A switch cabinet insulation deterioration early warning method, device and computer program product

Abstract

This invention discloses a method, device, and computer program product for early warning of insulation degradation in switchgear. The method simultaneously acquires transient magnetic vector signals and characteristic gas concentration signals within the switchgear, and obtains magnetic and gas feature sequences after preprocessing. These sequences are then constructed into a high-order fusion tensor. Cross-modal correlation features are extracted using a tensor decomposition algorithm, and an attention alignment mechanism is used to adaptively compensate and fuse the magnetic and gas features in the temporal dimension, generating a cross-modal fusion feature sequence. A combined temporal prediction network is constructed. First, a first temporal feature extraction module captures local abrupt changes in the fusion feature sequence, and then a second temporal feature extraction module captures long-range temporal dependencies, outputting a future multi-time-step insulation health index prediction sequence. Based on the prediction sequence and a preset early warning threshold, a tiered early warning command is output. This invention effectively solves the problem of time-scale misalignment of magnetic-gas heterogeneous features through high-order tensor fusion, demonstrating strong engineering applicability.

Description

Technical Field

[0001] This invention relates to the field of power equipment monitoring technology, specifically to a method, device, and computer program product for early warning of insulation degradation in switchgear. Background Technology

[0002] Switchgear, as the most numerous and widely distributed core switching equipment in the power distribution network, is a key node for power distribution, control, and protection. Its insulation operation directly determines the reliability, safety, and continuity of the entire power supply system, playing an irreplaceable supporting role in the stable operation of the smart grid. Circuit breaker contacts, busbar joints, and cable joints within the switchgear are prone to localized overheating due to poor contact. Under the combined effects of electrical, thermal, and mechanical stresses, the insulation material can experience partial discharge and pyrolytic aging, ultimately leading to insulation breakdown, short-circuit tripping, or even combustion and explosion accidents.

[0003] Current monitoring technologies for latent faults in switchgear mainly fall into two categories: The first type is transient magnetic field detection technology based on tunneling magnetoresistive (TMR) sensing. Its core principle is to capture the nanosecond-level weak magnetic vector signal generated during partial discharge through TMR sensors. This technology has the advantages of fast response speed, high time resolution, and high sensitivity to early discharge defects. It can accurately capture the instantaneous characteristics of partial discharge and realize the rapid detection of early sudden discharge defects. However, it is significantly affected by the interference of complex electromagnetic environment on site, and can only reflect the instantaneous behavior of partial discharge. It cannot effectively characterize the cumulative degree of long-term thermal aging of insulation materials and cannot fully reflect the overall deterioration state of switchgear insulation system.

[0004] The second category is insulation pyrolysis monitoring technology with characteristic gas detection as its core. Its core logic is to directly reflect the degree of deterioration and pyrolysis cumulative effect of insulation materials by detecting the concentration changes of characteristic gases (such as CO, NO2, O3, etc.) generated during the pyrolysis aging process of insulation materials. This technology has the advantages of strong anti-electromagnetic interference ability and the ability to characterize the long-term cumulative state of insulation deterioration. However, it has obvious problems of gas generation lag and gas diffusion delay. It has low sensitivity and slow response speed to early sudden partial discharge defects and cannot capture short-term sudden fault characteristics in time.

[0005] In practical engineering applications, existing switchgear fault early warning technologies still have many core bottlenecks, making it difficult to meet the core requirements of smart grids for predictive maintenance of equipment conditions. These bottlenecks are manifested in the following four aspects: First, the fusion depth of multi-source heterogeneous features is insufficient, and the misalignment between magnetic and gas signals across time scales is prominent. Transient magnetic field signals are fast variables in the nanosecond to millisecond range, mainly reflecting the instantaneous burst behavior of partial discharge; characteristic gas signals are slow variables in the hour to day range, mainly reflecting the long-term cumulative effect of thermal aging of insulating materials. Existing multi-source feature fusion methods mostly remain at the level of simple feature splicing or weighted voting at the decision level, without deeply exploring the inherent physical relationship and dynamic evolution law between "partial discharge-insulation degradation-thermal gas generation". This makes it impossible to achieve adaptive alignment and deep fusion of heterogeneous features with two different time scales, resulting in weak fault characterization capabilities of the fused features and difficulty in improving the robustness and accuracy of fault diagnosis.

