A machine learning-based energy retail terminal intelligent equipment diagnosis method

Abstract

The application discloses an energy retail terminal intelligent equipment diagnosis method based on machine learning, comprising the following steps: collecting multi-source diagnosis original data of energy retail terminal equipment; generating a candidate fault event sequence through association processing; performing weak supervision label correction on the candidate fault event sequence; constructing a fault propagation relationship according to component association relationship, control calling relationship, state change sequence relationship and alarm triggering relationship; training a propagation constraint weak supervision neural Hough diagnosis model based on a fault diagnosis training sample, and obtaining a candidate fault result; performing propagation path constraint and main fault and accompanying abnormality screening on the candidate fault result based on the fault propagation relationship; and generating a target root cause diagnosis result according to a root cause candidate result. The application realizes terminal equipment root cause diagnosis by adopting propagation constraint and weak supervision correction.

Description

Technical Field

[0001] This invention relates to the field of intelligent fault diagnosis and data processing technology, and in particular to a diagnostic method for intelligent energy retail terminal equipment based on machine learning. Background Technology

[0002] During operation, energy retail terminal equipment continuously generates operational status data, alarm event data, maintenance work order data, component replacement data, and recovery operation data. Existing technologies typically rely on operational parameter monitoring, alarm rule matching, or historical work order retrieval to identify equipment anomalies and locate faults. Some solutions also employ machine learning models to classify and diagnose equipment status in order to improve fault detection efficiency and operation and maintenance efficiency.

[0003] However, existing technologies mostly rely on single data sources or static features for diagnosis, failing to adequately utilize the correlations between multi-source heterogeneous operation and maintenance data, making it difficult to form a complete fault event evolution process. Furthermore, labeling bias exists between historical maintenance records and alarm records, which can easily affect the quality of diagnostic samples when directly used for model training. On the other hand, existing solutions do not adequately characterize the propagation relationships between equipment components, control calls, the sequence of state changes, and alarm triggers, making it difficult to effectively distinguish between the primary fault and accompanying anomalies generated by its propagation. Therefore, there is still room for improvement in the accuracy of root cause localization and the consistency of diagnostic results.

[0004] Therefore, how to provide a machine learning-based diagnostic method for intelligent energy retail terminal equipment 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 machine learning-based intelligent equipment diagnostic method for energy retail terminals. This invention fully utilizes multi-source operation and maintenance data correlation analysis, weakly supervised label correction, fault propagation relationship modeling, and a propagation-constrained weakly supervised neural Hawkes diagnostic model to jointly process the operation status data, alarm event data, maintenance work order data, component replacement data, and recovery operation data generated during the operation of energy retail terminal equipment. This enables the construction of candidate fault events, generation of fault diagnosis training samples, identification of candidate faults, screening of main faults and associated anomalies, and output of target root cause diagnosis results.

[0006] A diagnostic method for intelligent energy retail terminal devices based on machine learning according to an embodiment of the present invention includes the following steps: Collect multi-source diagnostic raw data from energy retail terminal equipment to obtain a multi-source diagnostic raw data set; Based on the multi-source diagnostic raw data set, correlation processing is performed to generate a candidate fault event sequence; Weakly supervised label correction is performed on candidate fault event sequences to generate fault diagnosis training samples; Based on the component relationships, control call relationships, state change sequence relationships, and alarm trigger relationships of energy retail terminal equipment, a fault propagation relationship is constructed; The model is trained based on the fault diagnosis training samples and the propagation constraint weakly supervised neural Hawkes diagnostic model. The candidate fault event sequences are then processed by event embedding encoding, continuous time state memory update, and event condition strength calculation combined with fault propagation relationship to obtain the candidate fault results. Based on the fault propagation relationship, propagation path constraints and screening of primary faults and associated anomalies are performed on candidate fault results to obtain root cause candidate results; Based on the root cause candidate results, the consistency of the corresponding propagation path, the consistency of the event sequence, and the consistency of the recovery results, the target root cause diagnosis results are generated.

[0007] Optionally, the multi-source diagnostic raw data includes operating status data, alarm event data, maintenance work order data, component replacement data, and recovery operation data.

[0008] Optionally, the step of performing correlation processing based on the multi-source diagnostic raw data set to generate a candidate fault event sequence specifically includes: The multi-source diagnostic raw data set is grouped to obtain a subset of the device raw data. The raw data subset is then sorted to form a time-ordered event stream. In a time-ordered event stream, an initial associated time window is constructed using alarm event records as the trigger center. Window extraction processing is performed on the time-ordered event stream to obtain initial associated event fragments. Association determination processing is then performed on the initial associated event fragments to obtain fault associated event fragments. Perform event boundary determination processing on fault-related event fragments to obtain a single candidate fault event; For multiple single candidate fault events corresponding to the same unified equipment identifier, perform segment overlap detection and adjacent segment splicing processing to obtain continuous candidate fault events; The consecutive candidate fault events are sorted by their start time to obtain a sequence of candidate fault events. A summary process is performed on the candidate fault event sequences corresponding to each unified equipment identifier to obtain a set of candidate fault event sequences.

[0009] Optionally, the step of performing weakly supervised label correction on the candidate fault event sequence to generate fault diagnosis training samples specifically includes: Perform candidate event expansion processing on the candidate fault event sequence set to form a candidate fault event set; Alignment processing is performed on the candidate fault event set to obtain the candidate work order set, the candidate component replacement set, and the candidate operation recovery set. Perform fault recurrence detection processing on the candidate fault event set to obtain fault recurrence information; Weak supervision label generation processing is performed on the candidate fault event set to obtain initial weak supervision labels, and label correction processing is performed on the initial weak supervision labels to obtain corrected weak supervision labels; The corrected weakly supervised labels are subjected to sample screening to obtain fault diagnosis training samples.

[0010] Optionally, the construction of fault propagation relationships based on component associations, control call relationships, state change sequence relationships, and alarm trigger relationships of energy retail terminal equipment specifically includes: Perform associated object extraction processing on energy retail terminal equipment to obtain a set of equipment components, a set of operating statuses, and a set of alarm events; Based on the set of equipment components, component association relationships are constructed to obtain a set of component association relationships; Identify the control signal input relationships, control command output relationships, and control link transmission relationships between the set of equipment components and the set of operating states, and construct a set of control call relationships; Based on the collection time sequence corresponding to the running status records in the time-ordered event stream, the order of changes of each running status item is identified, and a set of relationships between the order of state changes is constructed. Match the abnormal states corresponding to each running status item with the alarm occurrence records corresponding to each alarm event item to construct an alarm trigger relationship set; The set of component association relationships, the set of control call relationships, the set of state change sequence relationships, and the set of alarm trigger relationships are mapped to propagation edges with propagation directions to obtain the set of fault propagation edges; The weight of each propagation edge is determined based on the set of fault propagation edges, thus obtaining the weighted fault propagation relationship. The set of equipment components, the set of operating states, and the set of alarm events are used as the propagation node set, and the set of fault propagation edges are used as the propagation edge set. The propagation edge weights corresponding to each propagation edge are used as the propagation edge weight set to construct a fault propagation graph, which is then used as the fault propagation relationship between equipment components, operating status, and alarm events.

