Photovoltaic equipment fault early warning method and system based on data driving

Abstract

The invention relates to a photovoltaic equipment fault early warning method and system based on data driving. The method comprises the following steps: acquiring historical fault reports, historical operation data and historical working audio data of multiple pieces of target equipment in a historical time period; determining a first fault condition of the target equipment based on a neural network prediction model according to the historical fault report and the historical operation data; determining a second fault condition in the target equipment based on a spectrogram analysis model according to the historical fault report and the historical working audio data; and acquiring real-time working data of the atomic component in the target equipment, performing fault early warning according to the first fault condition and the second fault condition, and storing the data in a fault analysis library. According to the method, the fault condition of the photovoltaic equipment can be comprehensively expressed, step-by-step transition of single-point warning, fault association and causal inference is realized, a photovoltaic power station can be ensured to quickly respond to fault hidden dangers, and the emergency processing capability of photovoltaic operation and maintenance management is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic operation and maintenance management, and in particular to a data-driven photovoltaic equipment fault early warning method and system. Background Art

[0002] In photovoltaic power generation systems, ensuring efficient system operation and timely troubleshooting are crucial to improving energy utilization and reducing operating costs. As the complexity and scale of photovoltaic systems increase, traditional fault diagnosis methods face challenges.

[0003] Existing photovoltaic equipment fault diagnosis and monitoring technologies integrate high-definition drone photography, temperature sensors, vibration sensors, and highly sensitive sound sensors. These technologies can simultaneously acquire multi-source data, including images, temperature, vibration, and sound, for data analysis and fault diagnosis. However, the scope of multi-source data analysis for photovoltaic field equipment in existing publicly available technologies is too broad, and the following technical drawbacks generally exist: First, the existing technology lacks authenticity verification in sample selection, which easily deviates from the actual results of actual diagnosis. The fault relationship between equipment and components is not considered in the fault analysis of cross-modal multi-source data, resulting in a lack of sufficient robustness in the comprehensive analysis results.

[0004] Secondly, the issue of scale analysis is not considered for single-modal audio data. The spectrum diagram in sound source fault diagnosis is obtained by direct calculation, and its samples are sparse. As a result, when the degree of fault difference quantified by the samples is used as the judgment basis for fault diagnosis, the accuracy of fault prediction will be reduced, and in-depth evaluation of equipment fault diagnosis cannot be achieved through abnormal sound detection. Especially when there are insufficient abnormal samples, the lack of effectiveness of sound source analysis will be particularly obvious.

[0005] Therefore, in view of the above-mentioned defects, there is an urgent need for a solution to solve the problems existing in the prior art. Summary of the Invention

[0006] The primary purpose of the present application is to solve at least one of the above problems and provide a data-driven photovoltaic equipment fault warning method and system.

[0007] In order to meet the various objectives of this application, this application adopts the following technical solutions: A data-driven photovoltaic equipment fault early warning method provided to meet one of the purposes of this application includes the following steps: Obtain historical fault reports, historical operation data, and historical working audio data of multiple target devices within a historical period; Based on historical fault reports and historical operating data, the neural network prediction model is used to determine the first fault condition of the target equipment; Determining a second fault condition in the target device based on a spectrogram analysis model according to historical fault reports and historical working audio data, the spectrogram analysis model comprising a core frame scanning network, a residual neural network, a pooling layer, and a fully connected layer, the core frame scanning network being provided with a core frame for scanning the spectrogram to extract features; The real-time working data of the atomic components in the target device is obtained, and fault warning is performed according to the first fault condition and the second fault condition, and stored in the fault analysis library.

[0008] On the other hand, a data-driven photovoltaic equipment fault early warning system is provided to meet one of the purposes of this application, including: A data collection module is used to obtain historical fault reports, historical operation data, and historical working audio data of multiple target devices within a historical period; A first fault prediction module, configured to determine a first fault condition of a target device based on historical fault reports and historical operating data and a neural network prediction model; a second fault prediction module, configured to determine a second fault condition in the target device based on a spectrogram analysis model according to historical fault reports and historical working audio data, the spectrogram analysis model comprising a core frame scanning network, a residual neural network, a pooling layer, and a fully connected layer, the core frame scanning network being provided with a core frame for scanning the spectrogram to extract features; The fault warning module is used to obtain real-time working data of atomic components in the target device, perform fault warning according to the first fault condition and the second fault condition, and store the data in the fault analysis library.

[0009] On the other hand, a data-driven photovoltaic equipment fault warning device is provided to meet one of the purposes of the present application, including a central processing unit and a memory, wherein the central processing unit is used to call and run a computer program stored in the memory to execute the steps of the data-driven photovoltaic equipment fault warning method described in the present application.

[0010] On the other hand, a computer-readable storage medium is provided to meet one of the purposes of the present application, wherein the computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to enable a computer to execute the data-driven photovoltaic equipment fault warning method as described in any one of the items disclosed in the first aspect of the present invention.

[0011] The technical solution of this application has many advantages, including but not limited to the following: This application has been approved. Firstly, for fault prediction and diagnosis based on cross-modal multi-source data in the field of photovoltaic operation and maintenance, it introduces sample authenticity verification and comprehensive analysis of fault relationships, which can improve the accuracy of fault prediction results based on samples. Secondly, the fault warning devices corresponding to multi-source data are atomically split. Based on the local fault analysis results of the atomic components, a neural network model is constructed to characterize the global fault situation that characterizes the relationship between the atomic components and the target equipment faults through the fusion of local fault features. At the same time, the fusion of local fault features provides a feasible solution for data alignment under different fault situation analyses, effectively improving the robustness of the model prediction results and enabling a better transition from single-point warnings to topological associations to causal inference. Next, the introduction of abnormal sound detection using multi-scale feature analysis can mine abnormal samples for fault diagnosis, achieving in-depth evaluation of fault diagnosis, fault warning, and operation and maintenance management. While retaining the feasibility of in-depth fault diagnosis of photovoltaic equipment through abnormal sound detection, the local and global fault conditions updated by the spectrum analysis model can more completely and comprehensively express the fault conditions of photovoltaic equipment. Finally, the data in the fault analysis library serves as a more accurate and solid data foundation. During the real-time warning process, it can provide fast and accurate photovoltaic equipment fault warning data search, ensuring a rapid response when there are potential fault hazards in photovoltaic power stations, and improving the emergency handling capabilities of photovoltaic operation and maintenance management. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which: Figure 1 This is a flow chart of an embodiment of the data-driven photovoltaic equipment fault early warning method of the present application; Figure 2 This is a schematic diagram of the data-driven photovoltaic equipment failure warning system used in this application; Figure 3 A schematic diagram of an exemplary spectrum graph analysis model framework of this application; Figure 4 This is a schematic diagram of a core frame scanning spectrum diagram in an exemplary core frame scanning network of the present application. DETAILED DESCRIPTION

[0013] The technical solution of the present application is applicable to the field of photovoltaic operation and maintenance management technology, and is particularly applicable to fault warning scenarios of photovoltaic equipment. In this context, a data-driven photovoltaic equipment fault warning method provided in an embodiment of the present application can be applied in a typical photovoltaic operation and maintenance system. The method is implemented in electronic devices, such as computer terminals, specifically operation and maintenance work computers.

[0014] The following describes in detail the technical solution of the present application and how the technical solution of the present application solves the above-mentioned technical problems with specific embodiments.

[0015] The following specific embodiments can be combined with each other, and the same or similar concepts or processes in some embodiments will not be described in detail. The following will describe the embodiments of the present application in conjunction with the accompanying drawings.

[0016] See also Figure 1 This application discloses a data-driven photovoltaic equipment fault early warning method. In its typical photovoltaic fault early warning embodiment, the method includes the following steps: Step 1100: Obtain historical fault reports, historical operation data, and historical working audio data of multiple target devices within a historical period; wherein the target devices include but are not limited to key equipment such as inverters, photovoltaic modules, junction boxes, transformers, and connectors in a photovoltaic system, wherein each target device includes a number of working parts. It is understandable that the working parts within the target device are differentiated according to their functions so as to achieve characterization in atomic units.

[0017] Optionally, the type of historical fault reports can be text, voice, image, video and other data, and corresponding data analysis methods can be further used to extract fault report data therein, such as fault type, fault description and parameters of the fault corresponding component.

[0018] Optionally, the historical operation data may include working data corresponding to atomic components in different target devices. Specifically, the working components may be power electronic components, capacitor components, cooling system components, control circuit components or other working components with inverter functions in the inverter. Data can be collected for the atomic components of different target devices through corresponding sensors (such as power sensors, temperature sensors, sound sensors, irradiance flux sensors and infrared sensors, etc.). For example, the inverter as the target device includes atomic components such as power electronic elements, capacitors, cooling systems and control circuits. For example, the operating temperature parameters and power parameters of the power electronic elements in the inverter are collected.