[0006] Secondly, long-cycle time-series prediction models have inherent limitations, failing to simultaneously address both short-term abrupt changes and long-term evolutionary characteristics. Switchgear insulation degradation is a long-cycle process, with condition monitoring data often covering long-series data spanning months or even years. Existing time-series prediction models struggle to meet the processing requirements of such long-series monitoring data: Gated recurrent unit (GRU)-based recurrent neural network models are prone to gradient vanishing and loss of long-range dependency information when processing long-series data, failing to effectively capture the gradual, long-cycle evolution of insulation degradation; while Transformer-type attention models can address long-range dependency issues to some extent, they suffer from high computational complexity, large inference delays, and difficulty in edge deployment. Furthermore, they lack the ability to capture short-term abrupt changes caused by partial discharge, making it impossible to achieve coordinated monitoring of short-term sudden faults and long-term degradation processes.

[0007] Third, the early warning mode is lagging behind and cannot achieve early warning of the evolution of latent faults. Existing switchgear fault early warning technologies mostly adopt a "threshold-based" post-event alarm mode, that is, an alarm signal is only issued when the fault characteristic parameters exceed a preset fixed threshold. This mode can only achieve passive alarm after the fault occurs, and cannot make multi-step advanced predictions of the future evolution trajectory of the switchgear insulation state. It is difficult to achieve early warning and proactive intervention for early latent faults, which is significantly different from the core requirement of "predictive maintenance" of smart grids.

[0008] Fourth, the model has poor generalization ability and field adaptability. Existing fault early warning models are mostly trained based on simulated data under ideal laboratory conditions, without incorporating the physical mechanism of switchgear insulation degradation. The distribution of model training data and actual field operation data differs significantly. Furthermore, the field operating environment is subject to complex interference factors such as fluctuations in ambient temperature and humidity, load changes, and strong electromagnetic interference. This results in the existing data-driven models having weak generalization ability under complex field conditions. The model accuracy decreases rapidly during long-term operation, failing to meet the needs of long-term stable monitoring in the field. Summary of the Invention

[0009] The technical problem to be solved by the present invention is to provide a method, device and computer program product for early warning of insulation degradation of switchgear, so as to realize high-precision quantitative inversion and advanced dynamic early warning of insulation degradation of switchgear.

[0010] To solve the above-mentioned technical problems, the present invention provides a method for early warning of insulation degradation in switchgear, comprising: Step S1: Synchronously acquire transient magnetic vector signals and characteristic gas concentration signals in the switch cabinet, and obtain magnetic feature sequences and gas feature sequences after preprocessing; Step S2: Construct the magnetic feature sequence and the gas feature sequence into a high-order fusion tensor, extract cross-modal correlation features based on the tensor decomposition algorithm, and perform adaptive compensation and fusion of magnetic and gas features in the temporal dimension through the attention alignment mechanism to generate a cross-modal fusion feature sequence. Step S3: Construct a combined temporal prediction network. First, capture the local mutation features of the fused feature sequence through the first temporal feature extraction module, and then capture the long-range temporal dependency through the second temporal feature extraction module. Based on the captured features, output the insulation health index prediction sequence for multiple future time steps. Step S4: Based on the predicted sequence and the preset warning threshold, output a graded warning instruction.

[0011] Preferably, step S1 specifically includes: The transient magnetic vector signal and the characteristic gas concentration signal are acquired by deploying a three-axis tunneling magnetoresistive (TMR) magnetic sensing array and a multi-component electrochemical gas sensing module, respectively, and the two types of signals are synchronized in time. The transient magnetic vector signal is processed by wavelet threshold denoising method. The db4 wavelet basis is selected for 5-level decomposition. The high-frequency coefficients are processed by soft threshold function to reconstruct the denoised magnetic signal. The time domain, frequency domain and time-frequency domain features are extracted to construct the magnetic feature matrix. The characteristic gas concentration signal is processed by moving average filtering and outlier removal algorithm based on the Laida criterion. Missing data is filled by linear interpolation, and trend features, rate of change features and cumulative gas production features are extracted to construct a gas feature matrix. The magnetic feature matrix and the gas feature matrix are respectively subjected to Z-Score normalization to obtain the normalized magnetic feature sequence and gas feature sequence.

[0012] Preferably, in step S2, the higher-order fusion tensor is a bimodal feature tensor obtained by reshaping the magnetic feature sequence and the gas feature sequence into third-order tensors and then concatenating them; the tensor decomposition algorithm is the Tucker decomposition algorithm based on alternating least squares, and the bimodal feature tensor is decomposed into a core tensor and three factor matrices corresponding to the temporal dimension, feature dimension, and modal dimension, respectively. The sum of the products of the moduli is used, and the optimization objective is to minimize the F-norm of the fitted residual tensor.

