A Deep Learning-Based Integrated Monitoring System and Method for Power Equipment Status

Abstract

This invention discloses a deep learning-based integrated monitoring system and method for power equipment status, relating to the field of power equipment maintenance technology. It constructs a structured feature set (MFD) by extracting behavioral information such as maintenance timestamps, operation type codes, and durations from original operation and maintenance records, and calculates the average cycle and rhythm deviation to generate a maintenance behavior rhythm feature (PTD). Simultaneously, it combines an operation and maintenance knowledge semantic model to generate a semantic vector set (EMB) for maintenance operations and constructs a comprehensive behavioral input sequence (SEQ). By inputting the behavioral input sequence into a deep sequence modeling network for temporal modeling, it outputs a behavioral state back-inference score (BSS) to quantify the potential risk level of maintenance behaviors, effectively realizing the "reverse perception" capability of equipment status. This improves the sensitivity, predictability, and response accuracy of power equipment operational health status judgment, effectively solving the problem of insufficient identification of "non-obvious degradation factors" in existing monitoring systems.

Description

Technical Field

[0001] This invention relates to the field of power equipment maintenance technology, specifically to a power equipment status integrated monitoring system and method based on deep learning. Background Technology

[0002] In today's information society, automated monitoring of power systems is becoming a crucial supporting area for industrial intelligence, falling under the broader category of industrial intelligent sensing and condition assessment technologies. This area encompasses the key capabilities for real-time sensing, remote monitoring, and intelligent assessment of equipment operating status across multiple industries, including energy, transportation, and manufacturing. In the power industry, particularly in the operation and maintenance of high-voltage transmission and transformation equipment, the accuracy and reliability of condition assessment directly impact the stability of the entire power grid and the security of power supply.

[0003] Focusing further on the internal aspects, key power equipment such as transformers, circuit breakers, cable joints, and switchgear, as core components of the power system, exhibit a complex temporal evolution network due to their aging, failures, and maintenance behaviors during long-term operation.

[0004] Currently, most mainstream condition monitoring methods rely on physical data such as electrical quantities, thermal imaging, and partial discharge detection collected by sensors, combined with manually defined threshold strategies or rule engines for judgment. While these methods are effective in identifying "manifested anomalies," their effectiveness and timeliness are significantly insufficient when facing latent degradation, precursors to sudden failures, or situations requiring repeated human intervention. Maintenance behavior data, in particular, is often only stored in information systems as maintenance documentation and is not used for model input or equipment condition modeling. This deprives equipment condition diagnosis of a crucial "human-machine interaction variable," making it impossible to infer the existence of potential anomalies from the "frequency of human maintenance." For example, a piece of equipment may undergo multiple repairs in a short period, but because sensor data does not significantly exceed limits, the system may still classify it as normal, leading to systemic biases in false alarms and misjudgments. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a deep learning-based integrated monitoring system and method for the status of power equipment, which solves the problems mentioned in the background section.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a deep learning-based integrated monitoring method for the status of power equipment, comprising the following steps:

[0007] S1. Collect equipment maintenance behavior information. By obtaining the maintenance timestamp, maintenance operation type code and operation duration from the operation and maintenance records, construct a structured feature set (MFD) of equipment maintenance behavior.

[0008] S2. Calculate the average cycle and fluctuation deviation of maintenance operations based on the feature set MFD, and generate the maintenance behavior rhythm feature PTD;

[0009] S3. Perform semantic embedding processing on the feature set MFD, generate the semantic vector set EMB of maintenance operations through the operation and maintenance knowledge semantic model, and construct the input sequence SEQ by combining the maintenance behavior rhythm feature PTD feature.

[0010] S4. Input the input sequence SEQ into the deep sequence modeling network to build an inference model between maintenance behavior and equipment health status, and output the behavior status back-inference score BSS.

[0011] S5. The behavioral state back-inference score BSS is fused with the health assessment score MON calculated based on the physical quantity acquisition model in the main monitoring channel of the equipment to generate a fused health status assessment value FUS, which determines the degree of health risk of the power equipment and triggers response actions.

[0012] Preferably, S1 includes S11;

[0013] S11. By extracting maintenance behavior records related to the target equipment from the database or log system of the power equipment operation and maintenance system, the records include the timestamp of the maintenance behavior, the operation type and the duration of the maintenance operation, and obtaining the original behavior field set RFD={tm(i),op(i),dur(i)}, where tm(i) represents the timestamp of the i-th maintenance operation in the record, op(i) represents the type code of the i-th maintenance operation in the record, and dur(i) represents the duration of the i-th maintenance operation in the record;

[0014] The operation types include preset operation types related to tightening, testing, dust removal, and reset.

[0015] Preferably, S1 includes S12;

[0016] S12. The original behavioral field set RFD is expressed in a structured vector form to construct a standard modeling input set MF. Then, the standard modeling input set MF is organized into a time series form to obtain the structured feature set MFD.

[0017] The structuring process employs a combination of field mapping, rule standardization, and unit unification. Field mapping involves extracting the timestamp tm(i), type code op(i), and duration dur(i) of the i-th maintenance operation from the original behavior field set RFD. Rule standardization enumerates labels for the type code op(i) of the i-th maintenance operation using a unified preset encoding table. Unit unification converts the timestamp tm(i) and duration dur(i) of the i-th maintenance operation into minutes, eliminating the format differences between hours and seconds.

