Power cabinet harmful gas monitoring and early warning method based on deep learning and pattern recognition

Abstract

This invention relates to the field of power equipment condition monitoring technology, and discloses a method for monitoring and early warning of harmful gases in power cabinets using deep learning and pattern recognition. The method includes: acquiring the time series of gas concentrations and operation event logs for each channel of a sensor array, performing time alignment and operation status labeling; generating channel-level time-series feature embedding vectors using gated cyclic units and calculating inter-channel coupling feature matrices; obtaining channel baseline drift factor vectors and cross-coupling drift factor matrices through joint non-negative matrix factorization; compensating and updating the transient baseline parameter set using drift factors; generating a fused feature vector after drift decoupling based on the operation status label; generating a confidence distribution vector using a multilayer spiking neural network dynamically modulated by a release threshold; performing counterfactual reasoning on a state-switching directed acyclic graph causal model after drift correction upon triggering; and performing a significance test on the causal effect vector and jointly determining and generating an early warning signal.

Description

Technical Field

[0001] The embodiments disclosed herein relate to the field of power equipment condition monitoring technology, and in particular to a method for monitoring and early warning of harmful gases in power cabinets using deep learning and pattern recognition. Background Technology

[0002] Sulfur hexafluoride ( Gas-insulated switchgear is widely used in medium and high voltage power cabinets due to its excellent insulation and arc-extinguishing properties. With the continuous expansion of the power grid and the increasing service life of the equipment, Gas-insulated switchgear generates gas during circuit breaker opening and closing, load switching, and other operations. , , , The decomposition products of these harmful gases, along with continuous leaks caused by insulation defects in the equipment, can also lead to the accumulation of these gaseous components inside the power cabinet. Real-time monitoring and accurate early warning of these harmful gases are crucial technical requirements for ensuring the safe operation of power equipment.

[0003] Existing gas monitoring solutions based on sensor arrays typically employ deep learning methods for feature extraction and pattern recognition of multi-channel concentration time series. However, with long-term service of the sensor array, the baseline and sensitivity of each channel drift independently, and the cross-sensitivity coefficient between channels changes asymmetrically with the degradation of sensor material performance. These drift phenomena cause a systematic shift in the distribution structure of the feature space, resulting in systematic biases in deep learning models based on fixed parameters. Simultaneously, the gas concentration changes generated during the transient period of normal operation of switching equipment and the concentration changes caused by fault leaks highly overlap in temporal patterns, further exacerbating the difficulty of pattern recognition.

[0004] Current gas monitoring and early warning solutions generally treat sensor drift as a one-time overall offset, failing to orthogonally decompose the physical causes and temporal scale differences of channel baseline drift and cross-coupling drift, and neglecting the asymmetric impact of these two types of drift on transient baseline reconstruction and causal inference models. Specifically, baseline drift causes a slow shift in the background signal before and after an operation. If this shift is not independently separated and tracked, transient baseline parameters cannot be dynamically updated, and short-term concentration pulses generated by normal operations such as circuit breaker opening and closing are easily misjudged as abnormal leaks. Furthermore, the asymmetric degradation of cross-sensitivity caused by sensor material degradation distorts the response ratios between various gas components, directly interfering with causal inference based on chemical reaction paths, leading to systematic biases in leak tracing and gradual trend identification. Under this combined drift, concentration fluctuations during the operational transient period are incorrectly amplified as fault characteristics, causing frequent false alarms. Simultaneously, the slowly increasing leak signal, momentarily masked by the operational event, is missed due to the failure of causal criteria.

[0005] Therefore, there is an urgent need for a gas monitoring and early warning method that integrates sensor drift orthogonal decomposition, transient baseline adaptive learning, and causal inference dynamic calibration to solve the problem of high false alarm and false negative rates in existing solutions under scenarios where operational events and progressive leaks are deeply coupled. Summary of the Invention

[0006] The purpose of this invention is to provide a method for monitoring and early warning of harmful gases in power cabinets based on deep learning and pattern recognition, which solves the technical problems of high false alarm rate and difficulty in distinguishing fault causes caused by baseline drift and transient interference during operation of power cabinet sensor arrays during long-term operation.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] As a first aspect of this application, the present invention discloses a method for monitoring and early warning of harmful gases in power cabinets using deep learning and pattern recognition, comprising the following steps:

[0009] Step S100: Obtain the gas concentration time series and operation event log of each channel of the sensor array in the power cabinet, and perform time alignment and operation status labeling.

[0010] Step S200: The concentration sequences of each channel are independently encoded by a gated recurrent unit to generate channel-level time-series feature embedding vectors, and the inter-channel coupling feature matrix is ​​generated based on the sliding window cross-correlation.

[0011] Step S300: Calculate the offset matrix and ratio matrix based on the characteristics of normal operating conditions during the static period, and obtain the channel baseline drift factor vector and cross-coupling drift factor matrix through joint non-negative matrix decomposition.

[0012] Step S400: Update the transient baseline parameter set by drift compensation using the channel baseline drift factor vector and the cross-coupling drift factor matrix;

[0013] Step S500: Select the corresponding processing method based on the operation status label at the current sampling time to generate the fused feature vector after drift decoupling;

[0014] Step S600: Input the fused feature vector into the multilayer spiking neural network with dynamic modulation of the firing threshold to generate a confidence distribution vector;

[0015] Step S700: When the probability value of any leakage category in the confidence distribution vector exceeds the trigger threshold, input the drift-corrected causal model for counterfactual reasoning.

[0016] Step S800: Obtain the fault causal effect vector and the operation causal effect vector, and perform a joint judgment based on the significance test to generate an early warning signal;

[0017] The joint nonnegative matrix decomposition decomposes the overall sensor distribution offset into two orthogonal subspace components: channel-independent baseline drift and inter-channel cross-sensitivity coupling drift.

[0018] Preferably, step S300, which involves obtaining the channel baseline drift factor vector and the cross-coupling drift factor matrix through joint nonnegative matrix decomposition, includes:

[0019] Step S301: The offset vectors of the centroids of each channel feature embedding vector relative to the centroids of the calibration stage are used to form an offset matrix. For the offset matrix Performing element-wise absolute value operations yields the non-negative offset matrix. And record the sign matrix of each element. ;

[0020] Step S302: Calculate the element-wise ratio matrix of the mean of the inter-channel coupling feature matrix of the normal operating condition frame to the inter-channel coupling feature matrix of the calibration stage. ;

[0021] Step S303, the non-negative offset matrix With the ratio matrix A joint nonnegative matrix is ​​formed by splicing along the column directions. For the joint nonnegative matrix Execution rank is Nonnegative matrix decomposition;

[0022] Step S304: Extract the non-negative basis matrix. The first column serves as the channel baseline drift factor vector. ;

[0023] Step S305, using the channel baseline drift factor vector For the ratio matrix The diagonal contribution components caused by independent channel drift are stripped to obtain the cross-coupling drift factor matrix.

[0024] Preferably, step S400 involves updating the transient baseline parameter set using the channel baseline drift factor vector and the cross-coupling drift factor matrix, including:

[0025] For each channel mean peak concentration increment The mean peak concentration increment after compensation is obtained by applying an inverse sensitivity scaling transform. ;

[0026] The mean vector of multi-channel peak concentration increments for each operation type Applying the inverse transformation correction to the cross-coupling drift factor matrix yields the corrected mean vector of peak concentration increments. ;

[0027] Among them, the average value of the peak concentration increment after compensation The corrected peak concentration increment mean vector is used for channel-by-channel calculation of transient baseline prediction values ​​for each channel. Used for reparameterization of the operational response path structure equation in causal models.

[0028] Preferably, in step S500, a fused feature vector after drift decoupling is generated based on the operation state label at the current sampling time. When the operation state label is in the operation transient period or the post-operation decay period:

[0029] Step S501: Calculate the predicted transient baseline value for each channel at the current time using the drift-compensated transient baseline parameter set according to the double exponential transient response curve function. ;

[0030] Step S502: Calculate the residual signal between the measured concentration values ​​of each channel and the predicted transient baseline values. ;

[0031] Step S503, using the channel baseline drift factor vector The corrected residuals are obtained by applying inverse drift-translation correction to the residual signals of each channel. ;

[0032] Step S504: The fused residual feature vector is generated by the component-sensing attention module based on the corrected residuals of each channel. .

