A deep learning-based building weak current equipment fault diagnosis method

Abstract

The application discloses a kind of based on deep learning's building weak current equipment fault diagnosis method, it is related to deep learning technical field, including: collection multi-source data, generate multi-source operation data;Preprocess multi-source data, generate credible state frame;Isolation pollution window, generate purification state sequence;Improved DLinear model is constructed, and initial prediction state is generated;Topological lag alignment is executed, and topological correction prediction state is generated;Calculate prediction residual, construct residual index and sample pool;RS-Hash algorithm is input, and direction separation bucket count is generated;Credible count correction, generate fault diagnosis result.The application is by weak current equipment credible state frame, improved DLinear model and RS-Hash algorithm, to weak current equipment operation exception is carried out prediction residual detection, topological propagation discrimination and credible score, improve source and send fault equipment identification precision, reduce the misjudgment rate of propagation affected equipment.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a method for diagnosing faults in building electrical equipment based on deep learning. Background Technology

[0002] Building low-voltage electrical equipment covers video surveillance, access control, intrusion alarms, public address systems, switches, low-voltage gateways, and building automation terminals. Operational data sources include online status, communication status, port status, control feedback, alarm records, maintenance records, and topology connection information. Existing building low-voltage electrical maintenance platforms mostly rely on fixed thresholds, manual inspections, and single-device alarm rules for fault diagnosis. While they can provide alerts for device offline, port interruptions, and alarm triggers, they lack sufficient handling of time alignment, field reliability, historical anomaly contamination, and topology propagation lag in multi-source data. This leads to issues such as missing, delayed, duplicate alarms, and low-reliability inputs in low-voltage electrical equipment status data.

[0003] Existing deep learning diagnostic methods often directly input device state sequences into prediction or classification models, lacking constraints on operational prototypes for building low-voltage electrical scenarios, sparse activation of key fields, and topological proximity state correction. This leads to the easy confusion between peak traffic, network load changes, and alarm linkage states and actual faults. Existing anomaly detection algorithms mostly perform isolated scoring on single device residuals, lacking topological lag adjacency residual indexing, residual direction diagnostic sample organization, and reliability count correction. This makes it difficult to distinguish between the originating fault device and the devices affected by propagation, resulting in insufficient accuracy in fault source localization and interpretability of diagnostic results.

[0004] Therefore, how to provide a deep learning-based method for diagnosing faults in building electrical systems is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a deep learning-based fault diagnosis method for building low-voltage electrical equipment. This invention utilizes trusted state frames of low-voltage electrical equipment, fault contamination window isolation operations, an improved DLinear model, and the RS-Hash algorithm to perform trusted modeling, topology correction prediction, residual direction detection, and trusted anomaly scoring on multi-source operating data of building low-voltage electrical equipment. This enables fault identification, fault source localization, and propagation anomaly judgment of low-voltage electrical equipment, and has the advantages of strong data adaptability, accurate fault location, low false alarm rate, and high interpretability of diagnostic results.

[0006] A method for diagnosing faults in building low-voltage electrical equipment based on deep learning according to an embodiment of the present invention includes: Collect operation monitoring data, alarm maintenance data, and topology data of low-voltage electrical equipment to generate multi-source operation data of building low-voltage electrical equipment; Preprocessing multi-source operation data of building low-voltage electrical equipment yields a reliable status frame for the low-voltage electrical equipment. Based on the trusted state frames of low-voltage equipment, a fault contamination window isolation operation is performed on the historical state sequence to generate a purified low-voltage equipment state sequence. An improved DLinear model is constructed, which includes a linear prediction layer for prototype working conditions, a sparse key field activation prediction layer, and a topological proximity state correction layer. The state sequence of the purification weak current equipment is input into the improved DLinear model to generate the initial predicted state of the weak current equipment. Based on the topology data and the initial predicted state of the weak current equipment, a topology propagation lag alignment operation is performed to generate lag aligned topology neighbor state data, and then a topology corrected predicted state is generated through the topology neighbor state correction layer. The predicted residuals of weak current equipment are calculated based on the reliable state frames of weak current equipment and the topology correction prediction state, and a topology lag adjacency residual index table and a residual direction diagnostic sample pool are constructed. Based on the prediction residuals of weak current equipment, the topology lag adjacency residual index table, and the residual direction diagnosis sample pool, a topology assembly residual vector is generated. The topology assembly residual vector is then input into the RS-Hash algorithm to generate a direction-separated hash bucket count. The trusted count correction is performed on the direction separation hash bucket count to obtain the trusted correction residual abnormal score, and the fault diagnosis results of building low-voltage equipment are generated.

[0007] Optionally, the generation of multi-source operating data for building low-voltage electrical equipment includes: The system collects operational monitoring data from building low-voltage electrical equipment and simultaneously records equipment identification, equipment type, collection time, and arrival time. The operational monitoring data includes online status, communication status, port status, control feedback status, and service status. Read alarm and maintenance data from the alarm record database and maintenance record database. The alarm and maintenance data includes alarm type, alarm time, recovery status, maintenance record and fault record. Read the topology data from the building's low-voltage topology configuration table. The topology data includes equipment identification, equipment type, network connection relationship, power supply connection relationship, control linkage relationship, and signal transmission relationship. By associating operation monitoring data, alarm maintenance data, and topology data according to equipment identifiers, multi-source operation data of building low-voltage electrical equipment is generated.

[0008] Optionally, obtaining the trusted state frame of the low-voltage equipment includes: Read the operation monitoring data, alarm maintenance data and topology data from the multi-source operation data of building low-voltage equipment, and extract the equipment identifier, equipment type, acquisition time, arrival time, status field name, status field value, alarm type and alarm time; Standardize the field format for status field names and values, and generate a unified diagnostic time based on the data acquisition time. Based on the unified diagnosis time, arrival time, status field name, status field value, alarm type, and alarm time, generate field missing status, delayed reporting status, missing data completion status, and duplicate alarm status. Based on the device identifier, the status field value at the unified diagnostic time is bound to the topology data to generate a low-voltage equipment status record. Data credibility is generated based on field missing status, delayed reporting status, missing data completion status, and repeated alarm status. Operating condition prototype markers, key field activation markers, and topology propagation association markers are generated based on device type, status field name, topology data, and unified diagnostic time. The weak current device status record, data credibility, operating condition prototype markers, key field activation markers, and topology propagation association markers are combined to generate a reliable status frame for the weak current device.

[0009] Optionally, generating the state sequence of the purification low-voltage electrical equipment includes: Read the device identifier, device type, unified diagnosis time, status field value, field missing status, delayed reporting status, missing data completion status, repeated alarm status, data credibility and topology propagation association marker from the trusted status frame of the weak current equipment, and form a historical status sequence according to the device identifier and unified diagnosis time. The historical status sequence is divided into candidate historical windows according to the continuous and unified diagnosis time. The number of low-confidence fields, missing fields, delayed reporting fields and repeated alarms in the candidate historical windows are counted to generate window status records. Based on window state records and topology propagation association markers, fault pollution states are identified, and candidate historical windows that meet the pollution determination criteria are marked as fault pollution windows, generating pollution window markers. Pollution isolation weights are generated based on pollution window markers, and the status field values ​​in the candidate historical windows are weighted and assembled to generate clean window status records. The status records of the purification window are summarized according to the unified diagnosis time, and a status sequence of the purification low-voltage equipment is generated.

[0010] Optionally, generating the initial predicted state of the low-voltage equipment includes: An improved DLinear model was constructed, integrating the linear prediction layer of the prototype working condition, the sparse key field activation prediction layer, and the topological proximity state correction layer into the DLinear prediction process. Input the state sequence of the purification low-voltage equipment into the improved DLinear model, read the equipment identifier, equipment type, unified diagnosis time, state field name, state field value, data credibility, working condition prototype mark and key field activation mark, and generate the model input state sequence; The model input state sequence is input into the working condition prototype condition linear prediction layer. Based on the working condition prototype label, the working condition condition linear prediction is performed on the state field value to generate the working condition condition prediction state. The model input state sequence and the predicted state under operating conditions are input into the sparse key field activation prediction layer. Based on the equipment type, data credibility and key field activation marker, the key state fields are determined, sparse activation prediction is performed, and the key field predicted state is generated. The predicted state of operating conditions and the predicted state of key fields are aligned according to the unified diagnostic time to generate the initial predicted state of weak current equipment. The initial predicted state of low-voltage equipment is associated with the equipment identifier, equipment type, and unified diagnostic time to generate an initial predicted state record; The improved DLinear model was trained by using a combination of operating condition prediction error, key field prediction error, topological proximity state correction error, and initial prediction state error as the joint optimization objective. The model parameters of the linear prediction layer of the operating condition prototype, the sparse key field activation prediction layer, and the topological proximity state correction layer were iteratively updated. When the change in the joint loss value in five consecutive training rounds was less than 0.001, the improved DLinear model was considered to have completed convergence training.

[0011] Optionally, generating the topology correction prediction state includes: Read the device identifier, device type, network connection relationship, power supply connection relationship, control linkage relationship and signal transmission relationship from the topology data, and read the initial predicted state of weak current equipment, unified diagnosis time and state field name from the initial predicted state record to generate topology edge state record; Based on the topology edge state record, determine the upstream adjacent devices, same link adjacent devices, same power supply branch devices, and control linkage devices corresponding to the current device, and generate topology edge type tags according to network connection relationship, power supply connection relationship, control linkage relationship, and signal transmission relationship respectively; The propagation lag retrieval window is determined based on the topological edge type label, and a candidate propagation state sequence of adjacent devices is generated; The candidate propagation state sequence of adjacent devices is matched with the initial predicted state of the weak current equipment of the current device by matching the state field correspondence, the predicted state change direction, and the unified diagnosis time offset, and a topology propagation lag matching record is generated. Based on the topology propagation lag matching record, the propagation alignment time of the adjacent device corresponding to the current device is determined, and the initial predicted state of the adjacent device at the propagation alignment time, the topology edge type label and the unified diagnosis time offset value are combined to generate lag alignment topology proximity state data. The lagging aligned topology neighbor state data and the initial predicted state of the weak current equipment are input into the topology neighbor state correction layer to generate the neighbor state correction amount, and the initial predicted state of the weak current equipment is topologically corrected to generate the topology corrected predicted state.

[0012] Optionally, the construction of the topological hysteresis adjacency residual index table and the residual direction diagnostic sample pool includes: Read the device identifier, device type, unified diagnosis time, status field name, status field value and data confidence from the trusted status frame of the weak current equipment, and read the prediction field name and prediction field value from the topology correction prediction status; Based on the equipment identification, unified diagnostic time, and status field name, the status field value and the prediction field value are used to calculate the predicted residual of the weak current equipment. The predicted residual of the weak current equipment includes the residual value, residual direction, residual field name, residual field category, and residual duration. Read the topology data and the lag-aligned topology neighbor status data, associate the current device prediction residual with the historical prediction residual of the neighboring device, calculate the residual magnitude matching value and topology propagation support, and generate a topology lag neighbor residual index table. Based on the residual values, residual directions, residual field categories, residual durations, and data reliability in the prediction residuals of low-voltage equipment, the residuals are classified and stored according to positive residuals, negative residuals, and zero-neighbor residuals, generating a residual direction diagnostic sample pool.

