Fault multidimensional diagnosis system and method based on cooperation of substation auxiliary control device and cloud platform

Abstract

The invention discloses a multi-dimensional fault diagnosis system based on cooperation of a transformer substation auxiliary control device and a cloud platform, and the system comprises the steps: collecting multi-modal sensor data through the transformer substation auxiliary control device, carrying out the unified format conversion and normalization processing, and extracting the feature parameters of each modal through a feature extraction algorithm; the model construction module is used for training the historical multi-modal sensor data marked with the fault type label through a cloud platform by utilizing a deep learning algorithm to obtain an initial fault diagnosis model, and optimizing the initial fault diagnosis model according to the preprocessed multi-modal sensor data; a fault diagnosis model of collaborative optimization of the cloud platform and the substation auxiliary control device is obtained; and the fault diagnosis module analyzes each modal characteristic parameter through a fault diagnosis model cooperatively optimized by the cloud platform and the substation auxiliary control device, and judges and identifies whether the substation equipment has a fault and the type and severity of the fault. According to the invention, the efficiency and accuracy of transformer substation fault diagnosis are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system intelligent operation and maintenance, in particular to a fault multi-dimensional diagnosis system and method based on substation auxiliary control device and cloud platform cooperation. BACKGROUND

[0002] As a key node of the power system, the normal operation of the equipment of the substation is crucial to the stability and reliability of the power system. With the increasing complexity of the power system and the number of equipment, the traditional fault diagnosis method has been unable to meet the high reliability requirements of modern substations. At present, when dealing with multi-modal sensor data, there is often a lack of effective data preprocessing and feature extraction methods, resulting in low data quality and affecting the accuracy of the diagnosis results; the model training focuses on a single data mode, lacking deep fusion of numerical data and image data; there are problems of insufficient multi-modal sensor data feature extraction and rigid task allocation mechanism; there are deficiencies in real-time fault diagnosis and cooperation between substation auxiliary control devices and cloud platforms, making it difficult to achieve rapid and accurate fault diagnosis and decision-making. SUMMARY

[0003] The purpose of the present application is to provide a fault multi-dimensional diagnosis system and method based on substation auxiliary control device and cloud platform cooperation, which improves the efficiency and accuracy of substation fault diagnosis.

[0004] To achieve this purpose, the fault multi-dimensional diagnosis system based on substation auxiliary control device and cloud platform cooperation designed by the present application comprises: The substation auxiliary control device is used to collect multi-modal sensor data in the substation, sequentially perform unified format conversion and normalization processing on the multi-modal sensor data to obtain preprocessed multi-modal sensor data, and extract feature parameters corresponding to each mode from the preprocessed multi-modal sensor data through a corresponding feature extraction algorithm; The model construction module is used to train historical multi-modal sensor data with labeled fault type labels through the cloud platform using a deep learning algorithm to obtain an initial fault diagnosis model, and optimize the initial fault diagnosis model according to the preprocessed multi-modal sensor data to obtain a fault diagnosis model optimized by the cloud platform and the substation auxiliary control device in cooperation; The fault diagnosis module is used to analyze and judge the feature parameters corresponding to each mode extracted from the preprocessed multi-modal sensor data through the fault diagnosis model optimized by the cloud platform and the substation auxiliary control device in cooperation to identify whether there is a fault in the substation equipment and the type and severity of the fault.

[0005] Preferably, the substation auxiliary control device is further configured to perform a regulation operation according to the characteristic parameters and upload a regulation decision basis to the cloud platform; wherein the regulation operation comprises linkage of a fan, an air conditioner, a dehumidifier and a water pump to complete regulation of the environment, and the regulation decision basis comprises characteristic parameter data, fault type determination results and corresponding historical regulation operation records.

[0006] Preferably, the multi-modal sensor data in the substation comprises numerical data and image data, wherein the numerical data comprises current data, voltage data, temperature data, partial discharge data and oil chromatographic monitoring data, and the image data comprises visible light images and thermal infrared images.

[0007] Preferably, the specific method for sequentially performing unified format conversion and normalization processing on the multi-modal sensor data to obtain preprocessed multi-modal sensor data comprises: unified format conversion of the numerical data and the image data through a unified conversion rule; normalization processing of the numerical data after unified format conversion by a Z-Score standardization method; and normalization processing of the image data after unified format conversion by a Min-Max scaling method.

[0008] Preferably, the specific method for extracting corresponding characteristic parameters of each mode from the preprocessed multi-modal sensor data by a corresponding feature extraction algorithm comprises: For the numerical data, frequency components of frequency domain signals are extracted by Fourier transform to convert time domain signals of the current data and the voltage data into the frequency domain signals, so as to extract amplitudes, phases and harmonics of the current data and the voltage data as characteristic parameters of the numerical data; temperature variation rates and hot spot temperatures of the temperature data are extracted by time domain statistical calculation of mean value, variance, peak value, peak-to-peak value and waveform factor of the temperature data, so as to extract the temperature variation rates and the hot spot temperatures as characteristic parameters of the numerical data; energy entropies of the partial discharge data are extracted by wavelet packet decomposition of original pulse signals of the partial discharge data by a Db4 wavelet base to convert the original pulse signals into a plurality of frequency band signals, so as to extract energy proportions of the frequency band signals and calculate the energy entropies as characteristic parameters of the numerical data; skewness, kurtosis and phase barycenter of phase distribution of the partial discharge data are extracted by phase angle and discharge amount of each discharge pulse of the partial discharge data extracted by a PRPD spectrum, so as to extract the skewness, the kurtosis and the phase barycenter as characteristic parameters of the numerical data; maximum discharge amount, adjacent pulse variation rate and pulse aggregation degree of the partial discharge data are extracted by pulse clustering analysis, so as to extract the maximum discharge amount, the adjacent pulse variation rate and the pulse aggregation degree as characteristic parameters of the numerical data; concentration ratio of methane to hydrogen, concentration ratio of ethylene to ethane and concentration ratio of acetylene to ethylene in the oil chromatographic monitoring data are calculated by IEC 60599 three-ratio encoding method, so as to extract the concentration ratios as characteristic parameters of the numerical data; gas production rate and gas production acceleration of the oil chromatographic monitoring data are extracted by sliding window derivative analysis, so as to extract the gas production rate and the gas production acceleration as characteristic parameters of the numerical data. For image data, the texture of the visible light image and the thermal infrared image is extracted by local binary pattern as a characteristic parameter of the image data; for the visible light image, the color of the visible light image is extracted by calculating the color distribution histogram as a characteristic parameter of the image data, and for the thermal infrared image, the color of the thermal infrared image is extracted by using the temperature color block distribution after pseudo-color mapping as a characteristic parameter of the image data; the appearance state of the substation equipment in the visible light image and the thermal infrared image is identified by the gradient direction histogram, the edge structural features of the substation equipment are captured, and the shape of the visible light image and the thermal infrared image is extracted by contour analysis as a characteristic parameter of the image data.

