A Substation Intelligent Auxiliary Monitoring Platform Based on Big Data Analysis

The substation intelligent auxiliary monitoring platform, which utilizes big data analytics and combines filter networks and TKAN dual-path feature extraction architecture, achieves deep fusion and intelligent diagnosis of multi-source data from substations. This solves the problems of data silos and insufficient intelligence in existing systems, improves the accuracy of fault root cause location and prediction confidence, and has self-evolution capabilities.

CN123121997APending Publication Date: 2026-11-17CHANGZHOU YONGBO MASCH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611288361.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-24
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing substation auxiliary monitoring systems lack multi-source data fusion capabilities, have limited intelligent analysis capabilities, cannot deeply mine massive amounts of multi-source data, rely on manual experience for fault diagnosis, and have to improve system reliability and real-time performance. Furthermore, they lack self-learning and evolution capabilities.

Method used

A substation intelligent auxiliary monitoring platform based on big data analysis is adopted, which combines filter network, TKAN dual-path feature extraction architecture, probability prediction head and causal inference engine to achieve deep fusion and intelligent diagnosis of multi-source heterogeneous data. Through real-time edge alarm, short-term fluctuation and long-term trend feature extraction, mean and variance quantification and causal knowledge base matching, the accuracy of fault root cause location and prediction confidence are improved.

Benefits of technology

It achieves deep fusion of multi-source heterogeneous data, improves the accuracy of fault root cause location and prediction confidence, reduces cloud computing load and communication latency, has self-evolution capability, supports collaborative decision-making for high-confidence automatic diagnosis and low-confidence safety fallback, and ensures the interpretability of diagnostic conclusions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123121997A_ABST
    Figure CN123121997A_ABST
Patent Text Reader

Abstract

This invention relates to the field of intelligent monitoring of substations, and discloses an intelligent auxiliary monitoring platform for substations based on big data analysis, comprising: a data acquisition module for acquiring multi-source data and outputting a time series matrix; an edge detection module for filtering and comparing with a threshold to generate alarm pushes; a distributed storage module for storing historical data and a causal knowledge base; a TKAN dual-path feature extraction module for outputting short-term features for short-term paths and long-term trend features for long-term gated step paths, concatenating them into a comprehensive feature; a TKAN probability prediction module for mapping the comprehensive feature to the mean and variance, with the variance activated by Softplus; a causal reasoning module for outputting fault types and handling suggestions; and an instruction output module for displaying fault types and handling suggestions, masking the mean and generating physical inspection instructions when the variance is greater than a second threshold. This invention achieves highly reliable automatic diagnosis and low-reliability safety inspection.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for substations, and in particular to an intelligent auxiliary monitoring platform for substations based on big data analysis. Background Technology

[0002] With the continuous expansion of the power grid and the constant improvement of its intelligence level, substations, as core nodes of the power system, are facing a prominent contradiction between the surge in the number of equipment and the insufficient allocation of maintenance personnel. The traditional model relying on manual inspections and on-site maintenance is no longer suitable for the needs of lean substation management. Therefore, the industry has begun to explore the introduction of intelligent technologies such as the Internet of Things, big data, and artificial intelligence into the field of substation monitoring, gradually promoting the transformation of the operation and maintenance model towards "intelligent inspection and remote operation and maintenance."

[0003] However, existing substation auxiliary monitoring systems still have the following shortcomings. First, most monitoring subsystems operate independently, lacking effective data fusion and collaborative analysis capabilities among video surveillance, environmental monitoring, online monitoring, and security and fire protection systems. Furthermore, the configuration tools and operating methods of different systems are inconsistent, resulting in numerous data silos. Second, the intelligent analysis level of existing systems is limited, generally remaining at the level of data acquisition and threshold alarms. They lack the ability to deeply mine and intelligently diagnose massive amounts of multi-source data, and fault judgment still relies heavily on human experience. The reliability and real-time performance of the systems also need improvement. In addition, most existing systems lack the ability to self-learn and evolve, failing to continuously improve analytical accuracy as operational data accumulates.

[0004] Therefore, there is an urgent need for a substation intelligent auxiliary monitoring platform that can integrate multi-source heterogeneous data, achieve deep intelligent analysis, and has self-evolution capabilities. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent auxiliary monitoring platform for substations based on big data analysis. This invention fully integrates filter networks, TKAN dual-path feature extraction architecture, probabilistic prediction head, and causal inference engine, and describes in detail the entire process of substation auxiliary monitoring from multi-source data acquisition, real-time edge alarm, dual-path feature extraction of short-term fluctuations and long-term trends, uncertainty quantification of mean and variance dual outputs, causal knowledge base matching diagnosis, to variance-driven physical inspection triggering. It has the advantages of strong deep fusion capability of multi-source heterogeneous data, high accuracy of fault root cause location, quantifiable evaluation of prediction confidence, and a sound decision-making safety fallback mechanism under complex operating conditions.

[0006] According to an embodiment of the present invention, a substation intelligent auxiliary monitoring platform based on big data analysis includes: The data acquisition module is used to collect data from various sensors and SCADA systems, forming raw data packets. It then interpolates, filters, aligns, and converts the raw data packets to different protocols, outputting a multivariate time series matrix. The edge detection module runs a filter network to filter the time series matrix, compares it with a threshold, and generates real-time alarms for local push notifications. The distributed storage module stores historical time series data and a causal knowledge base. The TKAN dual-path feature extraction module is used to receive a time series matrix, execute a short-term RKAN path to process local subsequences with a preset window length to output short-term features, execute a long-term gated step path to pass hidden states at preset time intervals to output long-term trend features, and splice the short-term features and long-term trend features into a comprehensive feature. The TKAN probability prediction module maps comprehensive features to the mean through a mean fully connected layer and to the variance through a variance fully connected layer, with the variance fully connected layer connected to the Softplus activation function. The causal inference module matches the mean with the causal knowledge base to output the fault type and handling suggestions when the variance is less than the high confidence threshold. The instruction output module displays the fault type and handling suggestions, and when the variance is greater than the low confidence threshold, it masks the mean and generates a physical inspection instruction to send to the inspection equipment.

[0007] Optionally, the output multivariate time series matrix of the data acquisition module specifically includes the following steps: acquiring voltage waveforms, current waveforms, temperature values, vibration spectra, partial discharge signals, and telemetry and tele-signaling status values ​​output by the SCADA system from multiple types of sensors; performing protocol parsing on each acquired raw data stream to convert non-standard protocol data into standard format data; performing missing value detection on each converted standard format data stream, locating the timestamp and data dimension of the missing value, and interpolating the missing value using a preset padding value; performing a preset denoising algorithm on each interpolated data stream to filter out noise components above a preset frequency threshold; aligning each data stream to the same time coordinate system according to the sampling timestamp, using a preset reference time axis; arranging the aligned data streams in chronological order into a two-dimensional matrix, where the matrix row vectors correspond to the values ​​of all data dimensions at the same timestamp, and the matrix column vectors correspond to the value sequence of a single data dimension on the entire time axis; the two-dimensional matrix is ​​the multivariate time series matrix.

[0008] Optionally, the operation of the edge detection module's filter network to filter the time series matrix and compare it with a threshold to generate a real-time alarm for local push includes the following steps: A multivariate time series matrix is ​​obtained and input into a general shaping filter layer of a filter network. The general shaping filter layer applies a first learnable frequency filter to each dimension of the multivariate time series matrix to extract the fundamental frequency component of each dimension. The fundamental frequency component output from the general shaping filter layer is input into a context shaping filter layer of the filter network. The context shaping filter layer applies a second learnable frequency filter to the fundamental frequency component to capture the residual dependency between the fundamental frequency component and the original input. The filtered feature output from the context shaping filter layer is compared with a preset alarm threshold. When the filtered feature is greater than the preset alarm threshold, a real-time alarm signal is generated and pushed locally to the substation's local monitoring station.