[0013] Optionally, the topological hysteresis adjacency residual index table includes: The topology lag adjacency residual index table consists of the current device identifier, adjacent device identifier, current device type, adjacent device type, topology edge type, current unified diagnosis time, propagation alignment time, unified diagnosis time offset value, current device prediction residual, adjacent device historical prediction residual, residual field name, residual direction, residual magnitude matching value, and topology propagation support. An index record is created between the current device's predicted residual and the historical predicted residual of adjacent devices, based on the device identifier, topology edge type, propagation alignment time, residual field name, and residual direction. A dynamic residual index structure with propagation lag, residual direction, and topology propagation support is formed based on the index record.

[0014] Optionally, the generation of direction-separated hash bucket counts includes: Read the prediction residuals of weak current equipment, the index table of topological lag adjacency residuals and the sample pool of residual direction diagnosis, and perform residual association according to equipment identifier, equipment type, unified diagnosis time, topological edge type and residual field name to generate topological residual association records. Based on the topological residual association record, the current device prediction residual, the historical prediction residual of adjacent devices, the residual direction, the residual field category and the topological propagation support are read, and the residuals are assembled according to the residual field name and the topological edge type to generate the topological assembly residual vector. The RS-Hash algorithm employs a random subspace selection mechanism for weak electrical equipment type constraints. Based on the topology assembly residual vector, equipment type, and residual field category, a candidate residual field set is determined, and random subspace selection is performed within the candidate residual field set to generate an equipment type constraint residual subspace. The RS-Hash algorithm employs a residual direction-separated hashing mechanism. Based on device type constraints, the residual subspace and residual direction are constrained, and hash mapping is performed according to positive residuals, negative residuals, and zero-neighbor residuals to generate a direction-separated hash bucket count.

[0015] Optionally, generating fault diagnosis results for building low-voltage electrical equipment includes: Read the direction separation hash bucket count, weak current equipment trusted status frame, topology lag adjacency residual index table and residual direction diagnosis sample pool, and extract data credibility, field missing status, delayed reporting status, missing completion status, duplicate alarm status, topology propagation support, residual direction and residual field category. Based on data credibility, field missing status, delayed reporting status, missing data completion status, and duplicate alarm status, a credibility count correction coefficient is generated to correct the direction separation hash bucket count and generate a credibility correction residual abnormal score. The fault diagnosis results of building low-voltage equipment are generated based on the reliable correction residual abnormality score, topology propagation support, residual direction and residual field category.

[0016] The beneficial effects of this invention are: This invention proposes a deep learning-based fault diagnosis method for building low-voltage electrical equipment. By collecting operational monitoring data, alarm maintenance data, and topology data of the low-voltage electrical equipment, it generates multi-source operational data and constructs a reliable status frame for the equipment. It performs reliable processing on missing fields, delayed reporting, missing data completion, and duplicate alarm states, improving the consistency and reliability of multi-source low-voltage electrical data before it enters the model. Compared to diagnostic methods relying on fixed thresholds and single-device alarm rules, this invention reduces the interference of low-reliability data, historical anomaly contamination, and duplicate alarms on fault judgment results, thus improving the stability of fault identification for building low-voltage electrical equipment.

[0017] This invention utilizes fault contamination window isolation, an improved DLinear model, topology propagation lag alignment, a topology lag adjacency residual index table, and the RS-Hash algorithm to complete the calculation of predicted residuals for low-voltage electrical equipment, topology correction prediction, direction separation hash bucket counting, and reliable correction residual anomaly scoring. Compared to methods that rely solely on device state sequences for anomaly detection, this invention can identify the propagation lag relationship between anomalies in upstream adjacent devices and current device anomalies, distinguish between the originating fault device and the affected devices, improve the accuracy of fault source location, reduce false alarm rates, and enhance the interpretability of fault diagnosis results and the efficiency of operation and maintenance troubleshooting for building low-voltage electrical equipment. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a deep learning-based fault diagnosis method for building low-voltage electrical equipment proposed in this invention; Figure 2 This is a schematic diagram of the improved DLinear model in the deep learning-based fault diagnosis method for building low-voltage electrical equipment proposed in this invention. Figure 3 This is a schematic diagram illustrating the process of generating directional separation hash bucket counts and performing trusted count correction using the RS-Hash algorithm in a deep learning-based fault diagnosis method for building low-voltage electrical equipment proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1 , Figure 2 and Figure 3 A deep learning-based fault diagnosis method for building low-voltage electrical equipment includes: Collect operation monitoring data, alarm maintenance data, and topology data of low-voltage electrical equipment to generate multi-source operation data of building low-voltage electrical equipment; Preprocessing multi-source operation data of building low-voltage electrical equipment yields a reliable status frame for the low-voltage electrical equipment. Based on the trusted state frames of low-voltage equipment, a fault contamination window isolation operation is performed on the historical state sequence to generate a purified low-voltage equipment state sequence. An improved DLinear model is constructed, which includes a linear prediction layer for prototype working conditions, a sparse key field activation prediction layer, and a topological proximity state correction layer. The state sequence of the purification weak current equipment is input into the improved DLinear model to generate the initial predicted state of the weak current equipment. Based on the topology data and the initial predicted state of the weak current equipment, a topology propagation lag alignment operation is performed to generate lag aligned topology neighbor state data, and then a topology corrected predicted state is generated through the topology neighbor state correction layer. The predicted residuals of weak current equipment are calculated based on the reliable state frames of weak current equipment and the topology correction prediction state, and a topology lag adjacency residual index table and a residual direction diagnostic sample pool are constructed. Based on the prediction residuals of weak current equipment, the topology lag adjacency residual index table, and the residual direction diagnosis sample pool, a topology assembly residual vector is generated. The topology assembly residual vector is then input into the RS-Hash algorithm to generate a direction-separated hash bucket count. The trusted count correction is performed on the direction separation hash bucket count to obtain the trusted correction residual abnormal score, and the fault diagnosis results of building low-voltage equipment are generated.

[0021] In this embodiment, generating multi-source operating data for building low-voltage electrical equipment includes: The system collects operational monitoring data from building low-voltage electrical equipment and simultaneously records equipment identification, equipment type, collection time, and arrival time. The operational monitoring data includes online status, communication status, port status, control feedback status, and service status. Read alarm and maintenance data from the alarm record database and maintenance record database. The alarm and maintenance data includes alarm type, alarm time, recovery status, maintenance record and fault record. Read the topology data from the building's low-voltage topology configuration table. The topology data includes equipment identification, equipment type, network connection relationship, power supply connection relationship, control linkage relationship, and signal transmission relationship. By associating operation monitoring data, alarm maintenance data, and topology data according to equipment identifiers, multi-source operation data of building low-voltage electrical equipment is generated.