[0009] Preferably, the fault diagnosis model comprises a classification model deployed in the substation auxiliary control device and a neural network model deployed in the cloud platform.

[0010] Preferably, the cloud platform uses a deep learning algorithm to train the historical multi-modal data labeled with fault type labels to obtain an initial fault diagnosis model, and optimizes the initial fault diagnosis model according to the preprocessed multi-modal sensor data, and the specific process of obtaining the fault diagnosis model is as follows: The cloud platform uses a long short-term memory network to train the historical multi-modal sensor data labeled with fault type labels to obtain an initial fault diagnosis model, and the initial fault diagnosis model is compressed and optimized to generate a classification model deployed in the substation auxiliary control device, which performs preliminary fault diagnosis on the preprocessed multi-modal sensor data obtained by the substation auxiliary control device; when the classification model deployed in the substation auxiliary control device outputs a fault probability lower than a confidence threshold value according to the preliminary fault diagnosis, the substation auxiliary control device uploads the corresponding preprocessed multi-modal sensor data to the cloud platform, the neural network model deployed in the cloud platform performs fault diagnosis on the uploaded preprocessed multi-modal sensor data, and feeds back the fault diagnosis result and fault decision information to the substation auxiliary control device, and the cloud platform and the substation auxiliary control device cooperatively optimize the fault diagnosis model through the cooperative dynamic optimization mechanism of the cloud platform and the substation auxiliary control device.

[0011] Preferably, the specific process of the fault diagnosis model cooperatively optimized by the cloud platform and the substation auxiliary control device for analyzing and judging the characteristic parameters of each modality extracted from the preprocessed multi-modal sensor data to identify whether the substation equipment has a fault and the type and severity of the fault includes: Inputting the characteristic parameters of each modality extracted from the preprocessed multi-modal sensor data into the fault diagnosis model cooperatively optimized by the cloud platform and the substation auxiliary control device; For the classification model deployed in the auxiliary control device of the substation, a mapping relationship between the feature parameters and the fault types is established by learning the feature parameters, and a classification area of each fault type is divided, when the input feature parameters are not contained in the classification area of any fault type, it is judged that the substation equipment does not exist fault, when the input feature parameters are contained in the classification area of any fault type, it is judged that the substation equipment exists the corresponding fault type, and the severity of the fault type is judged by the probability value of the fault type; For the neural network model deployed in the cloud platform, the mapping relationship between the feature parameters and the fault types is automatically learned through the nonlinear transformation of multiple neurons, the feature parameters are taken as the input of the input layer of the neural network model, the weight matrix and the activation function of the neural network model are calculated layer by layer, and then the fault probability and the activation value are output by the output layer, when the fault probability output by the neural network model does not exceed the set threshold, it is judged that the substation equipment does not exist fault, when the fault probability output by the neural network model exceeds the set threshold, it is judged that the substation equipment exists fault, and the fault type is output, and the severity of the fault type is represented by the size of the activation value output by the output layer of the neural network model.

[0012] A multi-dimensional fault diagnosis method based on cooperation of substation auxiliary control device and cloud platform, comprising the following steps: Collecting multi-modal sensor data in the substation, sequentially performing unified format conversion and normalization processing on the multi-modal sensor data to obtain preprocessed multi-modal sensor data, and extracting feature parameters corresponding to each mode from the preprocessed multi-modal sensor data through a corresponding feature extraction algorithm; The cloud platform trains the historical multi-modal sensor data labeled with fault type labels by using a deep learning algorithm to obtain an initial fault diagnosis model, and optimizes the initial fault diagnosis model according to the preprocessed multi-modal sensor data to obtain a fault diagnosis model optimized by the cloud platform and the substation auxiliary control device in cooperation; The fault diagnosis model optimized by the cloud platform and the substation auxiliary control device in cooperation is used to analyze and judge the feature parameters corresponding to each mode extracted from the preprocessed multi-modal sensor data to identify whether the substation equipment exists fault, and the type and severity of the fault.

[0013] A computer program product comprising a computer program, which, when executed by a processor, implements the steps of the above method.

[0014] The beneficial effects of the present application are: The application can distribute multi-modal sensor data to a cloud platform or a sub-control device of a transformer substation for processing according to the importance and real-time requirements of the multi-modal sensor data, reduce unnecessary data transmission, and improve fault diagnosis efficiency and real-time performance; can organically integrate multi-modal sensor data, mine potential correlations between multi-modal sensor data, more comprehensively reflect the operation state of the transformer substation equipment, and provide rich data support for accurate fault diagnosis; through a deep learning algorithm, the generalization ability and adaptability of the fault diagnosis model of the cloud platform and the transformer substation sub-control device collaborative optimization can be improved, so that different types and degrees of faults can be accurately identified, and the transformer substation equipment has strong robustness to various operation conditions and environmental changes of the transformer substation equipment. Through accurate and timely fault diagnosis, the downtime and maintenance cost of equipment failure can be reduced, equipment damage and maintenance work caused by misjudgment and missed judgment can be avoided, and operation and maintenance costs can be reduced. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a structural schematic diagram of the application; Figure 2 is a flowchart of the application. DETAILED DESCRIPTION