[0019] Optionally, the historical working audio data is an audio sample collected by a highly sensitive sound sensor, which can align the fault anomalies in the historical fault report according to the historical time period, so as to obtain the abnormal samples required for fault diagnosis corresponding to the sound source data. It can be understood that there are dimensional differences between the sound source data and the historical operation data, and data fusion diagnosis is required to effectively utilize the abnormal samples. If the sound source data is directly fused with the equipment operation data, the accuracy and effectiveness of the fault diagnosis will be reduced.

[0020] Step 2100: Determine the first fault condition of the target device based on the neural network prediction model according to historical fault reports and historical operation data; wherein, a neural network model is selected corresponding to the atomic components of different target devices, such as a long short-term memory network or a graph neural network can be used to pre-train through a data set to output a fault prediction result.

[0021] In a specific implementation, the fault report data corresponding to each atomic component in the target device is determined based on the historical fault report, wherein the historical fault report includes the fault type, fault time, environmental conditions, fault cause and historical fault audio; Next, based on the historical operating data, the operating data corresponding to each atomic component is determined, wherein the operating data includes electrical parameters, environmental parameters, and device status parameters; Through the above implementation method, based on the fault report data and operation data corresponding to each atomic component, the fault condition of each atomic component is determined as a fault sample through the neural network model, and the fault time recorded in the fault report can be used to connect the fault correlation of different atomic components in the same target device within the historical period, and the fault cause can be used as a verification basis for fault correlation analysis, thereby realizing more intelligent and comprehensive component-level maintenance and fault monitoring of the target device.

[0022] Furthermore, the electrical parameters in the operating data are mapped according to the equipment type. For example, overvoltage and undervoltage phenomena are detected in the voltage of photovoltaic modules, inverters and combiner boxes. Abnormal current fluctuations or short circuits are identified based on the current in the photovoltaic modules, inverters and combiner boxes. The output power of photovoltaic modules and inverters is used to evaluate the efficiency and performance of the photovoltaic system. For inverters, the stability of the output current is judged based on the frequency. Parameters such as temperature and irradiation in environmental parameters are used to describe the equipment life and power generation efficiency, and equipment status parameters such as insulation resistance, contact resistance and aging degree are used to describe equipment short circuit, leakage, overheating and equipment life.

[0023] Furthermore, the atomic components serve as working components of the target equipment, and their operating parameters may be the IV characteristic curve, temperature coefficient, efficiency change rate, etc. of the cell in the photovoltaic module; the environmental parameters may be the operating temperature of the power electronic components in the inverter; the equipment status parameters may be the contact resistance of the junction box in the photovoltaic module; and it is easy for technicians in this field to know that when collecting and confirming the operating data of the target equipment, the corresponding equipment and corresponding components can be sampled according to the fault prediction requirements.

[0024] In a specific embodiment, the authenticity of the historical fault report is determined based on the fault report data and operation data corresponding to each of the atomic components and based on the authenticity prediction neural network model; It should be noted that if the authenticity of the fault report data is not verified, false alarm warnings may be generated due to sensor failure, data transmission errors or other non-actual fault reasons, resulting in waste of resources and unnecessary emergency responses. If the actual fault is ignored, the accuracy of the subsequent fault prediction results will be affected because the original fault report has not been verified. By verifying the authenticity of the fault report, it can be ensured that the early warning system makes predictions based on real data and event samples, which helps to reduce false alarms and missed alarms, thereby improving the overall performance of the early warning system.

[0025] The authenticity verification of the sample includes the following steps in its specific implementation: For each atomic component, each fault report data corresponding to the atomic component is input into a first authenticity prediction neural network model, and a first authenticity prediction probability is output for all input fault report data; the first authenticity prediction neural network model is obtained by training a local fault training data set including multiple training fault report data sets and corresponding authenticity annotations; For each atomic component, the fault report data and operation data of the same historical period are input into a second authenticity prediction neural network model, and a second authenticity prediction probability corresponding to the fault report data is output, where the second authenticity prediction neural network model is obtained by using multiple training fault report data sets, atomic component training operation data, and authenticity-labeled training data sets; Calculating a weighted sum of the first authenticity prediction probability and the second authenticity prediction probability to obtain an authenticity parameter corresponding to each of the fault report data; the calculation of the weighted sum includes a first weight and a second weight, the first weight representing the authenticity prediction accuracy of the first authenticity prediction neural network model, and the second weight representing the historical proportion of the fault report data to all true fault report data of the atomic component; The authenticity threshold is set according to the authenticity parameter to filter the historical fault report to determine the authenticity of the historical fault report, and a true fault report is output.

[0026] Through the above embodiment, the authenticity of the fault report data can be comprehensively determined based on the two neural network models, and then the real fault report data can be screened to obtain the real fault report data, so that the fault condition of each atomic component can be accurately determined based on the real problem report in the future, and more intelligent and comprehensive fault warning and operation and maintenance management of components and target equipment in the photovoltaic system can be achieved.

[0027] Furthermore, the local fault of each atomic component is determined based on the fault prediction neural network model according to the real fault report data; It should be noted that for each atomic component, its impact on the global fault within the target device should be considered. It can be understood that for atomic components in different target devices or atomic components in the same target device, there is cross-modal data as operating parameters for fault analysis.

[0028] Therefore, before comprehensively analyzing the global faults of the target equipment, the local faults reflected by each atomic component can be considered separately, and the characteristic description of the local faults and the fault prediction value for risk warning can be determined before data fusion.

[0029] Furthermore, compared with the fusion of input parameter features and comprehensive analysis, after independently analyzing the local fault features, feature fusion is performed based on the relationship between faults to generate a universal sequence composed of local fault sequences to describe global faults. The universal sequence can unify the data type in time and space, which can improve the robustness of the model for comprehensive analysis of cross-modal data.

[0030] In specific implementation, determining a local fault includes the following steps: For each atomic component, the determined actual fault report data and the atomic component's operating data are input into a fault prediction neural network model to output the local fault and predicted failure probability corresponding to each atomic component. The fault prediction neural network model includes a feature extraction network, a classification prediction network, and a prediction verification network. It can be seen that through the above embodiments, the fault prediction neural network model corresponding to the atomic component type can be determined based on the real fault report data and the atomic component operation data collected by the sensor, so as to achieve accurate detection of local defects, and subsequently facilitate early warning and maintenance of the global faults of the target equipment, and assist in realizing the global description of the photovoltaic system based on local faults, providing a more accurate data basis for fault warning, and improving the robustness of comprehensive analysis of cross-modal data.

[0031] As an optional embodiment, the fault prediction neural network model is trained by the following steps: First, a training dataset of multiple data types in the operation data of multiple atomic components, a fault report dataset of multiple atomic component types, and a training dataset of fault annotations are used as a unified dataset. The feature extraction network and the classification prediction network are trained using the unified dataset until convergence. Secondly, the trained feature extraction network and classification prediction network are distilled based on the three-way combination of the type of atomic component, the data type corresponding to the atomic component, and the local fault corresponding to the atomic component, to obtain the feature extraction network and classification prediction network in the fault prediction neural network model corresponding to the three-way combination; Next, multiple other associated training data sets for the same local fault in the unified data set are determined based on the three-way combination relationship, and a label correction result is determined based on the labeling results of the multiple associated training data sets for the same local fault, so as to obtain a corrected training data set for training the prediction verification network; The prediction verification network is trained according to the corrected training data set until convergence. The prediction verification network is used to compare the actual fault report data as input with multiple predicted local faults, and to compare the component operation data with multiple predicted local faults to determine the result correction probability corresponding to the prediction result, and each predicted local fault probability is corrected by multiplying the correction probability.

[0032] Specifically, through the above-mentioned optional embodiments, the feature network and the classification network can be uniformly trained through a unified training data set of multiple different components and data types, and then distilled and trained separately based on different three-way combination relationships to obtain a model with better prediction effect. In particular, the use of a real fault report data set that has passed the authenticity test can make the results of the model fault prediction closer to the real diagnosis. For the fault prediction results, the classification and verification output network are trained based on the labeling correction results between multiple training data sets, so that the final output of the local fault prediction results of the atomic components is more accurate and reliable, which facilitates the subsequent analysis and early warning of the comprehensive description of the global fault based on the local fault.

[0033] In a specific embodiment, according to the local failure of each of the atomic components, a global failure of the target device is determined based on a neural network prediction model fused with local failure features.