[0009] Optionally, the distributed storage module's storage of historical time-series data and causal knowledge base specifically includes the following steps: Define a columnar storage format, dividing the multivariate time series matrix into multiple column families according to data dimensions. Each column family corresponds to the full time-axis sequence value of a single data dimension. Establish a storage structure for each column family with device number as the primary index and timestamp as the secondary index. Store the causal knowledge base as a set of key-value pairs, where the key represents the fault type identifier, and the value represents the set of multivariate time-series symptom patterns corresponding to this fault type, as well as the corresponding handling suggestion text. Each time a new multivariate time series matrix is ​​received, append it to the corresponding device number and timestamp storage location in the columnar storage structure. After each causal inference module outputs a matching fault type and handling suggestion, write the fault type and the corresponding set of multivariate time-series symptom patterns as new records into the key-value pair set of the causal knowledge base. After each full write operation, update the metadata information of the distributed storage cluster, which includes the data volume of each column family, the latest timestamp, and the oldest timestamp.

[0010] Optionally, the short-term RKAN path processing of the TKAN dual-path feature extraction module outputs short-term features from local subsequences of a preset window length, specifically including the following steps: Receiving a multivariate time series matrix, and extracting a local subsequence of a preset window length with the current time as the endpoint. The local subsequence contains multivariate observations from a preset number of consecutive time steps; Inputting the local subsequence into a first-layer RKAN unit, which includes a B-spline basis function layer and a recursive connection layer. The B-spline basis function layer maps each dimension of the input vector to B-spline basis function expansion coefficients, and the recursive connection layer nonlinearly fuses the B-spline expansion coefficients of the current time step with the hidden state of the previous time step, outputting a first-layer hidden state sequence; Inputting the first-layer hidden state sequence into a second-layer RKAN unit, which has the same structure as the first-layer RKAN unit. The second-layer RKAN unit performs the same B-spline basis function expansion and recursive connection operations on the first-layer hidden state sequence, outputting a second-layer hidden state sequence. The hidden state output by the second-layer RKAN unit at the end of the local subsequence is used as the summary representation of the local subsequence, and the summary representation is output as the short-term feature vector to the concatenation operation.

[0011] Optionally, the gated straddle path of the TKAN dual-path feature extraction module transmits the hidden state output long-term trend features at a preset time interval, specifically including the following steps: Receiving a multivariate time series matrix, using a preset time interval as the gate interval, dividing the multivariate time series matrix along the time axis into several gated intervals, each gated interval containing a preset number of consecutive time steps; At the end of each gated interval, extracting the current hidden state from the current RKAN unit, multiplying the current hidden state by a learnable gate coefficient, and directly transmitting the hidden state multiplied by the gate coefficient to the input layer of the RKAN unit in the preset number of gated intervals through straddle connections, skipping the recursive calculation of multiple layers of RKAN units; RKAN units in time steps not covered by straddle connections calculate the hidden state layer by layer according to the standard recursive method; At the end of each gated interval, the hidden state transmitted by the straddle connection is added to the hidden state calculated layer by layer by the current RKAN unit and then normalized to serve as the fused hidden state at the end of the gated interval; The fused hidden states at the end time of all departmental control intervals are concatenated in chronological order to form a long-term trend feature vector, which is then output to the concatenation operation.

[0012] Optionally, mapping the comprehensive features to a mean through a mean fully connected layer specifically includes the following steps: Receiving the comprehensive feature vector output by the TKAN dual-path feature extraction module, the comprehensive feature vector contains short-term feature components and long-term trend feature components; Inputting the comprehensive feature vector into a mean fully connected layer, the mean fully connected layer contains input nodes and output nodes, the input nodes and output nodes are connected in a fully connected manner, each connection corresponds to a trainable weight parameter, and each output node corresponds to a trainable bias parameter; The mean fully connected layer performs a linear transformation on the comprehensive feature vector, and calculates the value of each output node; Combining the values ​​of all output nodes of the mean fully connected layer into a mean vector, each component of the mean vector corresponds to the center estimate of one dimension of the predicted target, and outputting the mean vector to the causal inference module.

[0013] Optionally, the step of mapping variance to variance through a variance fully connected layer, followed by a Softplus activation function, specifically includes the following steps: The system receives the comprehensive feature vector output from the TKAN dual-path feature extraction module. This comprehensive feature vector is then input into a variance-fully connected layer. The variance-fully connected layer contains input and output nodes, connected in a fully connected manner. Each connection corresponds to a trainable weight parameter, and each output node corresponds to a trainable bias parameter. The variance-fully connected layer performs a linear transformation on the comprehensive feature vector, calculating the value of each output node as the sum of the products of all input node values ​​and the corresponding weight parameter, plus the corresponding bias parameter. The values ​​of all output nodes in the variance-fully connected layer are combined to form an initial variance vector. Each component of the initial variance vector corresponds to an initial variance estimate for one dimension of the predicted target. Each component of the initial variance vector is then input into a Softplus activation function. The Softplus activation function performs a nonlinear transformation on each component, mapping any real-valued input to a positive real-valued output. This mapping is achieved by performing a natural exponential operation on the input value to obtain an exponent value, which is then processed to obtain the transformed output value. All output components after transformation by the Softplus activation function are combined into a final variance vector. Each component of the final variance vector corresponds to a positive real variance value of one dimension of the predicted target. The final variance vector is then output to the causal inference module and the instruction output module.

[0014] Optionally, when the variance is less than the high confidence threshold, the causal inference module matches the mean with the causal knowledge base to output the fault type and handling suggestions, specifically including the following steps: Reading the causal knowledge base from the distributed storage module, which includes a fault type identifier field, a multivariate time-series symptom pattern field, and a handling suggestion text field; Receiving the mean vector and the final variance vector from the TKAN probability prediction module, and calculating the standard deviation vector for each component of the final variance vector; Adding each component of the mean vector to the corresponding standard deviation vector component to obtain the upper bound of the confidence interval, and subtracting each component of the mean vector from the corresponding standard deviation vector component to obtain the lower bound of the confidence interval; Constructing the current observation pattern vector interval with the mean vector as the center and the lower and upper bounds of the confidence interval as boundaries; Calculating the interval overlap between the current observation pattern vector interval and the multivariate time-series symptom pattern field of each record in the causal knowledge base; Select the causal knowledge base record with the highest matching degree as the matching result, and read the values ​​of the fault type identifier field and the handling suggestion text field from the matching record; output the values ​​of the fault type identifier field and the handling suggestion text field to the instruction output module.

[0015] Optionally, the process of generating physical inspection commands and sending them to the inspection equipment via the command output module includes the following steps: The final variance vector is received from the TKAN probability prediction module, and each component of the final variance vector is compared with the second confidence threshold. When the variance component of any dimension exceeds the second confidence threshold, the mean vector output by the causal inference module is masked, and a variance exceeding limit warning is displayed in the visualization module. The physical inspection instruction header is formed by combining the current timestamp, the device number corresponding to the variance exceeding limit dimension, and the data dimension name. The optimal inspection path and inspection point set of the target device are read from the preset inspection path database using the device number as an index. The physical inspection instruction header is concatenated with the optimal inspection path and inspection point set to form a complete physical inspection instruction. The complete physical inspection instruction includes an instruction type field, a target device field, a path field, a point field, and a timestamp field. The instruction type field is assigned the value "physical inspection". The complete physical inspection instruction is sent to the inspection robot and inspection drone in data frame format through the wireless communication channel.