[0018] Vectorization is performed by concatenating fixed-length numerical vectors. The timestamp tm(i), type code op(i), and duration dur(i) of the i-th maintenance operation after the structured processing are stacked sequentially to obtain a two-dimensional array matrix. The vectorization of the original behavior field set RFD is completed by traversing the original behavior field set RFD to obtain a three-dimensional matrix, and the standard modeling input set MF is obtained.

[0019] The organization is represented in time series form. Using the unique primary index order of the timestamp tm(i) of the i-th maintenance operation after vectorization, all maintenance records are sorted to obtain the structured feature set MFD.

[0020] Preferably, S2 includes S21 and S22;

[0021] S21. Based on the timestamp tm feature of the time series in the structured feature set MFD, calculate the time interval between each pair of adjacent maintenance behavior operations, and after statistically analyzing all vector interval values, determine the average maintenance behavior cycle (AMT) of the equipment in the current analysis period. This reflects the maintenance intervention rhythm reference cycle of the equipment under normal working conditions. When the actual maintenance cycle is near the average cycle (AMT), it indicates that the equipment maintenance behavior operation is normal. When the actual maintenance cycle is greater than or significantly shorter than the average cycle (AMT), it indicates that the equipment maintenance behavior operation is abnormal, indicating that the equipment frequently has problems and is subject to high-frequency intervention.

[0022] S22. Based on the average time period (AMT) and combined with the timestamp (tm) feature of the time series in the structured feature set (MFD), the deviation between the adjacent time intervals of all maintenance behaviors and the average time period (AMT) is quantified to construct a feature quantity to describe the degree of fluctuation in the rhythm of maintenance behaviors, which is marked as the maintenance behavior rhythm feature (PTD). This reflects the stable periodicity of maintenance behaviors in the time dimension, as well as whether there are rhythmic instabilities caused by abnormal equipment status, frequent failures, and intensive human intervention.

[0023] Preferably, S3 includes S31;

[0024] S31. Based on the type-encoded op features in the structured feature set MFD, semantic embedding modeling is performed on the type-encoded op features of each maintenance operation. The type-encoded op of the operation type is mapped to a semantic vector through a predefined operation and maintenance knowledge semantic representation model, forming an operation semantic vector set EMB. The operation semantic vector set EMB is used to represent the feature differences and inter-operation correlations of various maintenance behaviors in the semantic dimension.

[0025] The semantic representation model is constructed using the Act2Vec model, which is an operation and maintenance semantic embedding model trained based on the text description of maintenance behavior in the historical operation and maintenance logs of power equipment. It achieves implicit semantic modeling of maintenance operation types by learning the vector encoding of the contextual relationship of keywords in maintenance statements.

[0026] Preferably, S3 includes S32;

[0027] S32. Jointly model the operation semantic vector set EMB with the timestamp tm and duration dur in the structured feature set MFD, and combine the maintenance behavior rhythm feature PTD to construct the input sequence SEQ of the maintenance behavior. The input sequence SEQ is used to describe the structural attributes, semantic expression and rhythm changes of each maintenance behavior in time sequence.

[0028] Preferably, each time step in the input sequence SEQ consists of the following feature concatenation steps:

[0029] S321. Extract the type code op, timestamp tm, and duration dur features corresponding to each maintenance behavior record from the structured feature set MFD, and sort them in ascending order according to the timestamp tm. These features are used to construct the structural attribute features of each time step in the input sequence to ensure that the maintenance behavior has temporal logical continuity in time series modeling.

[0030] S322. Extract the operation semantic vector corresponding to each structured maintenance record from the operation semantic vector set EMB. The operation semantic vector is the vector obtained by converting the type encoding field op in the structured feature set MFD through the Act2Vec model in the previous step S31. It is used to represent the inherent category features and operation similarity information of the maintenance operation in the semantic dimension in the input sequence.

[0031] S323. The maintenance behavior rhythm feature PTD is broadcast as a unified global time rhythm feature, so that the scalar is consistent with the number of structured records in the input sequence, generating a rhythm vector sequence that is aligned one-to-one with the maintenance records. Then, it is concatenated with the output of steps S321 and S322 in the time step dimension to complete the feature fusion of the maintenance behavior input sequence SEQ.

[0032] Preferably, S4 includes S41;

[0033] S41. Input the input sequence SEQ into the deep sequence modeling network for modeling processing, construct an inference model for inferring the relationship between maintenance behavior and equipment health status, and output the behavior state back-inference score BSS for quantifying the potential operating risks of the equipment. The behavior state back-inference score BSS is used to quantify whether the current maintenance behavior implies a trend of equipment deterioration. The value range is between 0 and 1. The larger the value, the higher the potential risk.

[0034] When the behavior state back-inference score (BSS) approaches 1, it indicates that the current maintenance behavior sequence characteristics highly match the characteristics of historical abnormal samples, and the equipment has a potential trend of operational degradation or hidden risks.

[0035] When the behavioral state back-inference score (BSS) approaches 0, it indicates that the overall characteristics of the maintenance behavior are similar to the historical health state samples, and the equipment operation status is relatively stable.

[0036] The approximation range is defined by the user.

[0037] The deep sequence modeling network is implemented using a gated recurrent neural network structure. By modeling the time dimension of the input sequence SEQ, it extracts the dynamic change patterns in the maintenance behavior feature sequence, thereby completing the implicit back-inference modeling of the equipment status trend.

