Equipment fault diagnosis method and system for photovoltaic power station application

Abstract

The invention discloses an equipment fault diagnosis method and system for photovoltaic power station application. The method comprises the following steps: acquiring multi-source acquisition data related to operation of target equipment, performing pre-processing, performing multi-scale feature extraction, and inputting features into a sparse attention time sequence model to generate at least two types of fault evidences and evidence representations thereof; calculating the similarity between the evidences based on the evidence representation to construct an evidence similarity matrix, determining the conflict degree of the evidences through the similarity matrix, and executing the adaptive calibration processing of conflict suppression on the to-be-fused fault evidences through the conflict degree; and performing secondary evidence fusion on the calibrated fault evidence, and outputting a fault diagnosis result and confidence thereof. Through quantification of evidence consistency and conflict degree and self-adaptive suppression of conflict evidence influence before fusion, the robustness and credibility of a diagnosis result are improved in a multi-source heterogeneous and evidence inconsistent scene, the risk of misjudgment and missed judgment is reduced, and the engineering applicability is enhanced.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power plant operation and maintenance monitoring technology, and in particular to a method and system for diagnosing equipment faults for photovoltaic power plant applications. Background Technology

[0002] Photovoltaic power plants typically include various types of equipment such as inverters, combiner boxes, transformer / step-up transformers, switchgear, DC-side cables and connectors, and tracking bracket drive and control units. These devices operate outdoors under complex conditions for extended periods, making them susceptible to factors such as temperature, humidity, irradiance fluctuations, dust contamination, component aging, poor contact, and communication link fluctuations. This can lead to frequent alarms, reduced efficiency, and abnormal shutdowns. To reduce operation and maintenance costs and improve power generation reliability, engineering projects typically use monitoring systems to continuously collect and analyze equipment operating status. This generates multi-source data, including continuous monitoring information such as electrical parameters, power, temperature, and efficiency, as well as event and status information such as alarm codes, protection actions, switch status, communication status, work orders, and maintenance records. This data provides a basis for fault diagnosis.

[0003] Existing fault diagnosis technologies for photovoltaic power plants mainly include: threshold- and rule-based diagnostic methods, mechanistic model- and state estimation-based methods, machine learning / deep learning-based pattern recognition methods, and fusion diagnostic methods for multi-source information. Threshold- and rule-based methods are simple to implement but sensitive to changes in operating conditions and struggle to cover complex fault scenarios. Mechanistic model methods rely on equipment parameters and modeling accuracy, and are prone to misjudgments when equipment ages, operating conditions drift, or data is missing. Data-driven intelligent diagnostic methods can learn features from historical data, but in situations where photovoltaic power plant fault samples are scarce, noise levels are high, data is missing, time alignment is difficult, and multi-source heterogeneous data are coupled, model outputs often exhibit instability. Furthermore, inconsistencies in evidence often occur between multi-source data, such as abnormal continuous monitoring curves without alarm triggering, alarm triggering without significant deviation in operating parameters, and conflicting indications from different monitoring points. This leads to over-reliance on a particular type of data in the fusion diagnostic results or decreased reliability in conflicting situations.

[0004] Common approaches to multi-source evidence fusion methods include fixed-weighting, voting strategies, or the introduction of evidence theory / probabilistic inference. However, in practical applications, the credibility of multi-source evidence fluctuates dynamically with changes in operating conditions, sensor quality, communication status, and noise levels. When there is significant conflict between pieces of evidence, the lack of a quantification and adaptive processing mechanism for the degree of conflict can lead to biased results, distorted confidence levels, or difficulty in achieving stable convergence. Therefore, improving the robustness and credibility of fault diagnosis results under multi-source heterogeneous data conditions, particularly in scenarios with inconsistencies and conflicts in evidence, remains a key technical challenge that needs to be addressed in this field. Summary of the Invention

[0005] In response to the aforementioned technical problems commonly found in the fault diagnosis of existing photovoltaic power plant equipment, this invention proposes a fault diagnosis method and system for photovoltaic power plant applications, aiming to improve the robustness and credibility of diagnostic results in scenarios of conflicting evidence.

[0006] To achieve the above objectives, the present invention provides a method for diagnosing equipment faults in photovoltaic power plant applications, the method comprising: Acquire multi-source data related to the operation of target equipment in photovoltaic power plants; Multi-scale feature extraction is performed on the preprocessed multi-source acquisition data, and at least two types of fault evidence for equipment fault diagnosis are generated by the sparse attention time series model. Evidence representations corresponding to each fault evidence are generated, and the evidence representations are feature information used to characterize the corresponding fault evidence pattern features. The similarity between evidences is calculated based on the evidence representation to construct an evidence similarity matrix, and the evidence conflict degree is determined by the evidence similarity matrix. The evidence conflict degree is used to perform adaptive calibration processing for conflict suppression on the faulty evidence to be fused. The evidence conflict degree is a conflict index used to characterize the degree of inconsistency between different faulty evidences. A second evidence fusion is performed on the fault evidence after adaptive calibration to obtain the equipment fault diagnosis results and their confidence levels.

[0007] More preferably, the multi-source acquisition data includes at least one of the following: continuous monitoring information, which includes time-series data of any one or a combination of electrical parameters, power, temperature, and efficiency; and event and status information, which includes any one or a combination of alarm codes, protection actions, switch status, communication status, work order record fields, and maintenance record fields.

[0008] More preferably, the preprocessing includes at least one of the following operations: missing value processing, outlier processing, time alignment, noise reduction and smoothing, normalization, feature filtering, segmentation by working condition, and slicing by window; the feature filtering includes any one or a combination of correlation analysis, mutual information analysis, and distance correlation analysis.

[0009] More preferably, the multi-scale feature extraction process includes: using a parallel or cascaded multi-receptive-field one-dimensional convolutional structure or temporal filtering structure to extract features from the preprocessed multi-source acquisition data, so as to obtain multi-scale features representing short-term disturbances and long-term trends respectively, and fusing the multi-scale features to form combined features, and inputting the combined features into a sparse attention temporal model for temporal modeling.

[0010] More preferably, the sparse attention temporal model includes a sparse attention branch model and a dense attention branch model; The process of generating at least two types of fault evidence for equipment fault diagnosis by the sparse attention temporal model includes: performing temporal modeling on the combined features of the input by the sparse attention branch model and the dense attention branch model respectively to obtain the first fault evidence and the second fault evidence, and performing adaptive fusion processing to obtain fused evidence; and calculating the reliability index of the fused evidence based on the attention weight distribution generated by the sparse attention branch model and the dense attention branch model. The reliability metrics include any one or a combination of attention entropy, attention sparsity, and cross-scale consistency, used for adaptive calibration processing of evidence similarity matrix construction or conflict suppression.

[0011] More preferably, the evidence characterization is obtained synchronously when generating fault evidence and is associated with the corresponding fault evidence; The evidence representation includes at least one of the following data: intermediate hidden feature vectors of a sparse attention time series model; attention weight distribution vectors or matrices, including any one or a combination of time step weights, variable weights, and cross-scale weights; and data obtained by dimensionality reduction or normalization of intermediate hidden feature vectors or attention weight distribution vectors or matrices.

[0012] Further preferably, the generation of the evidence conflict degree includes: calculating the support degree of each faulty piece of evidence based on the similarity value between each faulty piece of evidence and the remaining faulty pieces of evidence in the evidence similarity matrix, wherein the support degree is used to characterize the degree of consistency between the current faulty piece of evidence and the remaining faulty pieces of evidence; generating the local conflict degree of each faulty piece of evidence based on the support degree, wherein the local conflict degree is negatively correlated with the support degree; and aggregating the local conflict degrees of each faulty piece of evidence to obtain the global conflict degree.