[0022] In this embodiment, obtaining the trusted state frame of the low-voltage equipment includes: Read the operation monitoring data, alarm maintenance data and topology data from the multi-source operation data of building low-voltage equipment, and extract the equipment identifier, equipment type, acquisition time, arrival time, status field name, status field value, alarm type and alarm time; Standardize the field format for status field names and values, and generate a unified diagnostic time based on the data acquisition time. Based on the unified diagnostic time, arrival time, status field name, status field value, alarm type, and alarm time, the system generates four statuses: missing field status, delayed reporting status, missing field completion status, and duplicate alarm status. Generate field missing status, delayed reporting status, missing data completion status, and duplicate alarm status, specifically as follows: In this embodiment, the device type, status field name, and status field value under the same device identifier and the same unified diagnostic time are read; The set of required status field names is determined based on the device type. The set of required status field names consists of fields corresponding to online status, communication status, port status, control feedback status, and service status. Compare the set of required status field names with the status field names that have already been read under the same device identifier and the same unified diagnostic time. When a field name in the required state field name set does not appear in the already read state field names, the corresponding field will be marked as missing. Read the collection time and arrival time, calculate the arrival time minus the collection time to obtain the reporting delay value, compare the reporting delay value with the preset reporting delay threshold, the preset reporting delay threshold is 3 seconds, when the reporting delay value is greater than 3 seconds, mark the corresponding status field value as delayed reporting status; Read the device identifier, status field name, and unified diagnosis time corresponding to the missing field status, and retrieve the valid status field values ​​of the same device identifier and the same status field name at adjacent unified diagnosis times. When there are valid status field values ​​at adjacent unified diagnosis times, the valid status field value with the closest time distance is written into the missing field position, and the corresponding field is marked as the adjacent diagnosis time to fill in the missing field. When there are no valid status field values ​​at adjacent unified diagnostic times, read the valid status field values ​​of the same device type and the same status field name at the same unified diagnostic time, calculate the reference field value, write the reference field value into the missing field position, and mark the corresponding field as a reference for the same type of device to fill in the missing field. Read the alarm type and alarm time under the same device identifier. When the same alarm type appears continuously within the preset repeated alarm interval, mark the corresponding alarm record as a repeated alarm state. The preset repeated alarm interval is 60 seconds. Write the field missing status, delayed reporting status, missing data completion status and repeated alarm status into the same status tag record according to the device identifier, unified diagnosis time and status field name; Based on the device identifier, the status field value at the unified diagnostic time is bound to the topology data to generate a low-voltage equipment status record. Data reliability is generated based on field missing status, delayed reporting status, missing data completion status, and duplicate alarm status. Operating condition prototype markers, key field activation markers, and topology propagation association markers are generated based on device type, status field name, topology data, and unified diagnostic time. The low-voltage equipment status record, data reliability, operating condition prototype markers, key field activation markers, and topology propagation association markers are combined to generate a reliable low-voltage equipment status frame, where: The credibility of the generated data is as follows: Read the field missing status, delayed reporting status, missing data completion status, and duplicate alarm status corresponding to the same device identifier, the same unified diagnostic time, and the same status field name; When the field is missing and the status is not triggered, the delayed reporting status is not triggered, the missing data completion status is not completed, and the duplicate alarm status is not triggered, the corresponding status field value will be determined as high-confidence data. When the field is missing, delayed reporting, or repeated alarm, or when the missing field is filled in at the time of the adjacent diagnosis or when the missing field is filled in by reference to the same type of device, the corresponding field value will be determined as de-reliable data. Read the field missing status, delayed reporting status, missing data completion status, and duplicate alarm status; When a field is missing, the number of de-trusted states increases by 1. When the delayed reporting status is triggered, the number of de-trusted statuses increases by 1; When the missing completion status is completed at a nearby diagnostic time or referenced by a similar device, the number of de-confidence statuses increases by 1. When the repeated alarm status is triggered, the number of de-trusted statuses increases by 1; The data credibility calculation value is equal to 1 minus the product of the number of decredible states and 0.15; When the calculated data credibility value is greater than or equal to 0.40, the calculated data credibility value is determined as the data credibility. When the calculated data credibility value is less than 0.40, 0.40 is determined as the data credibility. Write the data credibility along with the device identifier, unified diagnostic time, status field name, and status field value into the same trusted field record; Generate working condition prototype markers, key field activation markers, and topology propagation association markers, specifically as follows: Read the equipment identifier, equipment type, unified diagnosis time, status field name and status field value from the status record of low-voltage equipment; read the alarm type, alarm time, recovery status, maintenance record and fault record from the alarm maintenance data; read the preset working condition prototype table, which consists of equipment type, working condition prototype name, working condition judgment field combination and working condition field value range. Establish a combination of operating condition judgment fields according to equipment type. The combination of operating condition judgment fields consists of online status, communication status, port status, control feedback status, service status, alarm type and recovery status. The working condition judgment fields under the same equipment identifier and the same unified diagnosis time are matched with the preset working condition prototype table, and the matching results are written into the status frame record corresponding to the same equipment identifier and the same unified diagnosis time to generate a working condition prototype mark. A candidate set of key fields is established according to the device type. The candidate set of key fields for cameras includes online status, communication status and service status. The candidate set of key fields for switches includes port status, communication status and service status. The candidate set of key fields for access controllers includes online status, control feedback status and service status. The candidate set of key fields for alarm hosts includes communication status, control feedback status and alarm type. The status field name is compared with the key field candidate set. When the status field name belongs to the key field candidate set under the corresponding device type, the status field name is marked as a key status field and a key field activation tag is generated. The system reads the neighbor device identifier and topology connection type corresponding to the current device based on the topology data. The topology connection type includes network connection relationship, power supply connection relationship, control linkage relationship and signal transmission relationship. When there is at least one type of topology connection between the current device and its neighboring devices, the current device and its neighboring devices are written into the topology propagation association record. The topology propagation association record is then bound to the unified diagnostic time, device identifier, and status field name to generate a topology propagation association tag. The preset working condition prototype table includes: Extract communication latency and packet loss rate from communication status; extract port error packet rate and port traffic utilization from port status; extract control feedback success rate and control feedback latency from control feedback status; extract bitstream frame rate and card swipe record interval from service status. Normal operating conditions of the camera: online status is set to 1, communication latency is less than or equal to 200 milliseconds, packet loss rate is less than or equal to 2%, and bitstream frame rate is greater than or equal to 24 frames per second; Camera communication fluctuation conditions: online status is set to 1, communication latency is greater than 200 milliseconds and less than or equal to 1000 milliseconds, packet loss rate is greater than 2% and less than or equal to 10%, and bitstream frame rate is greater than or equal to 15 frames per second. Normal operating conditions of the switch: Communication status is normal, port status is connected, port error packet rate is less than or equal to 1%, and port traffic utilization is greater than or equal to 5% and less than or equal to 80%. Switch network congestion conditions: communication status is fluctuating, port status is connected, port error packet rate is greater than 1% and less than or equal to 5%, and port traffic utilization is greater than 80% and less than or equal to 95%. Normal operating conditions of the access control controller: online status is 1, control feedback success rate is greater than or equal to 95%, control feedback delay is less than or equal to 500 milliseconds, and card swipe record interval is greater than or equal to 1 second; Access controller control linkage conditions: online status is set to 1, control feedback success rate is greater than or equal to 90%, control feedback delay is greater than 500 milliseconds and less than or equal to 1500 milliseconds, alarm type is linkage trigger; Alarm recovery status of the alarm host: communication status is normal, alarm type is recovery alarm, recovery status is recovered, and the time interval between the alarm time and the corresponding recovery status time is less than or equal to 300 seconds. Post-maintenance recovery status: maintenance records are set to "completed", fault records are set to "processed", online status is set to 1, communication status is set to "normal", and recovery status is set to "recovered". Generate a trusted status frame for low-voltage equipment, specifically as follows: Read the status records of low-voltage equipment, data reliability, operating condition prototype markers, key field activation markers, and topology propagation association markers; Establish a status frame record according to the equipment identification and unified diagnosis time, and write the status field name, status field value, field missing status, delayed reporting status, missing data completion status, repeated alarm status, data credibility, operating condition prototype mark, key field activation mark and topology propagation association mark into the status frame record to generate a reliable status frame for weak current equipment.

[0023] In this embodiment, generating the state sequence of the purification low-voltage electrical equipment includes: Read the device identifier, device type, unified diagnosis time, status field value, field missing status, delayed reporting status, missing data completion status, repeated alarm status, data credibility and topology propagation association marker from the trusted status frame of the weak current equipment, and form a historical status sequence according to the device identifier and unified diagnosis time. The historical state sequence is divided into candidate historical windows according to the continuous and unified diagnosis time. The number of low-confidence fields, missing fields, delayed reporting fields, and duplicate alarms in the candidate historical windows are counted to generate window status records, where: The historical state sequence is divided into candidate historical windows according to the continuous and unified diagnostic time, specifically as follows: Read the unified diagnosis time in the historical state sequence, and use 5 consecutive unified diagnosis times as a candidate historical window length; Starting from the unified diagnosis time at the beginning of the historical state sequence, the candidate historical window is divided by sliding. Each sliding interval is one unified diagnosis time. The device identifier, device type, status field value, field missing status, delayed reporting status, missing data completion status, repeated alarm status, data reliability, and topology propagation association marker in the candidate historical window are written into the candidate historical window record. Generate window state records, specifically: Read the data credibility, field missing status, delayed reporting status, missing data completion status, and duplicate alarm status from the candidate history window records; Status fields with a data confidence level less than 0.70 are classified as low-confidence fields. The system counts the number of fields whose missing status is triggered, and generates the number of missing fields. Count the number of fields with a delayed reporting status that is triggered, and generate the number of fields with delayed reporting. Count the number of times the repeated alarm status is triggered, and generate the repeated alarm count; Write the number of low-confidence fields, the number of missing fields, the number of fields with delayed reporting, and the number of duplicate alarms into the candidate history window record to generate a window status record; Fault contamination status is identified based on window state records and topology propagation association markers. Candidate historical windows that meet the contamination determination criteria are marked as fault contamination windows, and contamination window markers are generated, where: The fault contamination state is identified based on window state records and topology propagation association markers, specifically as follows: Read the number of low-confidence fields, the number of missing fields, the number of delayed reported fields, the number of repeated alarms, and the topology propagation association markers in the window status record; When the number of low-confidence fields is greater than 2, the candidate historical window is marked as low-confidence contaminated. When the number of missing fields is greater than 1, the candidate historical window is marked as missing contaminated. When the number of delayed reporting fields is greater than 1, the candidate historical window is marked as delayed contaminated. When the number of repeated alarms is greater than 2, the candidate historical window is marked as alarm contaminated. When the topology propagation association is marked as triggered, the candidate historical window is marked as topology propagation contaminated. Summarize the low-reliability contamination state, missing contamination state, delayed contamination state, alarm contamination state, and topology propagation contamination state; Generate a faulty contamination state; The specific criteria for determining pollution are as follows: Read the fault pollution status. When at least one of the following states is triggered: low confidence pollution status, missing pollution status, delayed pollution status, alarm pollution status, and topology propagation pollution status, mark the candidate historical window as a fault pollution window. When the low-confidence pollution state, missing pollution state, delayed pollution state, alarm pollution state, and topology propagation pollution state are all not triggered, the candidate history window is marked as a normal history window. Write the faulty contaminated window and the normal history window into the contaminated window marker; Pollution isolation weights are generated based on pollution window markers. The status field values ​​in the candidate historical windows are then weighted and assembled to generate a cleanup window status record, where: Pollution isolation weights are generated based on pollution window markings, specifically as follows: Read the contamination window markers and fault contamination status; When a candidate historical window is marked as a normal historical window, the pollution isolation weight is set to 1. When a candidate historical window is marked as a low-confidence contamination state, the contamination isolation weight is set to 0.70; When a candidate historical window is marked as missing contamination, the contamination isolation weight is 0.60; When a candidate historical window is marked as a delayed contamination state, the contamination isolation weight is 0.75; When a candidate historical window is marked as an alarm pollution state, the pollution isolation weight is 0.65; When a candidate historical window is marked as being in a topology propagation contamination state, the contamination isolation weight is set to 0.50; When there are multiple fault contamination states in the same candidate historical window, the minimum value among the contamination isolation weights corresponding to each fault contamination state is taken as the contamination isolation weight. The status field values ​​in the candidate history window are weighted and assembled, specifically as follows: Read the status field value and pollution isolation weight from the candidate history window. For numerical status field values, multiply the status field value by the pollution isolation weight to generate the purification status field value. For both enumerated and binary status field values, retain the status field value as the cleanup status field value, and write the pollution isolation weight into the same cleanup window record to generate a cleanup window status record. Based on the unified diagnostic time, the cleanroom window status records are summarized to generate a cleanroom low-voltage equipment status sequence, including: Generate the state sequence of the purification low-voltage electrical equipment, specifically as follows: Read the purification window status records corresponding to each candidate historical window, sort the purification window status records according to the device identifier and unified diagnosis time, summarize the purification status field values ​​of the same device identifier under consecutive unified diagnosis times, and write the device identifier, device type, unified diagnosis time, status field name, purification status field value and pollution isolation weight into the same purification sequence record to generate a purification weak current equipment status sequence.