[0013] Preferably, in step S2, the adaptive compensation and fusion of magnetic and gas features in the temporal dimension through the attention alignment mechanism specifically includes: The feature dimension factor matrix and time sequence dimension factor matrix obtained by decomposing the magnetic feature sequence and the gas feature sequence are projected onto a unified latent feature space to obtain magnetic latent features and gas latent features, respectively. A learnable modal alignment weight matrix is ​​introduced, and an attention mechanism is used to calculate the intermodal temporal attention matrix between the magnetic latent features and the gas latent features; Based on the temporal attention matrix, the gas latent features are temporally reweighted to obtain gas latent features that are temporally aligned with the magnetic latent features; The aligned magnetic latent features and gas latent features are concatenated and combined with the core tensor obtained from decomposition. The cross-modal fusion feature sequence is generated by reconstructing the feature sequence through learnable fusion weights and biases.

[0014] Preferably, in step S3, the first temporal feature extraction module is a multi-layer causal dilated convolutional network, which is composed of multiple stacked residual blocks. Each residual block contains two layers of causal convolution and dilated convolution. The dilation factor increases exponentially with the number of network layers and forms residual connections through identity mapping. The second temporal feature extraction module is a multi-layer selective state space model. The discretization step size and state parameter matrix of this model are dynamically generated by linear projection of the current input features to selectively remember and forget temporal information.

[0015] Preferably, in step S3, the training of the combined time series prediction network adopts a joint loss function, which is composed of a weighted sum of numerical accuracy loss and trend consistency loss; the numerical accuracy loss is the mean square error between the predicted insulation health index and the actual insulation health index, and the trend consistency loss is calculated based on the first-order difference sign difference between the predicted insulation health index and the actual insulation health index.

[0016] Preferably, in step S4, the preset early warning threshold is dynamically adjusted based on the type of insulation material, service life, load conditions, and ambient temperature and humidity of the switchgear using a fuzzy inference algorithm; the graded early warning instructions include at least a first-level early warning, a second-level early warning, and a third-level early warning, with each level of early warning corresponding to different insulation health index ranges and pyrolysis conversion rate ranges, as well as corresponding operation and maintenance strategies.

[0017] Preferably, the method further includes an on-site engineering closed-loop implementation step: After model quantization and pruning, the trained high-order fusion tensor decomposition network and the combined temporal prediction network are lightweighted and deployed to the edge computing terminal to adapt to the computing power of the edge device. Regularly collect fault data confirmed by on-site inspections and supplementary experimental data from the laboratory, and use incremental learning methods to iteratively update the deployed model online.

[0018] The present invention also provides a switchgear insulation degradation early warning device, comprising: One or more processors; Memory; One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the one or more processors, and the one or more applications are configured to perform the switchgear insulation degradation early warning method.

[0019] The present invention also provides a computer program product, including computer instructions, which instruct a computer device to perform operations corresponding to the switchgear insulation degradation early warning method.

[0020] The beneficial effects of this invention are as follows: By constructing a cross-modal fusion network based on Tucker high-order tensor decomposition, this invention organically fuses high-frequency transient magnetic vector signals and low-frequency characteristic gas concentration signals into a unified high-order tensor representation. Furthermore, it introduces an intermodal temporal attention mechanism to adaptively compensate for the physical lag effect of "discharge-gas generation," fundamentally solving the problem of misalignment of magnetic-gas heterogeneous features on the time scale. This invention overcomes the technical limitations of traditional simple splicing and fusion, deeply explores the intrinsic correlation between the two physical quantities, and improves the characterization ability of the fused features for insulation degradation by more than 40%, achieving a fault diagnosis accuracy of 98.5%. Building upon this foundation, this invention constructs a combined time-series prediction network cascaded with TCN and Mamba. Utilizing the large receptive field of TCN's causal dilated convolution, it accurately captures short-term abrupt changes in insulation state caused by partial discharge. Simultaneously, leveraging the long-range dependency preservation capability of the Mamba selective state-space model, it effectively overcomes the long-range information forgetting problem of traditional recurrent neural networks. Furthermore, by combining a hardware-friendly parallel scanning algorithm, the computational complexity is reduced from O(n²) to O(n), and the inference speed is improved by more than 5 times. This achieves high-precision multi-step advance prediction of insulation health index up to 7 days in the future, with the root mean square error controlled within 1.2%. In addition, this invention deeply integrates the physical mechanism of insulation pyrolysis kinetics with a data-driven model, possessing excellent resistance to environmental interference and load fluctuations. It is adaptable to switchgear of different models and service lives. The model supports lightweight edge deployment, requiring no modification to primary equipment, posing no electrical safety risks, and meeting the dual needs of pre-installation in new switchgear and intelligent transformation of existing switchgear. It has extremely high engineering promotion value and application prospects. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating a switchgear insulation degradation early warning method according to an embodiment of the present invention.

[0023] Figure 2 This is a principle block diagram of a switchgear insulation degradation early warning method according to an embodiment of the present invention.