[0013] More preferably, the adaptive calibration process includes: determining corresponding calibration parameters for each piece of fault evidence based on global or local conflict degree, wherein the calibration parameters are used to adjust the contribution of fault evidence in secondary evidence fusion; Further, it includes: dividing the fault evidence to be fused into a high-similarity evidence set and a low-similarity evidence set based on the similarity value of the evidence similarity matrix; for the low-similarity evidence set or fault evidence whose local conflict degree meets the preset calibration conditions, performing any one or a combination of weight reduction, elimination, or reallocation of evidence quality to a determined candidate fault set according to the calibration parameters; the preset calibration conditions include any one or a combination of local conflict degree exceeding a preset conflict threshold or local conflict degree being located in a preset ranking interval; the determined candidate fault set is a set corresponding to a single candidate fault.

[0014] More preferably, the adaptive calibration process further includes: performing prior estimation and weighting processing on fault evidence based on global conflict degree or local conflict degree, and using Bayesian estimation to suppress the quality of evidence pointing to an uncertain candidate fault set or to recover it to a definite candidate fault set; the uncertain candidate fault set is a set containing at least two candidate faults.

[0015] More preferably, the secondary evidence fusion adopts a combination fusion rule based on DS evidence theory, including: In the first stage, highly similar evidence sets are fused to obtain intermediate fused evidence; In the second stage, the intermediate fused evidence is further fused with the remaining fault evidence after adaptive calibration to obtain the equipment fault diagnosis results and their confidence levels.

[0016] To achieve the above method, the present invention also provides an equipment fault diagnosis system for photovoltaic power plant applications, the system comprising: The data acquisition unit is used to acquire multi-source data related to the operation of target equipment in a photovoltaic power station. The preprocessing unit is used to preprocess multi-source acquired data; The feature extraction and evidence generation unit is used to perform multi-scale feature extraction on the preprocessed multi-source collected data, and input the extraction results into the sparse attention time series model to generate at least two types of fault evidence for equipment fault diagnosis by the sparse attention time series model, and generate evidence representations corresponding to each fault evidence, wherein the evidence representations are feature information used to characterize the corresponding fault evidence pattern features. The similarity matrix and conflict degree determination unit is used to calculate the similarity between evidences based on evidence characterization to construct an evidence similarity matrix, and to determine the evidence conflict degree from the evidence similarity matrix. The evidence conflict degree is a conflict index used to characterize the degree of inconsistency between different faulty evidence. An adaptive calibration unit is used to perform adaptive calibration processing to suppress conflict in the fault evidence to be fused, utilizing the degree of evidence conflict. The secondary evidence fusion unit is used to perform secondary evidence fusion on the fault evidence after adaptive calibration to obtain the equipment fault diagnosis results and their confidence levels. The result output unit is used to output the device fault diagnosis results and their confidence levels.

[0017] Compared with the prior art, the method and system of the present invention have at least the following beneficial effects: By constructing an evidence similarity matrix and generating evidence conflict degree, a quantitative characterization of the degree of inconsistency of multi-source evidence can be achieved, which can improve the controllability and stability of the fusion process in evidence conflict scenarios. By using conflict suppression adaptive calibration based on conflict degree, the contribution of the fault evidence to be fused can be dynamically adjusted, which can reduce the negative impact of abnormal evidence, noise, missing data and operating condition fluctuations on the diagnostic results and improve the robustness of the diagnosis. By combining multi-scale feature extraction and sparse attention temporal modeling, both short-term perturbations and long-term trend information are taken into account, thereby enhancing the ability to represent fault modes at different time scales. By using secondary evidence fusion to output diagnostic results and confidence levels, the results can be expressed with better credibility. Furthermore, in some implementations, the combination of Bayesian estimation and DS evidence theory fusion rules can be combined to further enhance conflict resolution and fusion convergence effects. The solution is highly modular, making it easy to integrate and deploy with existing monitoring and operation and maintenance systems of photovoltaic power plants, and has good engineering applicability. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the operation flow of the equipment fault diagnosis method provided in the embodiments of the present invention; Figure 2 This is a schematic diagram of the equipment fault diagnosis method and system structure provided in the embodiments of the present invention. Detailed Implementation

[0019] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention are only used to explain the present invention and not to limit the scope of protection of the present invention; equivalent substitutions or modifications made by those skilled in the art to the embodiments without departing from the concept of the present invention should all fall within the scope of protection of the present invention.

[0020] In actual operation and maintenance scenarios of photovoltaic power plants, the operating status of target equipment such as inverters, combiner boxes, transformer / step-up transformers, switchgear, strings, and their communication / acquisition units is usually characterized by both continuous monitoring information and event / status information. The former includes time-series data such as electrical parameters, power, temperature, and efficiency, while the latter includes alarm codes, protection actions, switch status, communication status, and work order / maintenance record fields. Due to differences in sensor accuracy, communication link jitter, fluctuations in operating conditions (irradiance, temperature and humidity, load changes, etc.), and factors such as missing data and noise, different data sources often provide inconsistent prompts for the same fault phenomenon. This manifests as abnormal continuous curves but no alarms triggered, alarms triggered but key electrical parameters not significantly deviating, and contradictory judgments from different monitoring points or different time windows. Such inconsistencies / conflicts in multi-source evidence lead to unstable diagnostic results from traditional fixed-weight fusion, simple voting, or single-model outputs. Especially under conditions of strong conflict, scarce samples, or operating condition drift, confidence distortion, misjudgment, or omissions are likely to occur, thereby reducing the efficiency of the operation and maintenance closed loop.

[0021] Based on the above problems, this invention proposes a method for equipment fault diagnosis in photovoltaic power plant applications. For example... Figure 1 As shown, the method includes: Acquire multi-source data related to the operation of target equipment in photovoltaic power plants; Multi-scale feature extraction is performed on the preprocessed multi-source acquisition data, and at least two types of fault evidence for equipment fault diagnosis are generated by the sparse attention time series model. Evidence representations corresponding to each fault evidence are generated, and the evidence representations are feature information used to characterize the corresponding fault evidence pattern features. The similarity between evidences is calculated based on the evidence representation to construct an evidence similarity matrix, and the evidence conflict degree is determined by the evidence similarity matrix. The evidence conflict degree is used to perform adaptive calibration processing for conflict suppression on the faulty evidence to be fused. The evidence conflict degree is a conflict index used to characterize the degree of inconsistency between different faulty evidences. A second evidence fusion is performed on the fault evidence after adaptive calibration to obtain the equipment fault diagnosis results and their confidence levels.

[0022] The multi-source acquired data is used to comprehensively reflect the operating status and event status of the target equipment. This data can be time-series data formed by continuous sampling, or event / status data describing equipment alarms, protection, switching, communication, and maintenance processes. Due to the different data sources, sampling frequencies, and varying characteristics of missing and noise levels, directly inputting this data into the model often introduces bias. This invention preprocesses the multi-source data before core modeling and fusion to improve data quality and ensure time consistency alignment, allowing subsequent feature extraction and evidence generation to be performed on a unified data basis, thereby reducing the impact of data defects on diagnostic stability.

[0023] Photovoltaic power plant equipment failures typically exhibit differences across time scales. Some failures manifest as short-term abrupt changes or disturbances, while others show long-term slow drifts or abnormal trends. To avoid feature omissions due to relying solely on a single time scale, this invention performs multi-scale feature extraction on preprocessed data. This allows the resulting features to simultaneously encompass both short-term disturbance information and long-term trend information, forming input features for subsequent time series modeling and thus improving adaptability to different types of failure modes.

