Big data-based transformer substation remote intelligent patrol method and system

Abstract

The invention relates to the technical field of transformer substations, and discloses a transformer substation remote intelligent patrol method and system based on big data, and the method comprises the steps: obtaining and preprocessing multi-source heterogeneous data; the data tokens are converted into a unified device state sequence; generating attention bias based on the causal knowledge base, guiding the cross-modal fusion network to analyze the sequence and outputting an abnormal token; constructing a dynamic causal map based on the abnormal token, identifying a fault root cause and forming a diagnosis conclusion; and generating a report according to the diagnosis conclusion, and performing closed-loop updating on the causal knowledge base. The system comprises a data acquisition module; a data processing module; an intelligent analysis module; and a decision support module. According to the method, causal reasoning is introduced, deep fusion of multi-source data and accurate positioning of fault root causes are achieved, the diagnosis conclusion interpretability is high, self-evolution can be achieved through closed-loop updating of the causal knowledge base, and the intelligent level of substation patrol is comprehensively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of transformer substations, and in particular to a transformer substation remote intelligent inspection method and system based on big data. BACKGROUND

[0002] As the core hub of the power system, the safe and stable operation of the transformer substation is crucial to the entire power grid. In order to replace the traditional inefficient and potentially dangerous manual inspection, the development of a remote intelligent inspection system using big data and artificial intelligence technology has become an important technical trend to improve operational efficiency and ensure the reliability of the power grid.

[0003] Currently, in the application of intelligent inspection of transformer substations, existing technical solutions generally collect massive operation data by deploying multiple types of sensors and make simple threshold judgments, and use data-driven models such as deep learning for analysis. These models can learn the statistical characteristics between data to some extent, enabling automatic identification and alarm of equipment abnormal states, and initially achieving the intelligentization of operational work.

[0004] However, the current technical solution stops at shallow association in data fusion, which easily loses key information during analysis. Therefore, when multiple abnormalities occur, only scattered alarms can be given, making it difficult to clarify the internal relationship between phenomena and locate the root cause of the fault, resulting in insufficient early warning capability. The ambiguity of this diagnosis makes the operational work more of a post-response, making it difficult to achieve true preventive maintenance. In addition, the analysis model is usually fixed, lacking a mechanism for autonomous learning and optimization from real fault cases, and the diagnostic capability is difficult to improve with the accumulation of operational experience.

[0005] Therefore, the present application proposes a transformer substation remote intelligent inspection method and system based on big data to solve the deficiencies of the prior art. SUMMARY

[0006] The purpose of the present application is to provide a transformer substation remote intelligent inspection method and system based on big data, which solves the problem of insufficient data fusion, lack of causal root cause positioning capability, and inability of the analysis model to self-evolve in the existing transformer substation remote inspection technology.

[0007] To achieve the above-mentioned purpose, the first aspect of the present application provides a transformer substation remote intelligent inspection method based on big data, comprising the following steps: Obtain multi-source heterogeneous data of a real-time transformer substation, and perform standardized preprocessing on the multi-source heterogeneous data.

[0008] Specifically, through the sensors deployed in the transformer substation, real-time collection is performed on the electrical data, visible light image and infrared spectrum data, key state monitoring data, and voiceprint data. After the data is acquired, data cleaning is performed on the multi-source heterogeneous data to correct or fill in missing values and eliminate outliers; then, the multi-source heterogeneous data after cleaning is subjected to format unification and time alignment processing based on timestamps to generate standardized data streams for subsequent analysis.

[0009] The multi-source heterogeneous data after standardization preprocessing is subjected to tokenization representation of unified spatiotemporal data streams to obtain a device state token sequence.

[0010] Specifically, for the electrical data and key state monitoring data, slicing is performed through a sliding window and the first preset encoder is used to encode them into first-type device state tokens; for the visible light image and infrared spectrum data, block processing is performed and the second preset encoder is used to encode each image block into second-type device state tokens; for the voiceprint data, it is first converted into a time-frequency spectrogram, then the time-frequency spectrogram is subjected to block processing, and the third preset encoder is used to encode each spectrogram block into third-type device state tokens. Finally, the first-type, second-type, and third-type device state tokens are combined to form a unified device state token sequence. Optionally, before the device state token sequence is input into a subsequent network, a corresponding position encoding vector is superimposed on each device state token in the device state token sequence to retain its original spatiotemporal position information.

[0011] An attention bias matrix is generated based on a preset causal knowledge base, and a cross-modal fusion attention network is used to perform fusion analysis on the device state token sequence. The core of this step is to bias the attention score using the attention bias matrix to output a state representation vector including an abnormal suspect token combination. The process of biasing the attention score using the attention bias matrix can be realized by the following formula: In the formula, A score is the attention score; Q is a query matrix generated by the device state token sequence; K is a key matrix generated by the device state token sequence; d k is the dimension of the key vector; M bias is the attention bias matrix.

[0012] The attention bias matrix M bias is generated in the following manner: the causal knowledge base is searched, and if there is a causal path in the causal knowledge base in which a physical phenomenon corresponding to a first device state token points to a physical phenomenon corresponding to a second device state token, the attention bias matrix M biasThe element values ​​at the positions corresponding to the second and first device status tokens are set to preset positive values; otherwise, they are set to zero. This allows the cross-modal fusion attention network to prioritize correlations that conform to existing physical laws or expert experience during the analysis process.