[0016] The beneficial effects of this invention are: (1) Frequency feature extraction and residual dependency capture of multivariate time series are performed at the edge through filter network to realize on-site alarm push for real-time anomaly detection, effectively reducing cloud computing load and communication latency. (2) The TKAN dual-path feature extraction architecture is adopted. The short-term RKAN path captures the instantaneous fluctuation features of the equipment status, and the long-term gated step path transmits the hidden state at a preset time interval to capture the slow degradation trend. The two are fused to form a full-scale representation of the equipment operating status, which significantly improves the prediction accuracy. (3) The mean fully connected layer and the variance fully connected layer output the prediction mean and variance respectively. The variance is mapped to a positive real number by the Softplus activation function to realize the uncertainty quantification assessment of the prediction results and provide confidence basis for subsequent diagnostic decisions. (4) A variance-driven dual threshold linkage mechanism is established. When the variance is within the first confidence threshold range, the causal reasoning module is started to match the knowledge base to output the fault type and handling suggestions. When the variance exceeds the second confidence threshold, the physical inspection command is triggered to realize the collaborative decision-making of high confidence automatic diagnosis and low confidence safety bottom line, avoiding the model blind output leading to misjudgment. (5) The causal reasoning module matches the current observation mode with the multivariate time-series symptom mode in the knowledge base by calculating the interval overlap, and selects the record with the highest matching degree as the diagnosis result. The diagnosis conclusion is interpretable. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the modules of a substation intelligent auxiliary monitoring platform based on big data analysis proposed in this invention. Figure 2 This is a schematic diagram of the dual-path memory mechanism of a substation intelligent auxiliary monitoring platform based on big data analysis proposed in this invention.

[0018] Figure 3 This is a schematic diagram illustrating the multidimensional uncertainty quantification output of a substation intelligent auxiliary monitoring platform based on big data analysis proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-3 A substation intelligent auxiliary monitoring platform based on big data analysis includes: The data acquisition module is used to collect data from various sensors and SCADA systems, forming raw data packets. It then interpolates, filters, aligns, and converts the raw data packets to different protocols, outputting a multivariate time series matrix. The edge detection module runs a filter network to filter the time series matrix, compares it with a threshold, and generates real-time alarms for local push notifications. The distributed storage module stores historical time series data and a causal knowledge base. The TKAN dual-path feature extraction module is used to receive a time series matrix, execute a short-term RKAN path to process local subsequences with a preset window length to output short-term features, execute a long-term gated step path to pass hidden states at preset time intervals to output long-term trend features, and splice the short-term features and long-term trend features into a comprehensive feature. The TKAN probability prediction module maps comprehensive features to the mean through a mean fully connected layer and to the variance through a variance fully connected layer, with the variance fully connected layer connected to the Softplus activation function. The causal inference module matches the mean with the causal knowledge base to output the fault type and handling suggestions when the variance is less than the high confidence threshold. The instruction output module displays the fault type and handling suggestions, and when the variance is greater than the low confidence threshold, it masks the mean and generates a physical inspection instruction to send to the inspection equipment.

[0021] In this embodiment, the output multivariate time series matrix of the data acquisition module specifically includes the following steps: Acquiring voltage waveforms, current waveforms, temperature values, vibration spectra, partial discharge signals, and telemetry and tele-signaling status values ​​output by the SCADA system from various sensors; Performing protocol parsing on each acquired raw data stream to convert non-standard protocol data into standard format data; Performing missing value detection on each converted standard format data stream, locating the timestamp and data dimension of the missing value, and interpolating the missing value using preset padding values; Performing a preset denoising algorithm on each interpolated data stream to filter out noise components above a preset frequency threshold; Using a preset reference time axis as a reference, aligning each data stream to the same time coordinate system according to the sampling timestamp; Arranging the aligned data streams in chronological order into a two-dimensional matrix, where the matrix row vectors correspond to the values ​​of all data dimensions at the same timestamp, and the matrix column vectors correspond to the value sequence of a single data dimension on the entire time axis; the two-dimensional matrix is ​​the multivariate time series matrix.

[0022] The process of parsing the acquired raw data from each source and converting non-standard protocol data into standard format data includes the following steps: The system identifies the communication link to which each data stream is attached and determines whether the channel carries standard or non-standard protocol frames based on configuration parameters. For data streams from standard power communication protocols, measurement values, status values, and time stamps are directly extracted by unpacking according to the protocol specifications. For non-standard protocol data, the frame header bytes are read as protocol feature codes and compared with templates in the private protocol feature library. Each template contains frame header flags, length field offsets, data type encoding tables, and verification algorithm identifiers. When a match is successful, the corresponding parsing script is called to extract the payload according to the specified offset and length, and convert it into physical quantity values ​​using the range conversion coefficient. When a match cannot be found, the self-identification process is triggered, which iterates through the byte combinations in the frame and performs a reasonableness check by combining them with the historical value range. The byte segments that pass the check are marked as valid data fields and stored in the feature library for subsequent use. All parsed data is uniformly packaged into an internal standard format message body, carrying the device identifier, sampling time, and quality bit.

[0023] The preset method for performing a preset denoising algorithm to filter out noise components above a preset frequency threshold in each of the interpolated data streams is as follows: A wavelet transform denoising algorithm is adopted, and the db4 wavelet in the Daubechies wavelet family is selected as the mother wavelet. The decomposition level is set to 5 levels. Multi-level wavelet decomposition is performed on the interpolated data of each path to decompose the signal into 1 approximate component and 5 detail components. The detail components correspond to noise components in different frequency ranges. Using a preset frequency threshold as the dividing point, the cutoff layer number of the threshold in the 1st to 5th detail components is determined. All detail components above the cutoff layer number are set to zero, and only the detail components and approximate components below the cutoff layer number are retained. Wavelet inverse transform is performed on the retained components to reconstruct the denoised signal sequence. The denoising threshold rule uses a soft threshold function, and the threshold value is calculated by multiplying the standard deviation of each layer of detail components by a coefficient. When the root mean square error between the reconstructed signal and the original signal is less than the preset error tolerance, the reconstructed signal is output as the effective signal after denoising.

[0024] In this embodiment, the edge detection module's operating filter network filters the time series matrix and compares it with a threshold to generate real-time alarms for on-site push. Specifically, this includes the following steps: A multivariate time series matrix is ​​obtained and input into a general shaping filter layer of a filter network. The general shaping filter layer applies a first learnable frequency filter to each dimension of the multivariate time series matrix to extract the fundamental frequency component of each dimension. The fundamental frequency component output from the general shaping filter layer is input into a context shaping filter layer of the filter network. The context shaping filter layer applies a second learnable frequency filter to the fundamental frequency component to capture the residual dependency between the fundamental frequency component and the original input. The filtered feature output from the context shaping filter layer is compared with a preset alarm threshold. When the filtered feature is greater than the preset alarm threshold, a real-time alarm signal is generated and pushed locally to the substation's local monitoring station.