[0024] Figure 3 This is a comparison chart of the insulation health index prediction results of the embodiments of the present invention and the traditional LSTM model. Detailed Implementation

[0025] The following description of the embodiments is taken with reference to the accompanying drawings, which illustrate specific embodiments in which the invention can be implemented.

[0026] Please refer to Figure 1 As shown, Embodiment 1 of the present invention provides a method for early warning of insulation degradation in switchgear, comprising: Step S1: Synchronously acquire transient magnetic vector signals and characteristic gas concentration signals in the switch cabinet, and obtain magnetic feature sequences and gas feature sequences after preprocessing; Step S2: Construct the magnetic feature sequence and the gas feature sequence into a high-order fusion tensor, extract cross-modal correlation features based on the tensor decomposition algorithm, and perform adaptive compensation and fusion of magnetic and gas features in the temporal dimension through the attention alignment mechanism to generate a cross-modal fusion feature sequence. Step S3: Construct a combined temporal prediction network. First, capture the local mutation features of the fused feature sequence through the first temporal feature extraction module, and then capture the long-range temporal dependency through the second temporal feature extraction module. Based on the captured features, output the insulation health index prediction sequence for multiple future time steps. Step S4: Based on the predicted sequence and the preset warning threshold, output a graded warning instruction.

[0027] As can be seen from the above steps, the embodiments of the present invention integrate the complementary advantages of magnetic sensing and characteristic gas detection, realize the physical correlation alignment of heterogeneous features through high-order tensor fusion, and realize multi-step advance prediction of insulation degradation trajectory by combining efficient time-series networks, and finally construct a mechanism-data dual-driven early warning system for switchgear insulation state evolution.

[0028] Please combine Figure 2 As shown, step S1 of this embodiment aims to construct a multi-physics field synchronous acquisition and preprocessing system for the partial discharge and insulation pyrolysis process in switchgear, solving the heterogeneity problem of magnetic signals and gas signals in terms of sampling frequency and time scale, and realizing synchronous acquisition, noise reduction, and standardized preprocessing of transient magnetic vector signals and characteristic gas concentration signals. Specifically, step S1 includes the following steps S11 to S13: Step S11: Construct a signal synchronous acquisition system.

[0029] A set of triaxial TMR magnetic sensors is deployed in the busbar compartment, circuit breaker compartment, and cable compartment of the switchgear to form a triaxial TMR magnetic sensing array. The magnetic field resolution of this sensor reaches [value missing]. Level, bandwidth coverage The sampling frequency is set to Simultaneously acquire transient magnetic vector signals along the X, Y, and Z axes. ; At the same time, carbon monoxide (CMO) is deployed in each independent gas chamber of the switchgear. ), nitrogen dioxide ( ) and ozone ( A three-channel electrochemical gas-sensitive module. Each channel has a detection resolution of [missing information]. The sampling frequency is set to Used for synchronous acquisition of characteristic gas concentration time series .

[0030] Furthermore, by using the edge data acquisition terminal and the IEEE 1588 precision time synchronization protocol, the timestamps of the magnetic signal and the gas signal are synchronized, and the local caching and preliminary preprocessing of the raw data are completed.

[0031] Step S12: Denoising and feature extraction preprocessing are performed on the magnetic signal and the gas signal, respectively.

[0032] For magnetic signal preprocessing: Wavelet thresholding denoising is used to denoise the original magnetic vector signal. A 5-level decomposition is performed using a db4 wavelet basis, and a soft thresholding function is applied to the high-frequency coefficients to reconstruct the denoised magnetic signal. Based on the denoised signal, the time-domain, frequency-domain, and time-frequency-domain features of partial discharge are extracted to construct a magnetic feature matrix. ,in, For timing length, The magnetic characteristic dimension includes: magnetic signal peak value, average discharge quantity, discharge repetition frequency, phase distribution characteristics, spectral center frequency, wavelet packet energy entropy, etc.

[0033] The expression for the soft threshold function is:

[0034] In the formula, For the first Layer Wavelet coefficients, For general thresholds, The standard deviation of noise. This represents the number of signal sampling points.

[0035] For gas signal preprocessing: a moving average filtering and outlier removal algorithm is used to process the original gas concentration sequence, according to the Laida criterion (…). (Criteria) Outliers were removed, and missing data was filled in using linear interpolation to obtain a smoothed gas concentration sequence. The trend characteristics, rate of change characteristics, and cumulative gas production characteristics of gas concentration are extracted to construct a gas feature matrix. ,in, The gas characteristic dimension includes: real-time concentration of each gas, hourly gas production rate, 24-hour cumulative concentration, concentration change gradient, Hurst index, etc.

