack voltage: 245.1V, single cell 001 voltage: 12.75V, single cell 002 voltage: 12.68V, ambient temperature: 25℃, estimated internal resistance of single cell 001: 1.05 milliohms, estimated internal resistance of single cell 002: 1.18 milliohms, load current: 5.3A, ripple voltage RMS value: 45mV].

[0040] The temporal feature vector stream described above is input into the first model, which then performs computational inference based on the input temporal feature vector stream according to the following steps: C1 identifies the device type corresponding to the state represented by the current input time-series feature vector stream, and then calls the computational inference model associated with that type of device; For example, the time-series feature vector stream mentioned above reflects the state of a battery pack. The model library may contain pre-set inference models for calculating battery internal resistance and capacity, and a pre-defined association relationship has been established between the battery pack and the battery internal resistance and capacity inference model. The first model first identifies the device type corresponding to the battery pack state reflected in the received time-series feature vector stream as a battery pack, and then, based on the pre-defined association relationship, matches the battery internal resistance and capacity calculation inference model bound to that battery pack.

[0041] C2, based on the invoked computational inference model, performs computational inference according to a preset strategy to generate the first diagnostic vector.

[0042] For example, for the battery internal resistance and capacity calculation inference model, the preset calculation inference strategy is as follows: First, extract the internal resistance of the individual battery in the time-series feature vector stream received in step C1, and simultaneously obtain the initial internal resistance and ambient temperature of the individual battery. Then, perform compensation calculation to deduce the ratio of the current estimated capacity of the individual battery to its rated capacity. Then, calculate the anomaly degree of the individual battery based on the ratio and infer the cause of the anomaly, and output the confidence degree of the calculated anomaly degree and the inferred cause of the anomaly.

[0043] Based on the preset computational reasoning strategy of the above example, the invoked computational reasoning model extracts, for example, the estimated internal resistance of the single cell 002 in the above example as 1.18 milliohms, the initial value of the single cell 002 as 1 milliohm, and the ambient temperature as 25°C. Then, these extracted parameters are substituted into the model formula to calculate the ratio of the current estimated capacity of the single cell 002 to its rated capacity. In this application, since the computational reasoning models bound to different types of devices are different, the preset computational reasoning strategies are also usually different, and the computational reasoning process is not within the scope of the claims of this application, it will not be specifically described.

[0044] Subsequently, the computational inference model preferably calculates the anomaly degree of the single cell 002 based on the ratio of its current estimated capacity to its rated capacity, as shown in the example above, and infers the possible causes of the anomaly. The anomaly degree and the confidence level of the cause of the anomaly are then formed as the first diagnostic vector.

[0045] For example, if the ratio of the current estimated capacity to the rated capacity of a single cell 002 is 92%, and the current state of the single cell 002 is inferred to be "capacity decay", the anomaly calculated for 8% capacity decay is, for example, 0.75 (range 0-1, 1 being the highest anomaly), the inferred cause of the anomaly is, for example, "severe sulfation or water loss", and the confidence of the anomaly and the cause of the anomaly is, for example, 0.8 (range 0-1, 1 being the highest confidence), then the generated first diagnostic vector is expressed as: [single cell 002, state: capacity decay, anomaly: 0.75, inferred cause: severe sulfation or water loss, anomaly confidence: 0, 8].

[0046] The temporal feature vector stream described above is synchronously input into the second model, which then performs anomaly detection based on the input as follows: D1, based on the pre-learned feature associations, extracts the feature array with associations from the temporal feature vector stream; D2, calculate the reconstruction error of the extracted feature array; D3 generates the second diagnostic vector based on the reconstruction error calculation results.

[0047] For example, the relationship between the operating voltage and internal resistance of a single battery cell under load is learned in advance. For instance, as the internal resistance of a single battery cell gradually increases, the operating voltage gradually decreases, and the two are negatively correlated. This negative correlation can be represented by a quadratic equation, with the estimated internal resistance as the independent variable and the operating voltage of the single battery cell as the dependent variable. The relationship between these two characteristics is characterized by solving the coefficients of the equation.