[0025] The ordinary shaping filter layer applies a first learnable frequency filter to each dimension of the multivariate time series matrix to extract the fundamental frequency component of each dimension, specifically including the following steps: Configure a set of learnable frequency parameters for each dimension of the multivariate time series matrix, with each set of frequency parameters including a center frequency value and a bandwidth value; The learnable frequency parameters for each dimension are combined with a preset filter kernel function to generate the first learnable frequency filter for that dimension. The preset filter kernel function is a set of learnable frequency parameters, including the center frequency and bandwidth, configured by the system for each dimension of the time series. Subsequently, these parameters are substituted into a preset kernel function with a bandpass filter shape, such as a Gaussian function, to generate a specific filter. Perform sliding window truncation on the multivariate time series data for each dimension, and truncate a subsequence segment of a preset window length each time; The subsequence fragments of each dimension are input into the first learnable frequency filter corresponding to that dimension. The first learnable frequency filter performs weighted processing on the subsequence fragments in the frequency domain. The weighting method is to assign positive weights to frequency components within a preset bandwidth centered on the center frequency value, and assign zero weights to frequency components outside the preset bandwidth. The weighted frequency domain result is converted back to the time domain signal by inverse Fourier transform, and the amplitude of the frequency component with the highest energy concentration in the time domain signal is taken as the fundamental frequency component output value of that dimension.

[0026] The difference between the first learnable frequency filter and the second learnable frequency filter is as follows: For the first learnable frequency filter, it uses a general, learnable frequency kernel. You can think of it as a filter that is directly simulated by initializing a weight parameter. The function of this filter is very direct: to perform signal filtering and time relationship modeling. The second learnable frequency filter is not a direct filter, but rather uses a learnable neural network to generate the filter. Its goal is to learn the dependency relationship by utilizing the compatibility between the filtered frequency and the original input signal.

[0027] The context shaping filter layer applies a second learnable frequency filter to the fundamental frequency component to capture the residual dependency between the fundamental frequency component and the original input, specifically including the following steps: Obtain the fundamental frequency component and the original multivariate time series matrix output from the ordinary shaping filter layer; concatenate the fundamental frequency component and the original input along the channel dimension to form a joint input tensor; feed the joint input tensor into a learnable filter generation network constructed from a multilayer perceptron. The learnable filter generation network contains an input layer, two hidden layers, and an output layer. Each hidden layer contains a predetermined number of neurons and uses an activation function. The output layer outputs a filter parameter vector with the same dimension as the fundamental frequency component; the filter parameter vector contains the gain coefficient and bias coefficient of each frequency channel. Multiply the gain coefficient by the amplitude of the corresponding frequency channel in the fundamental frequency component and add it to the bias coefficient to obtain the filtered intermediate features; transform the intermediate features back to the time domain using an inverse Fourier transform to obtain the time-domain reconstructed signal; subtract the time-domain reconstructed signal from the original input to obtain the residual signal; input the residual signal into a residual dependency learning network. The residual dependency learning network uses a multilayer temporal convolution structure, sliding convolution of the residual signal along the time dimension with a convolution kernel of a predetermined size to extract the local temporal dependencies between time steps in the residual signal; The dependency features output by the residual dependency learning network are added to the fundamental frequency component element by element and fused. The fusion result is used as the output feature of the context shaping filter layer.

[0028] In this embodiment, the distributed storage module's storage of historical time-series data and causal knowledge base specifically includes the following steps: Define a columnar storage format, dividing the multivariate time series matrix into multiple column families according to data dimensions. Each column family corresponds to the full time-axis sequence value of a single data dimension. Establish a storage structure for each column family with device number as the primary index and timestamp as the secondary index. Store the causal knowledge base as a set of key-value pairs, where the key represents the fault type identifier, and the value represents the set of multivariate time-series symptom patterns corresponding to this fault type, as well as the corresponding handling suggestion text. Each time a new multivariate time series matrix is ​​received, append it to the corresponding device number and timestamp storage location in the columnar storage structure. After each causal inference module outputs a matching fault type and handling suggestion, write the fault type and the corresponding set of multivariate time-series symptom patterns as new records into the key-value pair set of the causal knowledge base. After each full write operation, update the metadata information of the distributed storage cluster, which includes the data volume of each column family, the latest timestamp, and the oldest timestamp.

[0029] Wherein, the value represents the set of multivariate time-series symptom patterns corresponding to this fault type. Specifically, the set of multivariate time-series symptom patterns is as follows: A structured data structure bound to a specific fault type identifier, consisting of multiple symptom entries, each symptom entry corresponding to a monitoring dimension; Each symptom entry contains a dimension identifier, a symptom type label, and a mode parameter vector; the dimension identifier points to a specific column in the multivariate time series matrix, including transformer winding hot spot temperature, top oil temperature, load current, ambient temperature, vibration acceleration, and partial discharge quantity; the symptom type label includes gradual trend type, sudden pulse type, periodic fluctuation type, and continuous over-limit type. The model parameter vector includes a rate of change threshold, a duration threshold, an extreme value boundary, and an upper limit for the variance of fluctuation. Gradual trend-type symptoms record the state where the rate of increase or decrease of this dimension exceeds the rate of change threshold within a preset time window; sudden pulse-type symptoms record instantaneous pulse events in which the amplitude of this dimension exceeds the extreme value boundary within a single sampling interval; periodic fluctuation-type symptoms record the frequency and amplitude of the amplitude of this dimension fluctuating regularly with the load cycle or environmental cycle; and continuous limit-exceeding-type symptoms record the state where the amplitude of this dimension exceeds the extreme value boundary at multiple consecutive sampling points. The symptom entries are connected by temporal logic operators, which include sequential relationships, simultaneous occurrence relationships, and delayed repetition relationships. The entire multivariate temporal symptom pattern set is represented as a multidimensional vector interval, and the allowable value range of each dimension is jointly defined by the extreme value boundary and the upper limit of the fluctuation variance in the pattern parameter vector.

[0030] In this embodiment, the short-term RKAN path processing of the TKAN dual-path feature extraction module outputs short-term features from local subsequences of a preset window length. Specifically, it includes the following steps: Receiving a multivariate time series matrix, and using the current time as the endpoint, extracting a local subsequence of a preset window length. The local subsequence contains multivariate observations from a preset number of consecutive time steps. Inputting the local subsequence into a first-layer RKAN unit, which includes a B-spline basis function layer and a recursive connection layer. The B-spline basis function layer maps each dimension of the input vector to B-spline basis function expansion coefficients. The recursive connection layer nonlinearly fuses the B-spline expansion coefficients of the current time step with the hidden state of the previous time step, outputting a first-layer hidden state sequence. Inputting the first-layer hidden state sequence into a second-layer RKAN unit, which has the same structure as the first-layer RKAN unit. The second-layer RKAN unit performs the same B-spline basis function expansion and recursive connection operations on the first-layer hidden state sequence, outputting a second-layer hidden state sequence. The hidden state output by the second-layer RKAN unit at the end of the local subsequence is used as the summary representation of the local subsequence, and the summary representation is output as the short-term feature vector to the concatenation operation.

[0031] The preset method for truncating a local subsequence of a preset window length is as follows: Different window length values ​​are set according to the signal characteristics of each monitoring dimension; for temperature monitoring dimension, the window length is set to 60 time steps; for load current and voltage monitoring dimension, the window length is set to 15 time steps; for vibration and partial discharge monitoring dimension, the window length is set to 6 time steps; for different monitoring dimensions, local subsequences are independently truncated using their respective window lengths; when the total number of time steps in the time series matrix is ​​less than the preset window length, the total number of time steps is used as the actual window length, and zeros are padded to the front of the truncated subsequence to the preset window length.