[0033] Preferably, the component-aware attention module uses the current inter-channel coupling feature matrix. After being flattened row by row, the query vector is generated by projection through a linear mapping layer. After correction, the residuals of each channel are passed through a linear mapping layer to generate key vectors and value vectors. A multi-head scaling dot product attention operation is then used to weight and sum the value vectors of each channel. The multi-head outputs are then concatenated with the linear mapping layer to generate the fused residual feature vector. When the operation status label is in the pre-operation static period, the channel baseline drift factor vector is used. and symbol matrix Embedded vectors of time-series features for each channel Performing inverse drift translation yields the corrected temporal feature embedding vector. ,Will The inputs, serving as both key and value vectors, are processed by the component-aware attention module with the same structure to generate a static fusion feature vector. .

[0034] Preferably, the firing threshold of each neuron in the multilayer spiking neural network is dynamically modulated according to the operation state label:

[0035] When the operation state label is in the operation transient period, the firing threshold of neurons in each layer is changed from the standard threshold. Adjust to ;

[0036] When the operation status label is in the pre-operation static period or the post-operation decay period, the issuance threshold of each layer is restored to the standard threshold. ;

[0037] In this process, the firing sequence of the top-level spiking neurons is input into the reserve pool computation layer, and a high-dimensional state vector is generated through nonlinear dynamic mapping via fixed-random connection computation nodes. This high-dimensional state vector is then used by a linear readout layer to generate the confidence distribution vector. .

[0038] Preferably, the drift-corrected causal model in S600 is a state-switching directed acyclic graph causal model, which includes operation-driven nodes. Gas generation node Sensor observation nodes and faulty / leaking nodes The gas generation node The structural equation coefficients of the operational response path are reparameterized using the mean vector of peak concentration increments after multi-channel coupling correction, and the sensor observation node... The mapping function is corrected using the inverse of the cross-coupling drift factor matrix, and the counterfactual reasoning includes:

[0039] At the fault leakage node Under the intervention condition of zeroing out, the counterfactual predicted value is calculated, and the difference between the actual observed value and the counterfactual predicted value is used as the failure causality vector. ;

[0040] In the operation driving node The counterfactual predicted value is calculated under the intervention condition of zeroing, and the difference between the actual observed value and the counterfactual predicted value is used as the operational causal effect vector. .

[0041] Preferably, step S800, which involves performing a joint determination based on a significance test to generate an early warning signal, includes:

[0042] When only the fault causal effect vector is mentioned Upon passing the significance test, a fault leak warning signal containing the gas type and estimated concentration level is generated;

[0043] When the fault causal effect vector and the causal effect vector of the operation When both pass the significance test, according to the two Norm ratio determines the proportion of fault components and operation component proportion Generate a composite early warning signal that includes the proportion of the fault component and the proportion of the operation component;

[0044] When only the causal effect vector of the operation is mentioned When the significance test is passed, it is determined that the gas change is caused by normal operation, the transient characteristics of the operation are recorded, and the transient baseline parameter set is updated.

[0045] As a second aspect of this application, the present invention also discloses a deep learning and pattern recognition-based power cabinet hazardous gas monitoring and early warning system, comprising:

[0046] The data acquisition and alignment unit is configured to acquire the gas concentration time series and operation event logs of each channel of the sensor array in the power cabinet, and perform time alignment and operation status labeling.

[0047] The feature encoding unit is configured to independently encode the concentration sequence of each channel through a gated recurrent unit to generate channel-level temporal feature embedding vectors, and to generate inter-channel coupling feature matrices based on sliding window cross-correlation calculation;

[0048] The drift decomposition unit is configured to calculate the offset matrix and ratio matrix based on the characteristics of normal operating conditions during the static period, and obtain the channel baseline drift factor vector and cross-coupled drift factor matrix through joint non-negative matrix decomposition.

[0049] The drift compensation unit is configured to update the transient baseline parameter set by drift compensation using the channel baseline drift factor vector and the cross-coupling drift factor matrix;

[0050] The residual calculation and fusion unit is configured to generate a drift-decoupled fusion feature vector based on the operation state label at the current sampling time.

[0051] An abnormal pattern recognition unit is configured to input the fused feature vector into a multilayer spiking neural network with dynamic modulation of the firing threshold to generate a confidence distribution vector.

[0052] The causal reasoning unit is configured to input the current sensor observation and operation status label into the drift-corrected causal model to perform counterfactual reasoning when the probability value of any leakage category in the confidence distribution vector exceeds the trigger threshold, thereby obtaining the fault causal effect vector and the operation causal effect vector.

[0053] The joint judgment and early warning unit is configured to perform a significance test on the fault causal effect vector and the operation causal effect vector, and perform a joint judgment to generate an early warning signal based on the test results.

[0054] As a third aspect of this application, the present invention also discloses an electronic device comprising:

[0055] At least one processor, and a memory communicatively connected to said at least one processor;

[0056] The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to perform the system and steps of the above-described deep learning and pattern recognition method for monitoring and warning of hazardous gases in power cabinets.

[0057] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0058] This invention discloses a method for monitoring and early warning of hazardous gases in power cabinets using deep learning and pattern recognition. By decomposing the overall sensor distribution offset into two orthogonal subspace components—a channel baseline drift factor vector and a cross-coupling drift factor matrix—subsequent compensation can independently correct for two different physical drift sources: independent channel baseline translation and cross-sensitivity changes between channels. This reduces the systematic contamination of the feature distribution structure by long-term sensor drift. By using drift factors to compensate and update the transient baseline parameter set for each operation type, and selecting the corresponding residual calculation or embedding vector correction path based on the operation state label, the fused feature vector maintains a representation adapted to the drift state during both the transient and static periods, helping to reduce the aliasing between the transient response and fault leakage signals. By introducing dynamic modulation of the release threshold based on the operation state label into a multilayer spiking neural network, the network increases the release threshold during the transient period to suppress low-amplitude response interference caused by transient baseline compensation residues, and restores the standard threshold during the static and decay periods to maintain sensitivity to weak, gradual leakage signals, thus achieving adaptive sensitivity switching under both operating conditions. By introducing drift correction into the state-switching directed acyclic graph causal model, and performing counterfactual reasoning under two intervention conditions—zeroing the fault leakage node and zeroing the operation-driven node—the fault causal effect vector and the operation causal effect vector are obtained. After significance testing and joint judgment, an early warning signal is generated, so that the attribution of the fault leakage component and the operation response component has an interpretable causal basis, which helps to reduce misjudgment and missed judgment caused by the superposition of the two types of signals. Attached Figure Description

[0059] In the attached diagram:

[0060] Figure 1This is a flowchart illustrating the steps of the hazardous gas monitoring and early warning method for power cabinets in this embodiment of the invention.

[0061] Figure 2 This is a schematic diagram of the overall processing flow of the power cabinet gas anomaly detection system in an embodiment of the present invention;

[0062] Figure 3 This is a schematic diagram of the time series of gas concentration in each channel before and after the circuit breaker closing event in an embodiment of the present invention;

[0063] Figure 4 This is a schematic diagram comparing the current value and calibration value of the inter-channel coupling feature matrix in an embodiment of the present invention;

[0064] Figure 5 This is a schematic diagram of the baseline drift factor distribution of each channel in an embodiment of the present invention;

[0065] Figure 6 This is a schematic diagram comparing the peak increments before and after compensation of transient baseline parameters during circuit breaker closing operation in an embodiment of the present invention;

[0066] Figure 7 This is a schematic diagram of the residual decomposition and correction results of each channel at a certain moment in an embodiment of the present invention;

[0067] Figure 8 This is a schematic diagram comparing the counterfactual reasoning failure effect and the operational effect vector in an embodiment of the present invention;

[0068] Figure 9 This is a schematic diagram illustrating the ratio of fault component to operation component in the composite early warning signal in an embodiment of the present invention. Detailed Implementation

[0069] The application scenarios involved in the embodiments of this disclosure are those containing sulfur hexafluoride. This paper discloses a method for monitoring and early warning of hazardous gases inside the power cabinet of a gas-insulated switchgear. During the long-term operation of such power cabinets, the baseline and sensitivity of each channel of the sensor array drift independently, and the cross-sensitivity coefficient between channels changes asymmetrically with the degradation of sensor material performance. This drift phenomenon leads to systematic biases in existing deep learning-based feature extraction and pattern recognition methods, specifically manifested as frequent false alarms during operational transients and missed detections of gradual leaks masked by operational conditions. To address this, this paper provides a method for monitoring and early warning of hazardous gases in power cabinets based on deep learning and pattern recognition. This method decomposes the overall sensor distribution offset into two orthogonal subspace components: a channel baseline drift factor and a cross-coupling drift factor. The obtained drift factors are then used to adaptively correct transient baseline parameters, residual signals, and causal inference models, thereby achieving accurate causal separation between operational transient responses and fault leaks.