[0016] The application will be further described in detail below in combination with the drawings and specific embodiments: Embodiment 1 A multi-dimensional fault diagnosis system based on the cooperation of a transformer substation sub-control device and a cloud platform, as shown in Figure 1 , it includes: The transformer substation sub-control device (the transformer substation sub-control device is an edge layer, and once an anomaly is detected, the transformer substation sub-control device can immediately trigger a fault adjustment mechanism to cut off the fault source with millisecond-level response speed to ensure stable operation of the power system) is used to collect multi-modal sensor data in the transformer substation, sequentially perform unified format conversion and normalization processing on the multi-modal sensor data to obtain preprocessed multi-modal sensor data, and extract feature parameters corresponding to each mode from the preprocessed multi-modal sensor data through a corresponding feature extraction algorithm. This design can effectively integrate multi-modal data from different sensors through unified format conversion and normalization processing, so that the multi-modal sensor data has a unified format and dimension, improves data quality, and facilitates subsequent processing and analysis. Through feature extraction on multi-modal sensor data, the dimension of the data can be reduced, and the efficiency and accuracy of subsequent analysis can be improved. The power environment monitoring device is a core component of the substation auxiliary control device, mainly responsible for environmental parameter monitoring in the substation, substation equipment state perception, and linkage control execution; the power environment monitoring device provides preprocessed multi-modal sensor data for the classification model deployed in the substation auxiliary control device, the substation auxiliary control device performs localized analysis (such as preliminary fault diagnosis), and the power environment monitoring device receives instructions from the substation auxiliary control device and performs physical operations (such as when the substation equipment is short-circuited or overloaded, the classification model deployed in the substation auxiliary control device diagnoses the fault type determination result, and the substation auxiliary control device starts forced ventilation cooling through the fan control box instruction of the power environment monitoring device; in the fire scenario, the smoke sensor triggers the fire sprinkler system, and the air conditioner is closed to prevent the spread of fire); The model construction module is configured to train the historical multi-modal sensor data labeled with fault type labels using a deep learning algorithm (the deep learning algorithm is a long short-term memory network LSTM, English full name Long Short-Term Memory) through a cloud platform (the cloud platform is a central layer) to obtain an initial fault diagnosis model, and optimize the initial fault diagnosis model according to the preprocessed multi-modal sensor data to obtain a fault diagnosis model optimized by the cloud platform and the substation auxiliary control device. The design can automatically learn complex features and patterns in multi-modal sensor data through the cloud platform using a deep learning algorithm. Compared with traditional rule-based diagnostic methods, the deep learning algorithm can better capture changes and potential fault patterns in multi-modal sensor data. The cloud platform can fully utilize historical multi-modal sensor data to train an initial fault diagnosis model. The cloud platform can integrate multi-modal sensor data from different substations, allowing the initial fault diagnosis model to learn more fault patterns and substation equipment operating states. The initial fault diagnosis model is optimized in combination with preprocessed multi-modal sensor data to improve the adaptability and generalization ability of the fault diagnosis model to multi-modal sensor data. Through the fault diagnosis model optimized by the cloud platform and the substation auxiliary control device, the substation equipment fault can be accurately identified, and potential problems can be predicted based on historical multi-modal sensor data trend analysis, providing a scientific basis for predictive maintenance of substation equipment. The cloud platform can access preprocessed multi-modal sensor data and protection substation trip information data (the protection information substation system is referred to as the protection substation, which generates fault event records when the relay protection device acts, including trip reasons (such as short-circuit current value, overload multiple), action time, and fault phase); The aforementioned cloud platform can construct a large-scale fault root cause model by combining an electrical component aging mechanism model (used to establish the mathematical relationship between electrical parameters (current harmonic distortion rate, dielectric loss angle) and substation equipment aging (insulation degradation, contact oxidation), a chemical decomposition mechanism model (used to analyze the thermal or electrical decomposition products of insulating materials and diagnose latent faults), and an expert experience base model (the expert experience base includes historical fault cases (such as partial discharge spectrum feature library), power system industry standards, and power system operation and maintenance personnel operation logs). Through a neural network model deployed on the cloud platform, the platform outputs the fault type and severity, and combines the large-scale fault root cause model to generate fault root cause reports and health assessments. This enables the platform to assess the health of substation equipment and generate predictive maintenance strategies. The electrical component aging mechanism model is mainly based on the current values ​​obtained from various telemetry points in the substation to calculate the insulation impedance and determine whether the substation equipment is aging or the insulation is deteriorating. The main principle is to collect substation telemetry data (such as leakage current, real-time current, and grounding current), and through time-varying insulation impedance calculation and trend analysis, identify the insulation deterioration process of capacitive equipment (bushings, surge arresters) online to achieve early warning of aging. The chemical decomposition mechanism model is used to analyze the various gas components in transformer oil chromatography. The main principle is to identify the thermal decomposition of insulation materials, corona discharge, arc faults, and other deterioration modes by monitoring the concentration changes and gas generation rates of characteristic gases (H2, CH4, C2H2, etc.) and quantitatively assess the aging status of the transformer's internal insulation. The expert experience base model has two uses: first, after multiple substation auxiliary control devices send alarms, a comprehensive analysis is performed to determine the root cause of the alarm; second, when a substation auxiliary control device alarms and maintenance personnel initially determine that a fault has occurred, the expert model analyzes the fault cause, identifies the source equipment of the fault, and provides maintenance and repair suggestions. The algorithm process for generating fault root cause reports and health assessments by collaboratively optimizing the fault diagnosis model through a cloud platform and substation auxiliary control devices, and combining it with a large-scale fault root cause model, is as follows: Step 1: Feature vectorization of the small model The outputs of the chemical decomposition mechanism model, the electrical component aging mechanism model, and the expert experience base model are respectively processed and transformed into a unified vector form to prepare for subsequent fusion. (1) Output processing of chemical decomposition mechanism model First, obtain the fault probability vector: P_c = [p_overheating(0-1), p_discharge(0-1), p_arc(0-1)], which contains the probability of the equipment experiencing three faults: overheating, discharge, and arcing, with values ​​ranging from 0 to 1; Vector enhancement is performed: the gas production rate Δgas / Δt is added as a fourth-dimensional feature to the fault probability vector to obtain the processed chemical model feature vector V_c; (2) Electrical component aging mechanism model output processing First, the insulation state matrix S_e is obtained, which is in the form of S_e = [[device ID, tan delta, leakage current, predicted life],...], with a dimension of n x 4, where each row represents the insulation state-related information of a device, and tan delta refers to the dielectric loss angle tangent value, which is an important parameter for characterizing the insulation performance of substation equipment (such as bushings, arresters, and capacitive devices); The insulation state matrix FC (S_e) is compressed through a fully connected layer, i.e., FC (S_e) is compressed into V_e, obtaining a 32-dimensional electrical component feature vector V_e, where FC (S_e) represents the insulation state matrix; (3) Expert experience library model output processing For the text root cause description given by experts, such as "CT saturation causes misoperation", the BERT model is used for encoding to obtain a 768-dimensional vector BERT (R) in the form of [0.23, -0.56,...]; The encoded vector is concatenated with the confidence C_r of expert diagnosis to obtain the expert model feature vector V_r = Concat (BERT (R), C_r), where Concat represents the vector concatenation operation. Step 2: Cross-modal evidence fusion (GAT algorithm) The graph attention network (GAT) is used to fuse the feature vectors of different modalities to mine the correlation between features. Constructing the evidence graph: determine the nodes in the graph as the three feature vectors obtained in step 1, i.e., node = {V_c, V_e, V_r}, where V_c represents the feature vector output by the chemical decomposition mechanism model, V_e represents the feature vector output by the electrical component aging mechanism model, and V_r represents the feature vector output by the expert experience library model. Calculate the edge weight: the weight of the edge is determined by the cosine similarity (cosine_similarity) between the node features, which measures the similarity between different feature vectors. Graph attention learning: the node update follows the following formula: h_i' = σ(Σ j α_{ij} W h_j ), where α_{ij} is the attention coefficient, W is the weight matrix, and σ is the activation function, where h_i' represents the updated feature vector of the i-th node after graph attention learning. The attention coefficient a_{ij} is calculated in the manner a_{ij} = softmax ( LeakyReLU ( a^T [Wh_i|| Wh_j] ) ), where a is a learnable vector, [Wh_i || Wh_j] represents concatenating Wh_i and Wh_j, and LeakyReLU ( a^T [Wh_i || Wh_j] ) is an intermediate operation in the process of calculating the attention coefficient a_{ij} ; After graph attention learning, the output fusion evidence H = [h_c', h_e', h_r'] is obtained, with a dimension of 256, where h_c' represents the updated feature of the chemical decomposition mechanism model feature vector V_c after graph attention learning (integrating the chemical features (such as overheating, discharge failure probability, and gas production rate) carried by V_c itself and other modal features (electric component feature V_e, expert experience feature V_r) ); h_e' represents the updated feature of the electrical component aging mechanism model feature vector V_e after graph attention learning (integrating the electrical insulation state information (such as tan delta, leakage current, and predicted life) contained in V_e and the correlation information of chemical features V_c and expert experience features V_r); h_r' represents the updated feature of the expert experience library model feature vector V_r after graph attention learning (integrating the expert text diagnosis semantics (such as fault root cause description) and confidence information contained in V_r and the correlation information of chemical features V_c and electrical component features V_e); Step 3: DeepSeek-45B semantic generation (prompt engineering) Based on the fused evidence, the DeepSeek-45B large model is used to generate the final diagnosis report, and the accuracy and practicality of the generated results are ensured through prompt engineering and model fine-tuning; Input template design: The template includes device profile, multi-model evidence, and generation task; Device profile: such as "{Device ID: disconnecting switch GW5-126; commissioning date: 2020-05; last maintenance: 2023-08}", providing basic information of the device; Multi-model evidence: covering chemical diagnosis, electrical component diagnosis, expert diagnosis, and fusion evidence vector information, for example, "[Chemical Diagnosis] High temperature overheating probability 92% (C2H4 / C2H6=4.2); [Electrical Component Diagnosis] Tan delta of bushing exceeds standard (0.025→0.041, +64%); [Expert Diagnosis] Mechanism jam causes refusal to move (confidence 85%); [Fusion Evidence Vector] H=[0.92, 0.64, 0.85,...]"; The generation task clearly requires the fault root cause model to complete three tasks: locating the faulty equipment and its root cause, assessing the risk level (high / medium / low), and outputting maintenance strategies (emergency / planned / monitoring). Fine-tuning strategy through large-scale root cause model: Domain knowledge injection: The model is further pre-trained on a dataset containing 100,000 substation fault records to incorporate professional knowledge from the substation domain.