[0032] The specific steps for outputting the first-level hidden state sequence are as follows: The input local subsequence contains multivariate observations at a preset window number of consecutive time steps, with each time step corresponding to a multidimensional vector; The multidimensional vector of the first time step in the local subsequence is input into the B-spline basis function layer of the first RKAN unit. The B-spline basis function layer performs B-spline basis function expansion on each dimension of the multidimensional vector, mapping the scalar value of each dimension to the B-spline basis function expansion coefficient vector corresponding to that dimension. The expansion coefficient vectors of all dimensions are concatenated to form the expansion coefficient matrix of the current time step. The expansion coefficient matrix of the current time step and the hidden state of the previous time step initialized to zero vector are input into the recursive connection layer. The recursive connection layer concatenates the expansion coefficient matrix and the hidden state of the previous time step in the channel dimension, and outputs the hidden state of the current time step through the activation function after fully connected mapping. The hidden state of the current time step is used as the hidden state of the previous time step in the next time step. The multidimensional vector of the next time step in the local subsequence is input into the B-spline basis function layer. The B-spline expansion and recursive connection operations are repeated until the multidimensional vector of the last time step in the local subsequence is processed. The hidden states output by each time step are arranged in chronological order to form the first layer of hidden state sequence.

[0033] In this embodiment, the gated straddle path of the TKAN dual-path feature extraction module transmits the hidden state output long-term trend features at a preset time interval, specifically including the following steps: Receiving a multivariate time series matrix, using the preset time interval as the gate interval, dividing the multivariate time series matrix along the time axis into several gated intervals, each gated interval containing a preset number of consecutive time steps; At the end of each gated interval, extracting the current hidden state from the current RKAN unit, multiplying the current hidden state by a learnable gate coefficient, and directly transmitting the hidden state multiplied by the gate coefficient to the input layer of the RKAN unit in the preset number of gated intervals through straddle connections, skipping the recursive calculation of multiple layers of RKAN units; RKAN units in time steps not covered by straddle connections calculate the hidden state layer by layer according to the standard recursive method; At the end of each gated interval, the hidden state transmitted by the straddle connection is added to the hidden state calculated layer by layer by the current RKAN unit and then normalized to serve as the fused hidden state at the end of the gated interval; The fused hidden states at the end time of all departmental control intervals are concatenated in chronological order to form a long-term trend feature vector, which is then output to the concatenation operation.

[0034] Specifically, the hidden state after multiplication by the gating coefficient is directly transmitted to the input layer of the RKAN unit with a preset number of gating intervals via a step connection, which includes the following steps: At the end of the current gate interval, the current hidden state vector is extracted from the top layer of the current RKAN unit; the current hidden state vector is multiplied by a learnable gate coefficient, which is a scalar with a value between 0 and 1, and is dynamically calculated by the input data features of the current gate interval through the fully connected layer. The hidden state vector after multiplying by the gating coefficient is copied a preset number of times, and each copy corresponds to the input layer of the RKAN unit in the next gated interval; For each of the subsequent preset number of gated intervals, the original input of the RKAN unit input layer at the start time of the gated interval is added element by element to the corresponding copy, and the result of the addition is used as the actual input of the RKAN unit input layer at the start time of the gated interval. The step connection directly spans all intermediate time steps and all intermediate RKAN unit layers between the end of the current gated interval and the start of the subsequent gated interval, without going through any recursive calculations or nonlinear transformations of intermediate RKAN units during the transmission process; RKAN units in time steps not covered by cross-step connections within each gated interval still receive hidden states from the previous time step in the standard recursive manner, and do not receive cross-step transmissions.

[0035] Among them, the RKAN units in time steps not covered by cross-connects are calculated layer by layer in the standard recursive manner. The standard recursive method is as follows: At each time step within each gated interval that is not covered by the step connection, the multivariate observation vector of the current time step is input into the B-spline basis function layer of the first RKAN unit. The B-spline basis function layer maps the scalar value of each dimension in the vector to the B-spline basis function expansion coefficient corresponding to that dimension, and concatenates the expansion coefficients of all dimensions to form the expansion coefficient matrix of the current time step. The expansion coefficient matrix of the current time step is concatenated with the hidden state output from the previous time step in the channel dimension. The concatenated vector is then transformed by a fully connected linear transformation and the first hidden state of the current time step is output through an activation function. The first hidden state of the current time step is used as the input of the second RKAN unit. The second RKAN unit performs the same B-spline expansion and recursive connection operation as the first RKAN unit. That is, the second B-spline basis function layer performs B-spline expansion on the values ​​of each dimension in the first hidden state, and after concatenating the expansion coefficient matrix with the hidden state of the previous time step of the second layer, it outputs the second hidden state of the current time step through fully connected mapping and activation function. For structures containing two or more RKAN elements, each subsequent layer uses the output of the previous layer at the current time step as its input, repeatedly performing B-spline expansion and recursive connection operations, and recursively pushing to the top layer. The hidden states of each time step at each layer retain the hidden states of the previous time step as inputs for the recursive connection. The new hidden state of the current time step is calculated only based on the observations of the current time step itself and the hidden states of the previous time step, without jumping across time steps.

[0036] In this embodiment, the comprehensive features are mapped to a mean value through a mean fully connected layer, specifically including the following steps: receiving the comprehensive feature vector output by the TKAN dual-path feature extraction module, the comprehensive feature vector contains short-term feature components and long-term trend feature components; inputting the comprehensive feature vector into the mean fully connected layer, the mean fully connected layer contains input nodes and output nodes, the input nodes and output nodes are connected in a fully connected manner, each connection corresponds to a trainable weight parameter, and each output node corresponds to a trainable bias parameter; the mean fully connected layer performs a linear transformation on the comprehensive feature vector, calculating the value of each output node as the sum of the products of all input node values ​​and the corresponding weight parameter of that node, plus the corresponding bias parameter of that node; combining the values ​​of all output nodes of the mean fully connected layer into a mean vector, each component of the mean vector corresponds to the center estimate of one dimension of the predicted target, and outputting the mean vector to the causal inference module.

[0037] In this embodiment, variance is mapped to variance through a variance fully connected layer, and the variance fully connected layer is connected to the Softplus activation function. The specific steps include the following: The system receives the comprehensive feature vector output from the TKAN dual-path feature extraction module. This comprehensive feature vector is then input into a variance-fully connected layer. The variance-fully connected layer contains input and output nodes, connected in a fully connected manner. Each connection corresponds to a trainable weight parameter, and each output node corresponds to a trainable bias parameter. The variance-fully connected layer performs a linear transformation on the comprehensive feature vector, calculating the value of each output node as the sum of the products of all input node values ​​and the corresponding weight parameter, plus the corresponding bias parameter. The values ​​of all output nodes in the variance-fully connected layer are combined to form an initial variance vector. Each component of the initial variance vector corresponds to an initial variance estimate for one dimension of the predicted target. Each component of the initial variance vector is input into a Softplus activation function. The Softplus activation function performs a non-linear transformation on each component, mapping any real-valued input to a positive real-valued output. The mapping method involves performing a natural exponential operation on the input value to obtain an exponent value, adding one to the exponent value, and taking the natural logarithm as the transformed output value. All output components after transformation by the Softplus activation function are combined into a final variance vector. Each component of the final variance vector corresponds to a positive real variance value of one dimension of the predicted target. The final variance vector is then output to the causal inference module and the instruction output module.