[0024] In this embodiment, generating the initial predicted state of the low-voltage equipment includes: An improved DLinear model is constructed by integrating the linear prediction layer for prototype operating conditions, the sparse key field activation prediction layer, and the topological proximity state correction layer into the DLinear prediction process. The process of constructing the improved DLinear model is as follows: The DLinear model retains the trend decomposition structure, seasonal fluctuation branch, linear prediction branch, residual connection path, and prediction output structure backbone, and structural modifications are made to the input representation, operating condition prediction, key field activation, and topology correction around the fault diagnosis scenario of building low-voltage equipment. Add a linear prediction layer with working condition prototype conditions to the input end of the prediction structure, and transform the input of ordinary state field values ​​into an input form with working condition prototype tags, so that normal operation condition, communication fluctuation condition, control linkage condition, alarm recovery condition and post-maintenance recovery condition can participate in the state prediction process. A sparse key field activation prediction layer is added to the state field prediction position, transforming the method of uniformly participating in the prediction of all state fields into a key state field sparse activation prediction method, so that device type, data credibility and key field activation markers can participate in the key state field filtering. The linear prediction structure is improved to form a parallel processing structure for operating condition prediction and key field prediction, and the unified linear prediction method is transformed into a dual-path expression method for operating condition prediction state and key field prediction state. A topology proximity state correction layer is added to the predicted output position to increase the correction entry for lagging aligned topology proximity state data, so that network connection relationships, power supply connection relationships, control linkage relationships and signal transmission relationships can participate in the predicted state correction. The prediction output method is improved by expanding the prediction target to a hierarchical prediction result of the predicted state of operating conditions, the predicted state of key fields, the initial predicted state of weak current equipment, and the predicted state of topology correction. The state sequence of the purification low-voltage electrical equipment is input into the improved DLinear model. The model reads the equipment identifier, equipment type, unified diagnostic time, state field name, state field value, data reliability, operating condition prototype marker, and key field activation marker to generate the model input state sequence, where: The model input state sequence is generated as follows: Read the equipment identifier, equipment type, unified diagnostic time, status field name, status field value, data reliability, working condition prototype mark, and key field activation mark from the status sequence of the purification low-voltage equipment; Sort the status field values ​​by time according to the device identifier and the unified diagnostic time; Write the status field value, data credibility, working condition prototype flag, and key field activation flag into the same input status record according to the status field name; Summarize the input status records of the same device identifier at consecutive and unified diagnostic times; Generate the model input state sequence; The model input state sequence is fed into the load condition prototype linear prediction layer. Based on the load condition prototype labels, linear prediction of the state field values ​​is performed to generate the load condition predicted state, where: The linear prediction layer for prototype operating conditions specifically includes: Operating status reading area: Reads the device identifier, device type, unified diagnostic time, status field name, status field value and operating status prototype mark from the model input status sequence, and divides the status field value into running status field, communication status field, control feedback field and business status field according to the device type; Operating condition prototype matching slot: Read the operating condition prototype mark and the preset operating condition prototype table, match the status field value under the same equipment identifier and the same unified diagnosis time with the normal operation condition, communication fluctuation condition, control linkage condition, alarm recovery condition and maintenance recovery condition, and generate operating condition matching results. Conditional linear prediction link: Select the corresponding linear prediction parameters according to the working condition matching results, perform trend decomposition and linear prediction on the state field values, and generate the predicted field values ​​of each state field under the corresponding working condition. Operating condition prediction output: Summarizes the prediction field values ​​according to equipment identifier, equipment type, unified diagnosis time and status field name, and generates the operating condition prediction status; The predicted operating conditions are generated as follows: The device identifier, device type, unified diagnostic time, status field name, status field value, data reliability, and operating condition prototype mark are read from the model input status sequence through the operating condition status reading area. According to the equipment identification and unified diagnostic time, the status field values ​​of the same low-voltage equipment are assembled into a working condition status vector; The preset working condition prototype table is read through the working condition prototype matching slot, and the working condition state vector is matched with the value range of the working condition field corresponding to the working condition prototype mark to generate the working condition matching strength. The conditional linear prediction parameters are determined based on the working condition matching strength, and then written into the conditional linear prediction link. The conditional linear prediction link is used to perform trend decomposition on the condition state vector to generate the condition trend state and the condition fluctuation state. Perform linear prediction on the trend state and the fluctuation state of the operating condition respectively, and generate trend prediction field value and fluctuation prediction field value; By summarizing the trend prediction field values ​​and fluctuation prediction field values ​​according to the equipment identification, unified diagnosis time and status field name through the operating condition prediction output port, the operating condition prediction status is generated. The model input state sequence and the predicted state under operating conditions are input into the sparse key field activation prediction layer. Key state fields are determined based on equipment type, data reliability, and key field activation tags. Sparse activation prediction is then performed to generate the key field predicted state, where: Sparse key fields activate the prediction layer, specifically including: Field contribution reading area: Reads the equipment type, state field name, state field value, data confidence and key field activation flags from the model input state sequence, and reads the predicted field values ​​from the predicted state under the working condition; Key field candidate slots: Establish a key field candidate set according to equipment type, compare the status field name with the key field candidate set, and filter candidate status fields related to fault diagnosis; Sparse activation gating: Based on data credibility and key field activation tags, candidate state fields are activated, and candidate state fields that meet the activation conditions are identified as key state fields, while the prediction participation intensity of non-key state fields is reduced. Key field prediction output: Perform sparse activation prediction on key state fields, and associate the predicted field values ​​corresponding to the key state fields with the predicted state of the working conditions to generate the predicted state of the key fields. Key status fields are determined based on device type, data reliability, and key field activation flags, specifically: The device type, state field name, state field value, data credibility, and key field activation flags in the model input state sequence are read through the field contribution reading area. Read the predicted field values ​​in the predicted state of the working condition, and calculate the field prediction deviation between the state field value and the predicted field value; The key field candidate set corresponding to the device type is read through the key field candidate slot, and the status field name is matched with the key field candidate set to generate a field candidate matching tag; Field activation scores are generated based on field candidate matching tags, data credibility, key field activation tags, and field prediction deviations. When a field candidate match is marked as matched, a key field activation is marked as activated, data confidence is greater than or equal to 0.40, and the field activation score is greater than the preset field activation threshold, the status field name is determined as the key status field. The field activation score is equal to the field candidate match value multiplied by 0.30, plus the data confidence multiplied by 0.25, plus the key field activation value multiplied by 0.25, plus the field prediction deviation normalization value multiplied by 0.20. When the field candidate match is marked as matched, the field candidate match value is 1, and when it is not matched, it is 0. When the key field activation is marked as activated, the key field activation value is 1, and when it is not activated, it is 0. The preset field activation threshold is 0.60. Write the key status fields, along with the device identifier, device type, unified diagnostic time, and field activation score, into the key field record; Perform sparse activation prediction, specifically as follows: Key field records, model input state sequences, and operating condition prediction states are read using sparse activation gating. The predictive engagement strength of key state fields is determined based on field activation scores; Set high participation intensity for the status field values ​​corresponding to critical status fields, and low participation intensity for the status field values ​​corresponding to non-critical status fields. Based on the intensity of predicted participation, the predicted field values ​​in the predicted state of the working condition are sparsely corrected to generate sparsely corrected predicted field values. By summarizing and sparsely correcting the prediction field values ​​through the key field prediction output port according to the device identifier, unified diagnosis time, status field name and prediction participation intensity, the key field prediction status is generated. The predicted state of operating conditions and the predicted state of key fields are aligned according to the unified diagnostic time to generate the initial predicted state of weak current equipment. The initial predicted state of low-voltage equipment is associated with the equipment identifier, equipment type, and unified diagnostic time to generate an initial predicted state record, in which: Generate an initial prediction state record, specifically as follows: Read the initial predicted state, equipment identification, equipment type, unified diagnosis time, state field name and predicted field value of the weak current equipment; read the working condition prototype mark and key field activation mark in the model input state sequence; read the prediction participation intensity in the predicted state of the key field; and then write it into the initial predicted state record. Establish an initial prediction status record according to the equipment identification and unified diagnostic time, and write the equipment type, status field name, prediction field value, working condition prototype mark, key field activation mark and prediction participation intensity into the initial prediction status record; The initial predicted state records under the same device identifier are aggregated according to the unified diagnostic time sequence to generate initial predicted state records that can be called by the topology proximity state correction layer. The improved DLinear model was trained using a combination of prediction errors based on operating conditions, key field prediction errors, topological proximity state correction errors, and initial prediction state errors as the joint optimization objective. The model parameters of the linear prediction layer for operating conditions, the sparse key field activation prediction layer, and the topological proximity state correction layer were iteratively updated. The improved DLinear model was considered to have completed convergence training when the change in the joint loss value over five consecutive training rounds was less than 0.001. The improved DLinear model is trained as follows: Extract the device identifier, device type, unified diagnosis time, status field name, status field value, data credibility, operating condition prototype marker, and key field activation marker from the historical trusted status frames of low-voltage equipment. Extract the values ​​of labeled status fields, labeled key status fields, and labeled topology correction status from historical fault diagnosis results; A training sample set was constructed, which included 12,000 sets of low-voltage equipment status samples, and was divided into a training set, a validation set, and a test set in an 8:1:1 ratio. The training set is input into the improved DLinear model. The working condition prediction state is output by the linear prediction layer of the working condition prototype, the key field prediction state is output by the sparse key field activation prediction layer, and the topology correction prediction state is output by the topology neighbor state correction layer. The average of the squared differences between the predicted state of the operating condition and the labeled state field values ​​is used to obtain the operating condition prediction error. The average of the squared differences between the predicted state of the key field and the labeled key state field values ​​is used to obtain the key field prediction error. The average of the squared differences between the predicted state of the topology correction and the labeled topology correction state values ​​is used to obtain the topology proximity state correction error. The average of the squared differences between the predicted state of the weak current equipment and the labeled state field values ​​is used to obtain the initial prediction state error. The joint loss value is generated by multiplying the operating condition prediction error by 0.30, the key field prediction error by 0.25, the topology proximity state correction error by 0.25, and the initial prediction state error by 0.20. The model parameters of the linear prediction layer, sparse key field activation prediction layer, and topological proximity state correction layer under the prototype working condition are iteratively updated based on the joint loss value. The learning rate is 0.001. When the change of the joint loss value in 5 consecutive training rounds is less than 0.001, the improved DLinear model is considered to have completed convergence training.