[0038] Preferably, S5 includes S51;

[0039] S51. The behavioral state back-inference score BSS is fused with the health assessment score MON calculated by the physical quantity acquisition model in the main monitoring channel of the equipment to generate a fused health status assessment value FUS. The health assessment score MON is a status score output by the main monitoring model after feature extraction and risk assessment based on electrical, electrothermal and mechanical operating data collected in the online monitoring system of the equipment. Its data sources include current waveform, voltage distortion, partial discharge amplitude and infrared temperature distribution sensing signal. The health assessment score MON has an independent risk identification function in the main monitoring system, and its range is between 0 and 1. It is used to represent the degree of health risk of the equipment in the physical monitoring dimension. The fused health status assessment value FUS generated by combining the behavioral state back-inference score BSS with the health assessment score MON is used to comprehensively reflect the health risk of the equipment in the two dimensions of maintenance behavior performance and physical operation status.

[0040] When the fusion health status assessment value FUS is greater than the set threshold θ, the device is determined to enter the early warning state, triggering the corresponding operation and maintenance response action or risk control strategy.

[0041] When the fused health status assessment value FUS is less than or equal to the set threshold θ, it is determined that the device will not enter the warning state and will not trigger operation and maintenance response actions or risk control strategies.

[0042] A deep learning-based integrated monitoring system for power equipment status includes an equipment behavior data acquisition module, a deviation feature generation module, a semantic construction module, an equipment behavior status generation module, and a status assessment and decision-making module.

[0043] The device behavior data acquisition module collects device maintenance behavior information and constructs a structured feature set (MFD) of device maintenance behavior by obtaining maintenance timestamps, maintenance operation type codes, and operation durations from the operation and maintenance records.

[0044] The deviation feature generation module calculates the average cycle and fluctuation deviation of maintenance operations based on the feature set MFD, and generates the maintenance behavior rhythm feature PTD.

[0045] The semantic construction module performs semantic embedding processing on the feature set MFD, generates a semantic vector set EMB for maintenance operations through the operation and maintenance knowledge semantic model, and constructs the input sequence SEQ by combining the maintenance behavior rhythm feature PTD feature.

[0046] The device behavior state generation module inputs the input sequence SEQ into a deep sequence modeling network to construct an inference model between maintenance behavior and device health state, and outputs the behavior state back-inference score BSS.

[0047] The status assessment and decision-making module integrates the behavioral status back-inference score (BSS) with the health assessment score (MON) calculated based on the physical quantity acquisition model in the main monitoring channel of the equipment to generate a fused health status assessment value (FUS), which determines the degree of health risk of the power equipment and triggers response actions.

[0048] This invention provides a deep learning-based integrated monitoring system and method for the status of power equipment, which has the following beneficial effects:

[0049] (1) By extracting behavioral information such as maintenance timestamps, operation type codes and durations from the original operation and maintenance records, a structured feature set MFD is constructed, and the average cycle and rhythm deviation are calculated to generate the maintenance behavior rhythm feature PTD. At the same time, the semantic vector set EMB of maintenance operations is generated by combining the operation and maintenance knowledge semantic model, and a comprehensive behavior input sequence SEQ is constructed. By inputting the behavior input sequence into a deep sequence modeling network for time series modeling, the behavior state back-inference score BSS used to quantify the potential risk level of maintenance behavior is output, which effectively realizes the "reverse perception" capability of equipment status. The behavior state back-inference score BSS is fused with the health assessment score MON output by the main monitoring system based on the sensor model to generate a fused health status assessment value FUS. This enables the system to not only have the conventional detection capability based on hardware signals, but also the trend reasoning capability based on maintenance behavior patterns, thereby significantly improving the sensitivity, predictability and response accuracy of the power equipment operation health status judgment, and effectively solving the problem of insufficient identification of "non-obvious degradation factors" in the existing monitoring system.

[0050] (2) Based on the maintenance behavior timestamp field in the structured feature set MFD, the time interval between each pair of adjacent maintenance operations is first calculated, and the average maintenance behavior period (AMT) within the current cycle of the equipment is statistically obtained as a benchmark reference for the maintenance rhythm. Then, based on the average period (AMT), a maintenance behavior rhythm feature (PTD) is constructed to measure the fluctuation of the maintenance behavior rhythm. This feature can effectively reflect whether the equipment has potential abnormal trends such as large fluctuations in operating status, disordered maintenance rhythm, or high frequency of manual intervention. It can complete the encoding and fusion of multi-dimensional maintenance behavior information at the input stage, significantly enhancing the model's ability to dynamically understand the process of equipment operating status changes and its ability to identify non-explicit intervention behaviors. This effectively overcomes the limitations of existing technologies that rely on a single structured field to characterize the trend of behavior changes.

[0051] (3) Based on the pre-constructed maintenance behavior input sequence SEQ, a deep sequence modeling network is used to learn the dynamic evolution characteristics of maintenance behavior in the time dimension. The potential behavioral pattern changes in the sequence are extracted through a gated recurrent neural network structure, and the behavioral state back-inference score BSS is output to measure the risk trend of the equipment. By comparing the feature similarity between the maintenance behavior sequence features and historical abnormal state samples, the potential deterioration signs hidden in the equipment before obvious sensor anomalies can be effectively identified. The behavioral state back-inference score BSS is further fused with the health assessment score MON generated based on real-time operation data analysis in the main monitoring channel of the equipment to generate a fused health status assessment value FUS, realizing a comprehensive quantitative judgment of the risk characteristics of power equipment in the "behavioral performance dimension" and "physical perception dimension". When the fused health status assessment value FUS exceeds the health threshold set by the user, the system can automatically determine that the equipment is in a warning state and trigger a response operation. This fusion mechanism, while maintaining the stability of physical quantity perception, significantly improves the system's ability to proactively identify abnormal state trends. It enables monitoring strategies to move beyond relying on single, real-time physical parameter fluctuations and instead proactively identify early signs of faults by combining behavioral evolution trajectories. This enhances the overall performance of equipment health management in terms of timely early warning, reasonable intervention, and the accuracy of monitoring closed loops. Attached Figure Description

[0052] Figure 1 This is a schematic diagram illustrating the steps of a deep learning-based integrated monitoring method for the status of power equipment according to the present invention.