[0013] Based on the combination of abnormal suspect tokens in the state representation vector, a dynamic causal graph is constructed, and the root cause of the fault is identified according to the dynamic causal graph to obtain a diagnostic conclusion.

[0014] This step specifically includes: extracting an abnormal physical variable set from the abnormal suspect token combination; using a preset causal discovery algorithm, learning and constructing the dynamic causal graph that represents the causal relationship between variables by performing a series of conditional independence tests on the variables in the abnormal physical variable set; finally, identifying the nodes with an in-degree of zero in the dynamic causal graph as the root cause of the fault, and forming the diagnostic conclusion based on the root cause of the fault.

[0015] The causal knowledge base is updated based on the diagnostic conclusions.

[0016] Specifically, once the diagnostic conclusion is verified to be correct by human verification, the verified causal relationships are extracted from the dynamic causal graph included in the diagnostic conclusion, and the verified causal relationships are added to the causal knowledge base, thereby realizing the closed-loop update and self-evolution of the knowledge base.

[0017] A smart inspection and diagnostic report is generated based on the diagnostic conclusions.

[0018] The intelligent inspection and diagnostic report includes the root cause of the fault, and the dynamic causal graph characterizing the transmission path between the root cause of the fault and the abnormal suspected token combination, providing decision support for operation and maintenance personnel.

[0019] A second aspect of the present invention provides a remote intelligent inspection system for substations based on big data, the system being configured to perform the methods described in any of the above embodiments.

[0020] In one specific embodiment, the system includes: The data acquisition module is used to acquire multi-source heterogeneous data from the substation in real time. The data processing module is used to perform standardized preprocessing on the multi-source heterogeneous data; The intelligent analysis module is used for: The multi-source heterogeneous data, after standardization and preprocessing, is represented by a unified spatiotemporal data stream tokenization to obtain a device status token sequence. generate an attention bias matrix based on a preset causal knowledge base, and perform fusion analysis on the device state token sequence by using a cross-modal fusion attention network, wherein when performing attention calculation, the cross-modal fusion attention network biases the attention score by using the attention bias matrix, so as to output a state representation vector including an abnormal suspect token combination; based on the abnormal suspect token combination in the state representation vector, a dynamic causal graph is constructed, and a fault root cause is identified according to the dynamic causal graph, and a diagnosis conclusion is obtained; a decision support module, configured to: update the causal knowledge base according to the diagnosis conclusion; generate an intelligent patrol diagnosis report based on the diagnosis conclusion Based on the above, the present application has at least one of the following beneficial technical effects: 1. The present application realizes the deep fusion of different data types in a single model by tokenizing the unified representation of multi-source heterogeneous data and using a cross-modal fusion attention network for analysis. This method breaks through the limitations of independent analysis or simple splicing of data in traditional methods, can more comprehensively capture the complex correlations between different physical quantities, solves the problem of insufficient data fusion in the prior art, and significantly improves the accuracy and comprehensiveness of the comprehensive evaluation of the state of the substation equipment based on big data.

[0021] 2. The present application generates an attention bias matrix by introducing a preset causal knowledge base to guide the fusion analysis network to focus on device state features with potential causal correlation, and constructs a dynamic causal graph for fault tracing on this basis. This design not only identifies abnormalities from massive data, but also accurately locates the fault root cause with an in-degree of zero, solves the problem that traditional methods can only find correlations but lack the ability to locate causal root causes, and makes the diagnosis conclusion of remote intelligent patrol more in-depth and interpretable.

[0022] 3. The present application establishes a feedback update mechanism from the diagnosis conclusion to the causal knowledge base, supplements new causal relationships in the verified dynamic causal graph to the knowledge base, realizes the closed-loop evolution of the analysis model, and continuously enriches the knowledge base of the system as the running time increases and the number of processed cases increases. The attention bias matrix generated by the system will be more and more accurate, thereby continuously improving the diagnosis efficiency and accuracy of subsequent intelligent patrol tasks, and solving the problem that the analysis capability of the existing static model is fixed and cannot evolve. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 a flowchart of the substation remote intelligent patrol method based on big data of the present application; Figure 2 a flowchart of the data processing and fusion analysis process of the present application; Figure 3The schematic diagram of the diagnosis decision and closed-loop updating process of the application; Figure 4 The architecture diagram of the substation remote intelligent patrol system based on big data of the application.

[0024] Among them, 10 is a data acquisition module; 20 is a data processing module; 30 is an intelligent analysis module; and 40 is a decision support module. DETAILED DESCRIPTION

[0025] The following will be combined with the drawings Figure 1 -Appendix Figure 4 The application will be further described in detail.