[0036] Step S13, Data Standardization Processing: The Z-Score standardization method is used to standardize the magnetic feature matrix. Gas characteristic matrix Standardize them separately to eliminate dimensional differences:

[0037] In the formula, The mean of the characteristic matrix is... The standard deviation of the characteristic matrix, This is the standardized feature matrix. The final standardized magnetic feature sequence is obtained. Gas characteristic sequence .

[0038] Step S2 is used to construct a cross-modal tensor fusion network based on Tucker decomposition. Magnetic and gas features are constructed as high-order tensors. Through tensor decomposition and kernel tensor interaction, adaptive alignment and deep fusion of the two heterogeneous modal features in a unified high-dimensional latent space are achieved, solving the problem of temporal scale misalignment between magnetic and gas signals. In this embodiment, the high-order fusion tensor is specifically implemented as a three-dimensional bimodal feature tensor, and the tensor decomposition algorithm is specifically implemented as Tucker decomposition. Specifically, step S2 includes the following steps S21 to S23: Step S21: Construct a higher-order fusion tensor.

[0039] First, the magnetic feature sequences and gas feature sequences are reconstructed into third-order feature tensors: the standardized magnetic feature sequences and gas feature sequences are reconstructed into third-order feature tensors, preserving the structural information of the temporal dimension, feature dimension, and modal dimension. Magnetic characteristic tensor: This corresponds to the time sequence length, magnetic feature dimension, and modal identifier; Gas characteristic tensor: This corresponds to the time series length, gas characteristic dimension, and modal identifier; To unify dimensions, take For dimensions that are insufficient, zero-padding is used to fill in the gaps. Then, the components are concatenated to obtain a bimodal feature tensor: This ensures dimensionality matching between modalities. This tensor is the high-order fusion tensor constructed in this embodiment.

[0040] Step S22: Extract cross-modal correlation features based on tensor decomposition algorithm.

[0041] This embodiment uses the Tucker decomposition algorithm to process the above-mentioned bimodal feature tensor. Perform Tucker higher-order decomposition, decomposing it into the product of a core tensor and three modality factor matrices, and extract the eigenvalues ​​of each mode:

[0042] In the formula: This is the core tensor, used to represent the interaction and correlation between different modalities. These are the ranks of the temporal dimension, feature dimension, and modal dimension, respectively; These are factor matrices representing the time-series dimension, feature dimension, and modality dimension, respectively, and the principal component features corresponding to each dimension. For tensors Modular multiplication operation; To fit the residual tensor.

[0043] The decomposition process employs alternating least squares (ALS) to minimize the F-norm of the residual tensor as the optimization objective.

[0044] The constraint is that each factor matrix is ​​a column orthogonal matrix, i.e. .

[0045] This decomposition enables the extraction of cross-modal correlation features.

[0046] Step S23 involves adaptive compensation and fusion based on an attention alignment mechanism. This step addresses the timescale misalignment between magnetic and gas signals caused by the hysteresis effect of physical processes (discharge and gas generation).

[0047] First, modal feature projection is performed to project the magnetic and gas features onto a unified latent feature space, thus obtaining the magnetic latent features. Gas hidden features The projection utilizes the factor matrix obtained from the decomposition. and Finish:

[0048] In the formula, These are the projection matrices of the magnetic and gas features in the feature dimension, respectively. It was obtained by splitting it.

[0049] Secondly, a learnable modal alignment weight matrix is ​​introduced. The intermodal temporal attention matrix is ​​calculated using the attention mechanism. :

[0050] In the formula, This is the intermodal temporal attention matrix, characterizing the correlation strength between the magnetic signal and the gas signal at different time steps. This attention matrix... The correlation strength between the magnetic signal and the gas signal at any two time steps was quantified.

[0051] Then, based on the attention matrix A, the gas latent feature sequence is temporally reweighted to achieve adaptive compensation of the hysteresis effect by the attention alignment mechanism, thus obtaining gas latent features that are temporally aligned with the magnetic features. .

[0052] Finally, the aligned magnetic latent features With gas hidden features The components are spliced ​​together and then processed using learnable fusion weights. With learnable fusion bias (This can be obtained through end-to-end training and optimization) Feature fusion and reconstruction are performed to generate the final cross-modal fused feature sequence. ,in, To fuse feature dimensions:

[0053] The first temporal feature extraction module mentioned in step S3 is specifically implemented as a TCN module based on causal dilated convolution, and the second temporal feature extraction module is specifically implemented as a Mamba module based on a selective state-space model. Step S3 is used to construct a TCN-Mamba joint temporal prediction network. Taking the cross-modal fused feature sequence as input, it first extracts local short-term abrupt change features through TCN, and then captures ultra-long-range temporal dependencies through the Mamba module, realizing the dynamic inversion of the insulation health state of the switchgear and multi-step advance prediction of future degradation trajectories. Specifically, step S3 includes the following steps S31 to S34: Step S31: Define the target value for prediction – the insulation health index.