[0053] Figure 2 This is a schematic diagram of a deep learning-based integrated monitoring system for the status of power equipment, according to the present invention. Detailed Implementation

[0054] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0055] Example 1

[0056] This invention provides a deep learning-based integrated monitoring method for the status of power equipment. Please refer to [link / reference]. Figure 1 This includes the following steps:

[0057] S1. Collect equipment maintenance behavior information. By obtaining the maintenance timestamp, maintenance operation type code and operation duration from the operation and maintenance records, construct a structured feature set (MFD) of equipment maintenance behavior.

[0058] S2. Calculate the average cycle and fluctuation deviation of maintenance operations based on the feature set MFD, and generate the maintenance behavior rhythm feature PTD;

[0059] S3. Perform semantic embedding processing on the feature set MFD, generate the semantic vector set EMB of maintenance operations through the operation and maintenance knowledge semantic model, and construct the input sequence SEQ by combining the maintenance behavior rhythm feature PTD feature.

[0060] S4. Input the input sequence SEQ into the deep sequence modeling network to build an inference model between maintenance behavior and equipment health status, and output the behavior status back-inference score BSS.

[0061] S5. The behavioral state back-inference score BSS is fused with the health assessment score MON calculated based on the physical quantity acquisition model in the main monitoring channel of the equipment to generate a fused health status assessment value FUS, which determines the degree of health risk of the power equipment and triggers response actions.

[0062] In this embodiment, a structured feature set MFD is constructed by extracting behavioral information such as maintenance timestamps, operation type codes, and durations from the original operation and maintenance records. The average cycle and rhythm deviation are calculated to generate a maintenance behavior rhythm feature PTD. At the same time, a semantic vector set EMB for maintenance operations is generated by combining the operation and maintenance knowledge semantic model, and a comprehensive behavior input sequence SEQ is constructed. By inputting the behavior input sequence into a deep sequence modeling network for time series modeling, a behavior state back-inference score BSS is output to quantify the potential risk level of maintenance behavior, effectively realizing the "reverse perception" capability of equipment status. The behavior state back-inference score BSS is fused with the health assessment score MON output by the main monitoring system based on the sensor model to generate a fused health status assessment value FUS. This enables the system to not only have conventional detection capabilities based on hardware signals, but also trend reasoning capabilities based on maintenance behavior patterns, thereby significantly improving the sensitivity, predictability, and response accuracy of power equipment operation health status judgment, and effectively solving the problem of insufficient identification of "non-obvious degradation factors" in existing monitoring systems.

[0063] Example 2

[0064] This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1 Specifically: S1 includes S11;

[0065] S11. By extracting maintenance behavior records related to the target equipment from the database or log system of the power equipment operation and maintenance system, the records include the timestamp of the maintenance behavior, the operation type and the duration of the maintenance operation, and obtaining the original behavior field set RFD={tm(i),op(i),dur(i)}, where tm(i) represents the timestamp of the i-th maintenance operation in the record, op(i) represents the type code of the i-th maintenance operation in the record, and dur(i) represents the duration of the i-th maintenance operation in the record;

[0066] The operation types include preset operation types related to tightening, testing, dust removal, and reset.

[0067] S1 includes S12;

[0068] S12. The original behavioral field set RFD is expressed in a structured vector form to construct a standard modeling input set MF. Then, the standard modeling input set MF is organized into a time series form to obtain the structured feature set MFD.

[0069] The structuring process employs a combination of field mapping, rule standardization, and unit unification. Field mapping involves extracting the timestamp tm(i), type code op(i), and duration dur(i) of the i-th maintenance operation from the original behavior field set RFD. Rule standardization enumerates labels for the type code op(i) of the i-th maintenance operation using a unified preset encoding table. Unit unification converts the timestamp tm(i) and duration dur(i) of the i-th maintenance operation into minutes, eliminating the format differences between hours and seconds.

[0070] Vectorization is performed by concatenating fixed-length numerical vectors. The timestamp tm(i), type code op(i), and duration dur(i) of the i-th maintenance operation after the structured processing are stacked sequentially to obtain a two-dimensional array matrix. The vectorization of the original behavior field set RFD is completed by traversing the original behavior field set RFD to obtain a three-dimensional matrix, and the standard modeling input set MF is obtained.

[0071] The organization is represented in time series form. Using the unique primary index order of the timestamp tm(i) of the i-th maintenance operation after vectorization, all maintenance records are sorted to obtain the structured feature set MFD.