[0038] The specific steps involved in combining all the output components after the Softplus activation function transformation into the final variance vector are as follows: Read the output value of each node after it has been transformed by the Softplus activation function from all the output nodes of the variance fully connected layer. Each output value corresponds to one dimension of the prediction target. Obtain the list of node numbers of the output nodes of the variance fully connected layer. Establish a one-to-one mapping relationship between each number in the node number list and the data dimension name of the prediction target. The mapping relationship is pre-stored in the configuration file. Arrange the output values ​​corresponding to each node number into a one-dimensional sequence according to the order of the node number list; bind each output value in the one-dimensional sequence to its corresponding data dimension name to form a dimension-variance pairing unit in the form of key-value pairs; All dimension-variance paired units are reordered lexicographically according to the data dimension names and stored in a vector container, with each position in the vector container storing one paired unit; all paired units in the vector container are read out sequentially, and the output values ​​of each paired unit are extracted and rearranged into a pure numerical one-dimensional array according to the reading order. A one-dimensional array is used as the numerical carrier of the final variance vector, and the set of data dimension names corresponding to each position is used as the index label of the final variance vector. The numerical carrier and the index label together constitute the final variance vector. The final variance vector is output to the causal inference module and the instruction output module.

[0039] In this embodiment, when the variance is less than the high confidence threshold, the causal inference module matches the mean with the causal knowledge base to output the fault type and handling suggestions, specifically including the following steps: Reading the causal knowledge base from the distributed storage module, which includes a fault type identifier field, a multivariate time-series symptom pattern field, and a handling suggestion text field; Receiving the mean vector and the final variance vector from the TKAN probability prediction module, and obtaining the standard deviation vector by taking the square root of each component of the final variance vector; Adding each component of the mean vector to the corresponding standard deviation vector component to obtain the upper bound of the confidence interval, and subtracting each component of the mean vector from the corresponding standard deviation vector component to obtain the lower bound of the confidence interval; Constructing the current observation pattern vector interval with the mean vector as the center and the lower and upper bounds of the confidence interval as boundaries; The overlap between the current observation mode vector interval and the multivariate time-series symptom pattern field of each record in the causal knowledge base is calculated. The overlap is calculated by taking the ratio of the intersection length to the union length of the current observation mode vector interval and the symptom pattern vector interval in each dimension, and averaging the ratios in all dimensions as the matching degree of the record. The causal knowledge base record with the highest matching degree is selected as the matching result, and the values ​​of the fault type identifier field and the treatment suggestion text field in the matching record are read. The values ​​of the fault type identifier field and the treatment suggestion text field are output to the instruction output module.

[0040] In this embodiment, the process of the instruction output module generating physical inspection instructions and sending them to the inspection equipment includes the following steps: The final variance vector is received from the TKAN probability prediction module, and each component of the final variance vector is compared with the second confidence threshold. When the variance component of any dimension exceeds the second confidence threshold, the mean vector output by the causal inference module is masked, and a variance exceeding limit warning is displayed in the visualization module. The physical inspection instruction header is formed by combining the current timestamp, the device number corresponding to the variance exceeding limit dimension, and the data dimension name. The optimal inspection path and inspection point set of the target device are read from the preset inspection path database using the device number as an index. The physical inspection instruction header is concatenated with the optimal inspection path and inspection point set to form a complete physical inspection instruction. The complete physical inspection instruction includes an instruction type field, a target device field, a path field, a point field, and a timestamp field. The instruction type field is assigned the value "physical inspection". The complete physical inspection instruction is sent to the inspection robot and inspection drone in data frame format through the wireless communication channel.

[0041] In this process, each component of the final variance vector is compared with a second confidence threshold, which is: The second confidence threshold is a preset numerical parameter, greater than the first confidence threshold, used to determine whether the variance output by the probability prediction module exceeds the confidence limit. The value of the second confidence threshold is predetermined after statistical analysis of the variance distribution in historical operating data. The specific determination method is as follows: Historical multivariate time series data of the substation under normal operating conditions for several consecutive days were collected. The historical data were then input into the TKAN dual-path feature extraction module and the TKAN probability prediction module in sequence to obtain the variance output sequence at each historical moment. The mean of the variance output sequence was calculated and multiplied by a preset multiplier coefficient as the initial value of the second confidence threshold. The specific value of the second confidence threshold is greater than that of the first confidence threshold. The first confidence threshold is used to determine whether the predicted mean can be used for causal inference matching, and the second confidence threshold is used to determine whether the current prediction variance is too high to the point that the predicted mean is completely unreliable. The second confidence threshold is manually set by the operations and maintenance personnel through the configuration file during system initialization or automatically calculated by the system based on historical data. After the second confidence threshold is set, it is stored in the configuration parameter area of ​​the distributed storage module for the instruction output module to read and call when performing variance comparison operations.

[0042] Example 1: During the trial operation of a substation, this invention selected a 220kV oil-immersed transformer as the monitoring object. This transformer had been running continuously for approximately 11 months since its last overhaul. Recently, the background monitoring system occasionally triggered oil temperature exceeding the limit alarm, but maintenance personnel did not find any obvious abnormalities during on-site inspections. Traditional threshold methods struggled to distinguish between sensor interference, load fluctuations, and actual insulation aging precursors, leading to a passive situation where maintenance decisions required repeated on-site confirmation.

[0043] To address this scenario, this invention was deployed at the substation, utilizing existing oil temperature sensors, winding hotspot sensors, ambient temperature and humidity sensors, and load current and voltage data provided by the SCADA system—a total of six monitoring dimensions—with a sampling interval of one minute, continuously monitoring for 168 hours. The data acquisition module integrated this six-dimensional data into a multivariate time series matrix. The filter network in the edge detection module extracted frequency features from the original sequence, quickly filtering out environmental noise and communication jitter interference. The cleaned sequence was then simultaneously sent to the distributed storage module and the TKAN dual-path feature extraction module.

[0044] During the 168-hour monitoring period, the TKAN dual-path feature extraction module's short-term path captured the rapid response characteristics of oil temperature fluctuations with load current using a data window of the past 15 minutes. The long-term path, with a gating interval of 60 minutes, transmitted the latent state, capturing the slow upward trend of oil temperature under the superposition of diurnal temperature differences and load cycles. The features output from the two paths were concatenated and input into the TKAN probability prediction module, which outputs the predicted mean and variance of oil temperature for the next 10 minutes every 5 minutes. Taking the data from the 72nd hour as an example, at this moment, the load current increased from 320 amps to 480 amps. The short-term path quickly captured the oil temperature rising from 82.3 degrees Celsius to 86.7 degrees Celsius, while the long-term path identified that the oil temperature baseline had slowly climbed from 80.1 degrees Celsius to 83.5 degrees Celsius over the past 48 hours. After concatenating the two, the probability prediction module gave an average oil temperature of 89.2 degrees Celsius and a variance of 2.4 for the next 10 minutes. Since the variance is below the first confidence threshold of 3.0, the system determines that the prediction is credible. The causal reasoning module matches the mean of 89.2 degrees Celsius and the corresponding multidimensional symptom pattern with the causal knowledge base, and the matching degree reaches 0.87. The knowledge base gives a preliminary diagnosis of accelerated insulation aging and suggests strengthening infrared temperature measurement tracking.