[0026] The embodiment of the application provides a substation remote intelligent patrol method and system based on big data, Refer to the drawings Figure 1 , Figure 1 It is the flow chart of the substation remote intelligent patrol method based on big data according to one embodiment of the application. The application provides a substation remote intelligent patrol method based on big data, which can include the following steps: S101: acquiring multi-source heterogeneous data of a real-time substation, and performing standardized preprocessing on the multi-source heterogeneous data; S102: tokenizing and representing the multi-source heterogeneous data after the standardized preprocessing in a unified space-time data stream, to obtain a device state token sequence; S103: generating an attention bias matrix based on a preset causal knowledge base, and performing fusion analysis on the device state token sequence by using a cross-modal fusion attention network, to output a state representation vector including an abnormal suspicious token combination; S104: constructing a dynamic causal graph based on the abnormal suspicious token combination in the state representation vector, and identifying a fault root cause according to the dynamic causal graph, to obtain a diagnosis conclusion; S105: updating the causal knowledge base according to the diagnosis conclusion; S106: generating an intelligent patrol diagnosis report based on the diagnosis conclusion.

[0027] Refer to the drawings Figure 1 and Figure 2 , Figure 1 It is the flow chart of the substation remote intelligent patrol method based on big data according to one embodiment of the application. The initial step S101 of the method is to acquire multi-source heterogeneous data of a real-time substation, and to perform standardized preprocessing on the multi-source heterogeneous data.

[0028] This step first performs data acquisition. Through a sensor network deployed at different positions in the substation, multi-source heterogeneous data capable of representing the running state of the equipment is collected in real time. In one specific embodiment, the multi-source heterogeneous data includes: Electrical data, visible light image and infrared spectrum data, voiceprint data, and key state monitoring data Electrical data: collected by current transformers and voltage transformers installed on primary equipment lines, and digitized by a data acquisition unit at a preset high sampling frequency (e.g., 10 kHz) to accurately capture steady-state and transient signal characteristics in the power grid.

[0029] Visible light image and infrared spectrum data: collected by a dual-spectrum camera with a pan-tilt function deployed on a preset patrol path. The infrared spectrum is used to obtain the thermal distribution information of equipment (such as joints, knife switches, transformer bushings, etc.), and the visible light image is used to obtain the appearance of the equipment (such as rust, damage), indicator light status, switch position, and surrounding environment information.

[0030] Among them, the visible light data acquisition should meet the following requirements: a) Support the acquisition of meter readings such as cabinet switch action frequency counters, barometers, arrester leakage current meters, oil temperature meters, winding temperature meters, hydraulic pressure meters, on-load voltage regulation position meters, various oil level meters, indoor temperature and humidity meters, etc. b) Support the acquisition of position state indicators of primary equipment such as circuit breakers, disconnectors, grounding knife switches (switches), and secondary equipment such as switching handles, pressure plates, indicator lights, and air switches; c) Support the acquisition of the appearance of equipment and facilities, etc. d) Support the acquisition of substation environment and building facility appearance, etc. e) The target related to feature and state recognition should be in the center of the acquisition frame; f) The acquired image should have time, point name, etc. information superimposed; g) It has all-weather video and image acquisition function.

[0031] Infrared spectrum data acquisition: Support the acquisition of infrared spectrum data of key parts such as equipment body, joints, bushings, and lead wires. The infrared thermal imaging camera supports frame temperature measurement and point temperature measurement.

[0032] Voiceprint data: collected by acoustic sensors (e.g., high-fidelity microphones) installed near devices such as transformers and reactors that are prone to abnormal sound, and digitized at a preset audio sampling rate (e.g., 44.1 kHz) to monitor local discharge sound, mechanical vibration sound, or cooling system operation sound of the equipment.

[0033] Key state monitoring data: such data should be supported by the main and auxiliary equipment monitoring system to obtain transformer (oil chromatogram, core / clamp current, partial discharge), main transformer bushing, current transformer, voltage transformer, and other key state monitoring data. The state monitoring data of main and auxiliary equipment such as sensors, voltage transformers, coupling capacitors, arresters, circuit breakers, GIS combined electrical appliances, disconnectors, storage batteries, fire fighting systems, etc. are recorded at a preset lower frequency (for example, once per minute).

[0034] After obtaining the original multi-source heterogeneous data, standardization preprocessing is performed on the data, which ensures the consistency and effectiveness of the input data of the subsequent analysis model. The preprocessing process specifically includes data cleaning, format unification and time alignment.

[0035] The data cleaning step aims to correct or remove noise and errors in the original data.

[0036] Electrical data (current and voltage waveform) Data cleaning: the focus is on noise removal and abnormal point rejection. A digital filter (such as a low-pass filter) is used to filter out high-frequency interference noise; statistical methods (such as Z-score or 3σ principle) are used to identify and remove transient sharp pulses caused by external interference, and linear interpolation is used for filling.

[0037] Format unification: all high-frequency sampled waveform data points are uniformly converted to a standard numerical format, such as 32-bit floating point, to ensure consistent dimensions and precision.

[0038] Time alignment: since it is high-frequency data, a fixed length time window (for example, 100 milliseconds) is taken around the uniform analysis time reference point, and the data segment within the window is taken as a complete sample at that time.

[0039] Visible light image and infrared spectrum data data cleaning: the focus is on image enhancement and correction. For visible light images, histogram equalization and other methods are used to improve image quality problems caused by uneven lighting or weather effects; for infrared images, noise reduction processing is performed to eliminate the influence of sensor thermal noise.

[0040] Format unification: all images are uniformly scaled to a fixed resolution (for example, 224x224 pixels) by bilinear interpolation and other methods; at the same time, the pixel values are normalized to a uniform numerical interval (such as [0, 1] or [-1, 1]).