[0017] Instruction fine-tuning: Train the model using 5,000 sets of <evidence vector-diagnostic report> samples to enable the model to better generate compliant diagnostic reports based on the input evidence.

[0018] Constraint Decoding: When generating a diagnostic report, parameters are set to constrain the decoding process, such as `generate(report, max_length=500, forbid_words=["Possible", "Recommendation to check"], force_include=["Device ID", "Risk Level"])`. Here, `generate` refers to the core operation instruction function for generating the report; `report` refers to the content carrier of the diagnostic report to be generated; `max_length=500` limits the maximum character length of the generated report to 500 characters (including spaces and punctuation); `forbid_words` refers to the list of prohibited words during the generation process, prohibiting the use of uncertain expressions such as "possible" and "recommendation to check"; and `force_include` refers to the key fields that must be included in the report, forcibly including key information such as "Device ID" and "Risk Level" to ensure the certainty and completeness of the generated report.

[0019] In the aforementioned model construction module, the cloud platform possesses powerful computing and storage capabilities. It receives multimodal sensor data and fault diagnosis results uploaded by substation auxiliary control devices. The cloud platform can also perform comprehensive analysis of multimodal sensor data from multiple substations to discover potential fault trends and patterns. Furthermore, the cloud platform provides data storage, management, and visualization, facilitating maintenance personnel to view the operating status and fault information of substation equipment at any time. The substation auxiliary control device has data preprocessing, feature extraction, and preliminary diagnosis functions. It can perform unified format conversion and normalization processing on multimodal sensor data within the substation, reducing data noise and interference. It can utilize built-in feature extraction algorithms to extract feature parameters from the preprocessed multimodal sensor data, perform preliminary fault diagnosis, and determine whether there are potential fault signs. The substation auxiliary control device can also perform corresponding multimodal sensor data processing and analysis tasks according to instructions from the cloud platform, uploading the processed multimodal sensor data and fault diagnosis results to the cloud platform.

[0020] The fault diagnosis module is used for analyzing and judging the feature parameters corresponding to each mode extracted from the preprocessed multi-modal sensor data by the fault diagnosis model optimized by the cloud platform and the sub-control device of the transformer substation, to identify whether there is a fault in the transformer substation equipment and the type and severity of the fault. The design can fully exert the powerful computing capability of the cloud platform and the real-time data processing capability of the sub-control device of the transformer substation, improve the fault diagnosis efficiency, and more accurately identify the fault type and severity. Through long-term monitoring and analysis of historical multi-modal sensor data, the performance degradation trend of the transformer substation equipment is predicted, and the maintenance strategy is planned in advance, realizing the transformation from passive maintenance to active maintenance, effectively reducing the failure rate of the transformer substation equipment, and improving the operation reliability of the transformer substation.

[0021] In the above technical solution, the sub-control device of the transformer substation is further used to execute adjustment operations according to the feature parameters, and upload adjustment decision basis to the cloud platform; wherein the adjustment operations include linkage of fans, air conditioners, dehumidifiers and water pumps, to complete the adjustment of the environment, and the adjustment decision basis includes feature parameter data, fault type determination results (including the type and severity of the fault) and corresponding adjustment operation records (the adjustment operation record refers to the operation of automatically executing the opening and stopping of the fan, air conditioner, dehumidifier and water pump according to the linkage strategy); the above design can realize dynamic regulation and control of environmental parameters by the sub-control device of the transformer substation based on feature parameters, and improve the fault response speed; The adjustment decision based on the feature parameter data, the fault type determination result and the corresponding historical adjustment operation record is completed by the sub-control device and the cloud platform in cooperation: the sub-control device first matches the historical adjustment operation according to the fault type determination result, determines the operation intensity (such as linkage of fans, dehumidifiers, etc.) combined with the feature parameter data and executes, while uploading the related data to the cloud platform; the cloud platform optimizes the decision model through deep learning algorithm on the historical records and real-time data, when the diagnosis confidence of the sub-control device is lower than the confidence threshold (the confidence threshold is 70%), the cloud platform further analyzes and feeds back the result, forming a dynamic closed-loop decision mechanism, to improve the fault disposal efficiency and accuracy.

[0022] In the technical solution, the multi-modal sensor data in the transformer substation includes numerical data and image data, wherein the numerical data includes current data (the current data includes core monitoring data, and the core monitoring data is ground current (unit: mA)), voltage data (including current data and voltage data of transformers, circuit breakers, disconnectors and cable joints), temperature data (including optical fiber temperature measurement data (temperature gradient dT / dL (unit: ℃ / m), temperature rise rate AT / At (unit: ℃ / min)), temperature data of transformers, circuit breakers, cable joints and capacitors), partial discharge data (the partial discharge data includes discharge quantity (apparent discharge quantity Q (unit: pC)), discharge phase (0°-360°), discharge frequency (times / cycle) and discharge sound size (DB)) and oil chromatographic monitoring data (the oil chromatographic monitoring data includes characteristic gas concentration (H2, CH4, C2H2, C2H4, C2H6, CO, CO2 (unit: ppm)), gas production rate (PPM / DAY) and gas ratio (CH4 / H2, C2H4 / C2H6, C2H2 / C2H4 (Rogers three ratios))), and the image data includes visible light images (including visible light images of insulators, transformers, circuit breakers and arresters) and thermal infrared images (including thermal infrared images of transformers, circuit breakers, cable joints and capacitors). The above design can comprehensively monitor the running state of various devices in the transformer substation by collecting data of multiple modalities including current, voltage, temperature, visible light image and thermal infrared image, can evaluate the health status of the transformer substation equipment from the aspects of electrical performance, mechanical performance and appearance state, and can comprehensively detect the running state of the transformer substation equipment.