[0025] In this embodiment, generating the topology correction prediction state includes: Read device identifiers, device types, network connection relationships, power supply connection relationships, control linkage relationships, and signal transmission relationships from the topology data. Also, read the initial predicted state of low-voltage equipment, unified diagnostic time, and state field name from the initial predicted state record to generate a topology edge state record, where: Generate topological edge state records, specifically as follows: Read the device identifier, device type, network connection relationship, power supply connection relationship, control linkage relationship and signal transmission relationship from the topology data; read the initial predicted state of weak current equipment, unified diagnosis time and status field name from the initial predicted state record. The topology data is associated with the initial predicted state record according to the device identifier. The device identifier, device type, unified diagnostic time, status field name, initial predicted state of weak current equipment, network connection relationship, power supply connection relationship, control linkage relationship and signal transmission relationship are written into the same topology edge record to generate a topology edge state record. Based on the topology edge state record, the upstream adjacent devices, same-link adjacent devices, same-power-branch devices, and control linkage devices corresponding to the current device are determined, and topology edge type tags are generated according to network connection relationship, power supply connection relationship, control linkage relationship, and signal transmission relationship, respectively, where: Identify the upstream adjacent devices, on the same link, on the same power supply branch, and control linkage devices corresponding to the current device, specifically: Read the device identifier, network connection relationship, power supply connection relationship, control linkage relationship and signal transmission relationship from the topology edge status record; The device identifier located upstream in the network connection relationship is identified as the upstream adjacent device; Devices that share the same link identifier with the current device in network connection or signal transmission relationships are identified as adjacent devices on the same link; devices that share the same power supply branch identifier with the current device in power supply connection relationships are identified as devices on the same power supply branch; and devices that have control command transmission or feedback linkage relationships with the current device in control linkage relationships are identified as control linkage devices. Generate a set of topologically adjacent devices; Generate topological edge type markers, specifically: Read the set of topological adjacent devices and the topological edge state records; When a network connection exists between the current device and its neighboring devices, a network connection edge marker is generated; When a power supply connection exists between the current device and an adjacent device, a power supply connection edge marker is generated; When there is a control linkage relationship between the current device and its adjacent devices, a control linkage edge marker is generated; When there is a signal transmission relationship between the current device and its neighboring devices, a signal transmission edge marker is generated; Write the network connection edge marker, power supply connection edge marker, control linkage edge marker, and signal transmission edge marker into the same topology edge type record; The propagation lag retrieval window is determined based on the topological edge type label, and a candidate propagation state sequence of adjacent devices is generated, wherein: The propagation lag retrieval window is determined as follows: Read the topology edge type marker and unified diagnostic time; The propagation lag retrieval window corresponding to the network connection edge marker takes 1 to 3 unified diagnosis times prior to the current unified diagnosis time; The propagation lag retrieval window corresponding to the power supply connection edge mark takes 0 to 2 unified diagnosis times prior to the current unified diagnosis time; The propagation lag retrieval window corresponding to the control linkage edge marker takes 1 to 4 unified diagnosis times prior to the current unified diagnosis time; The propagation lag retrieval window corresponding to the signal transmission edge marker takes 1 to 3 unified diagnosis times prior to the current unified diagnosis time; Write the topological edge type label, the current unified diagnosis time, and the propagation lag retrieval window into the same lag window record; Generate a candidate propagation state sequence for adjacent devices, specifically as follows: Read the adjacent device identifier, topology edge type marker, and propagation hysteresis retrieval window from the hysteresis window record; Retrieve the initial predicted state record corresponding to the adjacent device within the propagation lag retrieval window; Read the status field names of adjacent devices and the initial predicted status of weak current equipment at each unified diagnostic time; The initial predicted states of the weak current equipment corresponding to adjacent equipment are arranged in the order of the unified diagnosis time. Generate a candidate propagation state sequence for adjacent devices; The candidate propagation state sequences of adjacent devices are matched with the initial predicted state of the current device's weak current equipment by matching state field correspondence, predicted state change direction, and unified diagnostic time offset, generating a topology propagation lag matching record, wherein: The status field matches accordingly, specifically: Read the state field names from the candidate propagation state sequence of adjacent devices, and read the state field names from the initial predicted state of the weak current equipment of the current device; The propagable field category is determined by the topology edge type label; The network connection edge is marked with the corresponding communication status, port status and service status; the power supply connection edge is marked with the corresponding online status and communication status; the control linkage edge is marked with the corresponding control feedback status and service status; and the signal transmission edge is marked with the corresponding communication status and service status. Mark the status field names that belong to the same propagable field category as the corresponding field matches; Matching the predicted direction of state change, specifically: Read the initial predicted state of weak current equipment at adjacent unified diagnosis times in the candidate propagation state sequence of adjacent devices; Calculate the direction of predicted state change of adjacent devices, and read the initial predicted state of the current device's weak current equipment at the adjacent unified diagnosis time. Calculate the predicted state change direction of the current device. When the predicted state change direction of the adjacent device is consistent with the predicted state change direction of the current device, generate a direction matching mark. When the predicted state change direction of an adjacent device is inconsistent with the predicted state change direction of the current device, a direction mismatch marker is generated. Unified diagnostic time offset matching, specifically: Read the unified diagnostic time corresponding to the current device, and read the candidate unified diagnostic time from the candidate propagation state sequence of adjacent devices; The unified diagnosis time offset value is obtained by subtracting the candidate unified diagnosis time from the unified diagnosis time corresponding to the current device. The unified diagnosis time offset value is compared with the propagation lag retrieval window. When the unified diagnosis time offset value is within the propagation lag retrieval window, an offset matching mark is generated. Write the field correspondence match, direction match mark, offset match mark, unified diagnosis time offset value and topology edge type mark into the same match record to generate a topology propagation lag match record; Based on the topology propagation lag matching record, the propagation alignment time of the adjacent device corresponding to the current device is determined. Then, the initial predicted state of the adjacent device at the propagation alignment time, the topology edge type label, and the unified diagnostic time offset value are combined to generate lag-aligned topology proximity state data, where: The specific time for determining propagation alignment is as follows: Read the field correspondence match, direction match flag, offset match flag, and unified diagnostic time offset value from the topology propagation lag match record; Candidate unified diagnostic moments that simultaneously have field correspondence matching, direction matching markers, and offset matching markers are determined as candidate propagation alignment moments. When there are multiple candidate propagation alignment moments, the candidate propagation alignment moment with the smallest unified diagnostic moment offset value is selected as the propagation alignment moment. Associate the propagation alignment time with the adjacent device identifier, the current device identifier, and the topology edge type marker; Generate hysteresis-aligned topological neighbor state data, specifically as follows: Read the initial predicted state of adjacent devices at the propagation alignment time; Read the propagation alignment time, topology edge type marker, and unified diagnostic time offset value, and establish a hysteresis alignment record according to the current device identifier, adjacent device identifier, and status field name; Write the initial predicted state of adjacent devices, propagation alignment time, topology edge type label and unified diagnosis time offset value into the hysteresis alignment record to generate hysteresis alignment topology proximity state data; The lagging aligned topology proximity state data and the initial predicted state of the weak current equipment are input into the topology proximity state correction layer to generate the proximity state correction value. The initial predicted state of the weak current equipment is then topologically corrected to generate the topology-corrected predicted state, where: The topology proximity state correction layer specifically includes: Proximity state alignment entry: Read the lag alignment topology proximity state data and the initial predicted state of weak current equipment, and extract the current device identifier, adjacent device identifier, topology edge type marker, propagation alignment time, and unified diagnosis time offset value; Edge type correction route: Divide the correction path according to network connection relationship, power supply connection relationship, control linkage relationship and signal transmission relationship, and write the initial predicted state of adjacent devices into the corresponding topology edge type path; Proximity state correction slot: Generates proximity state correction amount based on topology edge type label and unified diagnostic time offset value, and performs topology correction on the predicted field value in the initial predicted state of weak current equipment; Topology correction output port: Summarizes the corrected prediction field values ​​according to device identifier, device type, unified diagnostic time and status field name, and generates topology correction prediction status; The neighbor state correction is generated as follows: Read the hysteresis-aligned topology proximity state data and the initial predicted state of weak current equipment through the proximity state alignment entry; The adjacent state correction path is divided according to the network connection edge mark, power supply connection edge mark, control linkage edge mark and signal transmission edge mark by the edge type correction route; The initial predicted state of adjacent devices, the unified diagnostic time offset value, and the topology edge type mark are read through the proximity state correction slot. Edge type correction coefficients are generated according to the topology edge type markings. The edge type correction coefficient corresponding to the network connection edge marking is 0.80, the edge type correction coefficient corresponding to the power supply connection edge marking is 0.90, the edge type correction coefficient corresponding to the control linkage edge marking is 0.75, and the edge type correction coefficient corresponding to the signal transmission edge marking is 0.70. The proximity state correction is equal to the product of the initial predicted state of the adjacent device and the edge type correction coefficient, divided by 1 and the sum of the uniform diagnostic time offset. The topology correction prediction state is generated as follows: Read the initial predicted state and adjacent state correction values ​​of the low-voltage equipment; Establish calibration field records according to equipment identification, unified diagnostic time, and status field name; When the same state field name corresponds to a neighboring state correction amount, the corrected prediction field value is equal to the initial prediction field value multiplied by 0.70, plus the neighboring state correction amount multiplied by 0.30; When the same state field name corresponds to multiple neighboring state correction values, the average of the multiple neighboring state correction values ​​is taken to generate the average neighboring state correction value. The corrected prediction field value is equal to the initial prediction field value multiplied by 0.70, plus the average neighbor state correction multiplied by 0.30; The topology correction prediction status is generated by summarizing the correction prediction field values ​​through the topology correction output port according to the device identifier, device type, unified diagnostic time, and status field name.