[0041] Time alignment: images are discrete snapshot data, and the frame of image closest to the analysis reference point is directly selected as the sample at that time.

[0042] Voiceprint data (device running noise) Data cleaning: the focus is on background noise suppression. A band-pass filter is used to filter out background environmental sound (such as wind and rain) unrelated to device abnormalities, or a spectral subtraction algorithm is used to separate the running voiceprint of the target device.

[0043] Format unification: convert one-dimensional time-domain acoustic signals into two-dimensional time-frequency spectrograms through short-time Fourier transform (STFT). Then, scale the time-frequency spectrograms to a fixed size and normalize them, as if processing images.

[0044] Time alignment: similar to electrical data, take a fixed-length audio segment (e.g., 1 second) centered on the analysis time reference point and perform subsequent processing to ensure consistent sample duration.

[0045] Key state monitoring data (device temperature, ambient humidity, etc.) Data cleaning: focus on outlier removal and missing value filling. Remove obvious abnormal readings caused by sensor jumps; for data missing, use forward / backward filling or linear interpolation to complete it.

[0046] Format unification: ensure that all similar data use uniform physical units (e.g., temperature unified in Celsius) and convert numerical values to standard floating-point format.

[0047] Time alignment: as low-frequency sampling data, directly select the sampling value closest to the analysis reference point as the representative data at that time.

[0048] Through step S101, a set of strictly aligned multi-source heterogeneous data sets in the time dimension is obtained.

[0049] Refer to the attached Figure 1 and Figure 2 , Figure 2 is the data processing and fusion analysis process diagram of an embodiment of the present application. This process corresponds to step S102 in the method flow, and its purpose is to convert the standardized preprocessed multi-source heterogeneous data with different sources and structures into a unified spatio-temporal data stream, which is specifically a sequence of consistent-dimensional device state tokens for subsequent cross-modal fusion attention network processing.

[0050] The specific implementation of this step is as follows: For the electrical data and key state monitoring data preprocessed in step S101, generate the first type of device state token. These data are essentially one-dimensional time series. Use sliding window technology to slice these sequence data. In an embodiment, set the window length L ts and step size S ts . The sliding window moves along the time axis with step size S ts , and each time it intercepts a length of L tsThe data segment is then input to a first preset encoder. The encoder is a linear projection layer that flattens and linearly transforms each one-dimensional data segment into a preset fixed dimension D vector. The dimension D vector is the first type of device state token.

[0051] For the preprocessed visible light image and infrared spectrum data, the generation of the second type of device state token is performed. First, each image with a resolution of HxW is divided into N non-overlapping image patches with a size of Pxp. N=(HxW) / (Pxp). Then, each image patch is input to a second preset encoder. The encoder is also a linear projection layer that flattens each PxpC (C is the number of image channels, for example, 1 for grayscale images and 3 for color images) image patch into a one-dimensional vector and linearly transforms it into a dimension D vector. The dimension D vector is a second type of device state token.

[0052] For the preprocessed voiceprint data, the generation of the third type of device state token is performed. First, a short-time Fourier transform (STFT) is performed on the one-dimensional voiceprint signal waveform to convert it into a two-dimensional time-frequency spectrum. The transform process uses a preset window function (for example, a Hanning window), a window length, and a frame shift length to generate a two-dimensional matrix that reflects the change of signal frequency over time.

[0053] Then, the time-frequency spectrum is regarded as a single-channel image, and the same blocking method as processing the visible light image is used to divide it into M non-overlapping two-dimensional image patches. Finally, each image patch is input to a third preset encoder (which has the same structure as the second preset encoder) to be transformed into a dimension D vector, which is a third type of device state token.

[0054] After generating the above three types of device state tokens, they are combined in a preset order to form a complete device state token sequence. In order to enable the subsequent model to understand the original position information of each token in the sequence, a corresponding position encoding vector also needs to be added to each device state token in the sequence. The dimension of the position encoding vector is also D. In a specific embodiment, the numerical value of the position encoding vector is generated by a preset trigonometric function formula. For the token at position pos in the sequence, the value of the jth element in the D-dimensional position encoding vector PE is calculated as follows: When the dimension index j is even, i.e., j=2i: PE(pos,2i)=sin(pos / 10000 2i / D ); When the dimension index j is odd, i.e., j=2i+1: PE(pos,2i+1) = cos(pos / 10000 2i / D ); where pos is the position index of the token in the sequence, pos ∈ [0, N-1]; N is the total length of the sequence; i is the dimension index pair, i ∈ [0, D / 2-1]; D is the total dimension of the encoding vector. The formula uses different frequencies of sine and cosine functions to generate a unique encoding vector for the token at different positions in the sequence, which can represent the absolute and relative position information of the token.

[0055] The generated position encoding vector is element-wise added to the device state token vector at the corresponding position to obtain a device state token sequence of content information and position information, which is used as the input of step S103.

[0056] Referring to the accompanying Figure 1 and Figure 2 , Figure 2 is a data processing and fusion analysis process diagram of an embodiment of the present application. The process corresponds to step S103 in the method flow, and receives the device state token sequence with superimposed position encoding generated in step S102 as input. The core of the process is to generate an attention bias matrix based on a preset causal knowledge base, and to perform fusion analysis on the device state token sequence using a cross-modal fusion attention network. When performing attention calculation, the cross-modal fusion attention network uses the attention bias matrix to bias the attention score to output a state representation vector including abnormal suspect token combinations. This step first generates an attention bias matrix based on a preset causal knowledge base. The causal knowledge base is a structured database of universally recognized causal relationships between different physical phenomena of the device. These causal relationships come from the device operation mechanism, historical fault data analysis, and prior knowledge of operation and maintenance experts.