[0011] Optionally, the step of training a weakly supervised neural Hawkes diagnostic model with propagation constraints based on fault diagnosis training samples, and performing event embedding encoding, continuous-time state memory updates, and event condition strength calculations combining fault propagation relationships on candidate fault event sequences to obtain candidate fault results specifically includes: The training sample encoding process is performed on the fault diagnosis training samples to obtain the training event sequence input; A propagation-constrained weakly supervised neural Hawkes diagnostic model is constructed based on the training event sequence input, and the parameters of the propagation-constrained weakly supervised neural Hawkes diagnostic model are initialized. Each candidate fault event in the training event sequence is input into the event embedding encoding layer, and event embedding encoding processing is performed to obtain the event embedding representation; Embed the event into the continuous-time state memory update layer, perform continuous-time state memory update processing, and obtain the continuous-time hidden state; In the continuous-time state memory update layer, time decay processing is performed on the continuous-time hidden state corresponding to the previous candidate fault event to obtain the historical hidden state after time decay. The continuous-time hidden state is input into the propagation constraint fusion layer, and propagation constraint fusion processing is performed in combination with the fault propagation relationship to obtain the propagation constraint state representation; The propagation constraint state representation is input into the event condition strength calculation layer, and the event condition strength calculation process is performed to obtain the event condition strength; The event condition intensity is input into the result output layer, and the candidate fault score generation process is performed to obtain the candidate fault score. The candidate fault type with the highest candidate fault score is taken as the candidate fault result of the corresponding candidate fault event. Based on the corrected weak supervision labels corresponding to the candidate fault results and the fault diagnosis training samples, the propagation constraint weak supervision neural Hawkes diagnostic model is trained to obtain the trained propagation constraint weak supervision neural Hawkes diagnostic model. Collect the current operating status data of the device to be diagnosed, and combine it with the alarm event data of the device to be diagnosed to construct a sequence of candidate fault events to be diagnosed; The sequence of candidate fault events to be diagnosed is input into the trained propagation-constrained weakly supervised neural Hawkes diagnostic model to obtain the candidate fault results.

[0012] Optionally, the step of inputting the continuous-time hidden state into the propagation constraint fusion layer and performing propagation constraint fusion processing in conjunction with the fault propagation relationship to obtain the propagation constraint state representation specifically includes: Based on the multi-source diagnostic raw data set corresponding to the current candidate fault event, the propagation node set in the fault propagation relationship is mapped to the propagation node set, and the event-related node set is obtained. Based on the node connection relationship, propagation direction, and propagation edge weight of the event-related node set in the fault propagation relationship, a local propagation subgraph is extracted to obtain local propagation constraint information; Perform local propagation constraint vector generation processing on the local propagation subgraph to obtain the local propagation constraint vector; Perform weight mapping on the local propagation constraint vector to obtain the propagation constraint mapping result; Perform state mapping processing on the continuous-time hidden states to obtain the continuous-time hidden state mapping results; Perform fusion calculation on the continuous-time hidden state mapping results, propagation constraint mapping results, and propagation constraint fusion bias vector to obtain intermediate fusion results; The intermediate fusion results are processed by nonlinear mapping to obtain the propagation constraint state representation; According to the order of the start time of each candidate fault event in the training event sequence, the propagation constraint state representations corresponding to each candidate fault event are arranged to obtain the propagation constraint state representation sequence.

[0013] Optionally, the step of inputting the propagation constraint state representation into the event condition strength calculation layer and performing event condition strength calculation processing to obtain the event condition strength specifically includes: The propagation constraint state is represented by the input event condition strength calculation layer, and a candidate fault type set is extracted; For any candidate fault type, read the condition strength parameter vector and condition strength bias term corresponding to the candidate fault type, and match them with the propagation constraint state representation to obtain the condition strength calculation input; Perform linear mapping processing on the propagation constraint state representation corresponding to any candidate fault event to obtain the linear strength value; Conditional intensity activation processing is applied to the linear intensity value to obtain the event conditional intensity.

[0014] Optionally, the step of performing propagation path constraints and screening of primary faults and associated anomalies on candidate fault results based on fault propagation relationships to obtain root cause candidate results specifically includes: Mapping the candidate fault results to the fault propagation relationship yields a set of candidate propagation nodes; Based on the propagation direction, propagation edge type, and propagation edge weight of each propagation edge in the fault propagation relationship, a propagation path extraction process is performed on the candidate propagation node set to obtain a candidate propagation path set. Perform propagation path constraint processing on the candidate propagation path set to obtain the constrained propagation path set; Based on the constrained propagation path set, the path support is calculated for each candidate fault result to obtain the path support. Based on the candidate fault scores, path support, and path location relationships, a primary fault and associated anomaly screening process is performed to obtain the primary fault candidate results and associated anomaly candidate results. Based on the candidate results of the main fault and the candidate results of the associated anomalies, the candidate results of the root cause are obtained.

[0015] Optionally, generating the target root cause diagnosis result based on the root cause candidate results, the consistency of the corresponding propagation path, the consistency of the event sequence, and the consistency of the recovery result specifically includes: Based on the root cause candidate results and candidate fault results, a root cause diagnosis and determination dataset is constructed. Based on the constrained propagation path, a propagation path consistency calculation is performed to obtain the propagation path consistency result. Based on the primary fault candidate node, the accompanying anomaly candidate node and the corresponding event occurrence time, perform event timing consistency calculation to obtain the event timing consistency result; Based on the primary fault candidate node, the accompanying anomaly candidate node and the corresponding recovery status flag, perform recovery result consistency calculation processing to obtain the recovery result consistency result; Based on the consistency values ​​of the propagation path, the event sequence, and the recovery result, a comprehensive evaluation of the root cause candidate results is performed to obtain the root cause diagnosis confidence value. Based on the root cause diagnosis confidence value, the target root cause diagnosis result is obtained.

[0016] The beneficial effects of this invention are: This invention unifies and correlates operational status data, alarm event data, maintenance work order data, component replacement data, and recovery operation data to construct a candidate fault event sequence, and generates fault diagnosis training samples based on this sequence. By incorporating multi-source operation and maintenance data into the same diagnostic link, this invention improves the completeness of fault event representation, ensuring that the formation, processing, and recovery processes of equipment anomalies correspond within the same data framework, thereby addressing the problem of fragmented fault information under single data source conditions.

[0017] This invention introduces a weakly supervised label correction mechanism during the training phase. It utilizes maintenance work order data, component replacement data, recovery operation data, and fault recurrence information to correct candidate fault event sequences, forming fault diagnosis training samples. This approach establishes a correspondence between the training samples and actual operation and maintenance results, reducing the impact of false alarm triggers, recording deviations, or label inconsistencies. This improves the quality of the training samples and enhances the subsequent diagnostic model's ability to identify real fault modes.

[0018] This invention further constructs a fault propagation relationship between equipment components, operating states, and alarm events based on the component associations, control call relationships, state change sequence relationships, and alarm trigger relationships of energy retail terminal equipment. It then combines a propagation-constrained weakly supervised neural Hawkes diagnostic model to perform event embedding encoding, continuous-time state memory updates, and event condition strength calculations on candidate fault event sequences. By introducing the fault propagation relationship into the diagnostic process, this invention can represent the fault propagation path under continuous-time evolution conditions. This allows candidate fault results to reflect not only the occurrence of anomalies but also the propagation relationships between anomalies, thereby improving the temporal and propagation consistency of root cause identification in complex scenarios.

[0019] In the output phase, this invention performs propagation path constraints and primary fault and associated anomaly screening on candidate fault results based on fault propagation relationships, and generates target root cause diagnosis results by combining propagation path consistency, event sequence consistency, and recovery result consistency. In this way, even when multiple related anomalies occur simultaneously, the primary fault and associated anomalies can be distinguished, ensuring that the final output results correspond to the equipment fault propagation process and recovery status, thus improving the stability and usability of root cause localization results. Attached Figure Description

[0020] 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 machine learning-based intelligent device diagnostic method for energy retail terminals proposed in this invention. Figure 2 This is a schematic diagram illustrating the fault propagation relationship construction process in the intelligent equipment diagnosis method for energy retail terminals based on machine learning proposed in this invention. Figure 3 This is a schematic diagram of the propagation-constrained weakly supervised neural Hawkes diagnostic model in the intelligent energy retail terminal diagnostic method based on machine learning proposed in this invention. Detailed Implementation

[0021] 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.