[0072] In this embodiment, by explicitly extracting the timestamps, operation types, and durations of maintenance behaviors, an original behavior field set (RFD) is constructed. Based on a structured method of field mapping, rule standardization, and unit unification, the original unstructured log data is transformed into a standard modeling input set (MF) with a unified format. Furthermore, by organizing all maintenance behavior records in ascending order using the timestamp field (tm) as a master index, a structured feature set (MFD) with a clear behavioral evolution logic is generated. This entire process not only significantly improves the usability and model adaptability of the original maintenance behavior data but also effectively solves the problem that existing maintenance data is difficult to directly use for deep modeling due to semantic inconsistencies, irregular formats, and broken time relationships. Especially in situations where external labels or fault feedback data are not required, the structured feature set (MFD) constructed in this step can serve as the temporal input basis for subsequent state-based reverse modeling, enhancing the system's ability to model the evolution trend of maintenance behaviors from the source. This provides a high-quality, highly consistent original input guarantee for subsequent behavior-driven intelligent perception of equipment status.

[0073] Example 3

[0074] This embodiment is an explanation based on Embodiment 2. Please refer to it. Figure 1 Specifically: S2 includes S21 and S22;

[0075] S21. Based on the timestamp tm feature of the time series in the structured feature set MFD, calculate the time interval between each pair of adjacent maintenance behavior operations, and after statistically analyzing all vector interval values, determine the average maintenance behavior cycle (AMT) of the equipment in the current analysis period. This reflects the maintenance intervention rhythm reference cycle of the equipment under normal working conditions. When the actual maintenance cycle is near the average cycle (AMT), it indicates that the equipment maintenance behavior operation is normal. When the actual maintenance cycle is greater than or significantly shorter than the average cycle (AMT), it indicates that the equipment maintenance behavior operation is abnormal, indicating that the equipment frequently has problems and is subject to high-frequency intervention.

[0076] S22. Based on the average time period (AMT) and combined with the timestamp (tm) feature of the time series in the structured feature set (MFD), the deviation between the adjacent time intervals of all maintenance behaviors and the average time period (AMT) is quantified to construct a feature quantity to describe the degree of fluctuation in the rhythm of maintenance behaviors, which is marked as the maintenance behavior rhythm feature (PTD). This reflects the stable periodicity of maintenance behaviors in the time dimension, as well as whether there are rhythmic instabilities caused by abnormal equipment status, frequent failures, and intensive human intervention.

[0077] S3 includes S31;

[0078] S31. Based on the type-encoded op features in the structured feature set MFD, semantic embedding modeling is performed on the type-encoded op features of each maintenance operation. The type-encoded op of the operation type is mapped to a semantic vector through a predefined operation and maintenance knowledge semantic representation model, forming an operation semantic vector set EMB. The operation semantic vector set EMB is used to represent the feature differences and inter-operation correlations of various maintenance behaviors in the semantic dimension.

[0079] The semantic representation model is constructed using the Act2Vec model, which is an operation and maintenance semantic embedding model trained based on the text description of maintenance behavior in the historical operation and maintenance logs of power equipment. It achieves implicit semantic modeling of maintenance operation types by learning the vector encoding of the contextual relationship of keywords in maintenance statements.

[0080] S3 includes S32;

[0081] S32. Jointly model the operation semantic vector set EMB with the timestamp tm and duration dur in the structured feature set MFD, and combine the maintenance behavior rhythm feature PTD to construct the input sequence SEQ of the maintenance behavior. The input sequence SEQ is used to describe the structural attributes, semantic expression and rhythm changes of each maintenance behavior in time sequence.

[0082] Each time step in the input sequence SEQ consists of the following feature concatenation steps:

[0083] S321. Extract the type code op, timestamp tm, and duration dur features corresponding to each maintenance behavior record from the structured feature set MFD, and sort them in ascending order according to the timestamp tm. These features are used to construct the structural attribute features of each time step in the input sequence to ensure that the maintenance behavior has temporal logical continuity in time series modeling.

[0084] S322. Extract the operation semantic vector corresponding to each structured maintenance record from the operation semantic vector set EMB. The operation semantic vector is the vector obtained by converting the type encoding field op in the structured feature set MFD through the Act2Vec model in the previous step S31. It is used to represent the inherent category features and operation similarity information of the maintenance operation in the semantic dimension in the input sequence.

[0085] S323. The maintenance behavior rhythm feature PTD is broadcast as a unified global time rhythm feature, so that the scalar is consistent with the number of structured records in the input sequence, generating a rhythm vector sequence that is aligned one-to-one with the maintenance records. Then, it is concatenated with the output of steps S321 and S322 in the time step dimension to complete the feature fusion of the maintenance behavior input sequence SEQ.

[0086] In this embodiment, a deep abstract representation of maintenance operation types in the semantic dimension is achieved. A maintenance behavior input sequence SEQ is constructed by fusing structured features, semantic embedding features, and rhythm offset features. This provides a high-dimensional, highly expressive sequence input foundation for the subsequent state judgment model. Based on the maintenance behavior timestamp field in the structured feature set MFD, the time interval between each pair of adjacent maintenance operations is first calculated. Then, the average maintenance cycle (AMT) within the current cycle of the equipment is statistically obtained, serving as a benchmark reference for the maintenance rhythm. Subsequently, based on the average cycle (AMT), the deviation between each time interval and the benchmark cycle is further calculated, constructing a maintenance behavior rhythm feature PTD to measure the fluctuation of maintenance behavior rhythm. This feature can effectively reflect potential abnormal trends such as large fluctuations in the equipment's operating status, disordered maintenance rhythm, or high-frequency manual intervention. Based on the maintenance operation type encoding field in the structured feature set MFD, the operation and maintenance knowledge semantic representation model Act2Vec is used to perform embedding learning on the encoding of each type of operation. The encodings are then concatenated and fused along the time step dimension to construct a maintenance behavior input sequence SEQ containing structural attribute features, semantic representation features, and rhythm fluctuation features. Through the above modeling mechanism, the encoding and fusion of multi-dimensional maintenance behavior information can be completed at the input stage, which significantly enhances the model's ability to dynamically understand the process of equipment operation status changes and to identify non-explicit intervention behaviors. This effectively overcomes the limitations of existing technologies that rely on a single structural field to depict the trend of behavior changes.