[0034] Optionally, the local fault results of atomic components come from multi-source, cross-modal operating data. Direct data fusion of the operating data will lose the representation of the fault information by the single-modal data. Therefore, in this application, the characteristics describing the local fault are determined based on the unit atomization of the components within the target device, which will reduce the loss of fault information during data fusion.

[0035] Furthermore, according to the local fault of each atomic component, a global fault of the target device is determined based on a neural network prediction model fused with local fault features, including the following steps: First, the self-attention mechanism is used to dynamically calculate the correlation weights of different local fault features to capture global dependencies. Furthermore, the local attention mechanism is used to extract the variability of local fault features through an embedding layer, and a sliding window is used to align the local time of multi-source data of atomic components. This results in a neural network prediction model that integrates local fault features. Optionally, for the expression of global fault features based on local fault features, the specific implementation method of the present application chooses to perform sequence modeling of local fault features through the attention mechanism. For the operating data fed back by the sensor, the entity can be mapped to the vector space to generate a word vector. For the position encoding of the feature word vector, the sequence order information is introduced according to the spatial coordinates of the target device and the atomic component and the timestamp of the operating data to form the word sequence position. For the state correlation between the atomic component and the target device, the inter-word dependency is expressed through the sequence label according to the attention mechanism. For the case of multi-physical coupling analysis, different modes of local fault features are captured in parallel based on the multi-head mechanism. Therefore, the local fault feature fusion neural network prediction model corresponding to the sequence modeling of the global fault chooses the GNN-Transformer network architecture to implement it, so that the prediction result of the global fault can better realize the transition from single-point warning to topological association to causal inference, and provide decision-level support for the operation and maintenance of photovoltaic fault warning.

[0036] Furthermore, the Transformer model in the neural network prediction model of local fault feature fusion includes an encoder containing multiple encoder layers and a decoder containing multiple decoder layers, wherein the encoder layer includes a local enhanced multi-head self-attention module, a variable sliding window module and a feedforward network module, and the decoder layer includes a local enhanced multi-head self-attention module, a multi-head self-attention module, a variable sliding window module and a feedforward network module. The variable sliding window module selects different local enhanced receptive fields for different encoder layers and decoder layers through the embedding layer input.

[0037] Next, a graph neural network is used to extract features of the local faults associated with the three combinations to obtain local fault features. The local fault features of the target device at adjacent time points are linearly transformed and dot-producted using a variable sliding window to obtain the attention score of the current embedding layer. The mask matrix is ​​used to limit the interaction range to determine the first attention score of the local fault features fused into the encoder layer. It is understandable that before feature fusion, it is necessary to perform feature extraction for the prediction of local faults based on graph neural networks to obtain local fault features that characterize the fault information, where the fault information includes fault description, fault components and associated fault data. The feature fusion part is based on the attention mechanism of the spatiotemporal Transformer network as a modeling tool for the global fault sequence. Those skilled in the art can understand that Transformer is essentially a sequence-to-sequence architecture, which is independent of the data type. Therefore, it can not only be applied to the feature fusion of cross-modal data, but its output results can also be aligned with the abnormal sample faults corresponding to the historical working audio data output by the multi-scale kernel frame scanning network. It can reduce the impact of cross-modal data on the robustness of the comprehensive analysis, and assist in achieving more accurate and comprehensive fault warning analysis based on different local faults and global faults, providing strong data support for photovoltaic system operation and maintenance.

[0038] As an example, the local fault results of the atomic components in the target device are represented as a sequence of local fault features extracted by the graph neural network. , Indicates the length of the local fault feature, which can also be understood as a sequence Include elements, the sequence As the input of the Transformer model, the attention score of the encoder layer The expression is: ; in, represents the attention score of the encoder layer, w is a constant and each attention score is calculated The corresponding w values ​​are different, i represents the number of layers of the corresponding encoder layer, and T represents the matrix transpose. represents the scaling factor, The first one represents the scaling factor The serial number corresponding to the value, The representation masking technique is used to retain the characteristic elements within the range w on the diagonal of the matrix.

[0039] Next, the output results of each encoder layer are subjected to feature dimension upgrade through linear and / or nonlinear mapping, and a label sequence of the global contribution of the local fault features is output. The second attention score of the decoding layer is calculated based on the feature dimension upgrade results through mask matrix fusion, and the label association of the global features based on the label sequence is used to obtain the attention matrix of the decoding layer. In the embodiment of the present application, the decoder is used as an example. The variable sliding window module of the decoder layer takes the high-level feature sequence and label sequence output by the encoder as input, combines the masking technology to fuse and calculate the attention score of the decoding layer. The expression of the attention score of the decoding layer is: ; in, represents the attention score of the decoder layer, represents an element in a tag sequence, Represents the elements in the high-level features, w is a constant and calculates each attention score The corresponding w values ​​are different, T represents the matrix transpose, represents the number of layers corresponding to the decoder layer, represents the scaling factor, The representation masking technique is used to retain the characteristic elements within the range w on the diagonal of the matrix.

[0040] As a further example, the output of each encoding layer is calculated based on the output of the previous encoder layer and the attention score of each layer and the attention weight that characterizes the contribution of local fault features to the global fault, and then the feedforward network performs linear mapping and / or nonlinear mapping to output a high-level feature sequence. and the corresponding label sequence , Represents high-level features The number of elements, Indicates the length of the label sequence, that is, the sequence number of the element in the label sequence. It should be noted that the original input sequence Dimensionality reduction is performed to deal with redundancy or noise during feature extraction in the graph neural network model. However, in order to improve the feature expression capability of local fault features, when fusing multi-source cross-modal data, each layer of the model input is upgraded to find a more suitable feature representation in a higher-dimensional space. Appropriately increasing the feature dimension can improve the performance of the model.

[0041] As a further example, the feedforward network module performs linear mapping and / or nonlinear mapping on the output of the encoder layer, and the expression of the high-level features corresponding to each encoder layer in the output encoder is: ; in, represents the high-level features output by the current encoder layer, represents the high-level features output by the previous encoder layer, Indicates the attention score of the encoder layer corresponding to the output current high-level feature, Represents the linear change function and / or nonlinear change function of the feedforward network module, Represents the local enhanced multi-head self-attention function.

[0042] Next, based on the attention scores corresponding to the local fault features, the range control of global dependencies is determined based on the self-attention mechanism. The first attention weights of different local fault features in the encoder layer are determined by enhancing the local features of the encoder layer. The global contribution of the local fault features is evaluated based on the first attention weights. The encoder in the embodiment of the present application is used as an example, and the expression of the encoder layer attention weight is: ; in, represents the high-level features X output by the previous encoder layer, represents the attention score of the current encoder layer, represents the scaling factor, T represents the matrix transpose, () represents the attention mechanism, () represents layer normalization, and i represents the sequence number of the i-th attention mechanism corresponding to the i-th encoder layer.

[0043] Next, based on the attention score and the attention matrix, the local fault features of the decoder layer are enhanced through a self-attention mechanism and a variable sliding window to determine a second attention weight of the decoder layer. The local fault features are concatenated according to the second attention weight, and a global fault result and predicted fault probability of the target device are output through linear mapping and / or nonlinear mapping. The decoder in the embodiment of the present application is used as an example. The multi-head self-attention module of the decoder layer adopts the attention mechanism combined with the multi-head scheme of local fault features for parallel calculation, and the label sequence of the global contribution is obtained. The contribution of the global information of the local fault feature sequence is associated, and the expression of the second attention weight of the decoder is: ; ; ; ; Among them, the tag sequence Through different linear transformations, we can obtain the query vector Q and key vector respectively. Sum value vector V, , , , ,and , is the linear change matrix of the query vector Q, is the key vector Linear change matrix, is the linear change matrix of the value vector V, represents the scaling factor corresponding to the query vector, represents the scaling factor corresponding to the key vector, represents the scaling factor corresponding to the value vector, represents the scaling factor in the h attention spaces, Represented as a multi-head self-attention module attention spaces, which is convenient for capturing attention features in different spaces. T represents matrix transposition. Indicates the scaling factor used to reduce the Dot product calculation gives The impact of the function, Representing the multi-head self-attention function, it can be understood that a cross-modal query key-value structure is introduced in the cross-modal data, such as the query vector Q comes from electrical parameters, while the key vector K and value vector V come from environmental parameters or device state parameters.

[0044] The decoder in the embodiment of the present application is used as an example. The local enhanced multi-head self-attention module of the decoder layer combines the attention score output by the variable sliding window and mask technology. , to control the scope of contextual interaction in the attention mechanism, enhance the local features of the decoder layer, and output the weighted expression of the attention corresponding to the decoder layer: ; in, represents the global fault feature result Y output by the previous decoder layer, T represents the matrix transpose, represents the scaling factor, represents the attention score corresponding to the current decoder layer, Represents high-level features of the encoder output , j represents the jth decoder layer corresponding to the decoder layer attention mechanism.