[0026] In this embodiment, the construction of the topological hysteresis adjacency residual index table and the residual direction diagnostic sample pool includes: Read the device identifier, device type, unified diagnosis time, status field name, status field value and data confidence from the trusted status frame of the weak current equipment, and read the prediction field name and prediction field value from the topology correction prediction status; Based on the equipment identification, unified diagnostic time, and status field name, the status field value and the predicted field value are used to calculate the predicted residual of the low-voltage equipment. The predicted residual of the low-voltage equipment includes the residual value, residual direction, residual field name, residual field category, and residual duration, among which: The prediction residual for low-voltage equipment is calculated by combining the status field value and the prediction field value, specifically as follows: Read the status field values ​​and prediction field values ​​corresponding to the same device identifier, the same unified diagnostic time, and the same status field name; Read the field type corresponding to the status field name. The field type includes numeric fields, binary fields, and enumerated fields. When the field type is numeric, the residual value is generated by subtracting the prediction field value from the status field value. When the field type is a binary field, the state field value and the prediction field value are converted into 0 or 1 state code values, and the prediction field code value is subtracted from the state field code value to generate the residual value; When the field type is an enumeration field, the state field value and the prediction field value are mapped to the corresponding enumeration code value, and the state field enumeration code value is subtracted from the prediction field enumeration code value to generate the residual value; The zero-neighbor threshold is set to 0.05 for numeric fields, 0.50 for binary fields, and 0.50 for enumerated fields. When the residual value is greater than the zero neighbor threshold of the corresponding field type, the residual direction is marked as a positive residual; When the residual value is less than the negative of the zero-neighbor threshold of the corresponding field type, the residual direction is marked as a negative residual; When the residual value is greater than or equal to the negative of the zero-neighbor threshold of the corresponding field type, and less than or equal to the zero-neighbor threshold of the corresponding field type, the residual direction is marked as a zero-neighbor residual. The residual field category is determined based on the status field name. The number of times the same device identifier and the same status field name appear in the same residual direction at consecutive unified diagnosis times is counted, and the residual duration count is generated. The equipment identifier, equipment type, unified diagnosis time, status field name, residual value, residual direction, residual field category, residual duration, and data credibility are combined to generate the predicted residual of the low-voltage equipment. Read the topology data and the lag-aligned topology neighbor status data, correlate the current device's prediction residual with the historical prediction residual of neighboring devices, calculate the residual magnitude matching value and topology propagation support, and generate a topology lag-adjacent residual index table, where: The residual magnitude matching value and topology propagation support are calculated as follows: Read the current device identifier, current unified diagnosis time, residual field name, residual value and residual direction from the current device prediction residual; read the adjacent device identifier, propagation alignment time, residual field name, residual value and residual direction from the adjacent device historical prediction residual; read the topology edge type label and unified diagnosis time offset value from the lag alignment topology proximity state data. Calculate the absolute value of the residual value of the current equipment's predicted residual to generate the current residual amplitude; calculate the absolute value of the residual value of the historical predicted residual of adjacent equipment to generate the adjacent residual amplitude; calculate the absolute value of the difference between the current residual amplitude and the adjacent residual amplitude to generate the residual amplitude difference. The residual amplitude matching value is equal to 1 divided by the sum of 1 and the residual amplitude difference; When the direction of the residual of the current equipment prediction residual is consistent with the direction of the residual of the historical prediction residual of the adjacent equipment, the direction consistency coefficient is 1. When the direction of the residual predicted by the current device is inconsistent with the direction of the residual predicted by the adjacent device in the past, the direction consistency coefficient is 0.50. When the topology edge type is marked as a network connection edge, the edge type support coefficient is 0.80; when the topology edge type is marked as a power supply connection edge, the edge type support coefficient is 0.90; when the topology edge type is marked as a control linkage edge, the edge type support coefficient is 0.75; and when the topology edge type is marked as a signal transmission edge, the edge type support coefficient is 0.70. The hysteresis decay coefficient is equal to 1 divided by 1 and the sum of the uniform diagnostic time offset value. The topology propagation support is equal to the product of the edge type support coefficient, the direction consistency coefficient, the hysteresis decay coefficient and the residual amplitude matching value. Write the residual magnitude matching value and topology propagation support into the topology hysteresis adjacency residual index record corresponding to the current device identifier, adjacent device identifier, propagation alignment time, topology edge type flag, and residual field name; Based on the residual values, directions, field categories, durations, and reliability of the residuals predicted by low-voltage equipment, residuals are classified and stored according to positive, negative, and zero-neighbor residuals, generating a residual direction diagnostic sample pool, where: Generate a residual orientation diagnostic sample pool, specifically as follows: Read the equipment identifier, equipment type, unified diagnosis time, residual value, residual direction, residual field name, residual field category, residual duration, and data reliability from the predicted residuals of low-voltage equipment; Establish positive residual sample regions, negative residual sample regions, and zero-neighbor residual sample regions according to the residual direction; When the residual direction is positive, the corresponding weak current equipment prediction residual is written into the positive residual sample area; when the residual direction is negative, the corresponding weak current equipment prediction residual is written into the negative residual sample area; when the residual direction is zero neighbor residual, the corresponding weak current equipment prediction residual is written into the zero neighbor residual sample area. Establish a sample index based on equipment identification, equipment type, unified diagnostic time, residual field name, and residual direction; Write the residual value, residual field category, residual duration, and data confidence level into the residual sample record corresponding to the sample index; The positive residual sample area, negative residual sample area, zero-neighbor residual sample area, and residual sample records are summarized to generate a residual direction diagnostic sample pool.

[0027] In this embodiment, the topological hysteresis adjacency residual index table includes: The topology lag adjacency residual index table consists of the current device identifier, adjacent device identifier, current device type, adjacent device type, topology edge type, current unified diagnosis time, propagation alignment time, unified diagnosis time offset value, current device prediction residual, adjacent device historical prediction residual, residual field name, residual direction, residual magnitude matching value, and topology propagation support. An index record is created between the current device's predicted residual and the adjacent device's historical predicted residual, based on the device identifier, topology edge type, propagation alignment time, residual field name, and residual direction. A dynamic residual index structure with propagation lag, residual direction, and topology propagation support is formed based on the index record.

[0028] In this embodiment, generating the direction-separated hash bucket count includes: Read the predicted residuals of low-voltage equipment, the index table of topological lag adjacency residuals, and the sample pool of residual direction diagnosis. Perform residual association according to equipment identifier, equipment type, unified diagnosis time, topological edge type, and residual field name to generate topological residual association records, including: Generate topological residual association records, specifically as follows: Read the equipment identifier, equipment type, unified diagnosis time, residual field name, residual value, residual direction, and residual field category from the predicted residuals of low-voltage equipment; Read the current device identifier, neighboring device identifier, topology edge type, propagation alignment time, historical prediction residuals of neighboring devices, and topology propagation support from the topology lag adjacency residual index table; Read the sample index, residual channel label, and data confidence level from the residual direction diagnostic sample pool; Establish the current equipment residual association based on the conditions that the equipment identifier is consistent with the current equipment identifier, the unified diagnosis time is consistent with the current unified diagnosis time, and the residual field name is consistent. Establish historical residual associations for adjacent devices based on their identifiers, propagation alignment times, and residual field names; Write the current device's prediction residual, the adjacent device's historical prediction residual, the residual channel label, the data confidence level, and the topology propagation support into the same association record to generate a topology residual association record; Based on the topological residual association record, the current device's predicted residual, the adjacent device's historical predicted residual, residual direction, residual field category, and topology propagation support are read. Residuals are then assembled according to the residual field name and topology edge type to generate a topology assembled residual vector, where: Residual assembly is performed according to the residual field name and topology edge type, specifically as follows: Read the current device prediction residual, adjacent device historical prediction residual, residual direction, residual field category, topology edge type, and topology propagation support from the topology residual association record; Establish field assembly locations based on residual field names; Establish network connection residual segments, power supply connection residual segments, control linkage residual segments, and signal transmission residual segments according to the topology edge type; Write the current device prediction residual into the current device residual segment, write the adjacent device historical prediction residual into the corresponding topology edge residual segment according to the topology edge type, write the topology propagation support into the propagation weight bit of the same field assembly position, and write the residual direction and residual field category into the direction category bit of the same field assembly position. Summarize the current device residual segments, topology edge residual segments, propagation weight bits, and direction category bits to generate a topology assembly residual vector; The RS-Hash algorithm employs a random subspace selection mechanism for weak current equipment type constraints. Based on the topology assembly residual vector, equipment type, and residual field category, a candidate residual field set is determined. Random subspace selection is then performed within this candidate residual field set to generate the equipment type-constrained residual subspace, where: The random subspace selection mechanism for low-voltage equipment type constraints is as follows: Read the device type, residual field category, and residual field name from the topology assembly residual vector; When the device type is a camera, the candidate residual field set includes online status residual, communication status residual, and service status residual. When the device type is a switch, the candidate residual field set includes port status residual, communication status residual, and service status residual. When the device type is an access control controller, the candidate residual field set includes online status residual, control feedback status residual, and service status residual. When the device type is an alarm host, the candidate residual field set includes communication status residual, control feedback status residual, and alarm type residual. The final candidate residual field set is generated by intersecting the candidate residual field set with the actual residual field names in the topology assembly residual vector. The number of fields in the random subspace is the square root of the number of fields in the final candidate residual field set, rounded up. Generate random subspace numbers, and select a corresponding number of residual fields from the final candidate residual field set based on the random subspace numbers; Read the residual value, residual direction, residual field category and topology propagation support corresponding to the selected residual field, arrange the selected residual fields according to the residual field name, and generate the device type constraint residual subspace; The RS-Hash algorithm employs a residual direction-separated hashing mechanism. Based on device type constraints, the residual subspace and residual direction are defined. Hash mapping is performed separately for positive residuals, negative residuals, and zero-neighbor residuals to generate a direction-separated hash bucket count. The residual direction split hashing mechanism is as follows: Read the residual values, residual directions, residual field names, and topology propagation support from the device type constraint residual subspace; Establish positive residual hash channels, negative residual hash channels, and zero-neighbor residual hash channels according to the residual direction. Write the residual values ​​with the residual direction of positive residual into the positive residual hash channel, write the residual values ​​with the residual direction of negative residual into the negative residual hash channel, and write the residual values ​​with the residual direction of zero-neighbor residual into the zero-neighbor residual hash channel. Read the maximum and minimum residual values ​​within the same residual hash channel; The residual range is equal to the maximum residual value minus the minimum residual value; When the residual range value is greater than 0, the hash bucket width is equal to the residual range value divided by 8; when the residual range value is equal to 0, the hash bucket width is 0.10. The random offset is determined by generating a random seed from the device identifier, unified diagnostic time and residual field name. The random offset value range is greater than or equal to 0 and less than the hash bucket width. The hash bucket number is equal to the sum of the residual value and the random offset, divided by the hash bucket width and rounded down. The number of samples in the same hash channel is counted according to the hash bucket number, generating positive hash bucket count, negative hash bucket count and zero neighbor hash bucket count. Summarize the positive hash bucket count, negative hash bucket count, and zero-neighbor hash bucket count to generate a direction-separated hash bucket count.