[0054] First, based on the pyrolysis kinetics theory, the insulation health index (HI) of the switchgear is defined as the target value for model prediction, which quantitatively characterizes the degree of insulation degradation:

[0055] In the formula, for The pyrolysis conversion rate of the insulating material at any given time was obtained from pyrolysis kinetic experiments and in-situ inversion. The critical conversion rate for insulation failure is set to a value of [value missing]. This corresponds to the critical state of insulation breakdown. The calculation method is as follows:

[0056] in, The initial mass of the insulating material. for Remaining mass at any given moment The residual mass after complete pyrolysis; The critical conversion rate for insulation failure is set to a value of [value missing]. This corresponds to the critical state of insulation breakdown. The closer the value is to 1, the healthier the insulation condition; the closer the value is to 0, the more severe the insulation degradation.

[0057] Step S32, the first temporal feature extraction module processes the data to capture local mutation features.

[0058] The first temporal feature extraction module in this embodiment is composed of stacked residual blocks of a multi-layer causal dilated convolutional network (TCN).

[0059] Specifically, a multi-layer causal dilated convolution is used to construct the TCN module, which extracts local short-term mutation features of the fused feature sequence to capture the insulation state mutation behavior caused by partial discharge. The core of the TCN module is causal convolution and dilated convolution. Causal convolution: guarantees temporal causality, i.e. The output at time depends only on Given inputs up to a given time and prior, with no future information leaked, the expression for a one-dimensional causal convolution is:

[0060] In the formula, For the first Layer convolution in Output at any moment The kernel size is the convolution kernel size. For the first Each convolutional kernel weight, For the first Layer in Input at any moment.

[0061] Dilated convolution: using the dilation factor Expanding the receptive field of convolution allows for the coverage of longer temporal sequences without increasing the kernel size. The expression for dilated convolution is:

[0062] In the formula, As the expansion factor, the first The dilation factor of the convolution layer is set to The receptive field is expanded exponentially through multi-layer stacking.

[0063] Residual Connections: Each TCN residual block contains two layers of dilated causal convolution with weight normalization, ReLU activation function, and Dropout layer. Simultaneously, residual connections between input and output are achieved through identity mapping, addressing the gradient vanishing problem in deep networks.

[0064] In the formula, Input for residual blocks, For residual mapping, This is the output for the residual block.

[0065] By stacking four layers of TCN residual blocks, the final output is a time sequence from which local mutation features are extracted. ,in For the hidden layer dimension.

[0066] Step S33, the second temporal feature extraction module processes – capturing long-range temporal dependencies.

[0067] The second temporal feature extraction module in this embodiment is composed of stacked multi-layer selective state-space models (Mamba), used to extract features from... It accurately captures the ultra-long-range evolution trend during the insulation degradation process.

[0068] Specifically, the feature sequence output by TCN is input into the Mamba module. Based on the Selective State Space Model (SSM), a hardware-aware selective scanning mechanism is used to accurately capture the ultra-long-range temporal dependencies of insulation degradation while maintaining linear computational complexity.

[0069] The discretized linear state-space model expression is as follows:

[0070] In the formula, The hidden state at any given time. For the state dimension; This is the state-space parameter matrix; for Input at time (i.e.) Output features ); Output for the model.

[0071] Mamba's core innovation lies in its input-related selectivity parameter matrix, namely... All are from the current input Selective memorization and forgetting of temporal information can be achieved through linear projection:

[0072]

[0073] In the formula, To input the relevant discretization step size, It is a time constant. This is the discretized state parameter matrix.

[0074] The final discretized selective state-space update formula is:

[0075] Step S34: Multi-step prediction and model training.

[0076] The parallel prefix scan algorithm is used to implement the parallel computation of the above recursive process, reducing the serial computation complexity of traditional RNNs from... It reduces to hardware-friendly parallel linear computation, while selectively filtering irrelevant timing information and focusing on preserving long-range evolution characteristics related to insulation degradation.

[0077] A six-layer stacked Mamba block architecture is used, with each layer containing a normalization layer, a Mamba selective SSM module, a feedforward network (FFN), and residual connections. The final output is a time series sequence containing long-range evolutionary characteristics. .

[0078] Multi-step prediction head: Inputs the output features of the Mamba module into a fully connected prediction head to realize future predictions. Multi-step advance prediction of insulation health index at each time step, outputting the prediction sequence. ,in The prediction step size is set to 24 / 72 / 168 steps (corresponding to the next 1 / 3 / 7 days).

[0079] During model training, a joint loss function is used, which consists of numerical accuracy loss. And trend consistency loss The weighted summation method is used to simultaneously ensure the numerical accuracy of the predicted values ​​and the accuracy of the deterioration trend.