[0045] The decoder in the embodiment of the present application is further exemplified. The feedforward network module of the decoder layer performs linear mapping and / or nonlinear mapping on the output of the decoder layer, and outputs a global fault result that integrates local fault features and includes the fault prediction probability. , Indicates that the global fault result contains M elements, and its expression is: ; in, represents the attention score of the decoder layer corresponding to the current high-level feature X, MHA represents the multi-head self-attention function, represents the local enhanced multi-head self-attention function, represents the feedforward network function, represents the high-level feature X output by the encoder, M represents the total number of layers of the corresponding decoder, and N represents the total number of layers of the corresponding encoder.

[0046] Finally, according to the local fault of the atomic component under the three-way combination relationship and the global fault of the target device under the fusion of the local fault features, the first fault condition of the target device is determined.

[0047] It can be understood that the first fault situation includes a global description of the fault situation of the photovoltaic equipment based on historical fault reports and atomic component operation data. At the same time, the fault situation of the atomic components in each target device can be traced and warned based on the combination of the three, so as to achieve more efficient and accurate photovoltaic fault warning based on local faults and global faults. In order to further improve the accuracy of fault warning, in addition to historical fault reports and equipment operation data, other sample parameters need to be further introduced as the basis for evaluating the warning results, so as to achieve the purpose of in-depth evaluation of fault diagnosis, fault warning and operation and maintenance management.

[0048] Step 3100: Determine a second fault condition in the target device based on a spectrogram analysis model according to historical fault reports and historical working audio data, wherein the spectrogram analysis model includes a core frame scanning network, a residual neural network, a pooling layer, and a fully connected layer, wherein the core frame scanning network is provided with a core frame for scanning the spectrogram to extract features; It should be noted that the multi-source data sources for fault samples obtained through historical fault reports and training data mainly include electrical parameters, environmental parameters, and equipment parameters, which are commonly used feature parameters for photovoltaic equipment fault diagnosis. In order to more accurately and comprehensively describe the operating status and fault conditions of photovoltaic equipment, the existing technology has introduced abnormal sound detection technology, such as the new energy power generation monitoring system provided by Publication No. CN119209891A. However, the above-mentioned existing technology is relatively superficial in analyzing spectrograms. Currently, abnormal sound detection still faces the problem of scarce abnormal samples. To address the problem of data scarcity, the existing technology has also explored data augmentation and pre-training model strategies, such as introducing anomalies in the spectrogram or generating false samples in the latent space to generate synthetic samples. However, the quality of synthetic samples is difficult to judge and may affect model performance. The pre-training model strategy uses large-scale data for pre-training and then fine-tunes it to machine audio. However, the fixed-architecture Transformer model lacks flexibility and is difficult to adapt to the diversity of spectrogram feature patterns. However, using spectrogram analysis when modeling machine sound is effective for abnormal sound detection. Therefore, it is necessary to address the problem of insufficient analysis of sound detection technology applied to photovoltaic equipment fault diagnosis.

[0049] Furthermore, the local fault feature fusion neural network prediction model proposed in this application is composed of a non-fixed architecture CNN-Transformer model. First, it supplements samples for the detection of fault abnormality samples. Second, the feature fusion result is obtained by aligning the local time output of multi-source data of atomic components under the CNN-Transformer model, which can be used as data support in the samples of the second fault situation analysis. While retaining the feasibility of in-depth photovoltaic equipment fault diagnosis, it comprehensively describes the fault warning status of the photovoltaic equipment based on the first fault situation and the second fault situation, providing favorable data support for subsequent operation and maintenance management.

[0050] See also Figure 3 When analyzing the second fault condition, in the embodiment of the present application, a composite network of a core frame scanning network, a residual neural network, a pooling layer, and a fully connected layer is used to construct a spectrum graph analysis model. In specific implementation, determining the second fault condition includes the following steps: The audio signal of the historical working audio data is converted into a historical spectrogram through short-time Fourier transform; the audio signal can be converted into a time-frequency domain representation through a preprocessing step, which facilitates the subsequent scanning of the spectrogram through the kernel frame.

[0051] Setting the target kernel frame size, and dynamically calculating the step size of the kernel frame in the time dimension and the step size of the frequency dimension based on the size of the target kernel frame and the number of target features; The historical spectrum is scanned by a target core frame, and features extracted by the scan are stacked into multi-channel features according to the scale. The multi-channel features are weighted according to the channel dimension to form a multi-channel feature map. The weighted processing is used to screen multiple features scanned by the same target core frame; Optionally, in addition to the core frame scanning network provided in this application, other spectrum graph feature extraction networks can also be used to obtain spectrum graph features. However, the core frame scanning network proposed in this application based on multiple core frames of different sizes can perform multi-scale analysis on the spectrum graph to obtain spectrum graph features at different scales, so that the scanning network can focus more on the target scale, and at the same time can capture fine-grained and coarse-grained features in the spectrum graph, which has a better recognition effect on the second fault situation and further improves the accuracy of photovoltaic equipment fault warning.

[0052] Optionally, you can customize kernel frames of different scales to scan the spectrum according to actual needs. The kernel frame is represented by , where h represents the height of the kernel frame and w represents the width of the kernel frame.

[0053] Preferably, when the kernel frame of the custom scale is higher than the width, the kernel frame is a rectangular kernel frame, which can match the significant frequency characteristics in the working audio of the target device while maintaining the fine granularity of capturing the time characteristics in the time dimension.

[0054] Preferably, the diversity of scanning is increased by setting multiple kernel frames of different sizes, so that the scanning network can not only focus on the target scale, but also capture the fine-grained and coarse-grained features in the spectrum graph.

[0055] Specifically, in order to ensure that the entire spectrum can be covered, the step size is determined and The dynamic calculation expression is: ; ; Among them, F represents the frequency dimension on the spectrum graph, T represents the time dimension on the spectrum graph, represents the number of features required in the frequency dimension, It represents the number of features required in the time dimension, h represents the kernel frame height and corresponds to the frequency dimension, and w represents the kernel frame width and corresponds to the time dimension.

[0056] For further information, see Figure 4 The dynamic step size calculation can adapt to the size of the spectrum graph and the setting of the target kernel frame, so as to extract the fault features of the spectrum structure at each scale, especially in the scenario of different target devices. It can enhance the adaptability of the scanning network for processing spectrum graphs with different resolutions, so that the scanning network can extract consistent fault features in different data sets corresponding to different atomic components of the target device. Since the required number of features are extracted through the kernel frame at different scales, in order to further obtain the significant features of the fault, a corresponding channel is set for each feature, that is, the number of channels is equal to the number of features. Then, the features can be stacked according to the kernel frame size and the number of features to obtain N channel feature maps. It can be understood that the size of the channel feature map obtained by weighting in the dimension of N channels is still the size determined according to the kernel frame size and the number of features before weighting.

[0057] Next, the output features of each core frame can be processed by a convolutional network spliced ​​with the scanning network. In order to further reduce computational loss and cost, a lightweight convolutional network can be selected. The lightweight convolutional network is used to extract information about fault features at the target scale, which can reduce the significant delay introduced by the embedding process. By integrating the fault features, a comprehensive representation of the faults in the input spectrum graph can be achieved, which can be used to supplement the insufficient samples of local faults and global faults in the first fault situation prediction results.

[0058] In a specific implementation, a residual neural network and a statistical pooling layer are used to generate an embedded representation based on the multi-channel feature map. The residual neural network is used to encode the nuclear frame scanning results of different scales to learn and adapt to the acoustic spectrum characteristics of the atomic components at different scales. The statistical pooling layer is used to calculate the mean and standard deviation of the output channel. As an example, the residual neural network and statistical pooling layer are used to extract multi-scale features and generate a unified embedding. The residual neural network can be set as a residual block to continuously stack two residual blocks of the same structure, doubling the number of channels from 32 to 64, corresponding to the size of each inner convolution kernel. , the first step residual block stride is 2 to halve the size of the feature map.

[0059] Furthermore, the statistical pooling layer calculates the mean and standard deviation of the channel to generate a 256-dimensional embedding feature. The lightweight residual network encodes the fault feature results of kernel frame scans at different scales, which can automatically learn and adapt to the fault features of target devices and atomic components at different granularities. Finally, the fault features at different scales obtained from the spectrum graphs of kernel frame scans at different sizes are integrated into a unified embedding representation, thereby providing a scalable fault feature expression for the first fault condition. By splicing the network, the complexity of the model can be reduced, and the work efficiency of operation and maintenance managers in fault diagnosis, detection, and early warning can be improved.