[0057] In a specific embodiment, the knowledge base is constructed as a directed graph G=(V, E). where each node V in the node set V represents a specific physical phenomenon or device state, for example, abnormal temperature rise of the main transformer winding, excessive ethylene content in the insulating oil chromatographic analysis, and local discharge sound signal in the switch cabinet.

[0058] Each directed edge (u, v) in the directed edge set E represents that the physical phenomenon corresponding to the node u is the direct cause of the physical phenomenon corresponding to the node V. For example, an edge from the abnormal temperature rise of the main transformer winding to the appearance of local high temperature points in the infrared spectrum indicates that the former is the cause of the latter.

[0059] Before generating the attention bias matrix, it is necessary to establish a mapping relationship between the device state tokens generated in step S102 and the physical phenomenon nodes in the causal knowledge base. This mapping is pre-defined and uniquely associates each type of token (e.g., a token originating from a specific temperature sensor, a token originating from a specific image patch of a specific camera) with a physical phenomenon node V in the knowledge base graph G.

[0060] Attention bias matrix M bias The generation process follows these steps: Initialize a zero matrix of dimension N×N, where N is the total length of the input device status token sequence.

[0061] For each element M in the matrix bias The values ​​(i,j) (where i is the row index and j is the column index, i and j range from 1 to N) are assigned. Indices i and j correspond to the i-th and j-th tokens in the device status token sequence, respectively.

[0062] By using a pre-defined mapping relationship, find the physical phenomenon node v corresponding to the i-th token. i The physical phenomenon node v corresponding to the j-th token j .

[0063] Perform a path search in the directed graph G of the causal knowledge base to determine whether there exists a path from node v. j to node v i A directed path. The existence of this path implies that the physical phenomenon v j It is a physical phenomenon v i The direct or indirect cause.

[0064] If such a directed path exists, then the matrix element M... bias The value of (i,j) is set to a preset positive value (e.g., a large positive number, such as 10.0); if no such path exists, the value of the element remains zero.

[0065] The device state token sequence is then fed into a cross-modal fusion attention network. This network can consist of multiple stacked, structurally identical encoder layers. At the core of each layer is a multi-head self-attention mechanism with causal bias.

[0066] Specifically, the mechanism first linearly transforms the input token sequence matrix to obtain the query matrix Q, the key matrix K, and the value matrix V′. Then, the attention score is calculated using the following formula: The attention score A is calculated using the following formula. score : In the formula, A scoreis the attention score; Q is the query matrix; K is the key matrix; d k is the dimension of the key vector; M bias is the attention bias matrix.

[0067] M bias is directly added to the dot product result of the query matrix and the key matrix, so that when the attention weight is calculated by the Softmax function later, the attention score between those token pairs with prior causal relationship (i.e. the position where M bias (i,j) is positive) is significantly improved, so that the model can give priority to the strong correlation information that conforms to the physical law or expert experience when fusing the analysis. The attention weight is then applied to the value matrix V' to generate the context vector after attention weighted fusion. The context vector is then processed by the feedforward network, residual connection and layer normalization to form the final output of the encoder layer.

[0068] After the network goes through the calculation of all the encoder layers, the output is a set of updated state representation vectors. Each vector in the vector sequence has fused the context information from all the input tokens and has been guided by the causal knowledge. By analyzing the state representation vector sequence (for example, by an additional classification head or according to the deviation of the vector from the normal state benchmark), abnormal suspect token combinations can be identified.

[0069] Referring to the accompanying Figure 1 and Figure 3 , Figure 3 is the diagnostic decision and closed-loop update process diagram of an embodiment of the present application. The process corresponds to step S104 in the method flow, and its purpose is to build a dynamic causal graph reflecting the internal transmission logic of the current fault event based on the abnormal suspect token combinations identified in step S103, and accurately identify the initial fault root cause that leads to a series of abnormal phenomena from it.

[0070] The specific implementation of this step is as follows: First, from the abnormal suspect token combinations identified in the output of step S103, an abnormal physical variable set is extracted. This extraction process uses the mapping relationship between tokens and physical phenomenon nodes established in step S103.

[0071] Specifically, each device state token that is judged to be abnormal is reversely mapped to the physical variable it represents (for example, the temperature of the A-phase winding of the main transformer, the amplitude of the acoustic signal inside the switch cabinet B3, etc.). All these identified abnormal physical variables are collected to form a set, which is used as the analysis object of causal discovery in this stage.

[0072] Then, the aforementioned set of abnormal physical variables is processed by using a preset causal discovery algorithm to learn and construct a dynamic causal graph. In a specific embodiment, the causal discovery algorithm used is the PC algorithm (Peter-Clark algorithm) based on conditional independence test. The specific execution process of the algorithm includes: Constructing an initial skeleton: a fully connected undirected graph is constructed with each variable in the set of abnormal physical variables as a node.