[0023] In the technical solution, the specific method for sequentially performing unified format conversion and normalization processing on the multi-modal sensor data to obtain the preprocessed multi-modal sensor data includes: The numerical data and the image data are converted into a unified format by a unified conversion rule (the unified conversion rule refers to converting the numerical data into a structured table, each column corresponds to a sensor parameter, each row corresponds to a time stamp sampling point, and all numerical data is sampled according to a unified time reference; the visible light image in the image data is uniformly converted into an RGB three-channel format, and the thermal infrared image in the image data is uniformly converted into an RGB three-channel format through pseudo-color mapping). The numerical data converted into the unified format is normalized by a Z-Score standardization method, and the calculation formula is: Wherein, x1 represents the original data value of the transformer substation equipment, μ represents the data mean of the transformer substation equipment, a standard deviation of data of the substation equipment, and z represents a normalized data value of the substation equipment; The normalized processing is performed on the image data converted into the unified format by using a Min-Max scaling method, and a calculation formula is as follows: wherein x represents an original pixel value, min represents a minimum pixel value in the image, max represents a maximum pixel value in the image, and x' represents a normalized pixel value; the above design can ensure time consistency and structuring of the numerical data by converting the numerical data into a structured table, can make the image data have a unified format by uniformly converting the visible light image and the thermal infrared image into an RGB three-channel format, and is convenient for subsequent processing; the numerical data and the image data are standardized and normalized by the normalization processing, the complexity of data preprocessing is reduced, and the efficiency of data processing is improved.

[0024] In the technical solution, the specific method for extracting the feature parameters corresponding to each modality from the preprocessed multi-modal sensor data by using a corresponding feature extraction algorithm includes: For the numerical data, the time domain signals of the current data and the voltage data are converted into frequency domain signals by using Fourier transform, the frequency components of the frequency domain signals are extracted, and the amplitude, the phase and the harmonics of the current data and the voltage data are extracted as the feature parameters of the numerical data (which can assist the substation auxiliary control device in identifying arc faults); the mean value, the variance, the peak value, the peak-to-peak value and the waveform factor of the temperature data are calculated by using time domain statistical calculation, and the temperature variation rate and the hotspot temperature of the temperature data are extracted as the feature parameters of the numerical data; the original pulse signal of the partial discharge data is converted into a plurality of frequency band signals (32 frequency band signals) by using wavelet packet decomposition based on a Db4 wavelet base, the energy proportion of each frequency band is extracted, and the energy entropy of the partial discharge data is calculated and extracted as the feature parameter of the numerical data; the phase angle and the discharge amount of each discharge pulse of the partial discharge data are extracted by using a PRPD spectrum, and the skewness, the kurtosis and the phase center of the phase distribution of the partial discharge data are extracted as the feature parameters of the numerical data; the maximum discharge amount, the adjacent pulse variation rate and the pulse aggregation degree of the partial discharge data are extracted as the feature parameters of the numerical data by using pulse clustering analysis; the concentration ratio of methane to hydrogen, the concentration ratio of ethylene to ethane, and the concentration ratio of acetylene to ethylene in the oil chromatographic monitoring data are calculated by using IEC 60599 three-ratio encoding method as the feature parameters of the numerical data, and the gas production rate and the gas production acceleration of the oil chromatographic monitoring data are extracted as the feature parameters of the numerical data by using sliding window derivative analysis. For image data, the texture of the visible light image and the thermal infrared image is extracted by local binary pattern as a feature parameter of the image data; for the visible light image, the color of the visible light image is extracted by calculating the color distribution histogram as a feature parameter of the image data, and for the thermal infrared image, the color of the thermal infrared image is extracted by using the temperature color block distribution after pseudo-color mapping as a feature parameter of the image data; the appearance state of the substation equipment in the visible light image and the thermal infrared image is identified by the gradient direction histogram, the edge structure feature of the substation equipment is captured, and the shape of the visible light image and the thermal infrared image is extracted by combining contour analysis as a feature parameter of the image data (which can support the substation auxiliary control device to adjust the overheated substation equipment); the above design can comprehensively evaluate the operation state of the substation equipment from multiple angles by extracting various feature parameters from numerical data and image data, reduce the misjudgment and missed judgment that may be caused by single modal data, reduce the risk of overfitting, and improve the stability and reliability of the fault diagnosis model of the cloud platform and the substation auxiliary control device in actual application.

[0025] In the above technical solution, the fault diagnosis model includes a classification model deployed in the substation auxiliary control device and a neural network model deployed in the cloud platform; the above design can perform preliminary fault diagnosis on the real-time collected preprocessed multi-modal sensor data through the classification model deployed in the substation auxiliary control device, discover potential faults in the first time, take timely measures, and reduce the impact of faults on the operation of the substation equipment; through the cooperative diagnosis mechanism of the classification model deployed in the substation auxiliary control device and the neural network model deployed in the cloud platform, the computing burden of the cloud platform can be effectively reduced, and the overall operation efficiency can be improved.

[0026] In the above technical solution, the cloud platform trains the historical multi-modal data labeled with fault type labels by using a deep learning algorithm to obtain an initial fault diagnosis model, and optimizes the initial fault diagnosis model according to the preprocessed multi-modal sensor data, and the specific process of obtaining the fault diagnosis model is as follows: The initial fault diagnosis model is obtained by training historical multi-modal sensor data labeled with fault type labels through a cloud platform using a long short-term memory network (LSTM), the classification model deployed in the substation auxiliary control device is generated by compressing and optimizing the initial fault diagnosis model, and the preprocessed multi-modal sensor data obtained by the substation auxiliary control device is preliminarily diagnosed through the classification model; when the classification model deployed in the substation auxiliary control device outputs a fault probability lower than a confidence threshold (the confidence threshold is 70%) according to the preliminary fault diagnosis, the substation auxiliary control device uploads the corresponding preprocessed multi-modal sensor data to the cloud platform (the cloud platform integrates the substation auxiliary control device data and the substation auxiliary control device trip signal through a conflict detection algorithm (such as D-S evidence theory) to output a final fault decision), the preprocessed multi-modal sensor data uploaded is diagnosed through the neural network model deployed in the cloud platform, and the fault diagnosis result and the fault decision information are fed back to the substation auxiliary control device, and a fault diagnosis model optimized by the cloud platform and the substation auxiliary control device is constructed through the dynamic optimization mechanism of the cloud platform and the substation auxiliary control device; the above design optimizes the initial fault diagnosis model through the preprocessed multi-modal sensor data, further improves the diagnosis performance of the fault diagnosis model, better adapts to the data characteristics in the actual running environment, and reduces misjudgment and omission; through the collaborative optimization mechanism of the cloud platform and the substation auxiliary control device, the fault diagnosis model can dynamically adjust parameters according to real-time data, ensuring high diagnosis efficiency and accuracy under different operating conditions. The conflict detection algorithm is mainly used to configure the confidence for each data source / model, such as the real trip confidence of SOE time sequence verification m1(A)=0.9, the interference false alarm confidence of SOE time sequence verification m1(B)=0.1, the real trip confidence of recording wave current mutation variable m2(A)=0.7, and the interference false alarm confidence of recording wave current mutation variable m2(B)=0.3 (interference false alarm means non-fault).