[0060] In a specific implementation, the embedded representation is clustered with the local fault samples and the global fault samples in the historical fault report and the first fault situation to generate abnormal samples, an abnormal score threshold is output based on the abnormal samples through distance measurement, the local fault of the atomic component and the global fault of the target device are updated based on the abnormal score threshold, and the second fault situation in the target device is output.

[0061] Optionally, the fault samples are subjected to multidimensional scaling analysis to adjust their feature dimensions to the same embedding representation as the feature dimensions in the first fault scenario. K-means clustering is then used to obtain several cluster centers representing the fault anomaly samples. For the test samples of atomic components in photovoltaic equipment, a cosine distance measurement can be performed between them and the generated anomaly samples. The minimum cosine distance is used as the outlier value for the test sample. Specifically, a preset anomaly score threshold can be used to classify test samples with anomaly scores above this threshold as anomalies.

[0062] It can be understood that since the application of the spectrum graph analysis model in the present application allows operation and maintenance personnel to obtain historical working audio when collecting fault reports, some abnormal samples can be determined from it. In the embodiment of the present application, the multi-source cross-modal fault data has been fused into output local features and global features through the non-fixed architecture GNN-Transformer network. The abnormal sample data of the working audio can be compatible in the Transformer network, that is, the first fault condition can be clustered and fused with the embedded representation output by the spectrum graph analysis model, which can enrich the fault samples and reduce the problem of sample sparsity. At the same time, the abnormal samples can improve the effectiveness of fault warning under multi-scale analysis. Therefore, the local fault conditions and global fault conditions after updating the spectrum graph analysis model can more completely and comprehensively express the fault conditions of the photovoltaic equipment. As a more accurate data basis, it can improve the efficiency and safety of the operation and maintenance management of the photovoltaic equipment.

[0063] 4100. Acquire real-time working data of atomic components in the target device, perform fault warning according to the first fault condition and the second fault condition, and store the data in a fault analysis library.

[0064] In a specific implementation, based on the local failure probability corresponding to each of the atomic components and the global failure probability corresponding to each of the target devices, the atomic components and associated target devices with failure probabilities higher than a preset probability threshold are screened to obtain multiple failure relationship combinations; Obtain real-time working data of atomic components in each fault relationship combination; In further implementation, the real-time working data is input into a fault warning neural network model for prediction to obtain an output target fault warning situation, and the fault warning neural network model is trained by the unified data set and the abnormal sample data set; the fault warning neural network model is preset with a fault association rule to learn the fault relationship combination, and the fault association rule is set by the fault identification results in the historical data and the fault set of the temporal and spatial correlation of the fault; Specifically, the preset steps of the fault association rule include: For each preset atomic component failure type in the target device, obtain multiple local fault results and corresponding historical fault times corresponding to the local fault type of the target device failure type in the historical fault report database; calculate the weighted sum of the similarities between each local fault result and all local fault results of the atomic components under the target device to obtain the similarity parameters corresponding to the atomic component failure type in the target device; screen out at least one centralized fault set from the multiple local fault results; specifically, the centralized fault set includes multiple local fault results whose time difference between the failure times is less than the time difference threshold; calculate the average fault time of all local fault results in each centralized fault set to obtain the collection time of each centralized fault set; calculate the average time difference between the collection time of each centralized fault set and the current time to obtain the time parameter corresponding to the atomic component failure type in the target device; calculate the product of the similarity parameter and the time parameter to obtain the priority parameter of the atomic component failure type in the target device; screen out at least one atomic component failure type whose priority parameter is greater than the parameter threshold from all atomic component failure types to obtain the target device failure association based on the atomic components for verifying the failure combination relationship.

[0065] It can be seen that through the above optional embodiments, local faults and global faults in the target device fault situation can be accurately determined based on the similarity of the fault type of the preset target device with the local fault in the historical data and the concentration and proximity of the historical fault time. The fault prediction results of the first fault situation and the second fault situation can be cross-verified, which facilitates accurate early warning of target device and atomic component failures and provides a more accurate data basis for photovoltaic operation and maintenance.

[0066] In a further implementation, the target fault warning situation is matched with the fault report data corresponding to the atomic component; in the case of a match, the target fault warning situation is displayed to the operation and maintenance personnel through a user graphical interface based on the fault classification and dynamic programming algorithm, and the target equipment is remotely controlled through the user collaborative operation interface; in the case of a mismatch, the target fault warning situation is determined to be untrue fault report data and is deleted from the fault analysis library.

[0067] In a further embodiment, the preset steps of the fault classification rule include: For each target device failure, determine the fault level parameter corresponding to the target device failure according to the preset fault level rule; determine the historical warning terminal record corresponding to the target device failure according to the historical fault report record corresponding to the failure type of the target device; calculate the total number of different types of warning terminals in the historical warning terminal records to obtain the terminal complexity parameter corresponding to the target device failure; calculate the product of the fault level parameter and the terminal complexity parameter to obtain the warning priority corresponding to the target device failure; determine the warning strategy corresponding to the target device based on the dynamic programming algorithm according to the historical warning terminal record and the warning priority corresponding to each target device failure.

[0068] It can be seen that through the above optional embodiments, the corresponding historical warning terminal records and warning priorities can be determined by analyzing the level of each target device fault and the historical warning records, so as to determine a more reasonable and accurate alarm strategy based on the dynamic programming algorithm, and realize more efficient and accurate target device warnings based on fault type and fault level, providing a more accurate data basis for photovoltaic operation and maintenance, and improving the safety of photovoltaic equipment operation.

[0069] In a further embodiment, the dynamic programming algorithm determines the early warning strategy of the target device, including the following steps: The objective function is set to maximize the number of target device failures determined to be warned in the warning strategy and minimize the total warning time corresponding to the warning strategy; wherein the total warning time is obtained by predicting the sum of the time taken to send each warning instruction in the warning strategy to the corresponding operation and maintenance user terminal device at the corresponding warning time; the time is obtained by inputting the warning instruction and the corresponding user terminal device into a trained time prediction model for prediction; the time prediction model is trained using a training data set including multiple training warning instructions, corresponding operation and maintenance user terminal devices, and sending time annotations; In specific implementation, the setting of restriction conditions includes: the higher the corresponding warning priority of the target device failure in the warning strategy, the earlier the corresponding warning time; no warning is issued for the target device failure whose corresponding warning priority in the warning strategy is lower than the preset first priority threshold; the average warning priority of all target device failures for which warnings are issued in the warning strategy is greater than the second priority threshold; the first priority threshold is greater than the second priority threshold; the operation and maintenance user terminal device corresponding to each target device failure in the warning strategy is the device that appears the most times in the corresponding historical warning terminal record; based on the dynamic programming algorithm, the warning strategies for all the target device failures are iteratively calculated according to the objective function and the restriction conditions until the optimal one is obtained to obtain the warning strategy corresponding to the target device.

[0070] It can be seen that through the above-mentioned optional embodiments, calculations can be performed based on dynamic programming algorithms and preset reasonable objective functions and constraints to determine a more reasonable and accurate fault warning strategy, and to achieve more efficient and accurate photovoltaic equipment fault warnings by reversely verifying fault relationship combinations and fault level classification processing based on fault types, providing a more accurate data basis for photovoltaic operation and maintenance, and improving the safety of photovoltaic equipment operation.

[0071] It should be noted that after the fault warning method of the present application provides the analysis results, the operation and maintenance system can trigger a corresponding warning signal to the operation and maintenance user end according to the target fault situation, and display the target fault situation through the user interface at the same time. Among them, the user graphical interface display can be displayed in a multi-level interface, including but not limited to thermal, equipment status details and real-time waveform diagrams, etc. It is easy for technical personnel in this field to know that the form of interface display can be adaptively set according to the target equipment and atomic components corresponding to the actual warning. After receiving the warning signal, the user collaborative operation interface remotely controls the equipment through KVM (such as keyboard, mouse and screen clicks, etc.), including but not limited to restarting the faulty equipment, switching to backup equipment or adjusting the system configuration, etc. Relying on the accuracy and real-time nature of the fault warning, the response time of the control process can be controlled within 100ms, ensuring that the photovoltaic power station can respond quickly when there is a hidden fault. The emergency processing capability of photovoltaic operation and maintenance management is improved.