[0048] In the aforementioned example of the time-series feature vector stream, the operating voltage of a single battery cell is 12.68V, and the estimated internal resistance is 1.18 milliohms. These two parameter values ​​from the time-series feature vector stream associated with each consecutive time window are used to inversely deduce the coefficients of the quadratic equation. If the deviation of the inversely deduced coefficients from the pre-learned coefficients (the absolute value of the difference between the same coefficients) is found to be greater than a preset deviation threshold, then the extracted feature array is determined to be abnormal, indicating that the state of the single battery cell 002 is abnormal. The reconstruction error is preferably the mean of the deviations of each coefficient.

[0049] Finally, the system calculates the probability (preferably the average of the reconstruction errors of each feature array extracted in step D1) and confidence level of the single cell based on the reconstruction error, and generates a textual description of the abnormal features, forming the second diagnostic vector.

[0050] The method for calculating the confidence level of the probability of abnormal equipment status, such as abnormality of a single cell, is not within the scope of the claims of this application, nor does it affect the solution of the technical problem of this application, so it will not be specifically explained.

[0051] The generated second diagnostic vector can be expressed as: [Single cell 002, State: Abnormal, Abnormal probability: 0.88, Abnormal feature: Deviation in correlation between voltage and internal resistance features, Abnormal confidence: 0.9].

[0052] After outputting the first and second diagnostic vectors from the input time-series feature vector stream using the first and second models, the collaborative intelligent diagnostic unit generates a fused diagnostic result. The fusion method is as follows: The first anomaly confidence score and the second anomaly confidence score are extracted from the first diagnostic vector and the second diagnostic vector, respectively. For example, in the example above, the first anomaly confidence score is 0.8 and the second anomaly confidence score is 0.9. Then, the first anomaly confidence score and the second anomaly confidence score are dynamically weighted and normalized. Finally, the normalized weights are multiplied by the corresponding anomaly probabilities and summed to obtain the fused anomaly diagnosis confidence score.

[0053] In this embodiment, the method for dynamically weighting the first anomaly confidence level is as follows: if the first anomaly confidence level is 0.8, and the preset weight for the first anomaly confidence level is, for example, 1, then the dynamic weight of the first anomaly confidence level is 0.8 × 1 = 0.8. The method for dynamically weighting the second anomaly confidence level is as follows: if the historical accuracy rate for judging similar faults in this battery pack is, for example, 85%, and the historical accuracy rate is used as the weighting factor, then the dynamic weight of the second anomaly confidence level is 0.9 × 0.85 = 0.765.

[0054] The dynamic weight normalization method is as follows: 0.765 / (0.765+0.8)=0.49, where 0.49 is the normalized weight of the dynamic weight of the second outlier confidence level, and the normalized weight of the dynamic weight of the first outlier confidence level is 0.51.

[0055] The anomaly diagnosis confidence obtained by fusing the first and second anomaly confidence scores is: 0.88 (anomaly probability carried in the second diagnostic vector) × 0.49 + 0.75 (anomaly degree carried in the first diagnostic vector) × 0.51 = 0.81. The generated fused diagnostic result can be expressed as follows: [The fault hypothesis for the battery pack is: early degradation of individual cell 002, fused confidence score is 0.81, supporting evidence is: the capacity of individual cell 002 has decayed to 92%, and the correlation between voltage and internal resistance has deviated].