[0024] This invention uses a sparse attention temporal model to perform temporal modeling of multi-scale features, generating at least two types of fault evidence for fault diagnosis. Here, fault evidence refers to output information that supports fault judgment, reflecting the confidence tendency, anomaly degree, or diagnostic support strength of candidate faults, and is used for subsequent fusion decision-making. Simultaneously, this invention generates evidence representations corresponding to each piece of fault evidence. These evidence representations are feature information characterizing the pattern features of the corresponding fault evidence, used to depict the internal pattern differences that lead to the judgment; they can be understood as a structured description of the fault evidence. The introduction of evidence representations allows for the measurement of consistency or difference between different pieces of evidence from the perspective of evidence pattern similarity, beyond the evidence content itself, providing a foundation for conflict identification and conflict governance.

[0025] This invention calculates the similarity between pieces of evidence based on evidence characterization and constructs an evidence similarity matrix. The evidence similarity matrix describes the degree of consistency among different faulty pieces of evidence at the level of pattern features. Higher similarity indicates a more consistent pattern among the different pieces of evidence; lower similarity indicates a greater difference in patterns between the evidence. Based on this, the degree of evidence conflict is determined. The degree of evidence conflict is a conflict index used to characterize the degree of inconsistency between different faulty pieces of evidence, quantifying the degree of divergence among current multi-source evidence. By transforming conflict from a phenomenological description into a calculable conflict index, this invention enables controllable handling of conflict before fusion, rather than passively bearing the fluctuations in results caused by conflict after fusion.

[0026] This invention utilizes the degree of evidence conflict to perform adaptive calibration for conflict suppression on faulty evidence to be fused. Adaptive calibration refers to dynamically adjusting the fusion contribution of different faulty pieces of evidence based on the degree of evidence conflict, enabling the fusion process to suppress the negative impact of inconsistent evidence when conflict is strong and maintain the effective contribution of evidence when conflict is weak. This calibration can be applied to the contribution of evidence, the allocation of evidence quality, or equivalent fusion inputs, making subsequent fusion more likely to converge to a stable and reliable diagnostic conclusion.

[0027] After completing the adaptive calibration process for conflict suppression, this invention performs secondary evidence fusion on the fault evidence to obtain the equipment fault diagnosis results and their confidence levels. Secondary fusion is used to form a phased fusion structure in the fusion strategy, ensuring that the fusion process balances the aggregation effect of consistent evidence with the supplementary effect of remaining evidence. This maintains the usability of the diagnostic conclusions while providing a confidence level output that can be used for operational and maintenance decisions. The confidence level characterizes the credibility of the diagnostic results, facilitating closed-loop processing by operations and maintenance personnel in conjunction with alarms, comparative analysis, and historical records.

[0028] In summary, the method of this invention, based on the preprocessing and multi-scale feature extraction of multi-source data, generates at least two types of fault evidence through a sparse attention temporal model and simultaneously generates evidence representations corresponding to each fault evidence; further, it constructs an evidence similarity matrix and generates evidence conflict degree through evidence representations, uses the evidence conflict degree to perform adaptive calibration processing for conflict suppression of the evidence to be fused, and finally performs secondary evidence fusion on the calibrated evidence to obtain the diagnostic results and their confidence levels, thereby improving the robustness and credibility of the diagnostic results in evidence conflict scenarios.

[0029] In one specific embodiment, the multi-source acquisition data obtained by the present invention includes at least continuous monitoring information and / or event and status information, which are used to characterize the operating status and fault symptoms of the target equipment of the photovoltaic power station from different dimensions.

[0030] Continuous monitoring information consists of time-series data collected continuously according to a preset sampling period, used to reflect the dynamic changes of the target equipment during operation. This continuous monitoring information may include, but is not limited to, any one or a combination of electrical parameter data, power data, temperature data, and efficiency data. For example, when the target equipment is an inverter, the continuous monitoring information may include time-series data such as DC-side voltage / current, AC-side voltage / current, output power, internal temperature, and conversion efficiency; when the target equipment is a combiner box or DC-side acquisition unit, the continuous monitoring information may include time-series data such as branch current, bus voltage, and cabinet temperature.

[0031] Event and status information consists of discrete information describing equipment operation events, status changes, and maintenance processes. It supplements the abnormal triggering causes and operational contexts that are difficult to express directly through continuous curves. This event and status information may include, but is not limited to, any one or a combination of alarm codes, protection actions, switch statuses, communication statuses, and work order record fields and maintenance record fields. For example, alarm codes can characterize the type of abnormality triggered by equipment self-testing or monitoring logic; protection actions can characterize the triggering of overvoltage, overcurrent, overtemperature, and other protection mechanisms; switch statuses can characterize status changes such as grid connection / disconnection and circuit breaker opening / closing; communication statuses can characterize whether there are disconnections, packet loss, or delay anomalies in the data acquisition link; and work order and maintenance record fields can characterize information such as manual handling results, component replacement, reset operations, and whether the fault was reproduced.

[0032] By simultaneously introducing continuous monitoring information and event / status information, this embodiment can comprehensively characterize the fault by combining alarm and handling context, in addition to curve trends and disturbance characteristics, thereby providing a more sufficient data foundation for subsequent evidence generation, similarity measurement, conflict suppression calibration and secondary evidence fusion.

[0033] In one specific embodiment, before performing multi-scale feature extraction and temporal modeling on the multi-source acquired data, preprocessing is performed on the acquired multi-source acquired data to improve data quality, eliminate differences in time and scale between data from different sources, and reduce the impact of noise and anomalies on subsequent evidence generation and fusion. Preprocessing may include at least one of the following operations, and may be combined and configured according to the target device type and data quality status.

[0034] For missing value handling, when there are missing values ​​in continuous monitoring information or event / status information, methods such as forward filling, linear interpolation, sliding window mean filling, or estimation filling based on adjacent devices / adjacent channels can be used to fill the missing values. For data segments that are continuously missing for more than a preset length, they can be marked as invalid segments and removed in subsequent slicing to avoid introducing uncontrollable biases.

[0035] Outlier handling: When outliers, sharp drops, sudden rises, or values ​​exceeding the physical reasonable range appear on the monitoring curve, they can be identified based on statistical thresholds such as quantile range and 3σ range or based on the rated range of the equipment, and then processed by methods such as cropping, replacing with the neighborhood median, and local regression correction. For records with obvious conflicts or repetitions in event / status information, deduplication and consistency correction can be performed based on timestamp and event code rules.

[0036] Time alignment is crucial because continuous monitoring information and event / state information have different sampling frequencies and triggering methods. To ensure that subsequent feature extraction and evidence generation are performed on a unified time axis, all data can be mapped to a unified time granularity. For example, resampling can be performed at fixed time steps, continuous monitoring information can be interpolated or aggregated, and event / state information can be counted within a time window, the most recent state can be maintained, or event triggering can be marked, thereby forming an alignable multi-source sequence.

[0037] Noise reduction and smoothing: To reduce the interference of sensor noise, communication jitter or measurement fluctuations on the identification of short-term disturbances, continuous monitoring curves can be processed by moving average, exponential moving average, median filtering or wavelet denoising; when it is necessary to retain the characteristics of sudden changes, an edge-preserving smoothing strategy can be adopted to avoid excessive smoothing that will flatten the fault peaks.

[0038] Normalization, in order to eliminate the influence of different dimensions (voltage, current, temperature, power, etc.) on model training and similarity calculation, can be achieved by using min-max normalization, standardization, or normalization according to the ratio of rated value / nominal value, so that multi-source features are on a comparable scale and the stability of subsequent evidence representation and similarity matrix construction can be improved.