[0073] Performing conditional independence test and removing edges: the algorithm iteratively tests whether there is a conditional independence relationship between any two nodes (variables) X and Y in the graph. For continuous variables, this test can be achieved by calculating the significance of the partial correlation coefficient. The test starts from zero-order conditional independence (i.e., the condition set is empty) and gradually increases the cardinality of the condition set. When testing the independence of variables X and Y given a variable set S, the sample partial correlation coefficient is first calculated. Subsequently, in order to make statistical inference, the partial correlation coefficient is converted into a test statistic approximately following a standard normal distribution by using Fisher's z-transformation, which is calculated as follows: where z XY·S is the test statistic, approximately following a standard normal distribution under the null hypothesis (X and Y are conditionally independent given S); r XY·S is the sample partial correlation coefficient of variables X and Y given the condition set S; and n is the number of data points (sample size).

[0074] The P-value is calculated according to the test statistic z XY·S . If the P-value is greater than a preset significance level a (such as 0.05), the null hypothesis of conditional independence is accepted, and the edge between X and Y is removed.

[0075] This process starts from k = 0, testing all connected node pairs, then k = 1, testing all still connected node pairs and all possible single-variable condition sets thereof, and so on, until there is no edge that can be removed. After this step, the remaining edges in the graph constitute the skeleton of the causal graph.

[0076] Determining the direction of the edges: Based on the skeleton graph, the direction of the edges is determined using specific rules, which is converted into a directed acyclic graph (DAG). The process first identifies a V structure in the skeleton. A V structure is composed of three nodes X, Z, Y, where there is an edge between X and Z, there is an edge between Y and Z, and there is no direct edge between X and Y. If in the previous inspection step, the variables X and Y are determined to be unconditionally independent, but become conditionally dependent when Z is the condition, then the direction of the edge between X and Z and the edge between Y and Z is determined to be directed to Z. After completing the orientation of all V structures, a set of specific direction propagation rules (for example, Meek rules) are applied to derive the direction of other edges, which is constrained to avoid generating new V structures or directed loops, until no more edges can be determined.

[0077] Through the above algorithm steps, a directed acyclic graph is finally constructed, which is the dynamic causal graph. It shows the causal transmission relationship between the current multiple abnormal physical variables in a data-driven manner.

[0078] Finally, the root cause of the fault is identified according to the dynamic causal graph, and a diagnosis conclusion is formed. In the dynamic causal graph, the in-degree (i.e. the number of edges pointing to the node) of each node is calculated. All nodes with an in-degree of zero in the graph are identified as the root cause of the current fault event. A node with an in-degree of zero means that it is a cause rather than a result in the current set of abnormal variables, and is the starting point of the entire abnormal chain. After identifying the root cause node of the fault, a clear diagnosis conclusion can be formed, which clearly states the root cause of the fault and can be accompanied by the transmission path of the root cause node to all other abnormal nodes, thereby completely explaining the occurrence and development process of the entire fault.

[0079] Referring to the accompanying Figure 1 and Figure 3 , Figure 3 diagnostic decision and closed-loop update process schematic diagram of an embodiment of the present application. The process corresponds to step S105 in the method flow, and the core purpose is to use the dynamic causal graph verified by the actual fault event to perform incremental learning and optimization on the prior causal knowledge base used in step S103, thereby forming a closed-loop intelligent system that can self-improve and continuously evolve.

[0080] The specific implementation of this step is as follows: Receive the diagnosis conclusion generated in step S104. The diagnosis conclusion completely includes the identified root cause of the fault and the dynamic causal graph describing the transmission relationship between the abnormal physical variables in the current fault event.

[0081] The diagnostic conclusion will be submitted for manual verification. The verification process aims to confirm whether the diagnostic conclusion (including fault root cause and causal conduction path) generated by the method is consistent with the actual fault situation of the physical world.

[0082] According to the results of manual verification, subsequent knowledge base updating operations are performed. This operation is conditional and only when the diagnostic conclusion is verified manually (the maintenance personnel go to the substation site, confirm whether the fault root cause and conduction path given by the system are completely consistent with the actual physical fault situation of the equipment through maintenance, review, etc.) is correct, the system will start the updating process of the knowledge base.

[0083] After obtaining the correct verification confirmation, all verified causal relationships are extracted from the dynamic causal graph contained in the diagnostic conclusion. Specifically, the dynamic causal graph is a directed acyclic graph, and each directed edge in it, such as an edge from a node representing a physical variable V a pointing to a node representing a physical variable V b , is considered as an exact causal relationship verified by this actual fault case.

[0084] All these causal relationships extracted from the verified dynamic causal graph are supplemented and stored in the prior causal knowledge base. When performing the supplement operation, in order to ensure the logical consistency and storage effectiveness of the prior causal knowledge base, the following rules are followed: Redundancy check: for each causal relationship (i.e. directed edge V a →V b ) to be supplemented, first query in the prior causal knowledge base to determine whether the same causal relationship exists. If it exists, do not add it again.

[0085] Consistency check: for a causal relationship V a →V b that needs to be added after redundancy check, perform a consistency check before adding it formally. The check aims to confirm that the addition of the edge will not form a directed loop in the directed graph structure of the prior causal knowledge base. This check can be performed by starting from node V h and performing graph traversal (e.g. depth-first search) in the current knowledge base to check whether it can access back to node V a . If it cannot access back to V a , it proves that no loop is formed.