[0070] Example 1

[0071] In this embodiment of the invention, a method for monitoring and early warning of harmful gases in power cabinets based on deep learning and pattern recognition is disclosed, such as... Figure 1 and Figure 2 As shown, the specific steps include:

[0072] Step S100: Obtain the gas concentration time series and operation event log of each channel of the sensor array in the power cabinet, and perform time alignment and operation status labeling.

[0073] Step S200: The concentration sequences of each channel are independently encoded by a gated recurrent unit to generate channel-level time-series feature embedding vectors, and the inter-channel coupling feature matrix is ​​generated based on the sliding window cross-correlation.

[0074] Step S300: Calculate the offset matrix and ratio matrix based on the characteristics of normal operating conditions during the static period, and obtain the channel baseline drift factor vector and cross-coupling drift factor matrix through joint non-negative matrix decomposition.

[0075] Step S400: Update the transient baseline parameter set by drift compensation using the channel baseline drift factor vector and the cross-coupling drift factor matrix;

[0076] Step S500: Select the corresponding processing method based on the operation status label at the current sampling time to generate the fused feature vector after drift decoupling;

[0077] Step S600: The confidence distribution vector is generated by the multilayer spiking neural network with fused feature vector input to the firing threshold dynamic modulation.

[0078] Step S700: When the probability value of any leakage category in the confidence distribution vector exceeds the trigger threshold, input the drift-corrected causal model for counterfactual reasoning.

[0079] Step S800: Obtain the fault causal effect vector and the operation causal effect vector, and perform a joint judgment based on the significance test to generate an early warning signal.

[0080] Specifically, step S100 involves acquiring the gas concentration time series and operation event logs of each channel of the sensor array within the power cabinet, and performing time alignment and operation status labeling. This includes acquiring the gas concentration time series and auxiliary temperature and humidity data for each channel of the sensor array within the power cabinet, and simultaneously acquiring the type identifier and precise timestamp of each operation event in the switchgear operation event log. The operation event timestamps are then aligned with the concentration time series, and operation status labels are added for each sampling moment based on the time alignment results. These operation status labels include the pre-operation static period, the operation transient period, and the post-operation decay period.

[0081] For example, a sensor array contains One channel, The total number of channels in the sensor array, respectively corresponding to For each target gas component, the gas concentration is synchronously collected by each channel at a preset sampling frequency, generating... A parallel concentration time series. Operational event logs are obtained from the protection and control devices of the switchgear via standard communication protocols. The type identifiers of the operational events include circuit breaker opening, circuit breaker closing, and load switching. The sensor array contains 5 channels, each monitoring... , , , The system samples five gas components, including CO, at a sampling frequency of 1 Hz. Operational event logs are acquired via the IEC 61850 protocol, with time alignment accuracy controlled to the millisecond level.

[0082] Step S200 involves using a gated recurrent unit (GRU) to independently encode the concentration sequences of each channel, generating channel-level temporal feature embedding vectors, and calculating the inter-channel coupling feature matrix based on sliding window cross-correlation. Before inputting the concentration time series of each channel into the GRU, Z-score normalization is applied to each channel to eliminate the influence of differences in the concentration dimensions of different gas components on the encoding process. A multi-level gated recurrent unit (GRU) is used to independently encode the normalized concentration sequences of each channel, generating channel-level temporal feature embedding vectors. The input to the GRU is the normalized concentration time series window for each channel, and the output is the fixed-dimensional feature embedding vector corresponding to each channel. This feature embedding vector represents the temporal evolution characteristics of the corresponding channel within the current time window. The training objective is to enable the feature embedding vectors of each channel to reconstruct the concentration sequence of the corresponding channel. A mean squared error loss function is used, and the Adam optimization algorithm is employed for parameter updates.

[0083] Simultaneously, the sliding window cross-correlation coefficient sequence between each channel pair is calculated. For One channel, The total number of channels in the sensor array is calculated. For the cross-correlation coefficient sequence, the cross-correlation coefficient sequence of all channel pairs is encoded by a one-dimensional convolutional layer to generate... Inter-channel coupling feature matrix Inter-channel coupling feature matrix The The element represents the first Channel and the The current cross-sensitivity coupling strength between channels, where For channel row index, This is the channel column index. For example, in one example, the multi-layer gated recurrent unit has a 2-layer structure, and each channel generates a 64-dimensional temporal feature embedding vector; the sliding window size is 30 sampling points, and the 1D convolutional layer has a kernel size of 3, generating 5 channels. Inter-channel coupling feature matrix .

[0084] Step S300 involves calculating the offset matrix and ratio matrix based on the characteristics of normal operating conditions during the static period, and obtaining the channel baseline drift factor vector and cross-coupling drift factor matrix through joint nonnegative matrix decomposition. The most recent continuous... The annotations are the channel-level feature embedding vectors and inter-channel coupling feature matrices corresponding to the normal operating condition frames during the pre-operation static period, where... This represents the number of normal operating condition frames used for drift estimation. For each channel... ( , (Total number of channels in the sensor array) Calculate The first normal operating condition frame Centroid of channel feature embedding vector And calculate the centroid. relative to the centroid of the calibration phase offset vector ,in and All dimensional vector, This is the output dimension of the hidden layer of the multi-layer gated loop unit. The offset matrix is ​​composed of the offset vectors of each channel. offset matrix for 3D matrix.

[0085] Simultaneously, calculation The mean of the inter-channel coupling feature matrix of a normal operating frame And calculate the mean. Relative to the inter-channel coupling characteristic matrix in the calibration stage element-wise ratio matrix The ratio matrix The The elements are , for The One element, For the calibration stage, the inter-channel coupling characteristic matrix The One element, For channel row index, Channel column index, ratio matrix for 3D matrix. Due to the ratio matrix Each element is a ratio of the coupling coefficients of the same type, a dimensionless pure numerical value, and is related to the offset matrix. Each element is a dimensionless value in the embedded space, and the two can be directly concatenated in step S300 to perform joint nonnegative matrix decomposition.

[0086] For the offset matrix Sum ratio matrix Perform joint nonnegative matrix decomposition to decompose the overall sensor distribution offset into two orthogonal subspace components: the channel baseline drift factor vector and the cross-coupling drift factor matrix.

[0087] It should be noted that the specific process of joint nonnegative matrix factorization includes:

[0088] Step S301, adjust the offset matrix Performing element-wise absolute value operations to satisfy the nonnegativity constraint yields a nonnegative offset matrix. ( (dimensional), while recording the offset matrix. Sign matrix of each element This is to allow for the recovery of orientation information during subsequent calibration.

[0089] Step S302, the non-negative offset matrix With ratio matrix By splicing along the column directions, a joint nonnegative matrix is ​​formed. Joint nonnegative matrix for 3D matrix.

[0090] Step S303, for the joint nonnegative matrix Execution rank is The nonnegative matrix decomposition yields... ,in To decompose the rank, for 3D nonnegative basis matrix, for A non-negative coefficient matrix.