[0080] Loss function design: A joint loss function combining weighted mean squared error (MSE) and trend consistency loss is adopted to ensure both numerical accuracy and trend accuracy of the predicted values.

[0081] Among them, the loss of numerical precision:

[0082] Trend consistency loss:

[0083] In the formula, The first difference of the true value, For symbolic functions, Weighting is applied to trend loss.

[0084] Model training: The AdamW optimization algorithm was used, and the learning rate was set to... The weight decays to The batch size is 32, the number of iterations is 100, and an early stopping mechanism is used to prevent overfitting. The training set and the test set are divided in an 8:2 ratio, and 5-fold cross-validation is used to optimize the model hyperparameters.

[0085] Step S4, based on the insulation health index prediction sequence obtained in step S3, and combined with the dynamically adjusted early warning threshold, constructs a three-level early warning mechanism to achieve early warning of latent faults.

[0086] Specifically, the warning threshold is not a fixed value, but is dynamically adjusted based on the type of insulation material of the switchgear, its service life, load conditions, and ambient temperature and humidity through a fuzzy inference algorithm. The basic warning threshold is set as follows: Level 1 Warning (Attention Level): Predicting the future In-step insulation health index And pyrolysis conversion rate This indicates that the insulation has undergone early and slight deterioration, and that inspections and data tracking should be strengthened. Level 2 Warning (Alert Level): Predicting the Future In-step insulation health index And pyrolysis conversion rate The message indicates significant insulation degradation, requiring a power outage for maintenance and defect repair within one week. Level 3 Warning (Emergency Level): Predicting the future In-step insulation health index 70, and the pyrolysis conversion rate This indicates a risk of insulation breakdown, requiring immediate power outage to prevent the fault from escalating.

[0087] When the model's prediction results reach the corresponding warning threshold, the system automatically triggers the corresponding level of warning instruction and outputs the following simultaneously: Current assessment results of switchgear insulation health status and future degradation trajectory curves; fault type inversion results (partial discharge / overheating defect / surface discharge) and fault probability; fault location area and defect severity assessment; corresponding level of operation and maintenance handling suggestions and maintenance priorities.

[0088] like Figure 3 As shown in the comparison curves of the Insulation Health Index (HI) prediction, the prediction curve of the method of this invention closely matches the actual decay curve of the HI throughout the entire prediction step size, maintaining a very small prediction error. In contrast, the error gap between the predicted and actual values ​​of the traditional LSTM model continues to widen as the prediction step size increases, resulting in significantly higher predictions. This model cannot accurately depict the decay law of switchgear insulation degradation and is prone to causing a lag in the prediction of the degradation process. This invention relies on the fusion of high-order tensors of magnetic and gas dual-mode signals and cross-modal attention alignment to complete feature fusion. Then, it uses a combined time-series network to simultaneously capture the local abrupt changes in degradation features and long-range time-series dependencies, accurately restoring the actual decay trend of the HI. It can adapt to the first, second, and third-level graded early warning thresholds, and promptly identify slight insulation degradation, significant degradation, and breakdown risks. This effectively solves the problems of large long-term prediction errors and slow degradation trend prediction caused by traditional LSTM models, resulting in late warnings and missed reports, and significantly improves the accuracy, timeliness, and reliability of switchgear insulation degradation early warning.

[0089] The following describes the deployment and closed-loop optimization process of a switchgear insulation degradation early warning method according to an embodiment of the present invention in an actual switchgear field.

[0090] The first step is on-site deployment and model lightweighting.

[0091] At the target switchgear site, the installation, debugging, and time synchronization of the three-axis TMR magnetic sensing array, multi-component electrochemical gas sensing module, and edge computing terminal were completed. Subsequently, through model quantization and pruning operations, the trained tensor fusion network and the TCN-Mamba combined prediction model were compressed and lightweightly deployed to the edge computing terminal, adapting it to the limited computing power of the edge device and enabling real-time inference at the edge.

[0092] Next comes real-time data processing and early warning triggering.

[0093] The edge terminal collects magnetic-aerodynamic signals in real time according to a preset sampling frequency, and executes the aforementioned steps S1 to S4 to complete preprocessing, cross-modal fusion, health status inversion, and degradation trajectory prediction online. Once the prediction result exceeds the dynamic early warning threshold, the corresponding level of early warning instruction is triggered.

[0094] Finally, the model is iteratively updated online.

[0095] Regularly collect fault data confirmed by on-site inspections and experimental data supplemented by the laboratory, and add them to the model training set. Employ incremental learning methods to fine-tune the parameters and update the deployed edge models online, so as to continuously improve the model's adaptability and prediction accuracy under complex on-site working conditions, forming a closed-loop optimization system of data-model-early warning.