[0072] It should be noted that when constructing a fault analysis library, in addition to recording and saving relevant data on faults, it is also necessary to consider the relationship between the fault conditions of target devices and atomic components during working hours and historical operating hours and changes over time. In addition, the fault analysis library construction process also needs to consider the classification and organization of historical fault conditions, including but not limited to classification and archiving according to different target devices, different atomic components, different fault types, fault samples, fault parameters and fault attribution, to facilitate data calls in subsequent control links, especially when facing the increase in system complexity caused by the addition of new equipment due to scale expansion, it can quickly respond to control needs, including parameter calibration of fault warning models in photovoltaic power station systems using fault analysis library data.

[0073] At the data storage level, the fault analysis library of the present application adopts a unified and structured data storage solution, which can encapsulate all information such as fault scenarios, equipment parameters, atomic component parameters, fault samples and corresponding fault sample warning solutions related to each fault type or each target device type in a standardized node database. Specifically, each node data can be classified and sorted according to the fault type or device type. Each node includes not only the corresponding attributes of the fault, but also other attribute scalars of the fault-related nodes, which can be one or a group of data, which can identify the specific location of the local fault in the fault classification, such as the fault ID or fault category; and the edge connection relationship defines the proximity or similarity between nodes, and is also applicable to fault scenarios, equipment parameters, atomic component parameters, fault samples and corresponding fault sample warning solutions as node storage and retrieval processing, providing a solid data foundation for achieving fast and accurate photovoltaic equipment fault warning data search.

[0074] Finally, the unique technical advantage of this application lies in the introduction of sample authenticity verification and comprehensive analysis of fault relationships for fault prediction and diagnosis based on cross-modal multi-source data in the field of photovoltaic operation and maintenance, which can improve the accuracy of fault prediction results based on samples. At the same time, the fault warning equipment corresponding to the multi-source data is atomically split, and a neural network model based on the local fault analysis results of the atomic components is constructed to characterize the global fault situation that characterizes the fault relationship between the atomic components and the target equipment through local fault feature fusion, effectively improving the robustness of the model prediction results, and can better achieve the transition from single-point warning to topological association to causal inference. In addition, the introduction of abnormal sound detection with multi-scale feature analysis can mine abnormal samples for fault diagnosis, so as to achieve in-depth evaluation of fault diagnosis, fault warning and operation and maintenance management. At the same time, through local fault feature fusion, a data alignment solution is provided under different fault situation analysis. While retaining the feasibility of in-depth photovoltaic equipment fault diagnosis through abnormal sound detection, the local fault situation and global fault situation updated by the spectrum analysis model can more completely and comprehensively express the fault situation of the photovoltaic equipment. As a more accurate data foundation, it can improve the efficiency and safety of photovoltaic equipment operation and maintenance management.

[0075] See also Figure 2According to one aspect of the present application, a data-driven photovoltaic equipment fault warning system is provided, the system comprising: a data acquisition module for acquiring historical fault reports, historical operation data and historical working audio data of multiple target devices within a historical period; a first fault prediction module for determining a first fault condition of the target device based on a neural network prediction model according to the historical fault reports and historical operation data; a second fault prediction module for determining a second fault condition in the target device based on a spectrum graph analysis model according to the historical fault reports and historical working audio data, the spectrum graph analysis model comprising a core frame scanning network, a residual neural network, a pooling layer and a fully connected layer, the core frame scanning network being provided with a core frame for scanning the spectrum graph to extract features; a fault warning module for acquiring real-time working data of atomic components in the target device, performing fault warning according to the first fault condition and the second fault condition, and storing the data in a fault analysis library.

[0076] On the basis of any embodiment of the system of the present application, in the system of the present application, the first fault prediction module is configured to determine the fault report data corresponding to each atomic component in the target device based on the historical fault report, the historical fault report including the fault type, fault time, environmental conditions, fault cause and historical fault audio; determine the operation data corresponding to each atomic component based on the historical operation data, the operation data including electrical parameters, environmental parameters and equipment status parameters; determine the authenticity of the historical fault report based on the authenticity prediction neural network model according to the fault report data and operation data corresponding to each of the atomic components, and determine the local fault of each atomic component based on the fault prediction neural network model according to the real fault report data; determine the global fault of the target device based on the neural network prediction model of the fusion of local fault features according to the local fault of each of the atomic components.

[0077] On the basis of any embodiment of the system of the present application, the system of the present application further includes: an authenticity verification module, which is configured to input each fault report data corresponding to each atomic component into a first authenticity prediction neural network model, and output a first authenticity prediction probability for all input fault report data; the first authenticity prediction neural network model is obtained by training a local fault training data set including multiple training fault report data sets and corresponding authenticity annotations; for each atomic component, the fault report data and operation data of the same historical period are input into a second authenticity prediction neural network model, and a second authenticity prediction probability corresponding to the fault report data is output, and the second authenticity prediction probability is obtained by training a local fault training data set including multiple training fault report data sets and corresponding authenticity annotations; for each atomic component, the fault report data and operation data of the same historical period are input into a second authenticity prediction neural network model, and a second authenticity prediction probability corresponding to the fault report data is output, and the second authenticity prediction probability is obtained by training a local fault training data set including multiple training fault report data sets and corresponding authenticity annotations. The neural network model is obtained through multiple training fault report data sets, atomic component training operation data and authenticity-labeled training data sets; the weighted sum of the first authenticity prediction probability and the second authenticity prediction probability is calculated to obtain the authenticity parameter corresponding to each of the fault report data; the calculation of the weighted sum value includes a first weight and a second weight, the first weight represents the authenticity prediction accuracy of the first authenticity prediction neural network model, and the second weight represents the historical proportion of the fault report data to all the real fault report data of the atomic component; the authenticity threshold is set according to the authenticity parameter to filter the historical fault report to determine the authenticity of the historical fault report, and the real fault report is output.

[0078] On the basis of any embodiment of the system of the present application, the system of the present application further includes: a local fault prediction module, which is configured to input the determined real fault report data and the operation data of the atomic component into a fault prediction neural network model for each atomic component to output the local fault and predicted fault probability corresponding to each atomic component, and the fault prediction neural network model includes a feature extraction network, a classification prediction network and a prediction verification network; wherein, the fault prediction neural network model is trained by the following steps: a training data set of multiple data types in the operation data of multiple atomic components, a fault report data set of multiple atomic component types and a training data set of fault annotations are used as a unified data set, and the feature extraction network and the classification prediction network are trained by the unified data set until convergence; a three-component of the type of atomic component, the data type corresponding to the atomic component and the local fault corresponding to the atomic component is trained. The feature extraction network and the classification prediction network that have been trained are distilled to obtain the feature extraction network and the classification prediction network in the fault prediction neural network model corresponding to the three-combination relationship; multiple other associated training data sets for the same local fault in the unified data set are determined according to the three-combination relationship, and the labeling correction results for the same local fault are determined according to the multiple associated training data sets to obtain the correction training data set for training the prediction verification network; the prediction verification network is trained according to the correction training data set until convergence, and the prediction verification network is used to compare the real fault report data as input with multiple predicted local faults, and to compare the component operation data with multiple predicted local faults to determine the result correction probability corresponding to the prediction result, and each predicted local fault probability is corrected by multiplying by the correction probability.

[0079] On the basis of any embodiment of the system of the present application, the system of the present application further includes: a global fault prediction module, which is configured to capture global dependencies by dynamically calculating the correlation weights of different local fault features based on a self-attention mechanism, and align the local time of multi-source data of atomic components through a variable sliding window that extracts local fault features through an embedding layer based on a local attention mechanism, and construct a neural network prediction model for local fault feature fusion; a graph neural network is used to extract features of local faults of the three combined relationships to obtain local fault features, and the local fault features of the target device at adjacent time points are linearly transformed and dot-producted through a variable sliding window to obtain the attention score of the current embedding layer, and the first attention score of the local fault feature fused into the encoder layer is determined by limiting the interaction range through a mask matrix; according to the attention score corresponding to the local fault feature, the range control of the global dependency is determined based on the self-attention mechanism, and the different local fault features in the encoder layer are determined by local feature enhancement of the encoder layer. The first attention weight of the feature is obtained, and the global contribution of the local fault feature is evaluated according to the first attention weight; the output result of each encoder layer is subjected to feature dimensionality upgrade through linear mapping and / or nonlinear mapping, and a label sequence of the global contribution of the local fault feature is output, the second attention score of the decoding layer is calculated through mask matrix fusion according to the feature dimensionality upgrade result, and the attention matrix of the decoding layer is obtained by labeling the global feature according to the label sequence; the local fault feature of the decoder layer is enhanced through the self-attention mechanism and the variable sliding window according to the attention score and the attention matrix to determine the second attention weight of the decoder layer, the local fault features are spliced ​​according to the second attention weight, and the global fault result and predicted fault probability of the target device are output through linear mapping and / or nonlinear mapping; the first fault condition of the target device is determined according to the local fault of the atomic component under the three-combination relationship and the global fault of the target device under the fusion of local fault features.