[0039] Feature selection, to reduce the dilution of the model's representational ability by redundant and noisy features, can be performed on candidate features during the preprocessing stage. Feature selection includes any one or a combination of correlation analysis, mutual information analysis, and distance correlation analysis. Correlation analysis filters features based on their correlation with target quantities (such as alarm triggering, fault labels, and anomaly scores) or the correlation between features. Mutual information analysis measures the contribution of features to reducing the uncertainty of fault states, retaining features with higher information content. Distance correlation analysis is used to identify nonlinear correlations, filtering out features weakly associated with fault states or eliminating highly redundant features. These methods can be used individually or in combination to obtain a feature subset for subsequent multi-scale feature extraction.

[0040] By segmenting the data according to operating conditions, considering that changes in operating conditions such as irradiation, temperature, and load can lead to changes in the distribution of normal operating characteristics, the data can be segmented based on operating condition indicators (such as irradiation range, temperature range, power range, grid-connected / disconnected status). This allows subsequent feature extraction and evidence generation to be carried out under relatively consistent operating conditions, thereby reducing the risk of misjudgment caused by operating condition drift.

[0041] By slicing by window, in order to adapt to time series modeling and multi-scale feature extraction, the aligned multi-source data can be sliced ​​by sliding window or fixed window to obtain multiple time window samples; the window length and step size can be determined according to the fault response time scale of the target device, so that both short-term disturbances and long-term trends can be covered at the sample granularity.

[0042] Through the above preprocessing operations, this embodiment achieves the control of missing and abnormal data, time axis unification, noise suppression and dimensional alignment before entering multi-scale feature extraction and sparse attention temporal modeling. When necessary, it forms a data representation that is more suitable for modeling and fusion through feature filtering, working condition segmentation and window slicing, providing a reliable input basis for subsequent fault evidence generation, evidence similarity matrix construction and conflict suppression calibration.

[0043] In one specific embodiment, to simultaneously capture the performance characteristics of target equipment faults at different time scales, after preprocessing the multi-source acquired data, multi-scale feature extraction is performed on the preprocessed multi-source acquired data, and the multi-scale features are fused to form combined features for time series modeling. Multi-scale feature extraction can be performed by constructing multiple one-dimensional convolutional structures or temporal filtering structures with different receptive fields in a parallel or cascaded manner to extract features from the aligned multi-source time series data. In the parallel mode, at least two scale branches can be set. The first scale branch uses a smaller receptive field to highlight short-term disturbance features, such as instantaneous electrical parameter jumps, spike noise, and local abnormal fluctuations. The second scale branch uses a larger receptive field to highlight long-term trend features, such as slow efficiency decay, temperature rise trends, and continuous power deviation. Each branch extracts features from the data within the same time window and outputs the corresponding multi-scale features. In the cascaded mode, a smaller receptive field can be used for preliminary feature extraction to retain fine-grained disturbance information, and then the output can be fed into a larger receptive field structure to integrate contextual information over a longer time range, thereby forming hierarchical multi-scale features.

[0044] After obtaining short-term perturbation features and long-term trend features, a fusion process is performed on the multi-scale features to form a combined feature. The fusion method can be concatenation, weighted summation, gated fusion, or other equivalent fusion methods to integrate key information from different time scales into a unified feature representation. This combined feature is used to simultaneously carry short-term anomaly and long-term drift information, facilitating subsequent time-series models to extract diagnostic evidence within a unified input space.

[0045] Using the aforementioned combined features as input to a sparse attention temporal model, temporal modeling is performed to learn the dynamic correlations and fault evolution patterns of multi-source information over time. Since the combined features have integrated multi-scale information, the sparse attention temporal model can consider both local critical time periods and global trend context during the modeling process, laying the input foundation for subsequent generation of fault evidence and its representation.

[0046] Through the above implementation methods, it is possible to simultaneously cover two typical fault manifestations, namely short-term disturbances and long-term trends, in the same diagnostic link, reduce the risk of missed detection caused by single-scale features, and provide more discriminative feature inputs for subsequent sparse attention time-series modeling, evidence generation, conflict governance, and fusion output.

[0047] In one specific embodiment, the aforementioned combined features are input into a sparse attention temporal model for temporal modeling to generate fault evidence for equipment fault diagnosis, and simultaneously generate a reliability index for subsequent conflict resolution and fusion. In this embodiment, the sparse attention temporal model includes a sparse attention branch model and a dense attention branch model; further, the process of the sparse attention temporal model generating at least two types of fault evidence for equipment fault diagnosis includes: performing temporal modeling on the input combined features by the sparse attention branch model and the dense attention branch model respectively to obtain first fault evidence and second fault evidence, and performing adaptive fusion processing to obtain fused evidence; based on the attention weight distribution generated by the sparse attention branch model and the dense attention branch model, calculating the reliability index of the fused evidence; the reliability index includes any one or a combination of attention entropy, attention sparsity, and cross-scale consistency, used for adaptive calibration processing of evidence similarity matrix construction or conflict suppression.

[0048] Sparse attention branching models are used to select a small number of key time segments and / or key variables for focused modeling over a longer time span, highlighting fault symptoms driven by these key segments. Dense attention branching models are used to more comprehensively model the relationships between time steps over the time series, preserving continuous evolutionary information and overall trend context. This dual-branch structure allows for a balance between focusing on key local information and preserving global temporal information during the same modeling process.

[0049] In one implementation, the sparse attention branch model outputs first fault evidence, and the dense attention branch model outputs second fault evidence. The fault evidence characterizes the degree of support the target device provides for candidate fault states within the current time window, and may be confidence tendency information of the candidate fault, anomaly scoring information, or an equivalent expression thereof.

[0050] Furthermore, adaptive fusion processing can be performed on the first and second fault evidence to obtain fused evidence, which integrates the complementary characterization of the fault mode by the two branches. The adaptive fusion processing can dynamically adjust the fusion ratio based on the consistency of the two branch outputs, attention distribution characteristics, or other equivalent information, thereby ensuring that the fused evidence maintains a more stable discriminative ability under different operating conditions and fault modes. To avoid confusion of fusion objects, the adaptive fusion processing in this embodiment is a branch output integration process within the sparse attention temporal model. The branch fusion evidence output is used to comprehensively characterize the complementary information of the two branches; simultaneously, the first and second fault evidence can still be retained as independent evidence and participate in subsequent conflict degree calculations and secondary evidence fusion.

[0051] To measure the credibility of fused evidence and provide a basis for subsequent evidence similarity matrix construction and conflict suppression calibration, in one embodiment, reliability indices are calculated based on the attention weight distributions generated by sparse attention branching models and dense attention branching models, respectively, corresponding to the first and second faulty evidence. Further reliability indices can be calculated for the branch-fused evidence. These reliability indices are used for subsequent evidence similarity matrix construction and / or conflict suppression calibration. The reliability indices may include any one or a combination of attention entropy, attention sparsity, and cross-scale consistency. Attention entropy characterizes the degree of concentration or dispersion of attention distribution, reflecting whether the model forms a clear focus of attention in the time or variable dimensions. Attention sparsity characterizes the degree of focus of attention weights on key time steps or key variables, reflecting whether the evidence is stably supported by a small number of key fragments. Cross-scale consistency characterizes the degree of consistency in attention focus across different scale features or different branch outputs, reflecting the collaborative stability of the evidence under multi-scale information.

[0052] In one specific embodiment, to enable subsequent similarity measurement and conflict identification of different fault evidence at the evidence pattern feature level, a corresponding evidence representation is obtained simultaneously with the generation of fault evidence, and an association is established between the evidence representation and the corresponding fault evidence. Further, the evidence representation includes at least one of the following data: an intermediate hidden feature vector of a sparse attention temporal model; an attention weight distribution vector or matrix, including any one or a combination of time step weights, variable weights, and cross-scale weights; and data obtained by dimensionality reduction or normalization processing from the intermediate hidden feature vector or attention weight distribution vector or matrix.