[0022] refer to Figures 1-3 A machine learning-based diagnostic method for intelligent energy retail terminal devices includes the following steps: Collect multi-source diagnostic raw data from energy retail terminal equipment to obtain a multi-source diagnostic raw data set; Based on the multi-source diagnostic raw data set, correlation processing is performed to generate a candidate fault event sequence; Weakly supervised label correction is performed on candidate fault event sequences to generate fault diagnosis training samples; Based on the component relationships, control call relationships, state change sequence relationships, and alarm trigger relationships of energy retail terminal equipment, a fault propagation relationship is constructed; The model is trained based on the fault diagnosis training samples and the propagation constraint weakly supervised neural Hawkes diagnostic model. The candidate fault event sequences are then processed by event embedding encoding, continuous time state memory update, and event condition strength calculation combined with fault propagation relationship to obtain the candidate fault results. Based on the fault propagation relationship, propagation path constraints and screening of primary faults and associated anomalies are performed on candidate fault results to obtain root cause candidate results; Based on the root cause candidate results, the consistency of the corresponding propagation path, the consistency of the event sequence, and the consistency of the recovery results, the target root cause diagnosis results are generated.

[0023] In this embodiment, the collection of multi-source diagnostic raw data from energy retail terminal equipment to obtain the multi-source diagnostic raw data set specifically includes: Fields are extracted from the running status data to obtain a set of running status records. Each running status record in the set includes a device identifier, collection time, and a running status feature vector at the corresponding time. The running status feature vector is composed of multiple running status feature values. Fields are extracted from the alarm event data to obtain an alarm event record set. Each alarm event record in the alarm event record set includes the device identifier, alarm occurrence time, alarm code, and alarm level. Fields are extracted from the maintenance work order data to obtain a set of maintenance work order records. Each maintenance work order record in the set includes equipment identifier, work order start time, work order end time, work order handling record, and work order fault conclusion marker. Fields are extracted from the component replacement data to obtain a set of component replacement records. Each component replacement record in the set includes the equipment identifier, replacement time, replacement component identifier, and replacement operation flag. Fields are extracted from the recovery operation data to obtain a recovery operation record set. Each recovery operation record in the recovery operation record set includes the device identifier, recovery time, and recovery status flag. The system performs missing field removal, duplicate record merging, and time format standardization on the operation status data, alarm event data, maintenance work order data, component replacement data, and recovery operation data to obtain standardized operation status data, standardized alarm event data, standardized maintenance work order data, standardized component replacement data, and standardized recovery operation data. According to the unified equipment identification mapping rules, the standardized operating status data, standardized alarm event data, standardized maintenance work order data, standardized component replacement data, and standardized recovery operation data are processed to normalize the equipment identification to obtain a unified equipment identification. The unified equipment identification mapping rules include extracting the original equipment identification information from operating status data, alarm event data, maintenance work order data, component replacement data, and recovery operation data, and matching it according to the preset equipment master index table. The rules also standardize the encoding prefix, separator, capitalization, site abbreviation, and historical version number, and merge multiple original equipment identifications that match the same energy retail terminal equipment into the same unified equipment identification. When multiple candidate mapping results exist, conflict resolution is performed based on site identifier, device type, time interval overlap, and component affiliation; when the original device identifier is missing, reverse completion is performed based on alarm source address, work order registration object, component installation location, and recovery record associated object; The standardized operational status data, standardized alarm event data, standardized maintenance work order data, standardized component replacement data, and standardized recovery operation data are sorted in ascending order according to a unified timeline. The sorted standardized operational status data, standardized alarm event data, standardized maintenance work order data, standardized component replacement data, and standardized recovery operation data are then merged to obtain a multi-source diagnostic raw data set.

[0024] In this embodiment, the process of associating the original multi-source diagnostic data set to generate a candidate fault event sequence specifically includes: The multi-source diagnostic raw data set is grouped and processed, and the operation status record, alarm event record, maintenance work order record, component replacement record and operation recovery record are classified according to the unified equipment identifier to obtain the equipment raw data subset corresponding to each energy retail terminal equipment; For each device's raw data subset, sorting is performed. The operating status records, alarm event records, maintenance work order records, component replacement records, and recovery operation records are sorted in ascending order according to the event occurrence time to obtain the time-ordered event stream of the corresponding energy retail terminal device. In a time-ordered event stream, an initial associated time window is constructed using alarm event records as the trigger center. The initial associated time window is determined by the preset forward time window length before the alarm occurrence time and the preset backward time window length after the alarm occurrence time. Perform window extraction processing on the time-ordered event stream to extract the running status records, alarm event records, maintenance work order records, component replacement records, and recovery operation records whose event occurrence time falls within the initial associated time window, and obtain the initial associated event fragments corresponding to the alarm event records; Perform association determination processing on the initial associated event fragments to obtain fault associated event fragments; The association determination process includes filtering the operation status records, alarm event records, maintenance work order records, component replacement records, and operation recovery records with consistent unified equipment identification; matching the filtered records according to time adjacency conditions; and establishing association relationships for records that meet the time adjacency conditions according to the handling association conditions. Among them, the time adjacency condition is that the absolute value of the difference between the time corresponding to any preceding event record and the time corresponding to any subsequent event record is less than or equal to the preset association time threshold. The handling association conditions include work order handling association between alarm event records and maintenance work order records, component maintenance association between maintenance work order records and component replacement records, and recovery handling association between maintenance work order records and recovery operation records. Perform event boundary determination processing on fault-related event fragments to obtain a single candidate fault event; The event boundary determination process includes taking the earliest event record time in the fault-related event segment as the event start time and taking the latest event record time in the fault-related event segment as the event end time. Merge the associated operational status records, associated alarm event records, associated maintenance work order records, associated component replacement records, and associated recovery operation records between the event start time and the event end time; For multiple single candidate fault events corresponding to the same unified equipment identifier, perform segment overlap detection and adjacent segment splicing processing to obtain continuous candidate fault events; The fragment overlap detection and adjacent fragment splicing process includes determining whether the difference between the start time of the latter event and the end time of the former event between two adjacent single candidate fault events is less than or equal to a preset fragment splicing time threshold; if the determination result is yes, the two adjacent single candidate fault events are merged into consecutive candidate fault events. The consecutive candidate fault events are sorted by their start time to obtain a sequence of candidate fault events under the same unified device identifier. A summary process is performed on the candidate fault event sequences corresponding to each unified equipment identifier to obtain a set of candidate fault event sequences.