[0091] Step S304: Extract the non-negative basis matrix The first column as 3D channel baseline drift factor vector ,in Indicates transpose. The first To the The baseline drift amplitude factor for each channel, with each element representing the independent baseline drift amplitude factor for the corresponding channel.

[0092] Step S305, from the nonnegative coefficient matrix Extract the corresponding ratio matrix from the middle From the column intervals, we obtain the submatrix. ( (dimensional); ratio matrix ( (Dimensional, with all elements being positive) serves as the cross-coupling drift factor matrix. The initialization basis, and using non-negative basis matrices The first column is the channel baseline drift factor vector. Contrast Matrix The diagonal contribution component caused by independent channel drift is stripped. Specifically: ,in This represents element-wise division. for A one-dimensional column vector for dimensional matrix and its 1st dimension matrix All elements in the row are equal to The resulting cross-coupling drift factor matrix for A 3D matrix, with each element being positive, representing the degree of pure coupling drift in the cross-sensitivity between channel pairs after stripping the independent drift contribution of each channel.

[0093] Joint nonnegative matrix factorization, by simultaneously processing embedded spatial offset information and coupling coefficient ratio information within the same decomposition framework, improves the extracted channel baseline drift factor vector. and cross-coupling drift factor matrix These correspond to two different physical drift sources: channel-independent baseline shift and inter-channel sensitivity cross-variation. The two components belong to orthogonal subspaces in the latent space of the joint decomposition, allowing subsequent compensation to independently correct for each type of drift. For example, in... Set to 10, decompose rank When set to 2. For a 5-channel sensor array, the joint nonnegative matrix... for dimension( , The 5-dimensional channel baseline drift factor vector was extracted through nonnegative matrix factorization. and Cross-coupling drift factor matrix .

[0094] Step S400 involves updating the transient baseline parameter set for each operation type using the drift factor for drift compensation. This is achieved using the channel baseline drift factor vector obtained in step S300. and cross-coupling drift factor matrix The transient baseline parameter set for each operation type is updated with drift compensation. The transient baseline parameter set includes the average peak concentration increment, rise time constant, and decay time constant for each channel. The transient baseline parameter set is determined and recorded during the initial system calibration phase.

[0095] For each channel ( , The average peak concentration increment (total number of channels in the sensor array) Applying an inverse sensitivity scaling transform corresponding to the channel drift factor, the mean of the compensated peak concentration increment is obtained. ,in Channel baseline drift factor vector The One element, The first phase determined for the calibration stage The average peak concentration increment of each channel. The physical meaning of this inverse sensitivity scaling transformation is: when sensor sensitivity increases or decreases due to aging, the sensor response amplitude caused by the same actual concentration change changes accordingly. By inversely scaling the drift factor, the average peak concentration increment is matched to the actual sensitivity of the current sensor. The rise time constant and decay time constant of each channel are kept constant because these two parameters characterize the time dynamics of gas diffusion and chemical reactions, and their physical characteristics are not affected by sensor drift.

[0096] For parameters involving multi-channel coupled responses, the mean vector of multi-channel peak concentration increments for each operation type is used. Apply inverse transform correction to the cross-coupling drift factor matrix: ,in Cross-coupling drift factor matrix The inverse matrix, The first To the The average peak concentration increment determined for each channel during the calibration phase. Due to the cross-coupling drift factor matrix. All elements are positive, cross-coupling drift factor matrix It is a positive matrix and invertible. The inverse transformation eliminates the impact of inter-channel cross-sensitivity drift on multi-channel joint baseline prediction.

[0097] It should be noted that the average peak concentration increment obtained by single-channel inverse sensitivity scaling in step S400 is the compensated peak concentration increment. The mean vector of peak concentration increments obtained by multi-channel coupling correction Used for different subsequent steps: mean of peak concentration increment after compensation As the average of the independent peak concentration increments for each channel, it is used as the transient baseline prediction value for each channel in the subsequent step S500. Channel-by-channel calculation; mean vector of peak concentration increments after multi-channel coupling correction. As a multi-channel joint parameter, it is used to reparameterize the operational response path structure equation of the state-switching directed acyclic graph causal model in step S700, so as to reflect the impact of inter-channel coupling drift on the joint response prediction. The above drift compensation update generates a drift-compensated transient baseline parameter set, which is used in subsequent steps for transient baseline prediction and causal model parameterization.

[0098] Step S500 involves selecting the corresponding processing method based on the operation status label at the current sampling time to generate the fused feature vector after drift decoupling. The corresponding processing method is selected based on the operation status label at the current sampling time as indicated in step S100. When the operation status label is in the operation transient period or the post-operation decay period, the following processing is performed:

[0099] Step S501: Using the transient baseline parameter set after drift compensation generated in step S400, calculate the parameters for each channel based on the current operation type identifier and the time elapsed since the operation. ( , (Total number of channels in the sensor array) at the current moment Transient baseline prediction It should be noted that the transient baseline prediction value... It is generated according to the double exponential transient response curve function, specifically as follows:

[0100] ;

[0101] in, For the first The static baseline concentration value of the channel prior to the occurrence of the operational event. The first generated in step S400 Mean of peak concentration increment after channel compensation The timestamp of the current operation event. For the first The rise time constant of the channel, For the first The decay time constant of the channel, This refers to the time elapsed since the operation. This represents an exponential function.

[0102] Step S502, calculate each channel actual concentration value Compared with transient baseline predictions residual signals between ,in For the first Channel at time The measured concentration value, This is the corresponding transient baseline prediction value.

[0103] Step S503, using the channel baseline drift factor vector Inverse drift-shift correction is applied to the residual signal of each channel. For each channel... According to the channel baseline drift factor Determine the residual domain offset of this channel caused by sensor baseline translation. ,in For the first The channel's static baseline concentration value before the operational event occurred, and from the residual signal Subtract residual domain offset The corrected residuals were obtained .

[0104] Step S504, will The corrected residuals of each channel form a residual vector. ,in The first To the The corrected residuals of each channel are input into the component-aware attention module to generate the fused residual feature vector after drift decoupling. .

[0105] It should be noted that the structure and operation process of the component-aware attention module are as follows: the current inter-channel coupling feature matrix is... ( (Wei) Flatten by row A 3D vector is generated by projection through a linear mapping layer. Dimensional query vector ,in For the query vector dimension. Corrected residuals of each channel Generate each through its own linear mapping layer Key vectors and Value vectors ,in The first To the The key vector corresponding to each channel The first To the Each channel corresponds to a value vector, and each key vector has a dimension of 1. The dimension is chosen to ensure compatibility with the dimension of the dot product operation of the query vector; [the following is used] Head attention computation, where The number of attention heads, and the query vector in each attention head. With each channel key vector Calculate the scaling dot product attention weights within the corresponding subspace, and apply these weights to the value vectors of each channel. Perform a weighted summation; The outputs of each attention point are concatenated and then processed through a linear mapping layer to generate a fused residual feature vector. The component-aware attention module uses the current inter-channel coupling feature matrix as the attention query condition, enabling the fusion weights to adaptively adjust based on the current state of inter-channel cross-sensitivity. This reduces the contribution of severely drifting channels to the fusion features and suppresses the contamination of the fusion feature distribution structure by inter-channel coupling drift. The component-aware attention module uses labeled normal operating condition frames and fault leakage frames as training data. The output layer is a fully connected classification layer, mapping the fusion residual feature vector to the probability distribution of each abnormal mode category. The cross-entropy loss between this probability distribution and the true label is used as the loss function, and the Adam optimization algorithm is employed for end-to-end training.

[0106] In at least one example of the embodiments of this disclosure, the component-aware attention module employs Head attention. When the operation state label is the pre-operation static period, drift correction is directly applied to the channel-level temporal feature embedding vectors generated in step S200: using the channel baseline drift factor vector. and symbol matrix The temporal feature embedding vectors of each channel are subjected to inverse drift and translation, specifically for each channel. Temporal feature embedding vector implement ,in Symbol matrix No. The corresponding row dimensional symbol vector, This represents element-wise multiplication. for One-dimensional vector The output dimension of the hidden layer of the multi-level gated loop unit; Corrected temporal feature embedding vectors for each channel These are used as inputs for the key vector and value vector, respectively. The first To the The corrected temporal feature embedding vectors of each channel are fused through a component-aware attention module with the same structure as the transient phase of the above operation to generate a static phase fused feature vector. At this point, the query vector is... Similarly, based on the current inter-channel coupling feature matrix The data is generated by flattening the rows and then projecting it through a linear mapping layer.