[0045] By the 120th hour, the load current jumped again from 350 amps to 520 amps. At this point, the short-term path detected an oil temperature jump from 85.1 degrees Celsius to 91.4 degrees Celsius, but the long-term path showed that the baseline oil temperature had been steadily climbing from 80.5 degrees Celsius to 87.2 degrees Celsius over the past 72 hours. The outputs of the two paths showed significant conflict. The probability prediction module gave an average oil temperature of 94.7 degrees Celsius for the next 10 minutes, but the variance increased sharply to 5.8, exceeding the second confidence threshold of 4.5. The output of the causal reasoning module was masked; the system did not push specific temperature values ​​or diagnostic conclusions, but instead directly generated a physical inspection command and sent it to the on-site inspection robot. The robot arrived at the transformer area 12 minutes later. Infrared thermal imaging showed an oil temperature of 89.2 degrees Celsius at the top of the transformer body, and a localized overheated area at the root of the A-phase bushing, reaching 96.5 degrees Celsius, which deviated from the conventional oil temperature measurement. This confirmed a poor contact defect at the bushing connection, rather than overall insulation aging. After the maintenance personnel arrived and addressed the issue, the overheating defect was eliminated, and the oil temperature returned to normal fluctuation range during subsequent monitoring.

[0046] During the trial operation of this invention, compared to another transformer at the same station using traditional threshold alarms, the total number of alarms decreased from 47 to 6, the number of invalid on-site confirmations decreased from 13 to 1, and the consistency rate between diagnostic conclusions and actual maintenance results reached 5 correct matches, with a correct matching rate of 83.3%. Traditional methods push the same level of alarm for all instances of exceeding limits, requiring maintenance personnel to screen and confirm each alarm, taking approximately 5.2 hours. This invention, through a variance linkage mechanism, reduces the number of manual interventions by 92.3%, requiring only one deployment of the inspection robot, and the time from triggering to on-site confirmation is only 12 minutes, significantly shortening the anomaly confirmation window. Simultaneously, because physical inspections are triggered by variance exceeding the threshold, unstable mean prediction outputs are shielded, avoiding the waste of maintenance resources caused by frequent false alarms during periods of severe signal fluctuations, as is common with traditional methods. This demonstrates that this invention has a significant advantage in distinguishing between sensor interference, load fluctuations, and actual equipment defects.

[0047] Table 1: Comparison of key indicators between the present invention and traditional threshold alarm methods during trial operation.

[0048] Table 1 lists the quantitative comparison results of the present invention with the traditional threshold alarm method during the trial operation. The traditional method triggered 47 alarms in 168 hours, of which 34 were false alarms caused by load fluctuations and instantaneous sensor interference. Maintenance personnel made 13 on-site visits for verification, but only one confirmed a genuine anomaly; the remaining 12 visits were ineffective, consuming approximately 5.2 hours of man-days. The present invention, by using a filter network to filter out high-frequency interference, reduced the total number of alarms to 6. Only one alarm triggered a physical inspection due to variance exceeding the threshold at the 120th hour, ultimately confirming a genuine defect of poor bushing contact. The remaining 5 alarms, after being matched with the knowledge base by the causal reasoning module, all provided diagnostic conclusions consistent with the maintenance records, achieving a correct matching rate of 83.3%. The variance-driven mechanism shielded the mean prediction output under high fluctuation conditions at the 120th and 156th hours, avoiding the pushing of unreliable values ​​when the signal changes drastically. Traditional methods, lacking variance assessment capabilities, triggered alarms during all seven periods of high load fluctuations, requiring manual verification for each. In contrast, this invention only required one deployment of inspection equipment and completed on-site verification within 12 minutes. The total time for manual intervention was reduced from 5.2 hours to 12 minutes, and the maintenance time consumed by false alarms was reduced from 4.1 hours to 0.2 hours. Overall, this invention reduces the on-site verification burden on maintenance personnel by approximately 90% while maintaining diagnostic quality. Simultaneously, it controls the average time from alarm triggering to conclusion output for a single anomaly to within 2 minutes, significantly improving the efficiency of substation anomaly handling.

[0049] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A substation intelligent auxiliary monitoring platform based on big data analysis, characterized in that, include: The data acquisition module is used to collect data from various sensors and SCADA systems, form raw data packets, interpolate, filter, align, and convert protocols for the raw data packets, and output a multivariate time series matrix. The edge detection module is used to run a filter network to filter the time series matrix, compare it with a threshold, generate real-time alarms, and push them locally; the distributed storage module is used to store historical time series data and a causal knowledge base. The TKAN dual-path feature extraction module is used to receive a time series matrix, execute a short-term RKAN path to process local subsequences with a preset window length to output short-term features, execute a long-term gated step path to pass hidden states at preset time intervals to output long-term trend features, and splice the short-term features and long-term trend features into a comprehensive feature. The TKAN probability prediction module maps comprehensive features to the mean through a mean fully connected layer and to the variance through a variance fully connected layer, with the variance fully connected layer connected to the Softplus activation function; the causal inference module matches the mean with the causal knowledge base and outputs the fault type and handling suggestions when the variance is less than the high confidence threshold. The instruction output module is used to display the fault type and handling suggestions. When the variance is greater than the low confidence threshold, the mean is masked and a physical inspection instruction is generated and sent to the inspection equipment. 2.The substation intelligent auxiliary monitoring platform based on big data analysis of claim 1, wherein, The output multivariate time series matrix of the data acquisition module specifically includes the following steps: acquiring voltage waveforms, current waveforms, temperature values, vibration spectra, partial discharge signals, and telemetry and tele-signaling status values ​​output by the SCADA system from multiple types of sensors; performing protocol parsing on each acquired raw data stream to convert non-standard protocol data into standard format data; performing missing value detection on each converted standard format data stream to locate the timestamp and data dimension of the missing value, and interpolating the missing value using a preset padding value; and performing a preset denoising algorithm on each interpolated data stream to filter out noise components above a preset frequency threshold. Using a preset baseline time axis as a reference, align all data streams to the same time coordinate system according to the sampling timestamp; The aligned data are arranged in chronological order into a two-dimensional matrix. The matrix row vectors correspond to the values ​​of all data dimensions at the same timestamp, and the matrix column vectors correspond to the value sequence of a single data dimension on the entire time axis. The two-dimensional matrix is ​​a multivariate time series matrix. 3.The substation intelligent auxiliary monitoring platform based on big data analysis of claim 1, wherein, The edge detection module's operating filter network filters the time-series matrix and compares it with a threshold to generate real-time alarms for local push notification. Specifically, this includes the following steps: Obtain the multivariate time series matrix, input the multivariate time series matrix into the ordinary shaping filter layer of the filter network, and apply the first learnable frequency filter to each dimension of the multivariate time series matrix to extract the fundamental frequency component of each dimension. The fundamental frequency component output from the ordinary shaping filter layer is input into the context shaping filter layer of the filter network. The context shaping filter layer applies a second learnable frequency filter to the fundamental frequency component to capture the residual dependency between the fundamental frequency component and the original input. The filtering characteristics output by the context shaping filter layer are compared with the preset alarm threshold. When the filtering characteristics exceed the preset alarm threshold, a real-time alarm signal is generated and pushed to the local monitoring station of the substation.