[0029] In this embodiment, generating fault diagnosis results for building low-voltage electrical equipment includes: Read the direction separation hash bucket count, weak current equipment trusted status frame, topology lag adjacency residual index table and residual direction diagnosis sample pool, extract data credibility, field missing status, delayed reporting status, missing completion status, repeated alarm status, topology propagation support, residual direction and residual field category, and write the direction separation hash bucket count, data credibility, field missing status, delayed reporting status, missing completion status, repeated alarm status, topology propagation support and residual field category into the same diagnosis scoring association record according to device identifier, unified diagnosis time, residual field name and residual direction; Based on data credibility, field missing status, delayed reporting status, missing data completion status, and duplicate alarm status, a credibility count correction coefficient is generated to correct the direction-separated hash bucket count, generating a credibility correction residual anomaly score, where: A reliability count correction coefficient is generated based on data reliability, field missing status, delayed reporting status, missing data completion status, and duplicate alarm status, specifically as follows: Read the data reliability, field missing status, delayed reporting status, missing data completion status, and duplicate alarm status from the diagnostic score associated records; When the field is missing and the status is "not triggered", the field completeness coefficient is 1. When the field missing status is triggered, the field integrity coefficient is 0.80; When the delay reporting status is not triggered, the delay correction coefficient is 1; When the delay reporting status is triggered, the delay correction factor is 0.85; When the missing padding status is "not padded", the padding correction factor is 1; When the missing information is filled in at a time close to the diagnosis time, the filling correction factor is 0.90. When the missing completion status is the reference completion for the same type of equipment, the completion correction factor is 0.85; When the repeated alarm status is not triggered, the repeated alarm correction coefficient is 1; When the repeated alarm status is triggered, the repeated alarm correction factor is 0.80; The confidence count correction factor is equal to the data confidence factor multiplied by the field integrity factor, then multiplied by the delay correction factor, then multiplied by the completion correction factor, and then multiplied by the duplicate alarm correction factor. Write the reliable count correction factor, along with the device identifier, unified diagnostic time, residual field name, and residual direction, into the same count correction record; The direction-separated hash bucket count is corrected to generate a reliable corrected residual anomaly score, specifically: The device identifier, unified diagnostic time, residual field name, residual direction, positive hash bucket count, negative hash bucket count, and zero neighbor hash bucket count are read from the hash bucket count separated by direction. Read the reliable count correction coefficient from the count correction record; When the residual direction is positive, read the positive hash bucket count and multiply the positive hash bucket count by the trusted count correction coefficient to generate the positive trusted correction bucket count; When the residual direction is negative, read the negative hash bucket count and multiply the negative hash bucket count by the trusted count correction coefficient to generate the negative trusted correction bucket count; When the residual direction is the zero-neighbor residual, read the zero-neighbor hash bucket count and multiply the zero-neighbor hash bucket count by the trusted count correction coefficient to generate the zero-neighbor trusted correction bucket count; When the residual direction is positive, the confidence correction residual abnormality score is equal to 1 divided by 1 and the sum of the positive confidence correction bucket count; When the residual direction is negative, the confidence correction residual abnormality score is equal to 1 divided by 1 and the sum of the negative confidence correction bucket count; When the residual direction is the zero-neighbor residual, the confidence correction residual abnormality score is equal to 1 divided by 1 and the sum of the zero-neighbor confidence correction bucket count; Write the credible correction residual anomaly score, along with the device identifier, unified diagnosis time, residual field name, residual direction, and credible count correction coefficient, into the same anomaly score record; The fault diagnosis results for building low-voltage electrical equipment are generated based on the reliable corrected residual anomaly score, topology propagation support, residual direction, and residual field category, where: Generate fault diagnosis results for building low-voltage electrical equipment, specifically as follows: Read the reliable correction residual anomaly score, reliable count correction coefficient, data reliability, topology propagation support, residual direction, and residual field category; When the confidence-corrected residual abnormality score is less than 0.60, it is marked as a normal diagnostic state; When the credible correction residual anomaly score is greater than or equal to 0.60 and the topology propagation support is less than 0.40, it is identified as the source of the fault. When the credible correction residual anomaly score is greater than or equal to 0.60 and the topology propagation support is greater than or equal to 0.40, the device is identified as being affected by propagation. Based on the residual field category, residual direction, and reliable corrected residual abnormality score, the fault type, fault change direction, and fault level are generated, and the fault diagnosis results of building low-voltage equipment are generated.

[0030] Example 1: This example applies to a comprehensive building low-voltage electrical operation monitoring scenario. The scenario includes video surveillance, access control, intrusion alarms, network switching, low-voltage gateways, and building automation equipment. The equipment is distributed across the integrated monitoring area, floor low-voltage rooms, and equipment cabinets. Problems encountered during low-voltage operation include communication delays, port congestion, decreased camera bitrate, delayed access control feedback, and repeated alarms from the alarm host. Traditional fixed threshold methods rely solely on the online status, alarm status, and port status of individual devices. When an upstream switch malfunction propagates the problem to downstream cameras and access controllers, it can easily lead to all affected devices being classified as independently faulty.

[0031] This embodiment constructs a batch of simulated operation data, continuously collected for 72 hours. The data collection targets include 96 cameras, 24 switches, 38 access controllers, 8 alarm hosts, 6 low-voltage gateways, and 14 building automation terminals. The sampling period is 1 second, resulting in 259,200 unified diagnostic moments. The original operation monitoring data includes online status, communication status, port status, control feedback status, and service status. The alarm maintenance data includes alarm type, alarm time, recovery status, maintenance records, and fault records. The topology data includes network connection relationships, power supply connection relationships, control linkage relationships, and signal transmission relationships. There are a total of 4,820,000 original field records, 32,860 records with missing fields, 21,740 records with delayed reporting, and 3,960 records with duplicate alarms.

[0032] The training sample set contains 12,000 sets of low-voltage equipment status samples: 9,600 sets for training, 1,200 sets for validation, and 1,200 sets for testing. Each training sample includes device identifier, device type, unified diagnostic time, status field name, status field value, data reliability, operating condition prototype marker, key field activation marker, and topology propagation association marker. Training sample A is a switch status sample, with the device type being a switch, the unified diagnostic time being the 138,421st diagnostic time, the port status being connected, the port error packet rate being 4.6%, the port traffic utilization rate being 91%, the communication status being fluctuating, the data reliability being 0.85, the operating condition prototype marker being network congestion, and the key field activation marker being activated. Training sample B is a camera status sample, with the unified diagnostic time being the 138,423rd diagnostic time, the online status being 1, the communication latency being 820 milliseconds, the packet loss rate being 8.4%, the bitrate frame rate being 17 frames per second, the data reliability being 0.70, and the topology propagation association marker being triggered. Training sample C is the access control controller status sample. The unified diagnosis time is the 156080th diagnosis time. The control feedback success rate is 88%, the control feedback delay is 1420 milliseconds, the alarm type is linkage trigger, and the data confidence level is 0.85.

[0033] After the data enters the processing flow, the system generates a reliable status frame for the low-voltage equipment based on the device identifier, acquisition time, and arrival time. The reporting delay threshold is set to 3 seconds, and the repeat alarm interval is set to 60 seconds. For missing fields, the system first reads the valid status field values ​​at adjacent unified diagnostic times to fill in the missing values; if no valid values ​​are available at adjacent unified diagnostic times, it reads the valid status field values ​​under the same device type and the same status field name to calculate the reference field value. After processing, the number of missing field records decreased from 32,860 to 4,210, the number of delayed reporting status markers increased to 21,740, the number of repeat alarm status markers increased to 3,960, and the proportion of highly reliable fields increased from 88.4% to 96.7%.

[0034] The system divides the trusted status frames of low-voltage equipment into candidate historical windows based on five consecutive unified diagnostic times. A total of 51,840 candidate historical windows are formed. Candidate historical windows with more than two low-confidence fields, more than one missing field, more than one delayed reporting field, more than two repeated alarms, or triggering topology propagation association are marked as fault-contaminated windows. After processing, 3,746 fault-contaminated windows are identified. The contamination isolation weights are set as follows: low-confidence contamination state 0.70, missing contamination state 0.60, delayed contamination state 0.75, alarm contamination state 0.65, and topology propagation contamination state 0.50. After purification, the abnormal peak value caused by repeated alarms to the model input is reduced from 3,960 to 620.

[0035] In the improved DLinear model training, the linear prediction layer for the prototype working conditions outputs the predicted state for the working conditions, the sparse key field activation prediction layer outputs the predicted state for the key fields, and the topological proximity state correction layer outputs the topological correction predicted state. The joint loss consists of the prediction error for the working conditions, the prediction error for the key fields, the topological proximity state correction error, and the initial prediction state error, with weights of 0.30, 0.25, 0.25, and 0.20, respectively, and a learning rate of 0.001. At the 12th epoch of training, the joint loss was 0.184, decreasing to 0.071 at the 25th epoch, and to 0.029 at the 41st epoch. After the 47th epoch, the joint loss value remained below 0.001 for five consecutive epochs, indicating model convergence. On the test set, the average prediction error without the topological proximity state correction layer was 0.083, while the average prediction error decreased to 0.056 after adding the topological proximity state correction layer.

[0036] During the simulated fault, the upstream switch experienced a port anomaly at the 138421st unified diagnostic time, with the port error packet rate increasing from 0.7% to 4.6% and the port traffic utilization increasing from 62% to 91%. At the 138423rd unified diagnostic time, the 17 downstream cameras experienced communication delays and a decrease in bitrate and frame rate, with the average communication delay increasing from 160 milliseconds to 820 milliseconds and the average bitrate and frame rate decreasing from 25 frames per second to 17 frames per second. Traditional fixed threshold methods would mark the 1 switch and 17 cameras as independent faults. Using the method of this invention, the system reads the network connection edge markings and sets the propagation lag retrieval window to 1 to 3 unified diagnostic times prior to the current time, obtaining the propagation alignment time at the 138421st unified diagnostic time. In the topology propagation support calculation, the network connection edge support coefficient is set to 0.80, the direction consistency coefficient is set to 1, the unified diagnostic time offset is 2, the lag attenuation coefficient is 0.333, and the residual amplitude matching value is 0.86, resulting in a topology propagation support of 0.229. The switch has a Trusted Correction Residual Anomaly Score of 0.82 and a Topology Propagation Support of less than 0.40, and is identified as the source of the fault. The downstream camera has a Trusted Correction Residual Anomaly Score between 0.61 and 0.74 and is marked as a device whose propagation is affected.

[0037] In another set of power supply branch anomaly samples, an malfunction in one power supply device caused increased control feedback delays in five access controllers. The power supply connection edge support coefficient was set to 0.90, the unified diagnostic time offset was 1, the residual amplitude matching value was 0.91, and the topology propagation support reached 0.410. The system identified the five access controllers as the devices affected by the propagation, and did not misjudge the access controllers as the source of the fault.

[0038] In the comparative experiment, the test sample included 42 devices with originating faults, 116 devices affected by propagation, and 1380 normal devices. The traditional fixed threshold method identified 184 faulty devices, with 47 false alarms, correctly located the originating fault in 31 devices, and correctly identified 82 devices affected by propagation. The traditional single-device residual scoring method identified 171 faulty devices, with 35 false alarms, correctly located the originating fault in 34 devices, and correctly identified 91 devices affected by propagation. The method of this invention identified 40 devices with originating faults, 109 devices affected by propagation, with 14 false alarms. The accuracy rate for locating originating faults reached 95.2%, and the accuracy rate for identifying devices affected by propagation reached 94.0%. The overall false alarm rate decreased from 3.4% of the traditional fixed threshold method to 1.0%. These comparative results demonstrate that this invention can improve the accuracy of fault identification for low-voltage equipment under conditions of missing multi-source data, delayed reporting, repeated alarms, and topology propagation interference, distinguishing between devices with originating faults and devices affected by propagation, and reducing the workload of operation and maintenance troubleshooting.