[0087] Example 4

[0088] This embodiment is an explanation based on Embodiment 3. Please refer to it. Figure 1 Specifically: S4 includes S41;

[0089] S41. Input the input sequence SEQ into the deep sequence modeling network for modeling processing, construct an inference model for inferring the relationship between maintenance behavior and equipment health status, and output the behavior state back-inference score BSS for quantifying the potential operating risks of the equipment. The behavior state back-inference score BSS is used to quantify whether the current maintenance behavior implies a trend of equipment deterioration. The value range is between 0 and 1. The larger the value, the higher the potential risk.

[0090] When the behavior state back-inference score (BSS) approaches 1, it indicates that the current maintenance behavior sequence characteristics highly match the characteristics of historical abnormal samples, and the equipment has a potential trend of operational degradation or hidden risks.

[0091] When the behavioral state back-inference score (BSS) approaches 0, it indicates that the overall characteristics of the maintenance behavior are similar to the historical health state samples, and the equipment operation status is relatively stable.

[0092] The approximation range is defined by the user.

[0093] The deep sequence modeling network is implemented using a gated recurrent neural network structure. By modeling the time dimension of the input sequence SEQ, it extracts the dynamic change patterns in the maintenance behavior feature sequence, thereby completing the implicit back-inference modeling of the equipment status trend.

[0094] S5 includes S51;

[0095] S51. The behavioral state back-inference score BSS is fused with the health assessment score MON calculated by the physical quantity acquisition model in the main monitoring channel of the equipment to generate a fused health status assessment value FUS. The health assessment score MON is a status score output by the main monitoring model after feature extraction and risk assessment based on electrical, electrothermal and mechanical operating data collected in the online monitoring system of the equipment. Its data sources include current waveform, voltage distortion, partial discharge amplitude and infrared temperature distribution sensing signal. The health assessment score MON has an independent risk identification function in the main monitoring system, and its range is between 0 and 1. It is used to represent the degree of health risk of the equipment in the physical monitoring dimension. The fused health status assessment value FUS generated by combining the behavioral state back-inference score BSS with the health assessment score MON is used to comprehensively reflect the health risk of the equipment in the two dimensions of maintenance behavior performance and physical operation status.

[0096] When the fusion health status assessment value FUS is greater than the set threshold θ, the device is determined to enter the early warning state, triggering the corresponding operation and maintenance response action or risk control strategy.

[0097] When the fused health status assessment value FUS is less than or equal to the set threshold θ, it is determined that the device will not enter the warning state and will not trigger operation and maintenance response actions or risk control strategies.

[0098] In this embodiment, a fusion judgment of behavioral risk and physical perception information is achieved, constructing a power equipment health assessment system with enhanced trend perception and response accuracy. Based on the previously constructed maintenance behavior input sequence SEQ, a deep sequence modeling network is used to learn the dynamic evolution characteristics of maintenance behavior over time. A gated recurrent neural network structure is used to extract potential behavioral pattern changes in the sequence, outputting a behavioral state back-inference score (BSS) to measure equipment risk trends. By comparing the feature similarity between the maintenance behavior sequence features and historical abnormal state samples, potential deterioration signs hidden in the equipment before obvious sensor anomalies can be effectively identified. Furthermore, the behavioral state back-inference score (BSS) is fused with the health assessment score (MON) generated from real-time operational data analysis in the equipment's main monitoring channel to generate a fused health status assessment value (FUS). This achieves a comprehensive quantitative judgment of the risk characteristics of power equipment under both the "behavioral performance dimension" and the "physical perception dimension." When the fused health status assessment value (FUS) exceeds the user-defined health threshold, the system automatically determines that the equipment is in a warning state and triggers a response operation. This fusion mechanism, while maintaining the stability of physical quantity perception, significantly improves the system's ability to proactively identify abnormal state trends. It enables monitoring strategies to move beyond relying on single, real-time physical parameter fluctuations and instead proactively identify early signs of faults by combining behavioral evolution trajectories. This enhances the overall performance of equipment health management in terms of timely early warning, reasonable intervention, and the accuracy of monitoring closed loops.

[0099] Example 5

[0100] A deep learning-based integrated monitoring system for the status of power equipment; please refer to [reference needed]. Figure 2 Specifically, it includes a device behavior data acquisition module, a deviation feature generation module, a semantic construction module, a device behavior status generation module, and a status assessment and decision-making module;

[0101] The device behavior data acquisition module collects device maintenance behavior information and constructs a structured feature set (MFD) of device maintenance behavior by obtaining maintenance timestamps, maintenance operation type codes, and operation durations from the operation and maintenance records.

[0102] The deviation feature generation module calculates the average cycle and fluctuation deviation of maintenance operations based on the feature set MFD, and generates the maintenance behavior rhythm feature PTD.

[0103] The semantic construction module performs semantic embedding processing on the feature set MFD, generates a semantic vector set EMB for maintenance operations through the operation and maintenance knowledge semantic model, and constructs the input sequence SEQ by combining the maintenance behavior rhythm feature PTD feature.

[0104] The device behavior state generation module inputs the input sequence SEQ into a deep sequence modeling network to construct an inference model between maintenance behavior and device health state, and outputs the behavior state back-inference score BSS.