4. The substation intelligent auxiliary monitoring platform based on big data analysis of claim 1, wherein, The distributed storage module's storage of historical time-series data and causal knowledge base specifically includes the following steps: Define a columnar storage format, dividing the multivariate time series matrix into multiple column families according to the data dimension. Each column family corresponds to the full time axis sequence value of a single data dimension. Establish a storage structure for each column family with the device number as the first-level index and the timestamp as the second-level index. Store the causal knowledge base as a set of key-value pairs, where the key represents the fault type identifier and the value represents the set of multivariate time series symptom patterns corresponding to this fault type and the corresponding handling suggestion text. Each time a new multivariate time series matrix is ​​received, the multivariate time series matrix is ​​appended to the storage location of the corresponding device number and timestamp in the columnar storage structure; Each time the causal reasoning module outputs a matched fault type and handling suggestion, the fault type and the corresponding set of multivariate time-series symptom patterns are written as a new record into the key-value pair set of the causal knowledge base; After each full write operation is completed, the metadata information of the distributed storage cluster is updated. The metadata information includes the data volume of each column family, the latest timestamp, and the oldest timestamp.

5. The substation intelligent auxiliary monitoring platform based on big data analysis of claim 1, wherein, The short-term RKAN path processing of the TKAN dual-path feature extraction module outputs short-term features from local subsequences of a preset window length, specifically including the following steps: receiving a multivariate time series matrix, and extracting a local subsequence of a preset window length with the current time as the endpoint. The local subsequence contains multivariate observations from a preset number of consecutive time steps; inputting the local subsequence into a first-layer RKAN unit, which includes a B-spline basis function layer and a recursive connection layer. The B-spline basis function layer maps each dimension of the input vector to the B-spline basis function expansion coefficients, and the recursive connection layer... The B-spline expansion coefficients of the current time step are nonlinearly fused with the hidden state of the previous time step to output the first-layer hidden state sequence. The first-layer hidden state sequence is then input into the second-layer RKAN unit, which has the same structure as the first-layer RKAN unit. The second-layer RKAN unit performs the same B-spline basis function expansion and recursive connection operations on the first-layer hidden state sequence to output the second-layer hidden state sequence. The hidden state output by the second-layer RKAN unit at the end of the local subsequence is used as the summary representation of the local subsequence, and the summary representation is output as the short-term feature vector to the concatenation operation. 6.The substation intelligent auxiliary monitoring platform based on big data analysis of claim 1, wherein, The gated step path of the TKAN dual-path feature extraction module transmits the long-term trend features of the hidden state output at a preset time interval. Specifically, it includes the following steps: receiving a multivariate time series matrix; using the preset time interval as the gate interval, dividing the multivariate time series matrix along the time axis into several gated intervals, each gated interval containing a preset number of consecutive time steps; at the end of each gated interval, extracting the current hidden state from the current RKAN unit; multiplying the current hidden state by a learnable gate coefficient; and connecting the hidden state after multiplying the gate coefficient through a step path. The input layer of RKAN units, which are passed to a preset number of gated intervals, skips the recursive calculation of multiple RKAN units by step connection. The hidden states of RKAN units in time steps not covered by step connection are calculated layer by layer according to the standard recursive method. At the end time of each gated interval, the hidden states passed by step connection are added to the hidden states calculated layer by layer by the current RKAN unit and then normalized to serve as the fused hidden states at the end time of the gated interval. The fused hidden states at the end time of all gated intervals are concatenated in chronological order to form a long-term trend feature vector and output to the concatenation operation. 7.The substation intelligent auxiliary monitoring platform based on big data analysis of claim 1, wherein, The TKAN probability prediction module maps the comprehensive features to the mean through a mean fully connected layer, specifically including the following steps: receiving the comprehensive feature vector output by the TKAN dual-path feature extraction module, the comprehensive feature vector contains short-term feature components and long-term trend feature components; inputting the comprehensive feature vector into the mean fully connected layer, the mean fully connected layer contains input nodes and output nodes, the input nodes and output nodes are connected in a fully connected manner, each connection corresponds to a trainable weight parameter, and each output node corresponds to a trainable bias parameter; The mean-based fully connected layer performs a linear transformation on the composite feature vector; The values ​​of all output nodes of the mean fully connected layer are combined into a mean vector. Each component of the mean vector corresponds to the center estimate of one dimension of the predicted target. The mean vector is then output to the causal inference module. 8.The substation intelligent auxiliary monitoring platform based on big data analysis of claim 1, wherein, The TKAN probability prediction module maps to variance through a variance fully connected layer, which is then connected to the Softplus activation function. Specifically, the module includes the following steps: Receive the comprehensive feature vector output by the TKAN dual-path feature extraction module; input the comprehensive feature vector into the variance fully connected layer, which contains input nodes and output nodes. The input nodes and output nodes are connected in a fully connected manner. Each connection corresponds to a trainable weight parameter, and each output node corresponds to a trainable bias parameter. The variance fully connected layer performs a linear transformation on the comprehensive feature vector, and calculates the value of each output node as the sum of the products of all input node values ​​and the corresponding weight parameters of the node, plus the corresponding bias parameters of the node. The values ​​of all output nodes of the variance fully connected layer are combined into an initial variance vector. Each component of the initial variance vector corresponds to the initial variance estimate of one dimension of the predicted target. Each component of the initial variance vector is input into the Softplus activation function. The Softplus activation function performs a nonlinear transformation on each component. This transformation maps any real-valued input to a positive real-valued output. The mapping method is to perform a natural exponential operation on the input value to obtain the exponential value, and then process the exponential value to obtain the transformed output value. All output components after transformation by the Softplus activation function are combined into a final variance vector. Each component of the final variance vector corresponds to a positive real variance value of one dimension of the predicted target. The final variance vector is then output to the causal inference module and the instruction output module. 9.The substation intelligent auxiliary monitoring platform based on big data analysis of claim 1, wherein, When the variance is less than the high confidence threshold, the causal inference module matches the mean with the causal knowledge base to output the fault type and handling suggestions. Specifically, this includes the following steps: reading the causal knowledge base from the distributed storage module, which contains a fault type identifier field, a multivariate time-series symptom pattern field, and a handling suggestion text field; receiving the mean vector and final variance vector from the TKAN probability prediction module, calculating the standard deviation vector for each component of the final variance vector; processing each component of the mean vector with its corresponding standard deviation vector component to obtain the upper bound of the confidence interval, and processing each component of the mean vector with its corresponding standard deviation vector component to obtain the lower bound of the confidence interval; constructing the current observation pattern vector interval with the mean vector as the center and the lower and upper bounds of the confidence interval as boundaries; calculating the interval overlap between the current observation pattern vector interval and the multivariate time-series symptom pattern field of each record in the causal knowledge base. Select the causal knowledge base record with the highest matching degree as the matching result, and read the values ​​of the fault type identifier field and the handling suggestion text field from the matching record; output the values ​​of the fault type identifier field and the handling suggestion text field to the instruction output module. 10.The substation intelligent auxiliary monitoring platform based on big data analysis of claim 1, wherein, The process of generating physical inspection commands and sending them to the inspection equipment via the command output module includes the following steps: The final variance vector is received from the TKAN probability prediction module. Each component of the final variance vector is compared with the second confidence threshold. When the variance component of any dimension exceeds the second confidence threshold, the mean vector output by the causal inference module is masked, and a variance exceeding limit warning is displayed in the visualization module. The physical inspection instruction header is formed by combining the current timestamp, the device number corresponding to the variance exceeding limit dimension, and the data dimension name. The optimal inspection path and inspection point set of the target device are read from the preset inspection path database using the device number as an index. The physical inspection instruction header is concatenated with the optimal inspection path and inspection point set to form a complete physical inspection instruction. The complete physical inspection instruction includes an instruction type field, a target device field, a path field, a point field, and a timestamp field. The instruction type field is assigned the value "physical inspection". The complete physical inspection instruction is sent to the inspection robot and the inspection drone in data frame format through the wireless communication channel.