[0056] In this application, the method for generating a preliminary set of fault hypotheses includes the following steps: E1 extracts abnormal correlation features from the second diagnostic vector generated by the second model. For example, the abnormal correlation feature "the voltage and internal resistance characteristics of single cell 002 deviate from the correlation" is extracted from the second diagnostic vector in the example above. Then, the main hypothesis corresponding to this abnormal correlation feature is directly matched from the pre-constructed knowledge graph. For example, if the knowledge graph records the corresponding hypothesis for "the voltage and internal resistance characteristics of single cell deviate from the correlation" as "the single cell has sulfation or water loss", then the main hypothesis matched from this abnormal correlation feature is expressed as: Hypothesis A: Single cell 002 has sulfation or water loss, and through calculation, the confidence level of hypothesis A is 0.81.

[0057] E2, based on the first diagnostic vector generated by the first model and the second diagnostic vector generated by the second model, identifies the associated hypotheses and / or derived hypotheses of the main hypothesis matched with E1, and calculates the confidence level of the associated hypotheses and / or derived hypotheses; For example, regarding hypothesis A in the above example, the knowledge graph records: if a single battery cell exhibits sulfation or water loss, further investigation is needed to determine if the local temperature of the battery rack containing the single battery cell is excessively high, leading to accelerated degradation. Based on this relationship, the related hypothesis B of hypothesis A is identified from the knowledge graph: the local temperature of the battery rack containing single battery cell 002 is excessively high. The confidence level of hypothesis B is then calculated based on the temperature of the battery rack carried in the first and second diagnostic vectors. The calculation principle for the confidence level of hypothesis B is the same as that for hypothesis A, and will not be repeated here. Similarly, derivative hypotheses of hypothesis A are further investigated. For example, if the graph contains a record stating "excessive charger ripple leads to accelerated deterioration of battery pack consistency," and the current ripple is found to be 45mV, which is considered high, then hypothesis C is proposed: excessive charger output ripple may be a potential cause of premature degradation of single battery cell 002. The confidence level of hypothesis C is then calculated, using the same principles as hypotheses A and B, and will not be repeated here.

[0058] Finally, the main hypothesis, related hypothesis, and derived hypothesis are arranged in descending order of confidence level to form a preliminary set of failure hypotheses for the same device, such as cell 002.

[0059] like Figure 1 As shown, the substation AC / DC power supply AI intelligent diagnostic system provided in this application also includes: The fault handling decision unit is connected to the collaborative intelligent diagnostic unit and is used to generate fault handling suggestions based on the fusion diagnostic results (including the fusion anomaly diagnostic confidence and preliminary fault hypothesis set).

[0060] The method for the fault handling decision unit to generate fault handling recommendations includes the following steps: F1 determines whether the confidence levels of all hypotheses in the initial fault hypothesis set are higher than the corresponding set confidence thresholds, and whether the principal hypothesis is ranked first in the set. If so, mark the currently diagnosed equipment as "urgent repair" and push the marking information and the location information of the equipment to the repair personnel, prompting the personnel to go to the site immediately to repair the equipment; If not, proceed to step F2; F2 determines whether the primary hypothesis is ranked first in the set and whether the confidence levels of all related and / or derived hypotheses are lower than the corresponding set confidence thresholds. If so, then based on the main assumption, a "maintenance suggestion" is generated for the equipment; for example, for the main assumption that "the individual battery cells have sulfation or water loss", the preset maintenance methods include two types: capacity testing of the individual battery cells and equalization charging activation. Then, these two methods are matched from the method library and pushed to the maintenance personnel. The maintenance personnel can consider going to the site for maintenance immediately, or they can adjust the maintenance plan according to the maintenance time requirements. If not, the equipment is marked as "Planned Maintenance," and the marking information and the equipment's location information are pushed to the maintenance personnel. Planned Maintenance indicates that the urgency of maintaining the equipment at the moment of the major assumption is not high, and more time is available for maintenance personnel.