[0107] It should be noted that in the static processing path, Corrected temporal feature embedding vectors for each channel Each for themselves The dimensional vectors, each channel's temporal feature embedding vector, are processed through their respective linear mapping layers to generate corresponding vectors. dimensional key vectors and value vectors, where each key vector has a dimension of 1. Dimension, and query vector Consistent dimensions ensure dimensionality compatibility in dot product operations; subsequently... Head attention operation, multi-head output concatenation, and linear mapping layer generate static fusion feature vectors. The process is exactly the same as the transient path of operation.

[0108] Step S600 involves inputting the fused feature vector into a multilayer spiking neural network with dynamic threshold modulation for anomaly pattern recognition, generating a confidence distribution vector. The fused residual feature vector generated in step S500 is then used... (When currently in the transient or post-operation decay period) or static period fusion feature vector (When currently in the pre-operation static period) Input a multilayer spiking neural network (SNN) for abnormal pattern recognition.

[0109] It should be noted that multilayer spiking neural networks include Layer of spiking neurons, in which The number of spiking neuron layers is defined, and the membrane potential of neurons in each layer iteratively evolves over time steps according to leak-integral firing (LIF) neuronal dynamics. The fused feature vector is within a preset range. Within a time step, it is encoded into an input pulse sequence, where This represents the total number of time steps for pulse coding, driving the accumulation of membrane potentials in each layer of neurons step by step. The firing threshold of each layer of neurons is dynamically modulated based on the operation state label at the current sampling time: when the operation state label is in the operation transient period, the firing threshold of each layer of neurons is adjusted from the standard threshold. According to the preset increase ratio Adjust to ,in The standard issuance threshold, To increase the issuance threshold ratio and suppress the interference of low-amplitude response caused by transient baseline compensation residue on anomaly identification; when the operation status label is in the pre-operation static period or the post-operation decay period, the issuance threshold of each layer is restored to the standard threshold. To maintain sensitivity to weak, progressive leakage signals, the threshold dynamic modulation enables the multilayer spiking neural network (MLN) to adaptively switch sensitivity between the operational transient and static periods, suppressing false alarms during operational transients while retaining the ability to detect progressive leakage. The MNN uses a labeled sequence of fused feature vectors as training data, employs the cross-entropy loss between the class probabilities mapped from the firing rate output at each time step through a linear readout layer and the true labels as the loss function, approximates the non-differentiable firing function using an alternative gradient method, and updates the parameters using the Adam optimization algorithm.

[0110] Top spiking neurons in The issuance sequence generated within a time step is input into the reserve pool computation layer, and then processed by fixed random connections in the reserve pool. Nonlinear dynamic mapping of computing nodes, generating A high-dimensional state vector, where Calculate the number of nodes for the reserve pool. The high-dimensional state vector is processed by a linear readout layer to generate the confidence distribution vector of each candidate hazardous gas anomaly pattern category. Confidence distribution vector Each element corresponds to the probability value of each candidate anomaly pattern category. Confidence distribution vector. The argmax operation yields the anomaly pattern category identifier corresponding to the highest confidence level. This category identifier, along with the probability value of the corresponding element, constitutes the input for triggering the judgment in the subsequent step S700. For example, in one example, a multilayer spiking neural network includes... Layered spiking neurons, time step During the operational transient period, the issuance threshold increase ratio Number of nodes for calculating the reserve pool .

[0111] Step S700 involves inputting the drift-corrected causal model for counterfactual inference when the confidence level exceeds the trigger threshold, thereby obtaining the fault causal effect vector and the operational causal effect vector. When the confidence distribution vector output in step S600... The probability value of any hazardous gas leak category exceeds the preset trigger threshold. At that time, the current sensor observations and operating status labels are input into the drift-corrected directed acyclic graph (DAG) causal model for counterfactual reasoning, where... This is the preset trigger threshold.

[0112] It should be noted that the state transition directed acyclic graph causal model includes the following four types of nodes and their directed causal connections: operation-driven nodes. Gas generation node Sensor observation nodes and faulty / leaking nodes ,in The first To the Sensor observations from each channel, This represents the total number of channels in the sensor array. Operation driver node. Characterizing the arc decomposition gas generation process triggered by the switching operation, the operation-driven node The value is determined by the current operation type identifier and the operation status label. (Fault / Leakage Node) The process of continuous hazardous gas leakage caused by equipment insulation defects is a latent variable to be inferred. Gas generation node. Simultaneously receive from the operation driver node and faulty / leaking nodes Directed causal input, gas generation node The structural equations linearly superimpose the contributions of the two causal paths, defining the generation relationships of each gas component under the operational response path and the fault leakage path. Sensor observation node. Receive gas generation node The output is used to generate the observation values ​​for each channel via a sensor response mapping function. The above four types of nodes constitute a directed acyclic graph structure: , , .

[0113] The specific method of drift correction is: gas generation node The structural equation coefficients of the operational response path are obtained by using the mean vector of peak concentration increments after multi-channel coupling correction generated in step S400. Reparameterization is performed to match the gas generation prediction of the operational response path with the actual response characteristics of the current sensor; sensor observation nodes The mapping function utilizes the cross-coupling drift factor matrix inverse matrix Correction is performed to map the observed values ​​from the drifted sensor response space back to the true concentration space during the calibration stage, in order to eliminate the influence of inter-channel cross-sensitivity drift on the inversion process from observed values ​​to true concentration.

[0114] The counterfactual reasoning process was performed under two different intervention conditions:

[0115] At the fault leak node Under the intervention condition of zeroing out, i.e. assuming no fault leakage, the counterfactual prediction value is calculated based on the drift-corrected state-switching directed acyclic graph causal model. , to the actual observed values Counterfactual predictions The difference is used as the fault causal effect vector. Fault causal effect vector Each element represents the concentration change component caused by fault leakage in each channel.

[0116] In operating the driver node Under the intervention condition of zeroing out, i.e. assuming no operational event occurs, counterfactual predictions are calculated based on the drift-corrected directed acyclic graph causal model of state transition. , to the actual observed values Counterfactual predictions The difference is used as the operational causal effect vector. , Operation causal effect vector Each element represents the concentration change component in each channel caused by the operational event.

[0117] For example, in one example, the trigger threshold Set to 0.8, intervention operations in counterfactual reasoning are implemented through do-calculus.

[0118] Step S800 involves performing a significance test on the fault causal effect vector and the operational causal effect vector, and generating a warning signal based on the test results through joint determination. The fault causal effect vector obtained in step S700 is then analyzed separately. and the operation of causal effect vector A significance test is performed. The significance test uses the Bootstrap resampling method, which takes the causal effect vectors as input and outputs the result determining whether each effect reaches the statistical significance level.

[0119] The joint determination method based on the significance test results is as follows:

[0120] When only the causal effect vector of the fault Upon passing the significance test, the current gas anomaly is determined to be caused by an equipment malfunction and leak, generating a fault leak warning signal that includes the gas type and estimated concentration level. The gas type is determined by the fault causality vector. The gas component corresponding to the channel with a significant positive value is determined, and the estimated concentration level is determined by the amplitude of the effect component of the corresponding channel.

[0121] When the causal effect vector of the fault and the operation of causal effect vector If both pass the significance test, the current gas anomaly is determined to contain both a fault leakage component and an operational response component. This is based on the fault causality vector. and the operation of causal effect vector of The ratio of norms determines the proportion of fault components. and operation component proportion ,in express Norm, The percentage of fault components. The proportion of operational components, and the fault causal effect vector. and the operation of causal effect vector All dimensional vector, This represents the total number of channels in the sensor array; both have the same dimension. The generation includes the proportion of fault components. and operation component proportion A composite early warning signal.