[0039] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for fault diagnosis of building low-voltage electrical equipment based on deep learning, characterized in that, include: Collect operation monitoring data, alarm maintenance data, and topology data of low-voltage electrical equipment to generate multi-source operation data of building low-voltage electrical equipment; Preprocessing multi-source operation data of building low-voltage electrical equipment yields a reliable status frame for the low-voltage electrical equipment. Based on the trusted state frames of low-voltage equipment, a fault contamination window isolation operation is performed on the historical state sequence to generate a purified low-voltage equipment state sequence. An improved DLinear model is constructed, which includes a linear prediction layer for prototype working conditions, a sparse key field activation prediction layer, and a topological proximity state correction layer. The state sequence of the purification weak current equipment is input into the improved DLinear model to generate the initial predicted state of the weak current equipment. Based on the topology data and the initial predicted state of the weak current equipment, a topology propagation lag alignment operation is performed to generate lag aligned topology neighbor state data, and then a topology corrected predicted state is generated through the topology neighbor state correction layer. The predicted residuals of weak current equipment are calculated based on the reliable state frames and topology correction prediction states of weak current equipment, and a topology lag adjacency residual index table and a residual direction diagnostic sample pool are constructed. Based on the prediction residuals of weak current equipment, the index table of topological lag adjacency residuals and the sample pool of residual direction diagnosis, a topological assembly residual vector is generated. The topological assembly residual vector is then input into the RS-Hash algorithm to generate a direction-separated hash bucket count. The trusted count correction is performed on the direction separation hash bucket count to obtain the trusted correction residual abnormal score, and the fault diagnosis results of building low-voltage equipment are generated.

2. The method for fault diagnosis of building low-voltage electrical equipment based on deep learning according to claim 1, characterized in that, The generated multi-source operation data of building low-voltage electrical equipment includes: The system collects operational monitoring data from building low-voltage electrical equipment and simultaneously records equipment identification, equipment type, collection time, and arrival time. The operational monitoring data includes online status, communication status, port status, control feedback status, and service status. Read alarm and maintenance data from the alarm record database and maintenance record database. The alarm and maintenance data includes alarm type, alarm time, recovery status, maintenance record and fault record. Read the topology data from the building's low-voltage topology configuration table. The topology data includes equipment identification, equipment type, network connection relationship, power supply connection relationship, control linkage relationship, and signal transmission relationship. By associating operation monitoring data, alarm maintenance data, and topology data according to equipment identifiers, multi-source operation data of building low-voltage electrical equipment is generated.

3. The method for fault diagnosis of building low-voltage electrical equipment based on deep learning according to claim 1, characterized in that, The process of obtaining the trusted state frame of the low-voltage equipment includes: Read the operation monitoring data, alarm maintenance data and topology data from the multi-source operation data of building low-voltage equipment, and extract the equipment identifier, equipment type, acquisition time, arrival time, status field name, status field value, alarm type and alarm time; Standardize the field format for status field names and values, and generate a unified diagnostic time based on the data acquisition time. Based on the unified diagnosis time, arrival time, status field name, status field value, alarm type and alarm time, generate field missing status, delayed reporting status, missing data completion status and duplicate alarm status. Based on the device identifier, the status field value at the unified diagnostic time is bound to the topology data to generate a low-voltage equipment status record. Data credibility is generated based on field missing status, delayed reporting status, missing data completion status, and repeated alarm status. Operating condition prototype markers, key field activation markers, and topology propagation association markers are generated based on device type, status field name, topology data, and unified diagnostic time. The weak current device status record, data credibility, operating condition prototype markers, key field activation markers, and topology propagation association markers are combined to generate a reliable status frame for the weak current device.

4. The method for fault diagnosis of building low-voltage electrical equipment based on deep learning according to claim 1, characterized in that, The generation of the purification low-voltage electrical equipment state sequence includes: Read the device identifier, device type, unified diagnosis time, status field value, field missing status, delayed reporting status, missing data completion status, repeated alarm status, data credibility and topology propagation association marker from the trusted status frame of the weak current equipment, and form a historical status sequence according to the device identifier and unified diagnosis time. The historical status sequence is divided into candidate historical windows according to the continuous and unified diagnosis time. The number of low-confidence fields, missing fields, delayed reporting fields and repeated alarms in the candidate historical windows are counted to generate window status records. Based on window state records and topology propagation association markers, fault pollution states are identified, and candidate historical windows that meet the pollution determination criteria are marked as fault pollution windows, generating pollution window markers. Pollution isolation weights are generated based on pollution window markers, and the status field values ​​in the candidate historical windows are weighted and assembled to generate clean window status records. The status records of the purification window are summarized according to the unified diagnosis time, and a status sequence of the purification low-voltage equipment is generated.

5. The method for fault diagnosis of building low-voltage electrical equipment based on deep learning according to claim 1, characterized in that, The generation of the initial predicted state of the low-voltage equipment includes: An improved DLinear model was constructed, integrating the linear prediction layer of the prototype working condition, the sparse key field activation prediction layer, and the topological proximity state correction layer into the DLinear prediction process. Input the state sequence of the purification low-voltage equipment into the improved DLinear model, read the equipment identifier, equipment type, unified diagnosis time, state field name, state field value, data credibility, working condition prototype mark and key field activation mark, and generate the model input state sequence; The model input state sequence is input into the working condition prototype condition linear prediction layer. Based on the working condition prototype label, the working condition condition linear prediction is performed on the state field value to generate the working condition condition prediction state. The model input state sequence and the predicted state under operating conditions are input into the sparse key field activation prediction layer. Based on the equipment type, data credibility and key field activation marker, the key state fields are determined, sparse activation prediction is performed, and the key field predicted state is generated. The predicted state of operating conditions and the predicted state of key fields are aligned according to the unified diagnostic time to generate the initial predicted state of weak current equipment. The initial predicted state of low-voltage equipment is associated with the equipment identifier, equipment type, and unified diagnostic time to generate an initial predicted state record; The improved DLinear model was trained by using a combination of operating condition prediction error, key field prediction error, topological proximity state correction error, and initial prediction state error as the joint optimization objective. The model parameters of the linear prediction layer of the operating condition prototype, the sparse key field activation prediction layer, and the topological proximity state correction layer were iteratively updated. When the change in the joint loss value in five consecutive training rounds was less than 0.001, the improved DLinear model was considered to have completed convergence training.

6. The method for fault diagnosis of building low-voltage electrical equipment based on deep learning according to claim 1, characterized in that, The generation of topology correction prediction states includes: Read the device identifier, device type, network connection relationship, power supply connection relationship, control linkage relationship and signal transmission relationship from the topology data, and read the initial predicted state of weak current equipment, unified diagnosis time and state field name from the initial predicted state record to generate topology edge state record; Based on the topology edge state record, determine the upstream adjacent devices, same link adjacent devices, same power supply branch devices, and control linkage devices corresponding to the current device, and generate topology edge type tags according to network connection relationship, power supply connection relationship, control linkage relationship, and signal transmission relationship respectively; The propagation lag retrieval window is determined based on the topological edge type label, and a candidate propagation state sequence of adjacent devices is generated; The candidate propagation state sequence of adjacent devices is matched with the initial predicted state of the weak current equipment of the current device by matching the state field correspondence, the predicted state change direction, and the unified diagnosis time offset, and a topology propagation lag matching record is generated. Based on the topology propagation lag matching record, the propagation alignment time of the adjacent device corresponding to the current device is determined, and the initial predicted state of the adjacent device at the propagation alignment time, the topology edge type label and the unified diagnosis time offset value are combined to generate lag alignment topology proximity state data. The lagging aligned topology neighbor state data and the initial predicted state of the weak current equipment are input into the topology neighbor state correction layer to generate the neighbor state correction amount, and the initial predicted state of the weak current equipment is topologically corrected to generate the topology corrected predicted state.

7. The method for fault diagnosis of building low-voltage electrical equipment based on deep learning according to claim 1, characterized in that, The construction of the topological hysteresis adjacency residual index table and the residual direction diagnostic sample pool includes: Read the device identifier, device type, unified diagnosis time, status field name, status field value and data confidence from the trusted status frame of the weak current equipment, and read the prediction field name and prediction field value from the topology correction prediction status; Based on the equipment identification, unified diagnostic time, and status field name, the status field value and the prediction field value are used to calculate the predicted residual of the weak current equipment. The predicted residual of the weak current equipment includes the residual value, residual direction, residual field name, residual field category, and residual duration. Read the topology data and the lag-aligned topology neighbor status data, associate the current device prediction residual with the historical prediction residual of the neighboring device, calculate the residual magnitude matching value and topology propagation support, and generate a topology lag neighbor residual index table. Based on the residual values, residual directions, residual field categories, residual durations, and data reliability in the prediction residuals of low-voltage equipment, the residuals are classified and stored according to positive residuals, negative residuals, and zero-neighbor residuals, generating a residual direction diagnostic sample pool.

8. A method for fault diagnosis of building low-voltage electrical equipment based on deep learning according to claim 7, characterized in that, The topological hysteresis adjacency residual index table includes: The topology lag adjacency residual index table consists of the current device identifier, adjacent device identifier, current device type, adjacent device type, topology edge type, current unified diagnosis time, propagation alignment time, unified diagnosis time offset value, current device prediction residual, adjacent device historical prediction residual, residual field name, residual direction, residual magnitude matching value, and topology propagation support. An index record is created between the current device's predicted residual and the historical predicted residual of adjacent devices, based on the device identifier, topology edge type, propagation alignment time, residual field name, and residual direction. A dynamic residual index structure with propagation lag, residual direction, and topology propagation support is formed based on the index record.

9. A method for fault diagnosis of building low-voltage electrical equipment based on deep learning according to claim 1, characterized in that, The generation of direction-separated hash bucket counts includes: Read the prediction residuals of weak current equipment, the index table of topological lag adjacency residuals and the sample pool of residual direction diagnosis, and perform residual association according to equipment identifier, equipment type, unified diagnosis time, topological edge type and residual field name to generate topological residual association records. Based on the topological residual association record, the current device prediction residual, the historical prediction residual of adjacent devices, the residual direction, the residual field category and the topological propagation support are read, and the residuals are assembled according to the residual field name and the topological edge type to generate the topological assembly residual vector. The RS-Hash algorithm employs a random subspace selection mechanism for weak electrical equipment type constraints. Based on the topology assembly residual vector, equipment type, and residual field category, a candidate residual field set is determined, and random subspace selection is performed within the candidate residual field set to generate an equipment type constraint residual subspace. The RS-Hash algorithm employs a residual direction-separated hashing mechanism. Based on device type constraints, the residual subspace and residual direction are constrained, and hash mapping is performed according to positive residuals, negative residuals, and zero-neighbor residuals to generate a direction-separated hash bucket count.

10. A method for fault diagnosis of building low-voltage electrical equipment based on deep learning according to claim 1, characterized in that, The generated fault diagnosis results for building low-voltage electrical equipment include: Read the direction separation hash bucket count, weak current equipment trusted status frame, topology lag adjacency residual index table and residual direction diagnosis sample pool, and extract data credibility, field missing status, delayed reporting status, missing completion status, duplicate alarm status, topology propagation support, residual direction and residual field category. Based on data credibility, field missing status, delayed reporting status, missing data completion status, and duplicate alarm status, a credibility count correction coefficient is generated to correct the direction separation hash bucket count and generate a credibility correction residual abnormal score. The fault diagnosis results of building low-voltage equipment are generated based on the reliable correction residual abnormality score, topology propagation support, residual direction and residual field category.

Images