[0053] In the process of temporal modeling of combined features and outputting fault evidence, the sparse attention temporal model internally forms feature representations and attention distributions to support the output. In this embodiment, when the model outputs the first fault evidence, the second fault evidence, and the fused evidence, it simultaneously outputs or extracts feature information corresponding to the fault evidence from within the model as evidence representation of the fault evidence.

[0054] The synchronous acquisition can be understood as follows: the evidence representation and the fault evidence originate from the same reasoning process, correspond to the same input time window / sample granularity, and are consistent with the fault evidence generation process in time, thereby ensuring that the representation information can truly reflect the pattern basis of the fault evidence.

[0055] To ensure that the subsequent similarity matrix construction and conflict degree generation accurately correspond to specific evidence, this embodiment establishes a corresponding evidence representation entry for each piece of fault evidence and binds them using a unified index or identifier. For example, fault evidence can be marked in the manner of device identifier—time window identifier—evidence type identifier, and the evidence representation indexed by the same identifier is associated with the fault evidence and stored, so that the corresponding evidence representation can be directly called when calculating the similarity between evidence.

[0056] The intermediate hidden feature vector refers to the intermediate layer feature representation formed by the sparse attention temporal model during the generation of fault evidence, used to characterize the high-level semantic or pattern features of the input combination features after temporal modeling; the attention weight distribution vector or matrix is ​​used to characterize the model's attention distribution to different information when generating fault evidence, and may include any one or a combination of time step weights, variable weights, and cross-scale weights, used to reflect the model's attention structure in the time dimension, variable dimension, or scale dimension; and the evidence representation obtained by processing the above data: to facilitate unified similarity calculation and matrix construction, the intermediate hidden feature vector or attention weight distribution vector / matrix can be subjected to dimensionality reduction and / or normalization processing to obtain evidence representation data for similarity calculation. For example, high-dimensional vectors can be mapped to fixed-dimensional representations, or attention matrices can be normalized to distribution representations at comparable scales.

[0057] Through the above implementation methods, evidence characterization can be obtained in a way that is from the same source, within the same window, and indexable with faulty evidence, and provides an input basis for the construction of subsequent evidence similarity matrices in a comparable data form. This allows the consistency and degree of conflict between evidence to be measured and managed at the pattern feature level, thereby supporting the robustness of subsequent conflict suppression calibration and secondary fusion output.

[0058] In one specific embodiment, to quantify the degree of inconsistency between different fault evidences, thereby providing a basis for subsequent adaptive calibration processing for conflict suppression, an evidence similarity matrix is ​​constructed based on evidence representation, and an evidence conflict degree is generated on this basis. The evidence conflict degree may include the local conflict degree corresponding to each fault evidence and the global conflict degree representing the overall conflict level.

[0059] For multiple pieces of fault evidence obtained within the same diagnostic time window, their corresponding evidence representations are first invoked, and the similarity between any two pieces of fault evidence is calculated to form an evidence similarity matrix. The similarity calculation can use similarity or distance metrics commonly used in the field, and when necessary, reliability indicators can be combined to weight or normalize the similarity, so that the similarity value can stably reflect the consistency of the evidence pattern characteristics.

[0060] In one optional implementation, similarity can be obtained by cosine similarity or similarity conversion based on Euclidean distance; support can be obtained by summing the similarity values ​​of the fault evidence and the other fault evidence, and can be combined with the corresponding reliability index for weighted summation; local conflict degree can be obtained from support through complementary mapping or scaling transformation to satisfy the monotonic relationship that the greater the support, the smaller the local conflict degree; global conflict degree can be obtained by aggregating the mean, weighted mean or maximum value of each local conflict degree to characterize the overall conflict level of the evidence in the current time window.

[0061] After obtaining the evidence similarity matrix, for each piece of faulty evidence, the support of that faulty evidence is calculated based on its similarity value with the other faulty evidence. The support is used to characterize the degree of consistency between the current faulty evidence and the other faulty evidence: when a faulty piece of evidence has a high similarity to the majority of evidence, its support is relatively higher; when a faulty piece of evidence has a low similarity to the majority of evidence, its support is relatively lower.

[0062] In one embodiment, the support can be obtained by summing the similarity values ​​of the row / column where the fault evidence is located, for example, by using the mean, weighted mean or other equivalent summarization methods; wherein the weighting can be determined by the reliability index and / or data quality information corresponding to each fault evidence, so as to make more credible evidence contribute more to the support calculation.

[0063] The local conflict level of each piece of faulty evidence is generated based on its support. This local conflict level is negatively correlated with the support level and is used to characterize the degree of inconsistency between the faulty evidence and the other evidence. Specifically, when the support level of a piece of faulty evidence is low, it indicates that its consistency with other evidence is poor, and the local conflict level of the faulty evidence is high; when the support level is high, the local conflict level of the faulty evidence is low.

[0064] In one embodiment, the local conflict degree can be obtained by monotonically decreasing support, for example by using the complement of support or by scaling support, to ensure that the relationship that the greater the support, the smaller the local conflict degree holds true.

[0065] After obtaining the local conflict degree of each fault evidence, the local conflict degrees of each fault evidence are aggregated to obtain the global conflict degree, which is used to characterize the overall conflict level of evidence within the current diagnostic time window. The aggregation method can be any one of summation, mean, weighted sum or maximum value or a combination thereof; among which, the weighting can be determined based on the reliability of evidence, evidence type or business importance, to adapt to the conflict sensitivity requirements of different devices and different scenarios.

[0066] Through the above implementation methods, the consistency relationship between multiple pieces of fault evidence can be transformed into calculable support and conflict indices without relying on the output of a single model, thereby providing an executable and adjustable quantitative basis for subsequent conflict suppression adaptive calibration.

[0067] In one specific embodiment, after obtaining the local and global conflict degrees, to reduce the disturbance of evidence conflict to the fusion diagnostic results, an adaptive calibration process for conflict suppression is performed on the fault evidence to be fused before performing secondary evidence fusion. This calibration process is driven by the evidence conflict degree, dynamically adjusting the contribution of different fault evidence in subsequent fusion, and suppressing or recovering conflicting evidence when necessary.

[0068] For multiple pieces of fault evidence within the same diagnostic time window, calibration parameters are determined for each piece of fault evidence based on the global conflict degree and / or the local conflict degree of each piece of fault evidence. These calibration parameters are used to adjust the contribution of fault evidence in subsequent secondary evidence fusion, for example, to scale the contribution allocation of fault evidence, adjust the equivalent weights during evidence fusion, or control the proportion of evidence quality allocated to different candidate fault sets.

[0069] In one implementation, fault evidence with higher local conflict level has a stronger suppression degree corresponding to its calibration parameters, thereby reducing the impact of the evidence on the final fusion result; conversely, fault evidence with lower local conflict level and consistent with the majority of evidence has a higher retention degree corresponding to its calibration parameters.

[0070] To distinguish between evidence groups with strong internal consistency and evidence groups that may introduce conflicts, the fault evidence to be fused is grouped based on the similarity value of the evidence similarity matrix, resulting in high-similarity evidence sets and low-similarity evidence sets.

[0071] In one embodiment, the grouping boundary can be determined based on the magnitude of the similarity value, the clustering result of the similarity distribution, or the similarity ranking result, so that the evidence in the high similarity evidence set is more mutually supportive, and the evidence in the low similarity evidence set is more likely to contradict other evidence.

[0072] Prioritize calibration for fault evidence with low similarity or whose local conflict levels meet preset calibration conditions. These preset calibration conditions may include: local conflict levels exceeding a preset conflict threshold, or local conflict levels falling within a preset ranking range (e.g., within the top percentages or top few items in the conflict level ranking), or any combination thereof. By introducing determinable calibration conditions, the system can adaptively identify conflict evidence that needs to be suppressed under different equipment, operating conditions, and data quality scenarios.