[0122] When only the causal effect vector is operated When the significance test is passed, it is determined that the current gas change is completely caused by normal operation. The transient characteristics of this operation are recorded and the transient baseline parameter set after drift compensation in step S400 is updated using the recorded transient characteristics of the operation, so that the subsequent transient baseline prediction can further adapt to the gradual characteristics of the sensor.

[0123] In at least one example of the embodiments of this disclosure, the confidence interval for the significance test is set to 95%.

[0124] This embodiment also discloses a power cabinet hazardous gas monitoring and early warning system based on deep learning and pattern recognition, including a data acquisition and alignment unit, a feature encoding unit, a drift decomposition unit, a drift compensation unit, a residual calculation and fusion unit, an abnormal pattern recognition unit, a causal reasoning unit, and a joint judgment and early warning unit.

[0125] The data acquisition and alignment unit is configured to acquire the gas concentration time series and temperature and humidity auxiliary data of each channel of the sensor array in the power cabinet, acquire the type identifier and precise timestamp of each operation event in the switchgear operation event log, align the operation event timestamps with the concentration time series, and label the operation status at each sampling moment. The feature encoding unit is configured to independently encode the concentration sequence of each channel using a multi-layer gated recurrent unit to generate channel-level temporal feature embedding vectors, and encode the inter-channel coupling feature matrix based on the sliding window cross-correlation coefficient sequence through a one-dimensional convolutional layer. The drift decomposition unit is configured to acquire the most recent continuous... The system calculates the offset matrix and ratio matrix for each normal operating frame using the channel-level feature embedding vector and inter-channel coupling feature matrix, and obtains the channel baseline drift factor vector and cross-coupling drift factor matrix through joint nonnegative matrix factorization. The drift compensation unit is configured to update the transient baseline parameter set for each operation type using the channel baseline drift factor vector and cross-coupling drift factor matrix, generating a drift-compensated transient baseline parameter set. The residual calculation and fusion unit is configured to select the corresponding processing method based on the operation state label at the current sampling time, calculate the residual signal during the operation transient or decay period, and apply inverse drift correction and component-aware attention fusion to generate a fused residual feature vector. During the static period, it directly applies drift correction and attention fusion to the channel-level temporal feature embedding vector to generate a static period fused feature vector. The abnormal pattern recognition unit is configured to input the fused feature vector into a multi-layer spiking neural network with dynamic modulation of the firing threshold, and generate a confidence distribution vector for each candidate abnormal pattern category through reservoir mapping and linear readout. The causal inference unit is configured to input the current sensor observations and operational status labels into a drift-corrected directed acyclic graph causal model when the probability value of any hazardous gas leakage category in the confidence distribution vector exceeds the trigger threshold. It then performs counterfactual inference under two intervention conditions: zeroing the fault leakage node and zeroing the operational driving node, to obtain the fault causal effect vector and the operational causal effect vector. The joint judgment and early warning unit is configured to perform significance tests on the fault causal effect vector and the operational causal effect vector respectively, and perform a joint judgment based on the test results, generating corresponding fault leakage early warning signals, composite early warning signals, or operational transient feature records and parameter update instructions.

[0126] Example 2

[0127] To verify the performance of the gas monitoring and early warning method described in Example 1, Example 2 provides specific experimental verification data. Example 2 adopts the method steps and parameter settings from Example 1, such as... Figures 3 to 9 As shown.

[0128] For step S100, the gas concentration time series and operation event logs of each channel of the sensor array in the power cabinet are obtained, and time alignment and operation status labeling are performed. (This is from a 220kV substation.) During the operation of the gas-insulated switchgear (No. GIS-07) in Year A, its built-in 5-channel sensor array continuously monitored... , HF The sampling frequency was 1Hz, and five gas components (including CO) were sampled. On March 14, 20XX, at 08:47:23, the protection control device reported a circuit breaker closing event via the IEC 61850 protocol, with a timestamp accuracy in the millisecond range. The data acquisition and alignment unit time-aligned this event timestamp with the concentration time series of the five channels and labeled each sampling moment with an operation status label based on the alignment result. The sampling moments from 08:47:00 to 08:47:22 were labeled as the pre-operation static period, the sampling moments from 08:47:23 to 08:49:53 (a total of 150 sampling points) were labeled as the operation transient period, and the post-operation decay period began from 08:49:54. As shown in Table 1, Table 1 shows the original sampling data of the sensor array and the operation status labels before and after the circuit breaker closing event.

[0129]

[0130] Table 1

[0131] For step S200, a gated cyclic unit is used to independently encode the concentration sequences of each channel to generate channel-level time-series feature embedding vectors, and a channel coupling feature matrix is ​​generated based on sliding window cross-correlation calculation. For the continuous observation data of GIS-07 cabinet after entering the operational transient period at 08:47:23, the feature encoding unit inputs the standardized concentration time series of the five channels into two layers of gated cyclic units for independent encoding, generating a 64-dimensional time-series feature embedding vector for each channel. Simultaneously, using 30 sampling points as a sliding window, the cross-correlation coefficient sequence for all channel pairs is calculated, totaling [number missing]. The cross-correlation coefficient sequence is encoded by a one-dimensional convolutional layer with a kernel size of 3 to generate... Inter-channel coupling feature matrix Due to the degradation of sensor material properties, Channel (Channel 1) and The coupling coefficient between channels (channel 4) has increased asymmetrically compared to the calibration stage, reflecting the cross-sensitivity drift characteristics between channels. Table 2 shows the coupling characteristic matrix between channels at the current time. With calibration stage matrix contrast.

[0132]

[0133] Table 2

[0134] For step S300, the offset matrix and ratio matrix are calculated based on the characteristics of normal operating conditions during the static period, and the channel baseline drift factor vector and cross-coupling drift factor matrix are obtained through joint nonnegative matrix decomposition. The drift decomposition unit extracts the most recent continuous drift factor of the GIS-07 cabinet. For each normal operating frame labeled as the pre-operation static period, the centroid of the 64-dimensional temporal feature embedding vector of each channel is calculated, and the difference between this centroid and the centroid of the initial calibration stage of the system in year A is obtained to obtain the offset vectors of the 5 channels, which are then used to form the offset matrix. ( (Dimension). Simultaneously, the mean of the inter-channel coupling feature matrix for 10 frames is calculated. With calibration stage matrix The element-wise ratios are used to obtain the ratio matrix. ( dimension).

[0135] For the offset matrix Performing element-wise absolute value yields Simultaneously record the symbol matrix ;Will and A joint nonnegative matrix is ​​formed by splicing along the column directions. ( (Wei); to Execution order The nonnegative matrix decomposition yields the nonnegative basis matrix. ( (dimensional) and non-negative coefficient matrix ( (Dimension). Extraction The first column yields the 5-dimensional channel baseline drift factor vector. .from Extract the corresponding part The column interval is obtained ( (Dimension), and strip the diagonal contribution components according to the following formula:

[0136] ;

[0137] in for A dimensional matrix, whose dimensional matrix is ​​the first dimensional matrix. All elements in the row are equal to , This represents element-wise division, from which the cross-coupling drift factor matrix is ​​obtained. ( (Dimension). As shown in Table 3, Table 3 is the channel baseline drift factor vector. And the drift characteristics of each channel.

[0138]

[0139] Table 3

[0140] For step S400, the transient baseline parameter set for each operation type is updated using a drift factor for drift compensation. The drift compensation unit, for the circuit breaker closing operation type, extracts the average peak concentration increment of each channel from the transient baseline parameter set recorded during the calibration phase of year A. Apply an inverse sensitivity scaling transform to channel 1 ( For example:

[0141] ;

[0142] Rise time constant of each channel and decay time constant Keep it unchanged. For multi-channel coupling correction, the mean vector of the 5-channel peak concentration increments will be used. Apply the inverse transformation of the cross-coupling drift factor matrix: The mean vector of peak concentration increments after multi-channel coupling correction is obtained. This is used for reparameterization of the causal model in subsequent step S700. As shown in Table 4, Table 4 compares the transient baseline parameters before and after compensation for the circuit breaker closing operation type.