[0086] Knowledge base update: only when the causal relationship to be supplemented passes the redundancy check and consistency check, the directed edge V a →V bThe consistency check is passed, the new knowledge is permanently added to the priori causal knowledge base. If the consistency check fails, the addition is rejected, and a system log can be generated to record the logical conflict.

[0087] Each time a successful diagnosis experience is manually verified, it is converted into an objective and reliable enhancement of the core knowledge base. This updated priori causal knowledge base will provide more accurate and comprehensive priori knowledge guidance for the model in subsequent diagnosis tasks (return to step S103), thereby realizing the closed-loop improvement and self-evolution of the entire diagnosis system capability.

[0088] Referring to the accompanying drawings Figure 1 and Figure 3 , Figure 3 The diagnostic decision and closed-loop update process schematic diagram of one embodiment of the present application. The process corresponds to step S106 in the method flow, and its purpose is to automatically convert the structured diagnostic conclusion formed in step S104 into an intelligent patrol diagnosis report with comprehensive content, clear logic and easy to understand, for subsequent equipment maintenance and decision use.

[0089] The specific implementation of this step is as follows: Receive the final diagnosis conclusion determined in step S104 as input. The diagnosis conclusion is a data structure that encapsulates all key analysis results of this fault event, and the core content includes: Fault root cause: the explicit identification of one or more physical variables identified as the starting point of the fault chain.

[0090] Dynamic causal graph: a directed acyclic graph that represents all related abnormal physical variables as nodes and their data-driven verified causal transmission relationships as directed edges. This graph contains the complete transmission path from the fault root cause to all other abnormal phenomena (i.e. the physical variables corresponding to the abnormal suspect token combination identified in step S103).

[0091] Based on the preset report template, the received diagnosis conclusion is parsed, and the key information is filled into the corresponding chapter of the report to generate the final intelligent patrol diagnosis report. The generation process specifically includes: Generate fault root cause summary: the system extracts the physical variable name corresponding to the root cause node from the diagnosis conclusion data structure. The name is directly and explicitly presented in the conclusion section of the report. For example, the report will generate the following text: The root cause of this patrol diagnosis is: [the internal connector of switch cabinet A13 is not in good contact].

[0092] Generating a visualized graph of fault propagation path: the system visualizes the whole dynamic causal graph contained in the diagnosis conclusion. This process converts the logical relationship between nodes and edges of the graph into graphical elements. In the generated graph: each node is represented by a box or circle, and the physical variable name it represents is labeled inside (for example, abnormal temperature rise of connector, generation of partial discharge sound signal, etc.).

[0093] The connection between nodes is drawn with an arrowed line according to the directed edge in the graph, and the direction of the arrow clearly indicates the direction of the causal relationship.

[0094] The node identified as the root cause will be highlighted in a special style (for example, using a different color or a thicker border), so that it is immediately apparent in the whole graph. This visualized graph is embedded in the report as an intuitive explanation of the overall picture of fault development, clearly showing how the fault root cause step by step triggers various phenomena represented by the abnormal suspect token combination.

[0095] Generating a list of abnormal evidence (optional enhancement): to enhance the traceability of the report, the system can also traverse each node in the dynamic causal graph and query its mapping relationship with the original device state token.

[0096] Thus, a detailed evidence list can be generated, which associates each identified abnormal physical phenomenon (node in the graph) with the specific data source that triggered the identification (for example, abnormal reading from a sensor at a certain time, or image from a certain area of a camera), providing a basis for subsequent fine review.

[0097] Finally, the above generated text, visualized graph and optional evidence list are integrated into a complete, structured document (for example, a PDF file or a web page), which is the intelligent patrol diagnosis report.

[0098] Refer to the accompanying Figure 4 , Figure 4 The figure is a schematic diagram of a substation remote intelligent patrol system based on big data according to an embodiment of the present application. The system specifically includes the following functional modules: Data acquisition module 10 for acquiring real-time substation multi-source heterogeneous data; Data processing module 20 for standardizing and preprocessing multi-source heterogeneous data; Intelligent analysis module 30 for: Tokenization representation of multi-source heterogeneous data after standardization preprocessing to obtain device state token sequence; The attention bias matrix is generated based on a preset causal knowledge base, and a cross-modal fusion attention network is used for fusion analysis on the device state token sequence. When performing attention calculation, the cross-modal fusion attention network uses the attention bias matrix to bias the attention score to output a state representation vector including an abnormal suspect token combination. Based on the abnormal suspect token combination in the state representation vector, a dynamic causal graph is constructed, and a fault root cause is identified according to the dynamic causal graph to obtain a diagnosis conclusion. The decision support module 40 is configured to: update the causal knowledge base according to the diagnosis conclusion; generate an intelligent patrol diagnosis report based on the diagnosis conclusion.

[0099] The system of the embodiment can be used to execute the method embodiments described above, and has similar principles and technical effects, which will not be described here.