[0080] On the basis of any embodiment of the system of the present application, the system of the present application further includes: a spectrum graph scanning module, which is configured to generate a historical spectrum graph from the audio signal of the historical working audio data through short-time Fourier transform; set the target kernel frame size, and dynamically calculate the step size of the kernel frame in the time dimension and the step size of the frequency dimension based on the size of the target kernel frame and the number of target features; scan the historical spectrum graph through the target kernel frame, and stack the features extracted by the scan into multi-channel features according to the scale, and the multi-channel features are weighted according to the channel dimension to form a multi-channel feature graph, and the weighted processing is used to screen multiple features scanned by the same target kernel frame; use residuals to perform The neural network and the statistical pooling layer generate an embedded representation based on the multi-channel feature map. The residual neural network is used to encode the kernel frame scanning results of different scales to learn and adapt to the acoustic spectrum characteristics of the atomic component at different scales. The statistical pooling layer is used to calculate the mean and standard deviation of the output channel. The embedded representation is clustered with the local fault samples and the global fault samples in the historical fault report and the first fault situation to generate abnormal samples. An abnormal score threshold is output based on the abnormal sample through distance measurement. The local fault of the atomic component and the global fault of the target device are updated based on the abnormal score threshold, and the second fault situation in the target device is output.

[0081] Based on any embodiment of the system of the present application, the system of the present application further includes: a fault warning operation and maintenance module configured to screen atomic components and associated target devices having a failure probability higher than a preset probability threshold based on the local failure probability corresponding to each of the atomic components and the global failure probability corresponding to each of the target devices, to obtain multiple fault relationship combinations; Acquire the real-time working data of the atomic components in each fault relationship combination; input the real-time working data into the fault warning neural network model for prediction to obtain the output target fault warning situation, and the fault warning neural network model is trained by the unified data set and the abnormal sample data set; the fault warning neural network model is preset with fault association rules to learn the fault relationship combination, and the fault association rules are set by the fault identification results in the historical data and the fault set of the fault temporal and spatial correlation; match the target fault warning situation with the fault report data corresponding to the atomic component; in the case of a match, display the target fault warning situation to the operation and maintenance personnel through the user graphical interface according to the fault classification and dynamic programming algorithm, and remotely control the target equipment through the user collaborative operation interface; in the case of a mismatch, determine the target fault warning situation as untrue fault report data and delete it from the fault analysis library.

[0082] Based on any embodiment of the system of the present application, the system of the present application further includes: a fault association rule module configured to obtain, for each preset atomic component fault type in a target device, multiple local fault prediction results and corresponding historical prediction times corresponding to the local fault type of the target device fault type in a historical fault report database; calculate a weighted sum of similarities between each local fault prediction result and all local fault results of the atomic components of the target device to obtain a similarity parameter corresponding to the atomic component fault type in the target device; screen out at least one concentrated fault set from the multiple local fault prediction results; specifically, the concentrated fault set includes multiple local fault results whose time difference between prediction times is less than a time difference threshold; calculate an average of the prediction times of all local fault results in each concentrated fault set to obtain a set time of each concentrated fault set; calculate an average of the time difference between the set time of each concentrated fault set and the current time to obtain a time parameter corresponding to the atomic component fault type in the target device; calculate the product of the similarity parameter and the time parameter to obtain a priority parameter of the atomic component fault type in the target device; screen out at least one atomic component fault type whose priority parameter is greater than the parameter threshold from all atomic component fault types to obtain a target device fault association based on the atomic components for verifying the fault combination relationship.

[0083] On the basis of any embodiment of the system of the present application, the system of the present application further includes: a fault classification module, which is configured to determine, for each of the target device faults, a fault level parameter corresponding to the target device fault according to a preset fault level rule; determine a historical early warning terminal record corresponding to the target device fault according to a historical fault report record corresponding to the fault type of the target device; calculate the total number of different types of early warning terminals in the historical early warning terminal records to obtain a terminal complexity parameter corresponding to the target device fault; calculate the product of the fault level parameter and the terminal complexity parameter to obtain a warning priority corresponding to the target device fault; determine a warning strategy corresponding to the target device based on a dynamic programming algorithm according to the historical early warning terminal record and the warning priority corresponding to each target device fault.

[0084] On the basis of any embodiment of the system of the present application, the system of the present application further includes: a dynamic programming early warning strategy module, which is configured to set the objective function as the maximum number of target equipment failures determined to be warned in the early warning strategy and the minimum total early warning time corresponding to the early warning strategy; wherein the total early warning time is obtained by predicting the sum of the time taken for each early warning instruction in the early warning strategy to be sent to the corresponding operation and maintenance user-end device at the corresponding early warning time; the time is obtained by predicting the time taken by inputting the early warning instruction and the corresponding user-end device into a trained time prediction model; the time prediction model is obtained by training a training data set including multiple training early warning instructions and corresponding operation and maintenance user-end devices and sending time annotations; wherein, the limit is set The control conditions include: the target device failure with a higher warning priority in the warning strategy has an earlier warning time; the target device failure with a corresponding warning priority lower than a preset first priority threshold in the warning strategy does not issue a warning; the average warning priority of all target device failures that are warned in the warning strategy is greater than the second priority threshold; the first priority threshold is greater than the second priority threshold; the operation and maintenance user terminal device corresponding to each target device failure in the warning strategy is the device that appears the most times in the corresponding historical warning terminal record; based on the dynamic programming algorithm, the warning strategies for all the target device failures are iteratively calculated according to the objective function and the constraint conditions until they are optimal, and the warning strategy corresponding to the target device is obtained.

[0085] Another embodiment of the present application provides a data-driven photovoltaic equipment fault warning device, comprising a processor, a computer-readable storage medium, a memory, and a network interface connected via a system bus. The computer-readable, non-volatile storage medium of the data-driven photovoltaic equipment fault warning device stores an operating system, a database, and computer-readable instructions. The database may store information sequences. When executed by the processor, the computer-readable instructions enable the processor to implement a data-driven photovoltaic equipment fault warning method.

[0086] The processor of this data-driven photovoltaic equipment fault warning device is used to provide computing and control capabilities, supporting the operation of the entire data-driven photovoltaic equipment fault warning device. The memory of this data-driven photovoltaic equipment fault warning device may store computer-readable instructions. When executed by the processor, these computer-readable instructions cause the processor to execute the data-driven photovoltaic equipment fault warning method of the present application. The network interface of this data-driven photovoltaic equipment fault warning device is used to connect and communicate with a terminal.

[0087] In this embodiment, the processor is used to execute Figure 2The memory stores the program code and various data required to execute the above modules or submodules. The network interface is used to realize data transmission between user terminals or servers.

[0088] The non-volatile readable storage medium in this embodiment stores the program code and data required to execute all modules in the data-driven photovoltaic equipment fault warning system of this application, and the server can call the program code and data of the server to execute the functions of all modules.

[0089] The present application also provides a non-volatile readable storage medium storing computer-readable instructions. When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the data-driven photovoltaic equipment fault warning method of any embodiment of the present application.

[0090] The present application also provides a computer program product, comprising a computer program / instruction, which implements the steps of the method described in any embodiment of the present application when executed by one or more processors.

Claims

1. A data-driven photovoltaic equipment fault early warning method, characterized in that: The steps include: Obtain historical fault reports, historical operation data, and historical working audio data of multiple target devices within a historical period; Based on historical fault reports and historical operating data, the neural network prediction model is used to determine the first fault condition of the target equipment; Determining a second fault condition in the target device based on a spectrogram analysis model according to historical fault reports and historical working audio data, the spectrogram analysis model comprising a core frame scanning network, a residual neural network, a pooling layer, and a fully connected layer, the core frame scanning network being provided with a core frame for scanning the spectrogram to extract features; The real-time working data of the atomic components in the target device is obtained, and fault warning is performed according to the first fault condition and the second fault condition, and stored in the fault analysis library.

2. The data-driven photovoltaic equipment fault early warning method according to claim 1, characterized in that: The method of determining the first fault condition of the target device based on the historical fault reports and historical operation data and the neural network prediction model comprises the following steps: Determining fault report data corresponding to each atomic component in the target device based on the historical fault reports, wherein the historical fault reports include fault type, fault time, environmental conditions, fault cause, and historical fault audio; Determine operating data corresponding to each atomic component based on the historical operating data, the operating data including electrical parameters, environmental parameters, and device status parameters; According to the fault report data and operation data corresponding to each of the atomic components, based on the authenticity prediction neural network model, the authenticity of the historical fault report is determined, and based on the fault prediction neural network model according to the actual fault report data, the local fault of each atomic component is determined; According to the local fault of each of the atomic components, a global fault of the target device is determined based on a neural network prediction model fused with local fault features.