[0025] In this embodiment, performing weakly supervised label correction on the candidate fault event sequence to generate fault diagnosis training samples specifically includes: The candidate event expansion process is performed on the candidate fault event sequence set. The candidate fault event sequence corresponding to each unified device identifier is read one by one, and a single candidate fault event is extracted from each candidate fault event sequence to form a candidate fault event set. Perform work order alignment processing on the candidate fault event set to obtain the candidate work order set; The work order alignment process includes extracting the start time and end time of each candidate fault event, matching the maintenance work order records in the maintenance work order data according to the unified equipment identifier, retaining the maintenance work order records whose start time and end time fall within the work order association time range of the corresponding candidate fault event, and obtaining the candidate work order set corresponding to each candidate fault event. Perform component replacement alignment processing on the candidate fault event set to obtain the candidate component replacement set; The component replacement alignment process includes extracting the start time and end time of each candidate fault event, matching the component replacement records in the component replacement data according to the unified equipment identifier, retaining the component replacement records whose replacement time falls within the component replacement association time range of the corresponding candidate fault event, and obtaining the candidate component replacement set corresponding to each candidate fault event. Perform recovery run alignment processing on the candidate fault event set to obtain the candidate recovery run set; The recovery operation alignment process includes extracting the event end time of each candidate fault event, matching the recovery operation records in the recovery operation data according to the unified device identifier, retaining the recovery operation records whose recovery time falls within the recovery operation association time range of the corresponding candidate fault event, and obtaining the candidate recovery operation set corresponding to each candidate fault event; Perform fault recurrence detection processing on the candidate fault event set to obtain fault recurrence information; The fault recurrence detection and processing includes sequentially scanning the sequence of candidate fault events corresponding to the same unified equipment identifier, and determining whether there are subsequent candidate fault events that fall within the recurrence judgment time range after the current candidate fault event; When there are subsequent candidate fault events and the alarm code set, component replacement record set, or maintenance work order fault conclusion mark of the current candidate fault event and the subsequent candidate fault events meet the preset similarity conditions, a fault recurrence mark is generated; when the above conditions are not met, a fault recurrence mark is generated. Weak supervision label generation processing is performed on the candidate fault event set to obtain initial weak supervision labels; The weak supervision label generation process includes generating work order support values ​​based on the candidate work order set, generating replacement support values ​​based on the candidate component replacement set, generating recovery support values ​​based on the candidate recovery operation set, and generating recurrence penalty values ​​based on the fault recurrence marker. The work order support value, replacement support value, and recovery support value are weighted and summed, and the penalty term corresponding to the recurrence penalty value is deducted to obtain the label confidence of the candidate fault event. The label confidence satisfies the following: the label confidence is the sum of the product of the work order support value and the work order support weight, the product of the component replacement support value and the component replacement support weight, and the product of the recovery support value and the recovery support weight, and then the product of the fault recurrence label and the recurrence penalty weight is subtracted. Work order support weight, component replacement support weight, recovery support weight, and recurrence penalty weight are used to represent the degree of influence of maintenance work order data, component replacement data, recovery operation data, and fault recurrence information on weakly supervised label generation, respectively. Perform label correction processing on the initial weakly supervised labels to obtain corrected weakly supervised labels; The label correction process includes classifying candidate fault events by label based on the confidence interval to which the label confidence level belongs. When the label confidence level falls into the positive sample confidence interval, a positive sample label is generated; when the label confidence level falls into the negative sample confidence interval, a negative sample label is generated; and when the label confidence level falls into the uncertain confidence interval, an uncertain label is generated. The corrected weak supervision labels are subjected to sample screening. Candidate fault events with positive or negative labels after correction are retained, while candidate fault events with uncertain labels after correction are removed. The retained candidate fault events are then combined with their corresponding corrected weak supervision labels to obtain fault diagnosis training samples.

[0026] In this embodiment, the fault propagation relationship is constructed based on the component associations, control call relationships, state change sequence relationships, and alarm trigger relationships of the energy retail terminal equipment, specifically including: Perform associated object extraction processing on energy retail terminal equipment to obtain a set of equipment components, a set of operating statuses, and a set of alarm events; The associated object extraction process includes extracting operating status items from operating status data to form an operating status set; extracting alarm event items from alarm event data to form an alarm event set; and extracting equipment component items from the equipment structure information, maintenance work order data, and component replacement data of energy retail terminal equipment to form an equipment component set. Perform component association relationship construction processing on the set of equipment components to obtain a set of component association relationships; The component association relationship construction process includes identifying the connection relationships, assembly relationships, and upstream and downstream transmission relationships between various equipment component items in the equipment component set; establishing component association edges between equipment component items with component association relationships to form a component association relationship set; The existence of component association relationships between each equipment component item is represented as a component association matrix. When there is a component association relationship between any two equipment component items, the corresponding matrix element takes the value of one, and when there is no component association relationship between any two equipment component items, the corresponding matrix element takes the value of zero. Identify the control signal input relationships, control command output relationships, and control link transmission relationships between the set of equipment components and the set of operating states, and construct a set of control call relationships; The existence of a control call relationship between each device component item and each operating status item is represented as a control call matrix. When there is a control call relationship between any device component item and any operating status item, the corresponding matrix element takes the value of one, and when there is no control call relationship between any device component item and any operating status item, the corresponding matrix element takes the value of zero. Based on the collection time sequence corresponding to the running status records in the time-ordered event stream, the order of changes of each running status item is identified, and a set of relationships between the order of state changes is constructed. The existence of a state change sequence relationship between each running state item is represented as a state change sequence matrix. When the change of any running state item occurs before the change of another running state item, the corresponding matrix element takes the value of one. When there is no state change sequence relationship between any running state item and another running state item, the corresponding matrix element takes the value of zero. Match the abnormal states corresponding to each running status item with the alarm occurrence records corresponding to each alarm event item to construct an alarm trigger relationship set; The existence of an alarm triggering relationship between each running status item and each alarm event item is represented as an alarm triggering matrix. When there is an alarm triggering relationship between any running status item and any alarm event item, the corresponding matrix element takes the value of one, and when there is no alarm triggering relationship between any running status item and any alarm event item, the corresponding matrix element takes the value of zero. The set of component association relationships, control call relationships, state change sequence relationships, and alarm trigger relationships are uniformly mapped to propagation edges with propagation directions to obtain the fault propagation edge set. Each propagation edge consists of a propagation starting node, a propagation destination node, a propagation edge weight, and a propagation edge type. The propagation edge type is used to distinguish component association relationships, control call relationships, state change sequence relationships, and alarm trigger relationships. The weight of each propagation edge is determined based on the set of fault propagation edges, thus obtaining the weighted fault propagation relationship. Specifically, this includes: the support values ​​for the co-occurrence of the statistical propagation initiation node and the propagation arrival node in the candidate fault event set, the support values ​​for their sequential occurrence, and the associated support values ​​for the corresponding maintenance work order data, component replacement data, and recovery operation data; The propagation edge weights are calculated based on co-occurrence support values, sequential occurrence support values, and associated support values. The propagation edge weights are the sum of the products of co-occurrence support values ​​and co-occurrence support weights, sequential occurrence support values ​​and sequential occurrence support weights, and associated support values ​​and associated support weights. Co-occurrence support weight, sequential occurrence support weight, and associated support weight are used to represent the degree of influence of co-occurrence support value, sequential occurrence support value, and associated support value on the calculation of propagation edge weights, respectively. The set of equipment components, the set of operating states, and the set of alarm events are used as the propagation node set, and the set of fault propagation edges are used as the propagation edge set. The propagation edge weights corresponding to each propagation edge are used as the propagation edge weight set to construct a fault propagation graph, which is then used as the fault propagation relationship between equipment components, operating status, and alarm events.