[0100] Although the embodiments of the present application have been shown and described, it can be understood by those skilled in the art that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A big data based remote intelligent patrol method for a substation, characterized in that, The method comprises the following steps: obtaining multi-source heterogeneous data of a real-time substation, and performing standardized preprocessing on the multi-source heterogeneous data; tokenizing and representing the multi-source heterogeneous data after standardized preprocessing in a unified space-time data stream to obtain a device state token sequence; generating an attention bias matrix based on a preset causal knowledge base, and performing fusion analysis on the device state token sequence by using a cross-modal fusion attention network, wherein when performing attention calculation, the attention bias matrix is used to bias the attention score to output a state representation vector including an abnormal suspect token combination; based on the abnormal suspect token combination in the state representation vector, constructing a dynamic causal graph, and identifying a fault root cause according to the dynamic causal graph to obtain a diagnosis conclusion; updating the causal knowledge base according to the diagnosis conclusion; generating an intelligent patrol diagnosis report based on the diagnosis conclusion.

2. The method of claim 1, wherein, The step of obtaining multi-source heterogeneous data of a real-time substation comprises: real-time collection of the multi-source heterogeneous data by sensors deployed in the substation, wherein the multi-source heterogeneous data includes electrical data, key state monitoring data, visible light image and infrared spectrum data, and voiceprint data; The step of performing standardized preprocessing on the multi-source heterogeneous data comprises: data cleaning of the multi-source heterogeneous data to correct or fill in missing values and eliminate outliers; format unification of the cleaned multi-source heterogeneous data, and time alignment processing based on timestamps.

3. The method of claim 1, wherein the method further comprises: The step of tokenizing and representing the multi-source heterogeneous data after standardized preprocessing in a unified space-time data stream to obtain a device state token sequence comprises: for the electrical data and key state monitoring data, slicing by a sliding window and encoding into a first type of device state token by using a first preset encoder; for the visible light image and infrared spectrum data, block processing and encoding each image block into a second type of device state token by using a second preset encoder; for the voiceprint data, first converting it into a time-frequency spectrum, then block processing the time-frequency spectrum, and encoding each frequency spectrum block into a third type of device state token by using a third preset encoder; combining the first type of device state token, the second type of device state token, and the third type of device state token to form the device state token sequence.

4. The method of claim 1, wherein, The biasing of the attention score by using the attention bias matrix specifically comprises: The attention score A is calculated by the following equation score : wherein Q is a query matrix generated from the sequence of device state tokens; K is a key matrix generated from the sequence of device state tokens; d k is the dimension of the key vector; M bias is the attention biasing matrix.

5. The method of claim 4, wherein, The attention bias matrix M bias is generated in the following way: If there exists a causal path in the causal knowledge base from the physical phenomenon corresponding to the first device state token to the physical phenomenon corresponding to the second device state token, the attention bias matrix M bias is set to a preset positive value for the element value of the position corresponding to the second device state token and the first device state token in the matrix, otherwise it is set to zero.

6. The method of claim 1, wherein the method is a big data based remote intelligent patrol method for a substation, characterized in that, The step of constructing a dynamic causal graph based on the abnormal suspect token combination in the state representation vector, and identifying a fault root cause according to the dynamic causal graph to obtain a diagnosis conclusion comprises: extracting an abnormal physical variable set from the abnormal suspect token combination; using a preset causal discovery algorithm to learn and construct the dynamic causal graph by performing conditional independence test on the variables in the abnormal physical variable set; identifying the nodes with in-degree of zero in the dynamic causal graph as the fault root cause, and forming the diagnosis conclusion based on the fault root cause.

7. The method of claim 1, wherein the method is a big data based remote intelligent patrol method for a substation, characterized in that, The step of updating the causal knowledge base according to the diagnostic conclusion comprises: extracting a verified causal relationship from the dynamic causal graph included in the diagnostic conclusion when the diagnostic conclusion is verified as correct by a human; storing the verified causal relationship into the causal knowledge base.

8. The method of claim 1, wherein the method is a big data based remote intelligent patrol method for a substation, characterized in that, The content of the intelligent patrol diagnosis report comprises: the fault root cause and the dynamic causal graph representing the transmission path between the fault root cause and the abnormal suspicion token combination.

9. The method of claim 1, wherein the method is a big data based remote intelligent patrol method for a substation, characterized in that, The method further comprises: stacking a corresponding position encoding vector for each device state token in the device state token sequence before inputting the device state token sequence into the cross-modal fusion attention network.

10. A substation remote intelligent patrol system based on big data, applied to the substation remote intelligent patrol method based on big data according to any one of claims 1-9, characterized in that, comprise: a data acquisition module configured to acquire multi-source heterogeneous data of a real-time substation; a data processing module configured to perform standardized preprocessing on the multi-source heterogeneous data; an intelligent analysis module configured to: tokenize and represent the multi-source heterogeneous data after standardized preprocessing in a unified spatiotemporal data stream to obtain a device state token sequence; generate an attention bias matrix based on a preset causal knowledge base and perform fusion analysis on the device state token sequence using a cross-modal fusion attention network, wherein the cross-modal fusion attention network uses the attention bias matrix to bias the attention score when performing attention calculation to output a state representation vector comprising an abnormal suspicion token combination; construct a dynamic causal graph based on the abnormal suspicion token combination in the state representation vector and identify a fault root cause according to the dynamic causal graph to obtain a diagnostic conclusion; a decision support module configured to: update the causal knowledge base according to the diagnostic conclusion; generate an intelligent patrol diagnosis report based on the diagnostic conclusion.

Images