[0073] For fault evidence that meets the calibration conditions, at least one of the following processing strategies is executed based on its corresponding calibration parameters: weighting reduction, which decreases the fusion contribution of the fault evidence, thus weakening its impact on the results in subsequent secondary evidence fusion; elimination, where fault evidence is removed from the current fusion set when there is significant conflict and insufficient reliability, to avoid abnormal evidence dominating the fusion; and evidence quality reallocation, which suppresses the contribution of the fault evidence to multiple candidates and difficult-to-determine aspects, and reallocates the recoverable evidence quality to the determined candidate fault set. The determined candidate fault set is the set corresponding to a single candidate fault, used to carry a more specific diagnostic indication.

[0074] The above calibration strategy can suppress the amplification effect of conflicting evidence while retaining valid evidence, so that subsequent secondary evidence fusion can be carried out on a less conflicting and more controllable evidence set, thereby improving the stability and credibility of the final fault diagnosis results and their confidence.

[0075] In one specific embodiment, to further enhance the controllability and interpretability of calibration processing in conflict scenarios, Bayesian estimation is introduced to perform prior estimation and weighting of fault evidence based on the determination of calibration parameters based on the degree of conflict and the completion of weight reduction / removal / quality redistribution. This is to achieve quality suppression of evidence pointing to uncertain candidate fault sets or recovery to certain candidate fault sets.

[0076] In this embodiment, the candidate fault set is used to represent the set of fault categories or fault locations that the diagnostic output may point to. A determined candidate fault set is the set corresponding to a single candidate fault; an uncertain candidate fault set is the set containing at least two candidate faults, used to describe situations where evidence supports multiple candidate faults simultaneously, but it is difficult to definitively attribute them to a single candidate fault.

[0077] After obtaining the global and local conflict levels, prior estimation and weighting are performed on each piece of faulty evidence. The prior estimation is used to initially characterize the credibility of the evidence before fusion, so that evidence with high conflict or high local conflict is given stronger uncertainty constraints when entering subsequent inference. The weighting is used to apply the above prior information to the allocation of evidence quality or the contribution of evidence, so that the impact of conflict on evidence is explicitly expressed in a probabilistic sense.

[0078] In one implementation, prior information can be represented by global conflict degree to characterize the overall inconsistency level of the current time window, and by local conflict degree to characterize the degree of deviation of a piece of evidence from the other evidence. The two can be used alone or in combination to determine the prior weight or prior confidence of the evidence in subsequent Bayesian estimation.

[0079] After completing the prior estimation and weighting, Bayesian estimation is used to update the quality of evidence: when the quality of a certain fault evidence points more to the set of uncertain candidate faults, that is, when it supports multiple candidate faults at the same time and is difficult to distinguish, Bayesian estimation is used to suppress the quality of this part of the evidence, so as to reduce its interference with the final fusion conclusion; at the same time, the quality of recoverable evidence is recovered to the set of definite candidate faults, so that the quality of evidence is more concentrated on the clear direction of a single candidate fault, thereby improving the determinability of the diagnostic conclusion.

[0080] In one embodiment, the magnitude of the suppression or recovery can be correlated with the global conflict level and the local conflict level: the higher the conflict level, the more inclined to suppress the uncertain candidate fault set and enhance the recovery to the certain candidate fault set; the lower the conflict level, the more conservative suppression of the uncertain set is maintained to avoid excessive contraction leading to information loss.

[0081] Through the above implementation methods, it is possible to suppress and recover the contributions of multiple candidates and difficult-to-determine evidence in conflict scenarios in a probabilistic sense. This means that the calibration process is not only reflected in the weight adjustment level, but also in reducing the diffusion of uncertainty at the evidence quality allocation level, thereby providing a more stable and clear evidence input basis for subsequent secondary evidence fusion.

[0082] In one specific embodiment, after completing the adaptive calibration process based on conflict degree and the suppression of uncertain sets and retrieval of certain sets based on Bayesian estimation, a secondary evidence fusion is performed on the calibrated fault evidence to output the equipment fault diagnosis result and its confidence level. This embodiment uses a combination fusion rule based on DS evidence theory to achieve secondary evidence fusion, and reduces the disturbance of evidence conflict on the fusion result in a staged manner.

[0083] In this embodiment, the evidence set used for secondary evidence fusion includes at least first fault evidence and second fault evidence, and may optionally include branch fusion evidence and supplementary evidence obtained by mapping event / state information; the secondary evidence fusion is performed after the conflict suppression calibration is completed, and is used to make a unified fusion decision on multiple calibrated pieces of evidence and output diagnostic results and their confidence levels.

[0084] First, a propositional framework for fault diagnosis is constructed to describe the set of candidate faults involved in the diagnosis. The basic propositions in the framework can be a single candidate fault (e.g., a specific fault type / location) or a set of multiple candidate faults, representing uncertainty. For each piece of fault evidence, it is converted into an evidence quality allocation within the propositional framework. This allocation characterizes the degree to which the evidence supports each candidate fault proposition or set of propositions. In one embodiment, when fault evidence clearly points to a single candidate fault, its evidence quality is allocated more to the corresponding single proposition; when fault evidence has multiple candidate uncertainties, its evidence quality can be allocated to the set of uncertain propositions containing multiple candidate faults; a certain proportion can also be reserved for the unknown / uncertain set to characterize the incompleteness or insufficient data quality of the evidence itself. This evidence quality allocation can be obtained from the confidence tendency output by the model, the normalized result of the anomaly score, or an equivalent mapping, and the allocation ratio can be adjusted during the mapping process by incorporating calibration parameters, reliability indices, or prior weighting results.

[0085] After allocating the quality of each piece of evidence, the combination and fusion rules of the DS evidence theory are used to fuse multiple pieces of evidence. The basic idea of ​​the fusion process is to enhance the common support of different pieces of evidence on the same proposition, while resolving conflicts and normalizing contradictory support, so that the fusion result reflects the conclusion jointly supported by multiple pieces of evidence.

[0086] In one embodiment, a measure of the conflict level can be obtained simultaneously during the fusion process to reflect the degree of inconsistency within the current evidence set. Combined with the aforementioned adaptive calibration process, this embodiment suppresses / recovers highly conflicting evidence before fusion, thus reducing instability caused by excessive conflict during the fusion stage. It should be noted that the conflict index used for adaptive calibration in this embodiment is the aforementioned global conflict degree and / or local conflict degree generated based on the evidence similarity matrix; the conflict measure obtained during the DS fusion process serves only as auxiliary information for consistency monitoring or result interpretation during the fusion process and does not replace the aforementioned conflict degree index.

[0087] To improve fusion stability, this embodiment first obtains a highly similar evidence set based on the evidence similarity matrix, and then performs a first-stage fusion on the highly similar evidence set. Because this evidence set has stronger internal consistency, it is easier to form stable intermediate fused evidence when using the DS combination fusion rule.

[0088] The output of the first stage is intermediate fusion evidence, which is still in the form of evidence quality allocation for the propositional framework, representing the intermediate judgment obtained by synthesizing a group of evidence with strong consistency.

[0089] After obtaining intermediate fused evidence, a second-stage fusion is performed with the remaining fault evidence after adaptive calibration. The remaining fault evidence includes evidence from the low-similarity evidence set or evidence suppressed by calibration parameters. Through a two-stage structure of first aggregating consistent evidence and then absorbing the remaining evidence, a stable foundation formed by consistent evidence is preserved on the one hand, and the boundary cases are supplemented and corrected using the calibrated remaining evidence on the other hand, thus balancing stability and information integrity.