[0027] Conflict factor K = m1(A)×m2(B) + m1(B)×m2(A) = 0.9×0.3 + 0.1×0.7 = 0.34 Real trip confidence m(A) = (0.9×0.7) / (1-0.34) = 0.63 / 0.66 ≈ 0.954 Interference false alarm confidence m(B) = (0.1×0.3) / 0.66 ≈ 0.045 Decision output: if m(real trip)>0.95, trigger the early warning strategy; if 0.6

[0028] In the technical solution, the specific process of analyzing and judging the feature parameters corresponding to each mode extracted from the preprocessed multi-modal sensor data by the fault diagnosis model optimized by the cloud platform and the sub-control device of the transformer substation to identify whether the transformer substation equipment has a fault and the type and severity of the fault includes: inputting the feature parameters corresponding to each mode extracted from the preprocessed multi-modal sensor data into the fault diagnosis model optimized by the cloud platform and the sub-control device of the transformer substation; for the classification model deployed in the sub-control device of the transformer substation, a mapping relationship between the feature parameters and the fault types is established by learning the feature parameters, and a classification region (a hypergeometric body (such as a multi-dimensional sphere, a cube, or a complex surface) in a feature space) of each fault type is divided, which is defined by the feature parameter distribution range of the fault type. When the input feature parameters are not included in any classification region of the fault type, it is determined that the transformer substation equipment does not have a fault; when the input feature parameters are included in any classification region of the fault type, it is determined that the transformer substation equipment has a corresponding fault type, and the severity of the fault type is determined by the probability value of the fault type, and the formula is as follows: wherein, the prediction probability represents the fault category probability output by the classification model, the confidence threshold represents the minimum probability of determining the existence of a fault, the maximum probability represents the maximum output value supported by the classification model, and the fault type reference severity represents the maximum severity value of the fault type; when the severity is greater than or equal to A% (A% is 0%) and less than B% (B% is 30%), the warning level is attention; when the severity is greater than or equal to B% and less than C% (C% is 60%), the warning level is warning; when the severity is greater than or equal to C% and less than D% (D% is 80%), the warning level is severe; and when the severity is greater than or equal to D% and less than or equal to E% (E% is 100%), the warning level is urgent; for the neural network model deployed in the cloud platform, the mapping relationship between the feature parameters and the fault types is automatically learned by the non-linear transformation of multiple neurons, the feature parameters are taken as the input of the input layer of the neural network model, the weight matrix and the activation function of the neural network model are calculated layer by layer, and the fault probability and the activation value are output by the output layer. When the fault probability output by the neural network model does not exceed the set threshold (the threshold can be set to 70%), it is determined that the transformer substation equipment does not have a fault; When the fault probability output by the neural network model exceeds the set threshold, it is judged that the substation equipment has a fault, and the type of the fault is output; the severity of the type of the fault is represented by the size of the activation value output by the output layer of the neural network model; when the activation value is greater than or equal to A% and less than B%, the warning level is attention; when the activation value is greater than or equal to B% and less than C%, the warning level is warning; when the activation value is greater than or equal to C% and less than D%, the warning level is serious; when the activation value is greater than or equal to D% and less than or equal to E%, the warning level is urgent; When the substation auxiliary control device or the cloud platform detects a fault, it timely sends a warning message to inform the operation and maintenance personnel to handle it, and provides the type and severity of the fault and the recommended treatment measures; the collected multi-modal sensor data, the extracted feature parameters corresponding to each mode in the pre-processed multi-modal sensor data, and the fault diagnosis results are stored and managed, a corresponding substation equipment database is established, and the fault related data is presented to the operation and maintenance personnel in a visual manner through the substation equipment database, the running state of the substation equipment (electrical parameters: current, voltage, power factor, harmonic content, etc. (displayed through line chart / instrument panel); mechanical parameters: vibration frequency, displacement, temperature (displayed through heat map / trend chart); environmental parameters: humidity, temperature, gas concentration (displayed through environmental monitoring instrument panel)) is displayed in the form of charts, graphs or maps, the location of the fault equipment (such as circuit breaker, transformer, bus) is marked on the topological map or GIS map; the number of faults is counted according to the substation area (such as high-voltage room, low-voltage room, outdoor equipment area), and the fault distribution is displayed through the heat map to show the fault density, so that the operation and maintenance personnel can intuitively understand the overall operation of the substation, timely find potential problems and make decisions; The classification model deployed in the substation auxiliary control device in the above design can more accurately identify the fault type and severity by learning the feature parameters to establish the mapping relationship between the feature parameters and the fault type; by setting different warning levels (attention, warning, serious, urgent), the power system can timely send corresponding warnings according to the severity of the fault to help the operation and maintenance personnel take appropriate measures, and improve the reliability and safety of the power system.

[0029] Embodiment 2 A multi-dimensional fault diagnosis method based on the cooperation of substation auxiliary control devices and cloud platforms, as shown in Figure 2As shown, the multi-modal sensor data is collected, converted and normalized by unified format, and the modal feature parameters are extracted by feature extraction algorithm; the historical multi-modal sensor data with fault type label is trained by cloud platform using deep learning algorithm, an initial fault diagnosis model is obtained, and the initial fault diagnosis model is optimized according to the preprocessed multi-modal sensor data, and a fault diagnosis model optimized by cloud platform and substation auxiliary control device is obtained; the modal feature parameters are analyzed and judged by the fault diagnosis model optimized by cloud platform and substation auxiliary control device, and whether the substation equipment exists fault and the type and severity of the fault are identified.

[0030] The specific method of multi-dimensional fault diagnosis includes the following steps: Collecting multi-modal sensor data in the substation, sequentially converting and normalizing the multi-modal sensor data to obtain preprocessed multi-modal sensor data, and extracting corresponding feature parameters of each mode from the preprocessed multi-modal sensor data by corresponding feature extraction algorithm; The historical multi-modal sensor data with fault type label is trained by cloud platform using deep learning algorithm, an initial fault diagnosis model is obtained, and the initial fault diagnosis model is optimized according to the preprocessed multi-modal sensor data, and a fault diagnosis model optimized by cloud platform and substation auxiliary control device is obtained; The modal feature parameters extracted from the preprocessed multi-modal sensor data are analyzed and judged by the fault diagnosis model optimized by cloud platform and substation auxiliary control device, and whether the substation equipment exists fault and the type and severity of the fault are identified.