[0061] In summary, by dynamically adjusting the data acquisition frequency and preprocessing algorithm parameters, a balance between data acquisition quality and resource consumption is achieved. Furthermore, the adjustment process, combined with real-time equipment operating conditions, makes the acquired data more valuable. Online dynamic fusion diagnosis of the first and second models overcomes the limitations of a single model, improves the accuracy of early fault diagnosis, and partially solves the problem of performance degradation of a single model over time. Both models use continuously generated time-series feature vector streams as diagnostic basis and continuously learn from diagnostic data and results as new samples, resulting in continuously improved diagnostic accuracy. By combining fusion diagnosis with a preliminary fault hypothesis set constructed for the equipment, and generating categorized handling suggestions through steps F1-F2, the false alarm rate for equipment and substation AC / DC power supply systems is reduced, and flexible control over the urgency of maintenance is achieved.

[0062] It should be stated that the above-described specific embodiments are merely preferred embodiments and technical principles applied in this application. Those skilled in the art should understand that various modifications, equivalent substitutions, and variations can be made to this application. However, such variations, as long as they do not depart from the spirit of this application, should be within the scope of protection of this application. Furthermore, some terminology used in this application's specification and claims is not limiting but merely for ease of description.

Claims

1. An AI-powered intelligent diagnostic system for AC / DC power supplies in substations, characterized in that, include: The data sensing and edge preprocessing unit is used to collect electrical quantities, temperature, internal resistance estimates and equipment image data of the substation power system in real time in a multimodal manner. During the acquisition process, based on the real-time operating conditions of the equipment and the preset elastic sampling strategy, the data acquisition frequency and preprocessing algorithm parameters are dynamically adjusted to generate a time-series feature vector stream. The collaborative intelligent diagnostic unit, connected to the data perception and edge preprocessing unit, is used to perform parallel computational reasoning based on the first model and anomaly detection based on the second model according to the received time-series feature vector stream, and generate a two-way fused anomaly diagnostic confidence and preliminary fault hypothesis set. The fault handling decision unit is connected to the collaborative intelligent diagnostic unit and is used to generate fault handling suggestions based on the fusion diagnostic results.

2. The substation AC / DC power supply AI intelligent diagnostic system according to claim 1, characterized in that, The method for dynamically adjusting the data acquisition frequency and preprocessing algorithm parameters based on the real-time operating conditions of the equipment and the preset flexible sampling strategy includes the following steps: A1, monitor whether the conditions for switching the data sampling strategy of the device are triggered. If so, adjust the data sampling frequency and preprocessing algorithm parameters according to the triggered sampling strategy to be switched, and proceed to step A2; If not, continue monitoring; A2, monitor whether the sampling strategy rollback condition for the device is triggered. If so, the device will revert to the initial data sampling strategy before the switch. If not, continue with the sampling strategy that was successfully switched.

3. The substation AC / DC power supply AI intelligent diagnostic system according to claim 1, characterized in that, The method for generating the time-series feature vector stream from the real-time acquired raw data includes the following steps: B1, divide the sliding time window, and calculate the corresponding standardized features for each type of preprocessed data within the same time window; B2, the standardized features calculated within the same time window are concatenated into a one-dimensional vector, and the one-dimensional vectors corresponding to each time window are sorted in chronological order to form the time-series feature vector stream; For the preprocessed voltage and current data, the standardization features include one or more of the following: mean, standard deviation, peak-to-peak value, and RMS ripple value of a specified frequency band. For temperature data, the standardized features include temperature gradient and / or spatial temperature difference; For equipment image data, the standardized features include one or more of the following: the coordinates and temperature value of the highest temperature point in the image, the area ratio of the high-temperature region, and the percentage of rusted pixels on the connecting strip.

4. The substation AC / DC power supply AI intelligent diagnostic system according to claim 1, characterized in that, The first model performs computational inference based on the input temporal feature vector stream according to the following steps: C1 identifies the device type corresponding to the state represented by the current input time-series feature vector stream, and then calls the computational inference model associated with that type of device; C2, based on the invoked computational reasoning model, performs computational reasoning according to a preset strategy to generate a first diagnostic vector.