[0027] In this embodiment, a weakly supervised neural Hawkes diagnostic model with propagation constraints is trained based on fault diagnosis training samples. The candidate fault event sequences are then subjected to event embedding encoding, continuous-time state memory updates, and event condition strength calculations incorporating fault propagation relationships to obtain the following specific candidate fault results: The training sample encoding process is performed on the fault diagnosis training samples to obtain the training event sequence input; The training sample encoding process includes extracting the set of operating status records, alarm event records, maintenance work order records, component replacement records, and recovery operation records corresponding to each candidate fault event in the fault diagnosis training sample, sorting each candidate fault event in ascending order according to the event start time to form a training event sequence, and representing each candidate fault event as an event feature vector; A propagation-constrained weakly supervised neural Hawkes diagnostic model is constructed based on the training event sequence input, and the parameters of the propagation-constrained weakly supervised neural Hawkes diagnostic model are initialized. The propagation-constrained weakly supervised neural Hawkes diagnostic model includes an event embedding encoding layer, a continuous-time state memory update layer, a propagation constraint fusion layer, an event condition strength calculation layer, and a result output layer. Each candidate fault event in the training event sequence is input into the event embedding encoding layer, and event embedding encoding processing is performed to obtain the event embedding representation corresponding to each candidate fault event. The event embedding corresponding to any candidate fault event is represented by multiplying the event feature vector corresponding to the candidate fault event with the event embedding weight matrix, adding the event embedding bias vector, and then processing it through a nonlinear mapping function. The event embedding weight matrix is ​​used to represent the mapping relationship from the event feature vector to the event embedding representation, and the event embedding bias vector is used to represent the bias adjustment amount of the event embedding encoding. The event embedding representation corresponding to each candidate fault event is input into the continuous-time state memory update layer, and continuous-time state memory update processing is performed to obtain the continuous-time hidden state corresponding to each candidate fault event. The time interval between the current candidate fault event and the previous candidate fault event is the event occurrence time of the current candidate fault event minus the event occurrence time of the previous candidate fault event. The continuous time hidden state corresponding to the current candidate fault event is obtained by multiplying the event embedding representation corresponding to the current candidate fault event with the current event embedding weight matrix, multiplying the historical hidden state of the previous candidate fault event after time decay with the historical state transfer weight matrix, multiplying the event time interval with the time interval mapping weight matrix, summing the above results with the state update bias vector, and then processing it with the hyperbolic tangent function. Among them, the current event embedding weight matrix is ​​used to represent the influence relationship of the current candidate fault event on the continuous time hidden state, the historical state transmission weight matrix is ​​used to represent the transmission relationship of the historical hidden state to the current continuous time hidden state, the time interval mapping weight matrix is ​​used to represent the influence relationship of the event time interval on the current continuous time hidden state, and the state update bias vector is used to represent the bias adjustment amount of the continuous time state memory update. In the continuous-time state memory update layer, time decay processing is performed on the continuous-time hidden state corresponding to the previous candidate fault event to obtain the historical hidden state after time decay. Among them, the historical hidden state after time decay is obtained by multiplying the continuous time hidden state corresponding to the previous candidate fault event with the time decay term element by element. The time decay term is the negative value of the product of the time decay coefficient and the event time interval as the natural exponential function value of the exponent. The time decay factor is used to represent the degree to which hidden historical states decay as the time interval between events increases; The continuous-time hidden state is input into the propagation constraint fusion layer, and propagation constraint fusion processing is performed in combination with the fault propagation relationship to obtain the propagation constraint state representation; The propagation constraint state representation is input into the event condition strength calculation layer, and the event condition strength calculation process is performed to obtain the event condition strength; Input the event condition intensity of each candidate fault event corresponding to each candidate fault type into the result output layer, perform candidate fault score generation processing, and obtain the candidate fault score of each candidate fault event corresponding to each candidate fault type. The candidate fault score for any candidate fault event belonging to any candidate fault type is the event condition strength of the candidate fault event corresponding to the candidate fault type divided by the sum of the event condition strengths of all candidate fault types corresponding to the candidate fault event; and the candidate fault type with the highest candidate fault score is taken as the candidate fault result of the corresponding candidate fault event. Based on the corrected weak supervision labels corresponding to the candidate fault results and the fault diagnosis training samples, the propagation constraint weak supervision neural Hawkes diagnostic model is trained to obtain the trained propagation constraint weak supervision neural Hawkes diagnostic model. Specifically, this includes: constructing a weak supervision label loss term based on the corrected weak supervision label, constructing a condition strength loss term based on the event condition strength, constructing a propagation constraint loss term based on the fault propagation relationship, and jointly optimizing the weak supervision label loss term, the condition strength loss term, and the propagation constraint loss term; The joint training loss function is the sum of the product of the weakly supervised label loss term, the conditional strength loss term and the corresponding conditional strength loss weight, and the propagation constraint loss term and the corresponding propagation constraint loss weight. The conditional strength loss weight is used to represent the influence of the conditional strength loss term in the joint training loss function, and the propagation constraint loss weight is used to represent the influence of the propagation constraint loss term in the joint training loss function. Collect the current operating status data of the device to be diagnosed, and combine it with the alarm event data of the device to be diagnosed to construct a sequence of candidate fault events to be diagnosed; The sequence of candidate fault events to be diagnosed is input into the trained propagation constraint weakly supervised neural Hawkes diagnostic model, and then processed sequentially through the event embedding encoding layer, the continuous time state memory update layer, the propagation constraint fusion layer, the event condition strength calculation layer, and the result output layer to obtain the candidate fault results.

[0028] In this embodiment, the continuous-time hidden state is input into the propagation constraint fusion layer, and propagation constraint fusion processing is performed in conjunction with the fault propagation relationship to obtain the propagation constraint state representation, specifically including: Map the set of multi-source diagnostic raw data corresponding to the current candidate fault event to the set of propagation nodes in the fault propagation relationship to obtain the set of event-related nodes corresponding to the current candidate fault event. Based on the node connection relationship, propagation direction, and propagation edge weight of the event-associated node set in the fault propagation relationship, a local propagation subgraph corresponding to the current candidate fault event is extracted to obtain local propagation constraint information; wherein, the local propagation constraint information includes the propagation starting node, propagation destination node, propagation edge type, and propagation edge weight corresponding to the event-associated node set; Perform local propagation constraint vector generation processing on the local propagation subgraph to obtain the local propagation constraint vector; The local propagation constraint vector generation process includes arranging each propagation edge in the local propagation subgraph in an ordered manner according to the propagation direction, generating edge constraint representations corresponding to each propagation edge according to the propagation edge weight, propagation edge type and node connection order, and aggregating each edge constraint representation to obtain the local propagation constraint vector. Perform weight mapping on the local propagation constraint vector to obtain the propagation constraint mapping result; The propagation constraint mapping result is obtained by multiplying the local propagation constraint vector by the propagation constraint mapping weight matrix; whereby the propagation constraint mapping weight matrix is ​​used to represent the mapping relationship from the local propagation constraint vector to the propagation constraint state representation; Perform state mapping processing on the continuous-time hidden state corresponding to the current candidate fault event to obtain the continuous-time hidden state mapping result; The continuous-time hidden state mapping result is obtained by multiplying the continuous-time hidden state corresponding to the current candidate fault event with the continuous-time hidden state mapping weight matrix; wherein, the continuous-time hidden state mapping weight matrix is ​​used to represent the mapping relationship from the continuous-time hidden state to the propagation constraint state representation; Perform fusion calculation on the continuous-time hidden state mapping results, propagation constraint mapping results, and propagation constraint fusion bias vector to obtain intermediate fusion results; The intermediate fusion result is the sum of the continuous-time hidden state mapping result, the propagation constraint mapping result, and the propagation constraint fusion bias vector; where the propagation constraint fusion bias vector is used to represent the bias adjustment amount of the propagation constraint fusion. A nonlinear mapping process is performed on the fusion intermediate results to obtain the propagation constraint state representation corresponding to the current candidate fault event; wherein, the propagation constraint state representation is the result after applying a hyperbolic tangent function to the fusion intermediate results; According to the order of the start time of each candidate fault event in the training event sequence, the propagation constraint state representations corresponding to each candidate fault event are arranged to obtain the propagation constraint state representation sequence.

[0029] In this embodiment, the propagation constraint state representation is input to the event condition strength calculation layer, and the event condition strength calculation process is performed to obtain the event condition strength, specifically including: The propagation constraint state representation corresponding to each candidate fault event is input into the event condition strength calculation layer, and the candidate fault type set is extracted. The candidate fault type set consists of all candidate fault types corresponding to the propagation constraint weakly supervised neural Hawkes diagnostic model. For any candidate fault type, read the condition strength parameter vector and condition strength bias term corresponding to the candidate fault type, and match the condition strength parameter vector with the propagation constraint state representation corresponding to each candidate fault event to obtain the condition strength calculation input; Perform linear mapping processing on the propagation constraint state representation corresponding to any candidate fault event to obtain the linear strength value of any candidate fault type corresponding to the candidate fault event. The linear mapping processing includes transposing the conditional strength parameter vector corresponding to any candidate fault type, multiplying it with the propagation constraint state representation corresponding to the candidate fault event, and adding the conditional strength bias term corresponding to the candidate fault type. Perform conditional strength activation processing on the linear strength value to obtain the event conditional strength of any candidate fault type corresponding to the candidate fault event. The conditional strength activation processing includes applying a conditional strength activation function to the linear strength value. For any candidate fault event, the linear mapping process and conditional strength activation process are repeatedly performed on all candidate fault types to obtain the event conditional strength of each candidate fault type.