[0090] After completing the second stage of fusion, the final fusion evidence is obtained, and the equipment fault diagnosis results and their confidence levels are output based on the final fusion evidence.

[0091] In one implementation, the diagnostic result can be determined based on the degree of support for a single candidate fault proposition in the final fused evidence, and the corresponding degree of support can be output as the confidence level. When the final fused evidence is still relatively scattered among multiple candidate fault propositions, the Top-K candidate faults and their confidence levels can be output, and the remaining quality allocation can be indicated as the uncertainty level to prompt maintenance personnel to conduct further verification.

[0092] In addition, the system can synchronously write the fusion output results into the historical alarm / work order field, and display the key evidence sources and their contribution changes in the comparative analysis module, providing a basis for subsequent tracing and model iteration.

[0093] Through the above implementation method, secondary evidence fusion is based on DS evidence theory. Under the phased structure of priority fusion of highly similar evidence sets, intermediate fused evidence, and then fusion with the remaining calibration evidence, conflict governance and evidence fusion are organically linked. It can obtain more stable fault diagnosis results in evidence conflict scenarios and output confidence characterization with engineering significance.

[0094] In one specific embodiment, the present invention also provides an equipment fault diagnosis system for photovoltaic power plant applications. This system can be deployed on a power plant-side server, edge computing gateway, or cloud-based operation and maintenance platform, and is linked with alarm centers, fault diagnosis, comparative analysis, historical alarms, and other business modules to achieve fault diagnosis and confidence level output for target equipment. Figure 2 As shown, the system includes at least the following functional units, each of which can be implemented by software, hardware, or a combination of both: The data acquisition unit is used to acquire multi-source data related to the operation of target equipment in a photovoltaic power station. The preprocessing unit is used to preprocess multi-source acquired data; The feature extraction and evidence generation unit is used to perform multi-scale feature extraction on the preprocessed multi-source collected data, and input the extraction results into the sparse attention time series model to generate at least two types of fault evidence for equipment fault diagnosis by the sparse attention time series model, and generate evidence representations corresponding to each fault evidence, wherein the evidence representations are feature information used to characterize the corresponding fault evidence pattern features. The similarity matrix and conflict degree determination unit is used to calculate the similarity between evidences based on evidence characterization to construct an evidence similarity matrix, and to determine the evidence conflict degree from the evidence similarity matrix. The evidence conflict degree is a conflict index used to characterize the degree of inconsistency between different faulty evidence. An adaptive calibration unit is used to perform adaptive calibration processing to suppress conflict in the fault evidence to be fused, utilizing the degree of evidence conflict. The secondary evidence fusion unit is used to perform secondary evidence fusion on the fault evidence after adaptive calibration to obtain the equipment fault diagnosis results and their confidence levels. The result output unit is used to output the device fault diagnosis results and their confidence levels.

[0095] In another specific embodiment, the method / system of the present invention can be integrated into the applicant's intelligent operation and maintenance platform for photovoltaic power plants. The front-end interface of the platform includes business modules such as an alarm center, fault diagnosis, comparative analysis, and historical alarms. These business modules are used to provide operation and maintenance personnel with interactive entry points, result display, and closed-loop archiving, and belong to the business carrying framework. The core diagnostic capability of the present invention is called as a background diagnostic engine to realize the output of equipment fault diagnosis results and their confidence levels, and to complete the closed loop in collaboration with the business modules.

[0096] The alarm center is used to aggregate real-time alarms and status change events from photovoltaic power plants. It supports filtering by power plant / region / equipment type / equipment number / alarm level / time range and displays information such as alarm code, trigger time, duration, associated measurement point, communication status, and current operating condition summary.

[0097] In one implementation, after the maintenance personnel select an alarm entry for a target device in the alarm center, the front end provides a button to initiate a diagnosis or perform a one-click diagnosis. When this operation is triggered, the platform sends the selected device identifier, alarm time point or time interval, alarm type, and relevant contextual parameters (such as grid connection status, sampling period, operating condition information, etc.) to the backend diagnostic engine as input conditions for the diagnostic request. Based on this, the backend diagnostic engine invokes the multi-source data acquisition, preprocessing, multi-scale feature extraction, evidence generation, conflict resolution, and secondary evidence fusion processes of the present invention to generate diagnostic conclusions and confidence levels.

[0098] The fault diagnosis module is used to display the diagnostic results output by this invention and support operation and maintenance decisions. The front-end provides the following interactive flow: Diagnostic task configuration: Operation and maintenance personnel select the target device and diagnostic time window (e.g., a front and back window centered on the alarm trigger time, or a custom start and end time), and select whether to include event / status information and work order fields; Start diagnosis: Operation and maintenance personnel click "Start Diagnosis," and the front-end sends the task configuration to the back-end diagnostic engine; Result presentation: After the back-end completes the diagnosis, the front-end displays the diagnostic results and their confidence level, conflict overview, key evidence summary, etc. The conflict overview is a summary of the degree of evidence conflict, such as conflict level indicators and evidence consistency prompts, used to indicate whether the evidence is consistent and whether manual review is required; the key evidence summary displays an overview of evidence sources that contribute significantly to the diagnosis, such as key time windows, key variables, and key alarm / status trigger information.

[0099] In one implementation, the front-end can present key evidence summaries in a visual format: for example, marking high-interest time segments on a timeline and highlighting monitoring quantities that contribute significantly in a variable list; it can also display evidence consistency / disagreement prompts to guide operations and maintenance personnel to quickly locate the root cause of anomalies. The above demonstration does not change the algorithm of this invention itself, but only serves as a front-end delivery method for the output results of this invention.

[0100] The comparative analysis module is used to assist in verifying diagnostic results. Its typical operation is as follows: Maintenance personnel click on comparative analysis on the fault diagnosis results page. The platform automatically imports parameters such as the target device, diagnostic time window, and key variable set. The front-end supports selecting comparison objects, such as: devices of the same model, adjacent devices under the same array / combiner box, or historical normal periods of the device. The front-end displays comparison curves and events, including: continuous curves for electrical parameters / power / temperature / efficiency, alarm codes and protection action timelines, and communication status and work order fields. Maintenance personnel can use this to verify whether the diagnostic conclusions are consistent with curve anomalies, alarm triggers, and changes in operating conditions, and can re-initiate the diagnosis when necessary, such as adjusting the time window or supplementing the data source.

[0101] In one implementation, the comparison analysis module can call the key evidence summary output by the diagnostic engine as the default comparison dimension, thereby reducing the cost of blindly screening a large number of test points by maintenance personnel and improving the positioning efficiency.

[0102] The historical alarm module is used to archive alarms and diagnostic handling processes, forming a traceable closed loop. Its typical process is as follows: After fault diagnosis is completed, the front end provides the option to generate handling records / associate work orders; maintenance personnel select the handling method, such as reset, derated operation, component replacement, on-site inspection, etc., and fill in the handling conclusion and remarks; the platform writes the diagnostic results, confidence level, diagnostic time window, key evidence summary, comparative analysis conclusion, and work order fields into the historical alarm record; historical alarms support retrieval and statistics by device, fault type, confidence interval, conflict level, and other dimensions for subsequent maintenance review and strategy optimization.

[0103] In one implementation, historical alarm records can serve as one of the data sources for subsequent diagnostic engines: for example, when a similar alarm occurs again, the platform can provide historical handling experience and recurrence patterns to assist maintenance personnel in quickly confirming and handling the issue. This closed-loop data accumulation belongs to the application method at the business system level and does not limit the specific implementation form of the algorithm of this invention.