[0143]

[0144] Table 4

[0145] For step S500, the corresponding processing method is selected based on the operation status label at the current sampling time, generating the fused feature vector after drift decoupling. The operation status label at the current sampling time 08:47:35 is the operation transient period, and the residual calculation and fusion unit executes the transient period processing path. Taking channel 1 ( For example, the static baseline concentration value before the operation event occurred. The operation has been completed after a certain period of time. Substituting into the double exponential transient response curve function, the peak increment after compensation is... Rise time constant decay time constant :Right now

[0146] ;

[0147] Measured concentration , the residual signal Further applying inverse drift translation correction, the residual domain offset of channel 1... Corrected residuals The corrected residuals from the five channels form a residual vector. ,enter The head group perceptual attention module uses the current inter-channel coupling feature matrix. After flattening, a query vector is generated through a linear mapping layer. After correction, the residuals of each channel are passed through a linear mapping layer to generate corresponding key vectors and value vectors. These vectors are then weighted and fused using a 4-head attention algorithm to generate a fused residual feature vector. As shown in Table 5, Table 5 shows the transient baseline prediction values, residual signals, and corrected residuals for each channel at 08:47:35.

[0148]

[0149] Table 5

[0150] In step S600, the multilayer spiking neural network with fused feature vectors and dynamic threshold modulation is used for anomaly pattern recognition to generate a confidence distribution vector. The anomaly pattern recognition unit fuses the residual feature vectors. enter Layered spiking neural networks, in Encodes the input pulse sequence within a time step. Since the current operational state label represents the operational transient period, the firing threshold of neurons in each layer is increased proportionally. Standard threshold Adjust to This is to suppress the interference of low-amplitude responses caused by transient baseline compensation residues on anomaly detection. The firing sequence input generated by the top-level spiking neurons over 20 time steps... The node's reserve pool computation layer generates a 500-dimensional high-dimensional state vector through a fixed-random connection nonlinear dynamic mapping, and then generates a confidence distribution vector through a linear readout layer. As shown in Table 6, Table 6 is the confidence distribution vector of the multilayer spiking neural network output.

[0151]

[0152] Table 6

[0153] In the confidence distribution vector, " The probability value for the "insulation defect leakage" category is 0.87, which exceeds the trigger threshold. This triggers the counterfactual reasoning process in step S700.

[0154] For step S700, when the confidence level exceeds the trigger threshold, the drift-corrected causal model is input for counterfactual inference to obtain the fault causal effect vector and the operational causal effect vector. The causal inference unit uses the 5-channel actual observations at 08:47:35. A directed acyclic graph causal model of state transitions, with input state labels (operational transient periods) and drift correction. Gas generation nodes in the model. The operational response path structure equation coefficients have been generated using step S400. Complete reparameterization, sensor observation node The mapping function has been utilized Correction complete.

[0155] At the fault leak node Under the intervention condition of zeroing out (i.e., assuming no fault leakage), the counterfactual prediction value is obtained through do-calculus intervention. Calculate the causal effect vector of the fault ; in the operation of the driving node Under the intervention condition of zero (i.e., assuming no operational event occurred), the counterfactual prediction value is obtained. Calculate the causal effect vector of the operation As shown in Table 7, Table 7 presents the counterfactual inference results of the effect vectors under the two intervention conditions.

[0156]

[0157] Table 7

[0158] For step S800, a significance test is performed on the fault causal effect vector and the operation causal effect vector, and a joint determination is executed based on the test results to generate an early warning signal. The joint determination early warning unit performs a significance test on the fault causal effect vector. and the operation of causal effect vector Significance tests for Bootstrap resampling were performed, with confidence intervals set at 95%. The test results showed that... and All passed the significance test, determining that the current gas anomaly simultaneously contains both a fault leakage component and an operational response component, and then executed the composite early warning signal generation process.

[0159] in accordance with and of Norm calculation of the proportion of each component, where , :

[0160]

[0161]

[0162] A composite early warning signal was generated for the GIS-07 cabinet, the content of which was: a detection was superimposed during the transient period of the circuit breaker closing operation. Insulation defects and leakage signs, percentage of fault components Operational component proportion The significant fault leakage channel is channel 1 ( (effect amplitude 12.39) and channel 4 ( The effect amplitude was 8.66, corresponding to a medium-level warning for gas concentration. It is recommended that maintenance personnel conduct further testing on the GIS-07 cabinet.

[0163] To implement the above embodiments, this application also discloses an electronic device. The electronic device may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) or a program loaded from a storage device into a random access memory (RAM). Various programs and data required for the operation of the electronic device are also stored in the RAM. The processing unit, ROM, and RAM are interconnected via a bus. An input / output (I / O) interface is also connected to the bus. Typically, the following devices can be connected to the I / O interface: input devices including, for example, a touchscreen, touchpad, keyboard, mouse, camera, microphone, accelerometer, gyroscope, etc.; output devices including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices including, for example, magnetic tape, hard disk, etc.; and communication devices. The communication device allows the electronic device to communicate wirelessly or wiredly with other devices to exchange data. Although electronic devices with various devices are shown, it should be understood that it is not required to implement or possess all of the shown devices. More or fewer devices may be implemented or possessed alternatively.

[0164] In particular, according to some embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, some embodiments of this disclosure include a computer program product comprising a computer program carried on a computer storage medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from a ROM. When the computer program is executed by a processing device, it performs the functions defined above in the methods of some embodiments of this disclosure.

[0165] It should be noted that the computer storage medium in some embodiments of this disclosure may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0166] In some embodiments of this disclosure, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device. In some embodiments of this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer storage medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer storage medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0167] In other implementations, clients and servers may communicate using any currently known or future-developed network protocol such as HTTP (Hypertext Transfer Protocol), and may interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet of Things), and end-to-end networks (e.g., ad hoc end-to-end networks), as well as any currently known or future-developed networks.

[0168] The aforementioned computer storage medium may be included within the aforementioned electronic device, or it may exist independently and not assembled into the electronic device. The aforementioned computer storage medium carries one or more programs, which, when executed by the electronic device, enable the electronic device to implement a multimodal large-model adversarial sample cleanup method based on a supervised diffusion model.

[0169] Computer program code for performing operations of some embodiments of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0170] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. Each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions. Units described in some embodiments of the present disclosure may be implemented in software or hardware. The described units may also be located in a processor, and the names of these units do not necessarily constitute a limitation on the unit itself.

[0171] The functions described above in this document can be performed, at least in part, by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: Field Programmable Gate Arrays (FPGAs), Application-Specific Integrated Circuits (ASICs), Application Standard Products (ASSPs), System-on-Chip (SoCs), Complex Programmable Logic Devices (CPLDs), and so on.

Claims

1. A method for monitoring and early warning of harmful gases in power cabinets using deep learning and pattern recognition, characterized in that, Includes the following steps: Step S100: Obtain the gas concentration time series and operation event log of each channel of the sensor array in the power cabinet, and perform time alignment and operation status labeling. Step S200: The concentration sequences of each channel are independently encoded by a gated recurrent unit to generate channel-level time-series feature embedding vectors, and the inter-channel coupling feature matrix is ​​generated based on the sliding window cross-correlation. Step S300: Calculate the offset matrix and ratio matrix based on the characteristics of normal operating conditions during the static period, and obtain the channel baseline drift factor vector and cross-coupling drift factor matrix through joint non-negative matrix decomposition. Step S400: Update the transient baseline parameter set by drift compensation using the channel baseline drift factor vector and the cross-coupling drift factor matrix; Step S500: Select the corresponding processing method based on the operation status label at the current sampling time to generate the fused feature vector after drift decoupling; Step S600: Input the fused feature vector into the multilayer spiking neural network with dynamic modulation of the firing threshold to generate a confidence distribution vector; Step S700: When the probability value of any leakage category in the confidence distribution vector exceeds the trigger threshold, input the drift-corrected causal model for counterfactual reasoning. Step S800: Obtain the fault causal effect vector and the operation causal effect vector, and perform a joint judgment based on the significance test to generate an early warning signal; The joint nonnegative matrix decomposition decomposes the overall sensor distribution offset into two orthogonal subspace components: channel-independent baseline drift and inter-channel cross-sensitivity coupling drift.