5. The substation AC / DC power supply AI intelligent diagnostic system according to claim 1, characterized in that, The second model performs anomaly detection based on the input temporal feature vector stream according to the following steps: D1, based on the pre-learned feature associations, extract the feature array with associations from the time-series feature vector stream; D2, calculate the reconstruction error of the feature array; D3 generates the second diagnostic vector based on the reconstruction error calculation results.

6. The substation AC / DC power supply AI intelligent diagnostic system according to claim 5, characterized in that, Using the first and second features in the feature array of each time window in a continuous time window as the independent and dependent variables of the higher-order equation, respectively, the coefficients of the higher-order equation are solved by reverse calculation. Then, the mean of the deviation from the corresponding coefficients of the higher-order equation learned in advance is calculated as the reconstruction error of the feature array.

7. The substation DC power supply AI intelligent diagnostic system according to claim 5, characterized in that, The feature elements in the second diagnostic vector include the anomaly probability calculated for the device, wherein the anomaly probability is the mean of the reconstruction error corresponding to each of the feature arrays extracted in step D1.

8. The substation AC / DC power supply AI intelligent diagnostic system according to claim 1, characterized in that, The fusion method for the anomaly diagnosis confidence is as follows: The first anomaly confidence and the first anomaly probability are extracted from the first diagnostic vector obtained from the first model inference, and the second anomaly confidence and the second anomaly probability are extracted from the second diagnostic vector obtained from the second model detection. Obtain the weights of the first anomaly confidence level and the second anomaly confidence level; the dynamic weight of the first anomaly confidence level is: the historical accuracy of judging similar faults of the device at a preset time interval from the current time; the weight of the second anomaly confidence level is set to 1; The dynamic weight of the anomaly confidence level is calculated as: the product of the first anomaly confidence level or the second confidence level and the corresponding weight obtained; The dynamic weights of the outlier confidence scores are normalized by calculating the ratio of the dynamic weight of the first or second outlier confidence score to the sum of the dynamic weights of the two outlier confidence scores as the normalized dynamic weights. The product of the first normalized dynamic weight and the first anomaly probability of the first anomaly confidence level is calculated, and the sum of the product of the second normalized dynamic weight and the second anomaly probability of the second anomaly confidence level is used as the fused anomaly diagnosis confidence level.

9. The substation AC / DC power supply AI intelligent diagnostic system according to claim 1, characterized in that, The method for generating the preliminary set of fault hypotheses includes the following steps: E1, extract abnormal correlation features from the second diagnostic vector generated by the second model, and then directly match the main hypothesis corresponding to the abnormal correlation features from the pre-built knowledge graph; E2, based on the first diagnostic vector generated by the first model and the second diagnostic vector generated by the second model, identify the associated hypotheses and / or derived hypotheses related to the main hypothesis, and calculate the confidence level of the associated hypotheses and / or the derived hypotheses; E3, after arranging the main hypothesis, the related hypothesis, and the derived hypothesis in descending order of confidence level, forms the preliminary fault hypothesis set.

10. The substation AC / DC power supply AI intelligent diagnostic system according to claim 1, characterized in that, The method by which the fault handling decision unit generates the fault handling recommendations includes the following steps: F1 determines whether the confidence levels of all hypotheses in the preliminary fault hypothesis set are higher than the corresponding set confidence thresholds, and whether the primary hypothesis is ranked first in the set. If so, mark the currently diagnosed device as "urgent repair" and push the marking information and the device's location information to the repair personnel; If not, proceed to step F2; F2 determines whether the primary hypothesis is ranked first in the set and whether the confidence levels of all the associated hypotheses and / or the derived hypotheses are lower than the corresponding set confidence thresholds. If so, then based on the main assumptions, a "maintenance suggestion" is generated for the equipment; If not, the device is marked as "planned maintenance," and the marking information and the device's location information are pushed to the maintenance personnel.

Images