[0104] The technical solution of the present invention has been described above with reference to specific embodiments. It should be understood that the above embodiments are only used to explain the technical concept of the present invention, and not to limit the scope of protection of the present invention; those skilled in the art can make equivalent substitutions or modifications to the specific implementation methods, parameter settings, network structure details and combination relationships of each unit without departing from the concept of the present invention, and all such substitutions or modifications should fall within the scope of protection of the present invention.

[0105] In summary, the embodiments described in this specification are merely illustrative of the technical solutions of this invention and are not intended to limit the scope of protection of this invention. Unless explicitly stated that they are mutually exclusive or incompatible, the technical features described in the embodiments of this invention can be combined, replaced, or deleted by those skilled in the art according to specific needs to form other implementation methods. Any equivalent substitutions or obvious modifications made to the above technical solutions without departing from the spirit and substance of this invention should be considered to fall within the scope of protection of this invention, which is defined by the appended claims.

Claims

1. A method for diagnosing equipment faults in photovoltaic power plant applications, characterized in that, The method includes: Acquire multi-source data related to the operation of target equipment in photovoltaic power plants; Multi-scale feature extraction is performed on the preprocessed multi-source acquisition data, and at least two types of fault evidence for equipment fault diagnosis are generated by the sparse attention time series model. Evidence representations corresponding to each fault evidence are generated, and the evidence representations are feature information used to characterize the corresponding fault evidence pattern features. The similarity between evidences is calculated based on the evidence representation to construct an evidence similarity matrix, and the evidence conflict degree is determined by the evidence similarity matrix. The evidence conflict degree is used to perform adaptive calibration processing for conflict suppression on the faulty evidence to be fused. The evidence conflict degree is a conflict index used to characterize the degree of inconsistency between different faulty evidences. A second evidence fusion is performed on the fault evidence after adaptive calibration to obtain the equipment fault diagnosis results and their confidence levels.

2. The equipment fault diagnosis method for photovoltaic power plant applications as described in claim 1, characterized in that, The multi-source acquisition data includes at least one of the following: continuous monitoring information, which includes time-series data of any one or a combination of electrical parameters, power, temperature, and efficiency; and event and status information, which includes any one or a combination of alarm codes, protection actions, switch status, communication status, work order record fields, and maintenance record fields. The preprocessing includes at least one of the following operations: missing value processing, outlier processing, time alignment, noise reduction and smoothing, normalization, feature filtering, segmentation by working condition, and slicing by window. The feature selection includes any one or a combination of correlation analysis, mutual information analysis, and distance correlation analysis.

3. The equipment fault diagnosis method for photovoltaic power plant applications as described in claim 1, characterized in that, The process of multi-scale feature extraction includes: using parallel or cascaded multi-receptive field one-dimensional convolutional structures or temporal filtering structures to extract features from preprocessed multi-source acquisition data to obtain multi-scale features that characterize short-term disturbances and long-term trends, and fusing the multi-scale features to form combined features, and inputting the combined features into a sparse attention temporal model for temporal modeling.

4. The equipment fault diagnosis method for photovoltaic power plant applications as described in claim 3, characterized in that, The sparse attention temporal model includes a sparse attention branch model and a dense attention branch model; The process of generating at least two types of fault evidence for equipment fault diagnosis by the sparse attention temporal model includes: performing temporal modeling on the combined features of the input by the sparse attention branch model and the dense attention branch model respectively to obtain the first fault evidence and the second fault evidence, and performing adaptive fusion processing to obtain fused evidence; and calculating the reliability index of the fused evidence based on the attention weight distribution generated by the sparse attention branch model and the dense attention branch model. The reliability metrics include any one or a combination of attention entropy, attention sparsity, and cross-scale consistency, used for adaptive calibration processing of evidence similarity matrix construction or conflict suppression.

5. The equipment fault diagnosis method for photovoltaic power plant applications as described in claim 1, characterized in that, The evidence representation is obtained synchronously when fault evidence is generated, and is associated with the corresponding fault evidence. The evidence representation includes at least one of the following data: the intermediate hidden feature vector of the sparse attention temporal model; Attention weight distribution vector or matrix, including any one or a combination of time step weights, variable weights, and cross-scale weights; Data obtained by dimensionality reduction or normalization of intermediate hidden feature vectors or attention weight distribution vectors or matrices.

6. The equipment fault diagnosis method for photovoltaic power plant applications as described in claim 1, characterized in that, The generation of the evidence conflict degree includes: calculating the support degree of each faulty piece of evidence based on the similarity value between each faulty piece of evidence and the remaining faulty pieces of evidence in the evidence similarity matrix, wherein the support degree is used to characterize the consistency between the current faulty piece of evidence and the remaining faulty pieces of evidence; generating the local conflict degree of each faulty piece of evidence based on the support degree, wherein the local conflict degree is negatively correlated with the support degree; and aggregating the local conflict degrees of each faulty piece of evidence to obtain the global conflict degree.

7. The equipment fault diagnosis method for photovoltaic power plant applications as described in claim 6, characterized in that, The adaptive calibration process includes: determining corresponding calibration parameters for each piece of fault evidence based on global or local conflict degree, wherein the calibration parameters are used to adjust the contribution of fault evidence in secondary evidence fusion. Further, it includes: dividing the fault evidence to be fused into a high-similarity evidence set and a low-similarity evidence set based on the similarity value of the evidence similarity matrix; for the low-similarity evidence set or fault evidence whose local conflict degree meets the preset calibration conditions, performing any one or a combination of weight reduction, elimination, or reallocation of evidence quality to a determined candidate fault set according to the calibration parameters; the preset calibration conditions include any one or a combination of local conflict degree exceeding a preset conflict threshold or local conflict degree being located in a preset ranking interval; the determined candidate fault set is a set corresponding to a single candidate fault.

8. The equipment fault diagnosis method for photovoltaic power plant applications as described in claim 7, characterized in that, The adaptive calibration process further includes: prior estimation and weighting of fault evidence based on global or local conflict degree, and using Bayesian estimation to suppress the quality of evidence pointing to an uncertain candidate fault set or to recover evidence pointing to a definite candidate fault set; the uncertain candidate fault set is a set containing at least two candidate faults.

9. The equipment fault diagnosis method for photovoltaic power plant applications as described in claim 7, characterized in that, The secondary evidence fusion adopts a combination fusion rule based on DS evidence theory, including: In the first stage, highly similar evidence sets are fused to obtain intermediate fused evidence; In the second stage, the intermediate fused evidence is further fused with the remaining fault evidence after adaptive calibration to obtain the equipment fault diagnosis results and their confidence levels.

10. A fault diagnosis system for photovoltaic power plant applications, characterized in that, The system includes: The data acquisition unit is used to acquire multi-source data related to the operation of target equipment in a photovoltaic power station. The preprocessing unit is used to preprocess multi-source acquired data; The feature extraction and evidence generation unit is used to perform multi-scale feature extraction on the preprocessed multi-source collected data, and input the extraction results into the sparse attention time series model to generate at least two types of fault evidence for equipment fault diagnosis by the sparse attention time series model, and generate evidence representations corresponding to each fault evidence, wherein the evidence representations are feature information used to characterize the corresponding fault evidence pattern features. The similarity matrix and conflict degree determination unit is used to calculate the similarity between evidences based on evidence characterization to construct an evidence similarity matrix, and to determine the evidence conflict degree from the evidence similarity matrix. The evidence conflict degree is a conflict index used to characterize the degree of inconsistency between different faulty evidence. An adaptive calibration unit is used to perform adaptive calibration processing to suppress conflict in the fault evidence to be fused, utilizing the degree of evidence conflict. The secondary evidence fusion unit is used to perform secondary evidence fusion on the fault evidence after adaptive calibration to obtain the equipment fault diagnosis results and their confidence levels. The result output unit is used to output the device fault diagnosis results and their confidence levels.

Images