[0031] Example 3 Xiaomi substation application Implementation process: 1. System deployment (1) Deploy the minimum perception terminal in the Xiaomi substation site: including 35 1080P infrared cameras (covering the main equipment area / switch room), 10 temperature and humidity sensors (installed on the key nodes of the transformer / switch cabinet), and access to the five-proof system locking logic and signal tripping substation through Modbus protocol; (2) Substation auxiliary control device: using embedded industrial computer (Intel Celeron J1900 / 4GB RAM), built-in lightweight diagnosis model (50MB CNN image recognition model + fault logic rule base).

[0032] 2. Substation auxiliary control device and cloud platform collaborative fault diagnosis process (1) When the substation triggers a trip event (such as reporting a "10kV outgoing line switch trip" event), immediately trigger the substation auxiliary control device to start collaborative diagnosis; (2) The substation auxiliary control device performs multi-source data synchronous collection: retrieves the five-prevention system locking state of the corresponding interval (checks the locking release timestamp); calls the camera to take an infrared image of the switch cabinet (such as resolution 1920x1080, temperature measurement range -20℃~550℃); reads the SOE event sequence of the substation (action timestamp error ≤2ms); (3) The substation auxiliary control device performs preliminary fault diagnosis: image model identifies switch mechanism mechanical position (outputs a position reliability of 92%); rule base matches trip event and five-prevention operation timing (if the locking release is later than the trip command by more than 0.5s, it is determined to be "mechanism jam"); (4) The substation auxiliary control device and the cloud platform cooperate in fault diagnosis and decision-making: (confidence 92%> confidence threshold 70%) The substation auxiliary control device directly triggers an early warning (pushes the "mechanism jam fault" to the operation and maintenance terminal); synchronously uploads the trip timing data to the cloud platform, which checks it through D-S evidence theory; the cloud platform automatically generates an intelligent diagnosis report containing positioning information and disposal suggestions, realizes a minute-level closed-loop disposal of "trip-recheck-diagnosis", and includes the maintenance suggestion: "check the switch mechanism connecting rod lubrication".

[0033] Finally, the substation auxiliary control device and the cloud platform cooperate in fault diagnosis within 10 seconds of the trip event, greatly reducing the misdiagnosis rate (which can be reduced to less than 5%).

[0034] Example 4 Application in long-term substation Implementation process: 1. Hierarchical deployment architecture (1) Edge layer 1 (dynamic environment / safety and security): auxiliary control host (ARM Cortex-A72) accesses fire sprinkler status, access control alarm signals, and runs temperature threshold rule base (threshold: transformer top oil temperature 85℃); (2) Edge layer 2 (electrical quantity diagnosis): special reasoning equipment (NVIDIA Jetson TX2) directly connects fault recorder (sampling rate 10kHz) and substation, and has built-in short circuit positioning algorithm (based on symmetric component method to calculate zero sequence current abruptness).

[0035] 2. Substation auxiliary control device and cloud platform collaborative fault diagnosis process (1) The edge layer 2 detects that the "220kV main transformer bushing B phase leakage current exceeds the threshold value (exceeds 200% for 3 minutes)", triggering the substation auxiliary control device and the cloud platform to cooperate: uploading multi-modal data: leakage current waveform, bushing infrared image, oil chromatogram online monitoring data (H2=120ppm, C2H2=15ppm); (2) The cloud platform fault root cause large model performs diagnosis: A. Electrochemical decomposition mechanism model: input oil chromatogram data, calculate gas production rate (CH4 / H2=1.2>1.0) according to IEC60599 standard, output "low temperature overheating" conclusion (confidence 88%); B. Electric component aging model: analyze the leakage current harmonic distortion rate (THD=8.5%), combined with the time-varying insulation impedance curve (decreased by 40% in 6 months), output "bushing insulation deterioration"; C. Expert experience rule base: SQL query historical fault library, match similar cases (case number F2023-047: bushing glue crack causes partial discharge); (3) Conflict detection and decision: D-S evidence theory integration: m1(overheating)=0.88, m2(deterioration)=0.75, K=0.88×0.25+0.12×0.75=0.31; final confidence m(fault)=0.88×0.75 / (1-0.31)=0.954; (4) Output predictive maintenance strategy: generate device health assessment report: bushing remaining life assessment is 6 months; push maintenance strategy: "carry out bushing dielectric loss test within 72 hours, replace within 3 months".

[0036] Finally, early warning of insulation deterioration is realized (can be advanced by 3 months), and the diagnosis accuracy is greatly improved (can be improved to more than 95%).

[0037] Embodiment 5 A computer program product comprising a computer program which, when executed by a processor, implements the steps of the method described in embodiment 2.

[0038] The contents not described in detail in the specification belong to the prior art known to those skilled in the art.

Claims

1. A multi-dimensional fault diagnosis system based on the cooperation of a substation auxiliary control device and a cloud platform, characterized in that, It comprises: The substation auxiliary control device is used for collecting multi-modal sensor data in the substation, sequentially performing unified format conversion and normalization processing on the multi-modal sensor data to obtain preprocessed multi-modal sensor data, and extracting corresponding feature parameters of each mode from the preprocessed multi-modal sensor data through a corresponding feature extraction algorithm; The model construction module is used for training historical multi-modal sensor data labeled with fault type labels through a deep learning algorithm on a cloud platform to obtain an initial fault diagnosis model, and optimizing the initial fault diagnosis model according to the preprocessed multi-modal sensor data to obtain a fault diagnosis model optimized by the cloud platform and the substation auxiliary control device in cooperation; The fault diagnosis module is used for analyzing and judging the corresponding feature parameters of each mode extracted from the preprocessed multi-modal sensor data through the fault diagnosis model optimized by the cloud platform and the substation auxiliary control device in cooperation to identify whether the substation equipment has a fault and the type and severity of the fault. 2.The substation auxiliary control device and cloud platform collaborative fault multi-dimensional diagnosis system according to claim 1, characterized in that: The substation auxiliary control device is also used for executing a regulation operation according to the feature parameters and uploading a regulation decision basis to the cloud platform; wherein the regulation operation includes a linkage of a fan, an air conditioner, a dehumidifier and a water pump to complete the regulation of the environment, and the regulation decision basis includes feature parameter data, fault type determination results and corresponding historical regulation operation records. 3.The substation auxiliary control device and cloud platform collaborative fault multi-dimensional diagnosis system according to claim 1, characterized in that: The multi-modal sensor data in the substation includes numerical data and image data, wherein the numerical data includes current data, voltage data, temperature data, partial discharge data and oil chromatographic monitoring data, and the image data includes visible light images and thermal infrared images.

4. The substation auxiliary control device and cloud platform collaborative fault multi-dimensional diagnosis system according to claim 3, characterized in that: The specific method of sequentially performing unified format conversion and normalization processing on the multi-modal sensor data to obtain preprocessed multi-modal sensor data includes: The numerical data and the image data are converted into unified format through a unified conversion rule to obtain unified format converted numerical data and image data; the unified format converted numerical data is normalized by a Z-Score standardization method; and the unified format converted image data is normalized by a Min-Max scaling method.