[0105] The status assessment and decision-making module integrates the behavioral status back-inference score (BSS) with the health assessment score (MON) calculated based on the physical quantity acquisition model in the main monitoring channel of the equipment to generate a fused health status assessment value (FUS), which determines the degree of health risk of the power equipment and triggers response actions.

[0106] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for integrated monitoring of the status of power equipment based on deep learning, characterized in that: Includes the following steps: S1. Collect equipment maintenance behavior information. By obtaining the maintenance timestamp, maintenance operation type code and operation duration from the operation and maintenance records, construct a structured feature set (MFD) of equipment maintenance behavior. S2. Calculate the average cycle and fluctuation deviation of maintenance operations based on the structured feature set MFD, and generate the maintenance behavior rhythm feature PTD; S3. Perform semantic embedding processing on the structured feature set MFD, generate the semantic vector set EMB of maintenance operations through the operation and maintenance knowledge semantic model, and construct the input sequence SEQ by combining the maintenance behavior rhythm feature PTD feature. S4. Input the input sequence SEQ into the deep sequence modeling network to build an inference model between maintenance behavior and equipment health status, and output the behavior status back-inference score BSS. S5. The behavioral state back-inference score BSS is fused with the health assessment score MON calculated based on the physical quantity acquisition model in the main monitoring channel of the equipment to generate a fused health status assessment value FUS, which determines the degree of health risk of the power equipment and triggers response actions.

2. The method for integrated monitoring of power equipment status based on deep learning according to claim 1, characterized in that: S1 includes S11; S11. By extracting maintenance behavior records related to the target equipment from the database or log system of the power equipment operation and maintenance system, the records include the timestamp of the maintenance behavior, the operation type and the duration of the maintenance operation, and obtaining the original behavior field set RFD={tm(i),op(i),dur(i)}, where tm(i) represents the timestamp of the i-th maintenance operation in the record, op(i) represents the type code of the i-th maintenance operation in the record, and dur(i) represents the duration of the i-th maintenance operation in the record; The operation types include preset operation types related to tightening, testing, dust removal, and reset.

3. The method for integrated monitoring of power equipment status based on deep learning according to claim 2, characterized in that: S1 includes S12; S12. The original behavioral field set RFD is expressed in a structured vector form to construct a standard modeling input set MF. Then, the standard modeling input set MF is organized into a time series form to obtain the structured feature set MFD. The structuring process employs a combination of field mapping, rule standardization, and unit unification. Field mapping involves extracting the timestamp tm(i), type code op(i), and duration dur(i) of the i-th maintenance operation from the original behavior field set RFD. Rule standardization enumerates labels for the type code op(i) of the i-th maintenance operation using a unified preset encoding table. Unit unification converts the timestamp tm(i) and duration dur(i) of the i-th maintenance operation into minutes, eliminating the format differences between hours and seconds. Vectorization is performed by concatenating fixed-length numerical vectors. The timestamp tm(i), type code op(i), and duration dur(i) of the i-th maintenance operation after the structured processing are stacked sequentially to obtain a two-dimensional array matrix. The vectorization of the original behavior field set RFD is completed by traversing the original behavior field set RFD to obtain a three-dimensional matrix, and the standard modeling input set MF is obtained. The organization is represented in time series form. Using the unique master index order of the timestamp tm(i) of the vectorized i-th maintenance operation, all maintenance records are sorted to obtain the structured feature set MFD.

4. The method for integrated monitoring of power equipment status based on deep learning according to claim 3, characterized in that: S2 includes S21 and S22; S21. Based on the timestamp tm feature of the time series in the structured feature set MFD, calculate the time interval between each pair of adjacent maintenance behavior operations, and after counting all vector interval values, determine the average maintenance behavior cycle AMT of the equipment in the current analysis cycle, which reflects the maintenance intervention rhythm reference cycle of the equipment under normal working conditions. When the actual maintenance cycle is near the average cycle AMT, it indicates that the equipment maintenance behavior operation is normal. S22. Based on the average time period (AMT) and combined with the timestamp (tm) feature of the time series in the structured feature set (MFD), the deviation between the adjacent time intervals of all maintenance behaviors and the average time period (AMT) is quantified to construct a feature quantity to describe the degree of fluctuation in the rhythm of maintenance behaviors, which is marked as the maintenance behavior rhythm feature (PTD). This reflects the stable periodicity of maintenance behaviors in the time dimension, as well as whether there are rhythmic instabilities caused by abnormal equipment status, frequent failures, and intensive human intervention.

5. The method for integrated monitoring of power equipment status based on deep learning according to claim 4, characterized in that: S3 includes S31; S31. Based on the type-encoded op features in the structured feature set MFD, semantic embedding modeling is performed on the type-encoded op features of each maintenance operation. The type-encoded op of the operation type is mapped to a semantic vector through a predefined operation and maintenance knowledge semantic representation model, forming an operation semantic vector set EMB. The operation semantic vector set EMB is used to represent the feature differences and inter-operation correlations of various maintenance behaviors in the semantic dimension. The semantic representation model is constructed using the Act2Vec model, which is an operation and maintenance semantic embedding model trained based on the text description of maintenance behavior in the historical operation and maintenance logs of power equipment. It achieves implicit semantic modeling of maintenance operation types by learning the vector encoding of the contextual relationship of keywords in maintenance statements.