2. The method for monitoring and early warning of harmful gases in power cabinets using deep learning and pattern recognition as described in claim 1, characterized in that, Step S300, which obtains the channel baseline drift factor vector and the cross-coupling drift factor matrix through joint nonnegative matrix decomposition, includes: Step S301: The offset vectors of the centroids of each channel feature embedding vector relative to the centroids of the calibration stage are used to form an offset matrix. For the offset matrix Performing element-wise absolute value operations yields the non-negative offset matrix. And record the sign matrix of each element. ; Step S302: Calculate the element-wise ratio matrix of the mean of the inter-channel coupling feature matrix of the normal operating condition frame to the inter-channel coupling feature matrix of the calibration stage. ; Step S303, the non-negative offset matrix With the ratio matrix A joint nonnegative matrix is ​​formed by splicing along the column directions. For the joint nonnegative matrix Execution rank is Nonnegative matrix decomposition; Step S304: Extract the non-negative basis matrix. The first column serves as the channel baseline drift factor vector. ; Step S305, using the channel baseline drift factor vector For the ratio matrix The diagonal contribution components caused by independent channel drift are stripped to obtain the cross-coupling drift factor matrix.

3. The method for monitoring and early warning of harmful gases in power cabinets using deep learning and pattern recognition as described in claim 1, characterized in that, Step S400 involves updating the transient baseline parameter set using the channel baseline drift factor vector and the cross-coupling drift factor matrix, including: For each channel mean peak concentration increment The mean peak concentration increment after compensation is obtained by applying an inverse sensitivity scaling transform. ; The mean vector of multi-channel peak concentration increments for each operation type Applying the inverse transformation correction to the cross-coupling drift factor matrix yields the corrected mean vector of peak concentration increments. ; Among them, the average value of the peak concentration increment after compensation The corrected peak concentration increment mean vector is used for channel-by-channel calculation of transient baseline prediction values ​​for each channel. Used for reparameterization of the operational response path structure equation in causal models.

4. The method for monitoring and early warning of harmful gases in power cabinets using deep learning and pattern recognition as described in claim 1, characterized in that, In step S500, a fused feature vector after drift decoupling is generated based on the operation state label at the current sampling time. When the operation state label is the operation transient period or the post-operation decay period: Step S501: Calculate the predicted transient baseline value for each channel at the current time using the drift-compensated transient baseline parameter set according to the double exponential transient response curve function. ; Step S502: Calculate the residual signal between the measured concentration values ​​of each channel and the predicted transient baseline values. ; Step S503, using the channel baseline drift factor vector The corrected residuals are obtained by applying inverse drift-translation correction to the residual signals of each channel. ; Step S504: The fused residual feature vector is generated by the component-sensing attention module based on the corrected residuals of each channel. .

5. The method for monitoring and early warning of hazardous gases in power cabinets using deep learning and pattern recognition as described in claim 4, characterized in that: The component-aware attention module uses the current inter-channel coupling feature matrix After being flattened row by row, the query vector is generated by projection through a linear mapping layer. After correction, the residuals of each channel are passed through a linear mapping layer to generate key vectors and value vectors. A multi-head scaling dot product attention operation is then used to weight and sum the value vectors of each channel. The multi-head outputs are then concatenated with the linear mapping layer to generate the fused residual feature vector. ; When the operation status label is in the pre-operation static period, the channel baseline drift factor vector is used. and symbol matrix Embedded vectors of time-series features for each channel Performing inverse drift translation yields the corrected temporal feature embedding vector. ,Will The inputs, serving as both key and value vectors, are processed by the component-aware attention module with the same structure to generate a static fusion feature vector. .

6. The method for monitoring and early warning of harmful gases in power cabinets using deep learning and pattern recognition according to claim 1, characterized in that, The firing threshold of each neuron in the multilayer spiking neural network is dynamically modulated according to the operation state label: When the operation state label is in the operation transient period, the firing threshold of neurons in each layer is changed from the standard threshold. Adjust to ; When the operation status label is in the pre-operation static period or the post-operation decay period, the issuance threshold of each layer is restored to the standard threshold. ; In this process, the firing sequence of the top-level spiking neurons is input into the reserve pool computation layer, and a high-dimensional state vector is generated through nonlinear dynamic mapping via fixed-random connection computation nodes. This high-dimensional state vector is then used by a linear readout layer to generate the confidence distribution vector. .

7. The method for monitoring and early warning of harmful gases in power cabinets using deep learning and pattern recognition as described in claim 1, characterized in that, The drift-corrected causal model in S600 is a state-switching directed acyclic graph causal model, which includes operation-driven nodes. Gas generation node Sensor observation nodes and faulty / leaking nodes The gas generation node The structural equation coefficients of the operational response path are reparameterized using the mean vector of peak concentration increments after multi-channel coupling correction, and the sensor observation node... The mapping function is corrected using the inverse of the cross-coupling drift factor matrix, and the counterfactual reasoning includes: At the fault leakage node Under the intervention condition of zeroing out, the counterfactual predicted value is calculated, and the difference between the actual observed value and the counterfactual predicted value is used as the failure causality vector. ; In the operation driving node The counterfactual predicted value is calculated under the intervention condition of zeroing, and the difference between the actual observed value and the counterfactual predicted value is used as the operational causal effect vector. .

8. The method for monitoring and early warning of hazardous gases in power cabinets using deep learning and pattern recognition according to claim 1, characterized in that, Step S800, which involves generating an early warning signal based on a joint determination using a significance test, includes: When only the fault causal effect vector is mentioned Upon passing the significance test, a fault leak warning signal containing the gas type and estimated concentration level is generated; When the fault causal effect vector and the causal effect vector of the operation When both pass the significance test, according to the two Norm ratio determines the proportion of fault components and operation component proportion Generate a composite early warning signal that includes the proportion of the fault component and the proportion of the operation component; When only the causal effect vector of the operation is mentioned When the significance test is passed, it is determined that the gas change is caused by normal operation, the transient characteristics of the operation are recorded, and the transient baseline parameter set is updated.

9. A deep learning and pattern recognition-based system for monitoring and warning of hazardous gases in power cabinets, used to execute the method for monitoring and warning of hazardous gases in power cabinets according to any one of claims 1 to 8, characterized in that, include: The data acquisition and alignment unit is configured to acquire the gas concentration time series and operation event logs of each channel of the sensor array in the power cabinet, and perform time alignment and operation status labeling. The feature encoding unit is configured to independently encode the concentration sequence of each channel through a gated recurrent unit to generate channel-level temporal feature embedding vectors, and to generate inter-channel coupling feature matrices based on sliding window cross-correlation calculation; The drift decomposition unit is configured to calculate the offset matrix and ratio matrix based on the characteristics of normal operating conditions during the static period, and obtain the channel baseline drift factor vector and cross-coupled drift factor matrix through joint non-negative matrix decomposition. The drift compensation unit is configured to update the transient baseline parameter set by drift compensation using the channel baseline drift factor vector and the cross-coupling drift factor matrix; The residual calculation and fusion unit is configured to generate a drift-decoupled fusion feature vector based on the operation state label at the current sampling time. An abnormal pattern recognition unit is configured to input the fused feature vector into a multilayer spiking neural network with dynamic modulation of the firing threshold to generate a confidence distribution vector. The causal reasoning unit is configured to input the current sensor observation and operation status label into the drift-corrected causal model to perform counterfactual reasoning when the probability value of any leakage category in the confidence distribution vector exceeds the trigger threshold, thereby obtaining the fault causal effect vector and the operation causal effect vector. The joint judgment and early warning unit is configured to perform a significance test on the fault causal effect vector and the operation causal effect vector, and perform a joint judgment to generate an early warning signal based on the test results.

10. An electronic device, characterized in that, include: At least one memory stores computer-executable instructions non-transiently; At least one processor, configured to run the computer-executable instructions, The computer-executable instructions are executed by the processor to implement the method according to any one of claims 1 to 8.

Images