[0030] In this embodiment, based on the fault propagation relationship, propagation path constraints and primary fault and associated anomaly screening are performed on the candidate fault results to obtain root cause candidate results, specifically including: Mapping the candidate fault results to the fault propagation relationship yields a set of candidate propagation nodes; The candidate propagation node set includes device component nodes, operating status nodes, and alarm event nodes corresponding to the candidate fault results; Based on the propagation direction, propagation edge type, and propagation edge weight of each propagation edge in the fault propagation relationship, a propagation path extraction process is performed on the candidate propagation node set to obtain a candidate propagation path set. Each candidate propagation path is formed by sequentially connecting the starting propagation node, the ending propagation node, and the directed propagation edge between the starting propagation node and the ending propagation node. Perform propagation path constraint processing on the candidate propagation path set to obtain the constrained propagation path set; The propagation path constraint processing includes retaining candidate propagation paths that meet the conditions of continuous propagation direction, connection of node types, and order of event occurrence, and eliminating candidate propagation paths that do not meet the conditions of continuous propagation direction, connection of node types, or order of event occurrence. Among them, the continuous propagation direction condition is used to limit the propagation arrival node of the previous propagation edge in the candidate propagation path to be consistent with the propagation start node of the next propagation edge; the node type connection condition is used to limit the node type matching relationship between adjacent propagation nodes in the candidate propagation path; and the event occurrence time sequence condition is used to limit the event occurrence time corresponding to the preceding propagation node in the candidate propagation path to be earlier than or equal to the event occurrence time corresponding to the following propagation node. Based on the constrained propagation path set, the path support calculation is performed on each candidate fault result to obtain the path support corresponding to each candidate fault result. The path support is calculated by dividing the sum of the propagation edge weights of each constrained propagation path containing the candidate fault result by the sum of the number of propagation edges contained in each constrained propagation path; the sum of the propagation edge weights is obtained by adding the propagation edge weights of all propagation edges in each constrained propagation path. Based on the candidate fault scores, path support, and path position relationships corresponding to each candidate fault result, a primary fault and associated anomaly screening process is performed to obtain the primary fault candidate results and associated anomaly candidate results. The main fault screening value is the sum of the product of the candidate fault score and the candidate fault score weight corresponding to the candidate fault result, the product of the path support and the path support weight corresponding to the candidate fault result, and then the product of the upstream dependency and the upstream dependency weight corresponding to the candidate fault result. The upstream dependency is obtained by accumulating the weights of the upstream propagation edges pointing to the propagation node corresponding to the candidate fault result; the main fault candidate result is the candidate fault result with the largest main fault screening value, and the accompanying anomaly candidate results are the other candidate fault results located downstream of the propagation path corresponding to the main fault candidate result; Based on the candidate results of the primary fault and the candidate results of the associated anomalies, the candidate results of the root cause are obtained. The candidate results of the root cause include the candidate nodes of the primary fault, the constrained propagation path corresponding to the candidate nodes of the primary fault, and the candidate nodes of the associated anomalies located on the constrained propagation path.

[0031] In this embodiment, generating the target root cause diagnosis result based on the root cause candidate results, the consistency of the corresponding propagation path, the consistency of the event sequence, and the consistency of the recovery result specifically includes: Extract the main fault candidate node, constrained propagation path, and associated anomaly candidate node from the root cause candidate results, and combine the event occurrence time, candidate fault score, and recovery status marker in the recovery operation data to construct a root cause diagnosis and judgment dataset. Based on the constrained propagation path, a propagation path consistency calculation is performed to obtain the propagation path consistency result. The propagation path consistency result is jointly determined by the continuity of propagation direction between adjacent propagation nodes in the constrained propagation path, the connection of propagation edge types, and the matching of propagation edge weights. The consistency value of the propagation path corresponding to any candidate result is the ratio of the sum of the weights of the propagation edges in the constrained propagation path that satisfy the conditions of continuous propagation direction and connection of node type to the sum of the weights of all propagation edges in the constrained propagation path. Based on the primary fault candidate node, the accompanying anomaly candidate node and the corresponding event occurrence time, perform event timing consistency calculation to obtain the event timing consistency result; Among them, the event timing consistency result is determined by the chronological relationship between the event occurrence time corresponding to the main fault candidate node and the event occurrence time corresponding to each accompanying anomaly candidate node; The event timing consistency value corresponding to any candidate result is the ratio of the number of node pairs that satisfy the condition that the event occurrence time of the primary fault candidate node is earlier than or equal to the event occurrence time of the accompanying abnormal candidate node to the total number of node pairs. Based on the primary fault candidate node, the accompanying anomaly candidate node and the corresponding recovery status flag, perform recovery result consistency calculation processing to obtain the recovery result consistency result; Among them, the consistency of the recovery result is determined by the recovery status of the accompanying abnormal candidate nodes after the fault corresponding to the main fault candidate node is eliminated; the consistency value of the recovery result corresponding to any root cause candidate result is the ratio of the number of accompanying abnormal candidate nodes that meet the recovery status mark as recovery completed to the total number of accompanying abnormal candidate nodes. Based on the consistency values ​​of the propagation path, the event time sequence, and the recovery result, a comprehensive evaluation of the root cause candidate results is performed to obtain the root cause diagnosis confidence value. The root cause diagnosis confidence value is the sum of the product of the propagation path consistency value and the propagation path consistency weight, the product of the event time sequence consistency value and the event time sequence consistency weight, and the product of the recovery result consistency value and the recovery result consistency weight. Based on the root cause diagnosis confidence value, the target root cause diagnosis result is obtained; wherein, the target root cause diagnosis result includes the target root cause node, the target propagation path, the target associated abnormal node, and the corresponding root cause diagnosis confidence value; The target root cause node is the primary fault candidate node with the highest root cause diagnosis confidence value. The target propagation path is the constrained propagation path corresponding to the target root cause node. The target associated anomaly node is the associated anomaly candidate node located downstream of the target propagation path and corresponding to the target root cause node.

[0032] Example 1: To verify the feasibility of this invention in practice, it was applied to an intelligent operation and maintenance scenario of an energy retail terminal equipment cluster. The equipment in this scenario includes refueling terminal control equipment, transaction control equipment, metering and data acquisition equipment, status monitoring equipment, and alarm linkage equipment. During continuous operation, these devices generate operational status data, alarm event data, maintenance work order data, component replacement data, and recovery operation data. Existing technologies typically rely on single alarm information, static operating parameters, or manual experience for fault diagnosis. When the same device triggers multiple related alarms consecutively within a short period, or when multiple component anomalies overlap during propagation, it is easy to encounter situations where the main fault is not clearly identified, accompanying anomalies are mixed into the root cause results, and recurring problems after maintenance are difficult to trace in a timely manner. This invention addresses these problems by uniformly associating multi-source diagnostic raw data to construct a candidate fault event sequence. It then utilizes maintenance closed-loop information to perform weakly supervised label correction, and combines equipment component associations, control call relationships, state change sequence relationships, and alarm trigger relationships to construct fault propagation relationships. Finally, it uses a propagation-constrained weakly supervised neural Hawkes diagnostic model to achieve root cause identification and propagation analysis.

[0033] In practical applications, the system first continuously collects multi-source diagnostic raw data from the target equipment cluster and integrates current, voltage, temperature, communication status, control response delay, alarm codes, work order handling records, component replacement records, and recovery status markers into the diagnostic platform. The platform correlates various data types according to unified equipment identifiers and event occurrence times, integrating abnormal states, alarm triggers, work order handling, component replacements, and recovery results generated by the same equipment within a continuous operating range to generate candidate fault event sequences. Then, based on maintenance work order data, component replacement data, recovery operation data, and fault recurrence information, the system performs weakly supervised label correction on the candidate fault event sequences to obtain fault diagnosis training samples. Furthermore, based on the connection and calling relationships between equipment components, the changes in operating states, and the triggering relationships between abnormal states and alarms, the system constructs a fault propagation relationship between equipment components, operating states, and alarm events, and embeds this fault propagation relationship into a propagation-constrained weakly supervised neural Hawkes diagnostic model. The model performs event embedding encoding, continuous-time state memory update, and event condition strength calculation on the candidate fault event sequence to output candidate fault results. Subsequently, based on the fault propagation relationship, the model performs propagation path constraints and screening of main faults and associated anomalies on the candidate fault results to obtain root cause candidate results. Finally, the model combines the consistency of propagation path, the consistency of event time sequence, and the consistency of recovery results to generate the target root cause diagnosis result.