6. The method for integrated monitoring of power equipment status based on deep learning according to claim 5, characterized in that: S3 includes S32; S32. Jointly model the operation semantic vector set EMB with the timestamp tm and duration dur in the structured feature set MFD, and combine the maintenance behavior rhythm feature PTD to construct the input sequence SEQ of the maintenance behavior. The input sequence SEQ is used to describe the structural attributes, semantic expression and rhythm changes of each maintenance behavior in time sequence.

7. The method for integrated monitoring of power equipment status based on deep learning according to claim 6, characterized in that: in, Each time step in the input sequence SEQ consists of the following feature concatenation steps: S321. Extract the type code op, timestamp tm, and duration dur features corresponding to each maintenance behavior record from the structured feature set MFD, and sort them in ascending order according to the timestamp tm. These features are used to construct the structural attribute features of each time step in the input sequence to ensure that the maintenance behavior has temporal logical continuity in time series modeling. S322. Extract the operation semantic vector corresponding to each structured maintenance record from the operation semantic vector set EMB. The operation semantic vector is the vector obtained by converting the type encoding field op in the structured feature set MFD through the Act2Vec model in the previous step S31. It is used to represent the inherent category features and operation similarity information of the maintenance operation in the semantic dimension in the input sequence. S323. The maintenance behavior rhythm feature PTD is broadcast as a unified global time rhythm feature, so that the maintenance behavior rhythm feature PTD is consistent with the number of structured records in the input sequence, generating a rhythm vector sequence that is aligned one-to-one with the maintenance records, and then concatenating it with the output of steps S321 and S322 in the time step dimension to complete the feature fusion of the maintenance behavior input sequence SEQ.

8. The method for integrated monitoring of power equipment status based on deep learning according to claim 7, characterized in that: S4 includes S41; S41. Input the input sequence SEQ into the deep sequence modeling network for modeling processing, construct an inference model for inferring the relationship between maintenance behavior and equipment health status, and output the behavior state back-inference score BSS for quantifying the potential operating risks of the equipment. The behavior state back-inference score BSS is used to quantify whether the current maintenance behavior implies a trend of equipment deterioration. The value range is between 0 and 1. The larger the value, the higher the potential risk. When the behavior state back-inference score (BSS) approaches 1, it indicates that the current maintenance behavior sequence characteristics highly match the characteristics of historical abnormal samples, and the equipment has a potential trend of operational degradation or hidden risks. When the behavioral state back-inference score (BSS) approaches 0, it indicates that the overall characteristics of the maintenance behavior are similar to the historical health state samples, and the equipment operation status is relatively stable. The approximation range is defined by the user. The deep sequence modeling network is implemented using a gated recurrent neural network structure. By modeling the time dimension of the input sequence SEQ, it extracts the dynamic change patterns in the maintenance behavior feature sequence, thereby completing the implicit back-inference modeling of the equipment status trend.

9. The method for integrated monitoring of power equipment status based on deep learning according to claim 8, characterized in that: S5 includes S51; S51. The behavioral state back-inference score BSS is fused with the health assessment score MON calculated by the physical quantity acquisition model in the main monitoring channel of the equipment to generate a fused health status assessment value FUS. The health assessment score MON is a status score output by the main monitoring model after feature extraction and risk assessment based on electrical, electrothermal and mechanical operating data collected in the online monitoring system of the equipment. Its data sources include current waveform, voltage distortion, partial discharge amplitude and infrared temperature distribution sensing signal. The health assessment score MON has an independent risk identification function in the main monitoring system, and its range is between 0 and 1. It is used to represent the degree of health risk of the equipment in the physical monitoring dimension. The fused health status assessment value FUS generated by combining the behavioral state back-inference score BSS with the health assessment score MON is used to comprehensively reflect the health risk of the equipment in the two dimensions of maintenance behavior performance and physical operation status. When the fusion health status assessment value FUS is greater than the set threshold θ, the device is determined to enter the early warning state, triggering the corresponding operation and maintenance response action or risk control strategy. When the fused health status assessment value FUS is less than or equal to the set threshold θ, it is determined that the device will not enter the warning state and will not trigger operation and maintenance response actions or risk control strategies.

10. A deep learning-based integrated monitoring system for the status of power equipment, applied to the deep learning-based integrated monitoring method for the status of power equipment as described in any one of claims 1 to 9, characterized in that: It includes a device behavior data acquisition module, a deviation feature generation module, a semantic construction module, a device behavior status generation module, and a status assessment and decision-making module; The device behavior data acquisition module collects device maintenance behavior information and constructs a structured feature set (MFD) of device maintenance behavior by obtaining maintenance timestamps, maintenance operation type codes, and operation durations from the operation and maintenance records. The deviation feature generation module calculates the average cycle and fluctuation deviation of maintenance operations based on the structured feature set MFD, and generates the maintenance behavior rhythm feature PTD. The semantic construction module performs semantic embedding processing on the structured feature set MFD, generates a semantic vector set EMB for maintenance operations through the operation and maintenance knowledge semantic model, and constructs the input sequence SEQ by combining the maintenance behavior rhythm feature PTD feature. The device behavior state generation module inputs the input sequence SEQ into a deep sequence modeling network to construct an inference model between maintenance behavior and device health state, and outputs the behavior state back-inference score BSS. The status assessment and decision-making module integrates the behavioral status back-inference score (BSS) with the health assessment score (MON) calculated based on the physical quantity acquisition model in the main monitoring channel of the equipment to generate a fused health status assessment value (FUS), which determines the degree of health risk of the power equipment and triggers response actions.

Images