5. The substation auxiliary control device and cloud platform collaborative fault multi-dimensional diagnosis system based on claim 1, characterized in that: The specific method of extracting corresponding feature parameters of each mode from the preprocessed multi-modal sensor data through a corresponding feature extraction algorithm includes: For numerical data, the time domain signals of current data and voltage data are converted into frequency domain signals by Fourier transform, the frequency components of the frequency domain signals are extracted, and the amplitude, phase and harmonics of the current data and voltage data are extracted as characteristic parameters of numerical data; the mean, variance, peak value, peak-to-peak value and waveform factor of temperature data are calculated by time domain statistical calculation, and the temperature change rate and hot spot temperature of the temperature data are extracted as characteristic parameters of numerical data; the original pulse signal of partial discharge data is converted into a plurality of frequency band signals by wavelet packet decomposition based on Db4 wavelet basis, the energy proportion of each frequency band is extracted, and the energy entropy of the partial discharge data is calculated and extracted as a characteristic parameter of numerical data, the phase angle and discharge amount of each discharge pulse of the partial discharge data are extracted by PRPD spectrum, and the skewness, kurtosis and phase center of the phase distribution of the partial discharge data are extracted as characteristic parameters of numerical data, the maximum discharge amount, adjacent pulse change rate and pulse aggregation degree of the partial discharge data are extracted by pulse clustering analysis as characteristic parameters of numerical data; the concentration ratio of methane to hydrogen, the concentration ratio of ethylene to ethane, and the concentration ratio of acetylene to ethylene in the oil chromatographic monitoring data are calculated by IEC 60599 three-ratio encoding method as characteristic parameters of numerical data, and the gas production rate and gas production acceleration of the oil chromatographic monitoring data are extracted by sliding window derivative analysis as characteristic parameters of numerical data; For image data, the texture of visible light images and thermal infrared images is extracted by local binary pattern as a characteristic parameter of image data; for visible light images, the color of visible light images is extracted by calculating the color distribution histogram as a characteristic parameter of image data, and for thermal infrared images, the color of thermal infrared images is extracted by using the temperature color block distribution after pseudo-color mapping as a characteristic parameter of image data; the appearance state of substation equipment in visible light images and thermal infrared images is identified by gradient direction histogram, and the edge structural features of substation equipment are captured, and the shape of visible light images and thermal infrared images is extracted by combining contour analysis as a characteristic parameter of image data. 6.The substation auxiliary control device and cloud platform collaborative fault multi-dimensional diagnosis system according to claim 1, characterized in that: The fault diagnosis model comprises a classification model deployed in a substation auxiliary control device and a neural network model deployed in a cloud platform.

7. The substation auxiliary control device and cloud platform collaborative fault multi-dimensional diagnosis system according to claim 6, characterized in that: The cloud platform trains historical multi-modal data labeled with fault type labels using a deep learning algorithm to obtain an initial fault diagnosis model, and optimizes the initial fault diagnosis model according to preprocessed multi-modal sensor data, and the specific process of obtaining the fault diagnosis model is as follows: The initial fault diagnosis model is obtained by training the historical multi-modal sensor data labeled with fault type labels through a cloud platform using a long short-term memory network, the classification model deployed in the sub-control device of the transformer substation is generated by compressing and optimizing the initial fault diagnosis model, and the classification model is used to preliminarily diagnose the fault of the preprocessed multi-modal sensor data obtained by the sub-control device of the transformer substation; when the classification model deployed in the sub-control device of the transformer substation outputs a fault probability lower than a confidence threshold value according to the preliminary fault diagnosis, the sub-control device of the transformer substation uploads the corresponding preprocessed multi-modal sensor data to the cloud platform, the neural network model deployed in the cloud platform is used to diagnose the fault of the uploaded preprocessed multi-modal sensor data, and the fault diagnosis result and fault decision information are fed back to the sub-control device of the transformer substation, and the cloud platform and the sub-control device of the transformer substation cooperatively construct a fault diagnosis model optimized by the cloud platform and the sub-control device of the transformer substation through a dynamic optimization mechanism. 8.The substation auxiliary control device and cloud platform collaborative fault multi-dimensional diagnosis system according to claim 7, characterized in that: The specific process of identifying whether the transformer substation equipment has a fault, the type of the fault and the severity of the fault by analyzing and judging the feature parameters corresponding to each modality extracted from the preprocessed multi-modal sensor data through the fault diagnosis model optimized by the cloud platform and the sub-control device of the transformer substation includes: inputting the feature parameters corresponding to each modality extracted from the preprocessed multi-modal sensor data into the fault diagnosis model optimized by the cloud platform and the sub-control device of the transformer substation; for the classification model deployed in the sub-control device of the transformer substation, a mapping relationship between the feature parameters and the fault types is established by learning the feature parameters, and a classification region of each fault type is divided, when the input feature parameters are not included in any classification region of the fault type, it is judged that the transformer substation equipment does not have a fault; when the input feature parameters are included in any classification region of the fault type, it is judged that the transformer substation equipment has a corresponding fault type, and the severity of the fault type is judged by the probability value of the fault type; for the neural network model deployed in the cloud platform, the mapping relationship between the feature parameters and the fault types is automatically learned through the non-linear transformation of multiple neurons, the feature parameters are taken as the input of the input layer of the neural network model, the weight matrix and the activation function of the neural network model are calculated layer by layer, and the fault probability and the activation value are output by the output layer, when the fault probability output by the neural network model does not exceed the set threshold value, it is judged that the transformer substation equipment does not have a fault; when the fault probability output by the neural network model exceeds the set threshold value, it is judged that the transformer substation equipment has a fault, and the fault type is output; the severity of the fault type is represented by the size of the activation value output by the output layer of the neural network model.

9. A multi-dimensional fault diagnosis method based on the cooperation of a substation auxiliary control device and a cloud platform, characterized in that, It includes the following steps: collecting multi-modal sensor data in the transformer substation, sequentially performing unified format conversion and normalization processing on the multi-modal sensor data to obtain preprocessed multi-modal sensor data, and extracting feature parameters corresponding to each modality from the preprocessed multi-modal sensor data through a corresponding feature extraction algorithm; The cloud platform uses a deep learning algorithm to train historical multi-modal sensor data labeled with fault type labels to obtain an initial fault diagnosis model, and optimizes the initial fault diagnosis model according to preprocessed multi-modal sensor data to obtain a fault diagnosis model optimized by the cloud platform and the sub-control device of the transformer substation. The fault diagnosis model optimized by the cloud platform and the sub-control device of the transformer substation analyzes and judges the feature parameters corresponding to each mode extracted from the preprocessed multi-modal sensor data to identify whether the transformer substation equipment has a fault and the type and severity of the fault.

10. A computer program product comprising a computer program, characterized in that, The computer program, when executed by a processor, implements the steps of the method described in claim 9.

Images