[0034] In this embodiment, an energy retail terminal equipment cluster within a certain operating cycle was selected as the verification object, including 120 terminal devices. A total of 2.4036 million operational status data entries, 13,300 alarm event data entries, 964 maintenance work order data entries, 318 component replacement data entries, and 1,126 recovery operation data entries were collected. After correlation processing, 2,642 candidate fault event sequences were formed, and after weak supervision label correction, 2,376 effective fault diagnosis training samples were formed. To verify the effectiveness of the invention, the method of the present invention was compared and analyzed with the single alarm rule method and the static feature classification method. The results show that the root cause localization accuracy of the method of the present invention reaches 94.8%, which is 21.3 percentage points higher than the single alarm rule method and 12.6 percentage points higher than the static feature classification method; the main fault identification accuracy reaches 96.1%, and the misjudgment rate of accompanying anomalies is reduced to 5.4%; the recurrence fault identification rate reaches 92.3%, and the consistency rate with actual maintenance conclusions reaches 95.4%. Meanwhile, in complex propagation alarm scenarios, the method of this invention can control the average single event diagnosis time to within 1.78 seconds and effectively reduce duplicate dispatching and misjudgment, indicating that the invention has good implementation effect in multi-source operation and maintenance data association, weak supervision label correction, propagation constraint diagnosis and root cause localization.

[0035] Table 1 Comparison of the Implementation Effects of Intelligent Diagnosis for Energy Retail Terminal Equipment

[0036] As can be seen from the table above, this invention demonstrates superior overall performance compared to single alarm rule methods and static feature classification methods in several key performance indicators of intelligent diagnosis of energy retail terminal equipment. Regarding root cause localization accuracy, the method of this invention achieves 94.8%, significantly higher than the 73.5% of the single alarm rule method and the 82.2% of the static feature classification method, representing improvements of 21.3 and 12.6 percentage points, respectively. This indicates that by uniformly associating operational status data, alarm event data, maintenance work order data, component replacement data, and recovery operation data, and combining this with a weakly supervised label correction mechanism, this invention can more accurately reconstruct the fault evolution process and reduce misjudgments caused by relying solely on single alarms or static features.

[0037] In terms of primary fault identification and differentiation of accompanying anomalies, the method of this invention also demonstrates strong identification capabilities. The table shows that the primary fault identification accuracy of the method of this invention reaches 96.1%, higher than the 78.4% of the single alarm rule method and the 85.7% of the static feature classification method; meanwhile, the false positive rate of accompanying anomalies is only 5.4%, significantly lower than the 22.8% and 14.9% of the comparative methods. This result reflects that the fault propagation relationship constructed by this invention based on the correlation between equipment components, control call relationships, the sequence of state changes, and alarm triggering relationships can more effectively distinguish between source anomalies and subsequent anomalies in the fault propagation chain, thereby making the screening of primary faults and accompanying anomalies more accurate and improving the reliability of the final root cause candidate results.

[0038] In terms of consistency between fault recurrence identification and diagnosis results, the method of this invention also achieves superior results. The recurrence fault identification rate reaches 92.3%, significantly higher than the 61.7% of the single alarm rule method and the 74.5% of the static feature classification method; the consistency rate with actual maintenance conclusions reaches 95.4%, while the comparative methods are 74.1% and 83.5%, respectively. These results indicate that by incorporating maintenance work order data, component replacement data, recovery operation data, and fault recurrence information into the weakly supervised label correction process, this invention can maintain a higher consistency between the training samples and the actual operation and maintenance closed-loop results, and further enhance the ability of the propagation constraint weakly supervised neural Hawkes diagnostic model to identify complex fault modes and recurrence fault modes.

[0039] Furthermore, in terms of engineering application indicators, the method of this invention controls the average diagnosis time to 1.78 seconds per event. Although slightly higher than the 1.21 seconds per event of the single alarm rule method, it is significantly lower than the 2.34 seconds per event of the static feature classification method. Meanwhile, the duplicate dispatch rate is only 6.2%, far lower than 19.7% and 11.8%. This indicates that the present invention maintains high diagnostic accuracy and consistency while still considering diagnostic efficiency, and effectively reduces the problem of duplicate dispatch caused by inaccurate root cause identification, demonstrating good practical application value.

[0040] The above are merely preferred embodiments 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 diagnostic method for intelligent energy retail terminal equipment based on machine learning, characterized in that, Includes the following steps: Collect multi-source diagnostic raw data from energy retail terminal equipment to obtain a multi-source diagnostic raw data set; Based on the multi-source diagnostic raw data set, correlation processing is performed to generate a candidate fault event sequence; Weakly supervised label correction is performed on candidate fault event sequences to generate fault diagnosis training samples; Based on the component relationships, control call relationships, state change sequence relationships, and alarm trigger relationships of energy retail terminal equipment, a fault propagation relationship is constructed; The model is trained based on the fault diagnosis training samples and the propagation constraint weakly supervised neural Hawkes diagnostic model. The candidate fault event sequences are then processed by event embedding encoding, continuous time state memory update, and event condition strength calculation combined with fault propagation relationship to obtain the candidate fault results. Based on the fault propagation relationship, propagation path constraints and screening of primary faults and associated anomalies are performed on candidate fault results to obtain root cause candidate results; Based on the root cause candidate results, the consistency of the corresponding propagation path, the consistency of the event sequence, and the consistency of the recovery results, the target root cause diagnosis results are generated.

2. The method for diagnosing intelligent energy retail terminal equipment based on machine learning according to claim 1, characterized in that, The multi-source diagnostic raw data includes operating status data, alarm event data, maintenance work order data, component replacement data, and recovery operation data.

3. The method for diagnosing intelligent energy retail terminal equipment based on machine learning according to claim 1, characterized in that, The step of performing correlation processing on the multi-source diagnostic raw data set to generate a candidate fault event sequence specifically includes: The multi-source diagnostic raw data set is grouped to obtain a subset of the device raw data. The raw data subset is then sorted to form a time-ordered event stream. In a time-ordered event stream, an initial associated time window is constructed using alarm event records as the trigger center. Window extraction processing is performed on the time-ordered event stream to obtain initial associated event fragments. Association determination processing is then performed on the initial associated event fragments to obtain fault associated event fragments. Perform event boundary determination processing on fault-related event fragments to obtain a single candidate fault event; For multiple single candidate fault events corresponding to the same unified equipment identifier, perform segment overlap detection and adjacent segment splicing processing to obtain continuous candidate fault events; The consecutive candidate fault events are sorted by their start time to obtain a sequence of candidate fault events. A summary process is performed on the candidate fault event sequences corresponding to each unified equipment identifier to obtain a set of candidate fault event sequences.

4. The method for diagnosing intelligent energy retail terminal equipment based on machine learning according to claim 1, characterized in that, The step of performing weakly supervised label correction on the candidate fault event sequence to generate fault diagnosis training samples specifically includes: Perform candidate event expansion processing on the candidate fault event sequence set to form a candidate fault event set; Alignment processing is performed on the candidate fault event set to obtain the candidate work order set, the candidate component replacement set, and the candidate operation recovery set. Perform fault recurrence detection processing on the candidate fault event set to obtain fault recurrence information; Weak supervision label generation processing is performed on the candidate fault event set to obtain initial weak supervision labels, and label correction processing is performed on the initial weak supervision labels to obtain corrected weak supervision labels; The corrected weakly supervised labels are subjected to sample screening to obtain fault diagnosis training samples.