3. The data-driven photovoltaic equipment fault early warning method according to claim 2, characterized in that: The method of determining the authenticity of historical fault reports based on the fault report data and operation data corresponding to each of the atomic components and the authenticity prediction neural network model comprises the following steps: For each atomic component, each fault report data corresponding to the atomic component is input into a first authenticity prediction neural network model, and a first authenticity prediction probability is output for all input fault report data; the first authenticity prediction neural network model is obtained by training a local fault training data set including multiple training fault report data sets and corresponding authenticity annotations; For each atomic component, the fault report data and operation data of the same historical period are input into a second authenticity prediction neural network model, and a second authenticity prediction probability corresponding to the fault report data is output, where the second authenticity prediction neural network model is obtained by using multiple training fault report data sets, atomic component training operation data, and authenticity-labeled training data sets; Calculating a weighted sum of the first authenticity prediction probability and the second authenticity prediction probability to obtain an authenticity parameter corresponding to each of the fault report data; the calculation of the weighted sum includes a first weight and a second weight, the first weight representing the authenticity prediction accuracy of the first authenticity prediction neural network model, and the second weight representing the historical proportion of the fault report data to all true fault report data of the atomic component; The authenticity threshold is set according to the authenticity parameter to filter the historical fault report to determine the authenticity of the historical fault report, and a true fault report is output.

4. The data-driven photovoltaic equipment failure early warning method according to claim 2, characterized in that: The method of determining the local fault of each atomic component based on the fault prediction neural network model according to the real fault report data includes the following steps: For each atomic component, the determined actual fault report data and the atomic component's operating data are input into a fault prediction neural network model to output the local fault and predicted failure probability corresponding to each atomic component. The fault prediction neural network model includes a feature extraction network, a classification prediction network, and a prediction verification network. The fault prediction neural network model is trained by the following steps: Using training data sets of multiple data types in the operation data of multiple atomic components, fault report data sets of multiple atomic component types, and training data sets of fault annotations as a unified data set, and training a feature extraction network and a classification prediction network using the unified data set until convergence; For a three-way combination of the type of atomic component, the data type corresponding to the atomic component, and the local fault corresponding to the atomic component, the trained feature extraction network and classification prediction network are distilled to obtain a feature extraction network and a classification prediction network in a fault prediction neural network model corresponding to the three-way combination; Determining multiple other associated training data sets for the same local fault in the unified data set based on the three-way combination relationship, and determining a labeling correction result based on the labeling results of the multiple associated training data sets for the same local fault to obtain a corrected training data set for training the prediction verification network; The prediction verification network is trained according to the corrected training data set until convergence. The prediction verification network is used to compare the actual fault report data as input with multiple predicted local faults, and to compare the component operation data with multiple predicted local faults to determine the result correction probability corresponding to the prediction result, and each predicted local fault probability is corrected by multiplying the correction probability.

5. The data-driven photovoltaic equipment failure early warning method according to claim 4, characterized in that: The method of determining the global fault of the target device based on the local fault of each atomic component and the neural network prediction model fused with the local fault features comprises the following steps: Based on the self-attention mechanism, the global dependency is captured by dynamically calculating the correlation weights of different local fault features. Based on the local attention mechanism, the variability of local fault features is extracted through the embedding layer, and the sliding window aligns the local time of multi-source data of atomic components to build a neural network prediction model for local fault feature fusion. A graph neural network is used to extract features of the local faults associated with the three combinations to obtain local fault features. The local fault features of the target device at adjacent time points are linearly transformed and dot-producted using a variable sliding window to obtain the attention score of the current embedding layer. The mask matrix is ​​used to limit the interaction range to determine the first attention score of the local fault features fused into the encoder layer. According to the attention scores corresponding to the local fault features, a range control of global dependencies is determined based on a self-attention mechanism, first attention weights of different local fault features in the encoder layer are determined by enhancing the local features of the encoder layer, and global contributions of the local fault features are evaluated based on the first attention weights; Perform feature dimension upgrade on each encoder layer output result through linear mapping and / or nonlinear mapping, and output a label sequence of the global contribution of the local fault feature. Calculate the second attention score of the decoding layer through mask matrix fusion based on the feature dimension upgrade result, and obtain the attention matrix of the decoding layer based on the label association of the global feature with the label sequence; Performing local fault feature enhancement on the decoder layer using a self-attention mechanism and a variable sliding window according to the attention score and the attention matrix to determine a second attention weight of the decoder layer, splicing the local fault features according to the second attention weight, and outputting a global fault result and predicted fault probability of the target device through linear mapping and / or nonlinear mapping; According to the local failure of the atomic components under the three-way combination relationship and the global failure of the target device under the fusion of the local failure characteristics, a first failure condition of the target device is determined.

6. The data-driven photovoltaic equipment fault early warning method according to claim 1, characterized in that: The method of determining the second fault condition of the target device based on the historical fault reports and the historical working audio data and the spectrogram analysis model comprises the following steps: Generate a historical spectrogram from the audio signal of the historical working audio data by short-time Fourier transform; Setting the target kernel frame size, and dynamically calculating the step size of the kernel frame in the time dimension and the step size of the frequency dimension based on the size of the target kernel frame and the number of target features; The historical spectrum is scanned by a target core frame, and features extracted by the scan are stacked into multi-channel features according to the scale. The multi-channel features are weighted according to the channel dimension to form a multi-channel feature map. The weighted processing is used to screen multiple features scanned by the same target core frame; A residual neural network and a statistical pooling layer are used to generate an embedded representation based on the multi-channel feature map. The residual neural network is used to encode the nuclear frame scanning results of different scales to learn and adapt to the acoustic spectrum characteristics of the atomic components at different scales. The statistical pooling layer is used to calculate the mean and standard deviation of the output channel. The embedded representation is clustered with the local fault samples and the global fault samples in the historical fault report and the first fault situation to generate abnormal samples, an abnormal score threshold is output according to the abnormal samples through a distance metric, the local fault of the atomic component and the global fault of the target device are updated based on the abnormal score threshold, and a second fault situation in the target device is output.

7. The data-driven photovoltaic equipment failure early warning method according to claim 4, characterized in that: The method of obtaining real-time working data of atomic components in the target device, performing fault warning according to the first fault condition and the second fault condition, and storing the data in the fault analysis library includes the following steps: According to the local failure probability corresponding to each of the atomic components and the global failure probability corresponding to each of the target devices, screening the atomic components and associated target devices whose failure probabilities are higher than a preset probability threshold, to obtain multiple failure relationship combinations; Obtain real-time working data of atomic components in each fault relationship combination; The real-time working data is input into a fault warning neural network model for prediction to obtain a target fault warning condition as an output, wherein the fault warning neural network model is trained by the unified data set and the abnormal sample data set; the fault warning neural network model is preset with fault association rules to learn the fault relationship combination, and the fault association rules are set by the fault identification results in the historical data and the fault set with temporal and spatial correlation; The target fault warning situation is matched with the fault report data corresponding to the atomic component; in the case of a match, the target fault warning situation is displayed to the operation and maintenance personnel through a user graphical interface according to the fault classification and dynamic programming algorithm, and the target equipment is remotely controlled through the user collaborative operation interface; in the case of a mismatch, the target fault warning situation is determined to be untrue fault report data and is deleted from the fault analysis library.

8. A data-driven photovoltaic equipment fault warning system, characterized in that: The system is used to execute the data-driven photovoltaic equipment fault early warning method according to any one of claims 1 to 7, and the system includes: A data collection module is used to obtain historical fault reports, historical operation data, and historical working audio data of multiple target devices within a historical period; A first fault prediction module, configured to determine a first fault condition of a target device based on historical fault reports and historical operating data and a neural network prediction model; a second fault prediction module, configured to determine a second fault condition in the target device based on a spectrogram analysis model according to historical fault reports and historical working audio data, the spectrogram analysis model comprising a core frame scanning network, a residual neural network, a pooling layer, and a fully connected layer, the core frame scanning network being provided with a core frame for scanning the spectrogram to extract features; The fault warning module is used to obtain real-time working data of atomic components in the target device, perform fault warning according to the first fault condition and the second fault condition, and store the data in the fault analysis library.

9. A data-driven photovoltaic equipment fault warning device, comprising: at least one processor, and A memory communicatively connected to the at least one processor; characterized in that the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the data-driven photovoltaic equipment fault early warning method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program implemented according to any one of claims 1 to 7 in the form of computer-readable instructions, and when the computer program is called and executed by a computer, the steps included in the corresponding method are executed.

Images