5. The method for diagnosing intelligent energy retail terminal equipment based on machine learning according to claim 1, characterized in that, The construction of fault propagation relationships based on the component relationships, control call relationships, state change sequence relationships, and alarm trigger relationships of energy retail terminal equipment specifically includes: Perform associated object extraction processing on energy retail terminal equipment to obtain a set of equipment components, a set of operating statuses, and a set of alarm events; Based on the set of equipment components, component association relationships are constructed to obtain a set of component association relationships; Identify the control signal input relationships, control command output relationships, and control link transmission relationships between the set of equipment components and the set of operating states, and construct a set of control call relationships; Based on the collection time sequence corresponding to the running status records in the time-ordered event stream, the order of changes of each running status item is identified, and a set of relationships between the order of state changes is constructed. Match the abnormal states corresponding to each running status item with the alarm occurrence records corresponding to each alarm event item to construct an alarm trigger relationship set; The set of component association relationships, the set of control call relationships, the set of state change sequence relationships, and the set of alarm trigger relationships are mapped to propagation edges with propagation directions to obtain the set of fault propagation edges; The weight of each propagation edge is determined based on the set of fault propagation edges, thus obtaining the weighted fault propagation relationship. The set of equipment components, the set of operating states, and the set of alarm events are used as the propagation node set, and the set of fault propagation edges are used as the propagation edge set. The propagation edge weights corresponding to each propagation edge are used as the propagation edge weight set to construct a fault propagation graph, which is then used as the fault propagation relationship between equipment components, operating status, and alarm events.

6. The method for diagnosing intelligent energy retail terminal equipment based on machine learning according to claim 1, characterized in that, The method involves training a weakly supervised neural Hawkes diagnostic model with propagation constraints based on fault diagnosis training samples, and performing event embedding encoding, continuous-time state memory updates, and event condition strength calculations in conjunction with fault propagation relationships on candidate fault event sequences to obtain candidate fault results. Specifically, this includes: The training sample encoding process is performed on the fault diagnosis training samples to obtain the training event sequence input; A propagation-constrained weakly supervised neural Hawkes diagnostic model is constructed based on the training event sequence input, and the parameters of the propagation-constrained weakly supervised neural Hawkes diagnostic model are initialized. Each candidate fault event in the training event sequence is input into the event embedding encoding layer, and event embedding encoding processing is performed to obtain the event embedding representation; Embed the event into the continuous-time state memory update layer, perform continuous-time state memory update processing, and obtain the continuous-time hidden state; In the continuous-time state memory update layer, time decay processing is performed on the continuous-time hidden state corresponding to the previous candidate fault event to obtain the historical hidden state after time decay. The continuous-time hidden state is input into the propagation constraint fusion layer, and propagation constraint fusion processing is performed in combination with the fault propagation relationship to obtain the propagation constraint state representation; The propagation constraint state representation is input into the event condition strength calculation layer, and the event condition strength calculation process is performed to obtain the event condition strength; The event condition intensity is input into the result output layer, and the candidate fault score generation process is performed to obtain the candidate fault score. The candidate fault type with the highest candidate fault score is taken as the candidate fault result of the corresponding candidate fault event. Based on the corrected weak supervision labels corresponding to the candidate fault results and the fault diagnosis training samples, the propagation constraint weak supervision neural Hawkes diagnostic model is trained to obtain the trained propagation constraint weak supervision neural Hawkes diagnostic model. Collect the current operating status data of the device to be diagnosed, and combine it with the alarm event data of the device to be diagnosed to construct a sequence of candidate fault events to be diagnosed; The sequence of candidate fault events to be diagnosed is input into the trained propagation-constrained weakly supervised neural Hawkes diagnostic model to obtain the candidate fault results.

7. The method for diagnosing intelligent energy retail terminal equipment based on machine learning according to claim 6, characterized in that, The step of inputting the continuous-time hidden state into the propagation constraint fusion layer and performing propagation constraint fusion processing in conjunction with the fault propagation relationship to obtain the propagation constraint state representation specifically includes: Based on the multi-source diagnostic raw data set corresponding to the current candidate fault event, the propagation node set in the fault propagation relationship is mapped to the propagation node set, and the event-related node set is obtained. Based on the node connection relationship, propagation direction, and propagation edge weight of the event-related node set in the fault propagation relationship, a local propagation subgraph is extracted to obtain local propagation constraint information; Perform local propagation constraint vector generation processing on the local propagation subgraph to obtain the local propagation constraint vector; Perform weight mapping on the local propagation constraint vector to obtain the propagation constraint mapping result; Perform state mapping processing on the continuous-time hidden states to obtain the continuous-time hidden state mapping results; Perform fusion calculation on the continuous-time hidden state mapping results, propagation constraint mapping results, and propagation constraint fusion bias vector to obtain intermediate fusion results; The intermediate fusion results are processed by nonlinear mapping to obtain the propagation constraint state representation; According to the order of the start time of each candidate fault event in the training event sequence, the propagation constraint state representations corresponding to each candidate fault event are arranged to obtain the propagation constraint state representation sequence.

8. The method for diagnosing intelligent energy retail terminal equipment based on machine learning according to claim 6, characterized in that, The process of inputting the propagation constraint state representation into the event condition strength calculation layer and performing event condition strength calculation processing to obtain the event condition strength specifically includes: The propagation constraint state is represented by the input event condition strength calculation layer, and a candidate fault type set is extracted; For any candidate fault type, read the condition strength parameter vector and condition strength bias term corresponding to the candidate fault type, and match them with the propagation constraint state representation to obtain the condition strength calculation input; Perform linear mapping processing on the propagation constraint state representation corresponding to any candidate fault event to obtain the linear strength value; Conditional intensity activation processing is applied to the linear intensity value to obtain the event conditional intensity.

9. The method for diagnosing intelligent energy retail terminal equipment based on machine learning according to claim 1, characterized in that, The process of performing propagation path constraints and screening of primary faults and associated anomalies on candidate fault results based on fault propagation relationships to obtain root cause candidate results specifically includes: Mapping the candidate fault results to the fault propagation relationship yields a set of candidate propagation nodes; Based on the propagation direction, propagation edge type, and propagation edge weight of each propagation edge in the fault propagation relationship, a propagation path extraction process is performed on the candidate propagation node set to obtain a candidate propagation path set. Perform propagation path constraint processing on the candidate propagation path set to obtain the constrained propagation path set; Based on the constrained propagation path set, the path support is calculated for each candidate fault result to obtain the path support. Based on the candidate fault scores, path support, and path location relationships, a primary fault and associated anomaly screening process is performed to obtain the primary fault candidate results and associated anomaly candidate results. Based on the candidate results of the main fault and the candidate results of the associated anomalies, the candidate results of the root cause are obtained.

10. The method for diagnosing intelligent energy retail terminal equipment based on machine learning according to claim 1, characterized in that, The process of generating the target root cause diagnosis result based on the root cause candidate results, the consistency of the corresponding propagation path, the consistency of the event sequence, and the consistency of the recovery result specifically includes: Based on the root cause candidate results and candidate fault results, a root cause diagnosis and determination dataset is constructed. Based on the constrained propagation path, a propagation path consistency calculation is performed to obtain the propagation path consistency result. Based on the primary fault candidate node, the accompanying anomaly candidate node and the corresponding event occurrence time, perform event timing consistency calculation to obtain the event timing consistency result; Based on the primary fault candidate node, the accompanying anomaly candidate node and the corresponding recovery status flag, perform recovery result consistency calculation processing to obtain the recovery result consistency result; Based on the consistency values ​​of the propagation path, the event sequence, and the recovery result, a comprehensive evaluation of the root cause candidate results is performed to obtain the root cause diagnosis confidence value. Based on the root cause diagnosis confidence value, the target root cause diagnosis result is obtained.

Images