[0096] The following is combined This paper presents a specific implementation case of an indoor AC metal-enclosed switchgear, providing a detailed explanation of the implementation process of a switchgear insulation degradation early warning method according to an embodiment of the present invention.

[0097] As an example, the target object is a substation. The KYN28A switchgear's internal insulation components mainly consist of epoxy resin contact boxes and busbar post insulators, while the cable joint insulation material is cross-linked polyethylene. The on-site deployment plan is as follows: one set of triaxial TMR magnetic sensors (Multidimensional Technology TMR3003) will be deployed in each of the switchgear busbar compartment, circuit breaker compartment, and cable compartment; one set of three-component electrochemical gas-sensitive modules will be deployed in each of the cable compartment and busbar compartment. Simultaneously, a temperature and humidity sensor and an edge computing terminal are configured, and the magnetic signal sampling frequency is set to... The gas signal sampling frequency is set to Multi-sensor time synchronization is achieved through the IEEE 1588 protocol. The specific implementation process is as follows: Step S1: Sample data collection: The operation data of the switchgear for 6 consecutive months were collected, including magnetic vector signal, characteristic gas concentration signal, ambient temperature and humidity, and load current data. At the same time, calibration data such as insulation dielectric loss, partial discharge, and pyrolysis conversion rate were obtained through regular power outage preventive tests. The insulation health index HI at the corresponding time was calculated, and a dataset containing 12,000 time series samples was constructed and divided into training set and test set in an 8:2 ratio.

[0098] Magnetic signal feature extraction: The original magnetic signal is subjected to 5-level db4 wavelet decomposition for noise reduction, and 16-dimensional magnetic features are extracted, including: peak values ​​of magnetic signals in three axes, average discharge quantity, discharge repetition frequency, power frequency phase distribution skewness, and spectral center frequency. Bandwidth, 5-layer wavelet packet energy entropy, signal kurtosis, margin factor.

[0099] Gas signal feature extraction: Applying moving average filtering to the original gas concentration sequence and... Outlier removal criteria were used to extract 8-dimensional gas features, including: Real-time concentration, hourly gas production rate, 24-hour cumulative concentration, concentration change gradient, Hurst trend index, and coefficient of variation of gas production rate.

[0100] Data standardization: The Z-Score method was used to standardize the magnetic and gas features separately, resulting in standardized feature sequences. The time series length was set to... (Corresponding to 24-hour historical data, with a time step of 10 minutes).

[0101] Step S2: Feature tensor construction: The standardized magnetic feature sequence ( ) and gas characteristic sequences ( Zero-filling is performed, and the build size is [size missing]. The bimodal feature tensor.

[0102] Tucker tensor decomposition: The feature tensor is decomposed using alternating least squares, and the temporal dimension rank is set. Rank of feature dimensions Modal dimension rank Solving for the core tensor yields the core tensor. With a factor matrix of three dimensions.

[0103] Modal alignment and fusion: Temporal correlation weights between magnetic and gas features are learned using a temporal attention matrix. Gas features are then temporally reweighted to compensate for gas generation lag effects, ultimately yielding... Dimensional cross-modal fusion feature sequences.

[0104] Step S3: Model hyperparameter settings: TCN module: 4-layer residual blocks, convolutional kernel size The inflation factors are 1, 2, 4, and 8, respectively, and the hidden layer dimensions are... ; Mamba modules: 6-layer Mamba blocks, state dimension Hidden layer dimensions Feedforward network expansion factor ; Prediction Header: Fully connected layer, outputs the insulation health index for the next 72 steps (corresponding to the next 3 days), prediction step size. ; Training parameters: AdamW optimizer, learning rate Weight decay Batch size 32, iterations 100, early stop patience value 10, trend loss weight. .

[0105] Step S4: Based on the type of insulation material and the service life of the switchgear, the three-level early warning threshold is dynamically set: Level 1 Warning: Forecast for the next 3 days This triggers a level-one warning, sends out daily status reports, and strengthens special patrols; Level 2 Warning: Forecast for the next 3 days 85 triggers an alarm-level warning, sends a maintenance work order, and arranges a power outage for maintenance within 7 days; Level 3 Alert: Forecast for the next 3 days 70, triggering an emergency warning, immediately report to dispatch and request a power outage.

[0106] Corresponding to the switchgear insulation degradation early warning method of Embodiment 1 of the present invention, Embodiment 2 of the present invention also provides a switchgear insulation degradation early warning device, comprising: One or more processors; Memory; One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the one or more processors, and the one or more applications are configured to perform the switchgear insulation degradation early warning method.

[0107] Corresponding to the switchgear insulation degradation early warning method in Embodiment 1 of the present invention, Embodiment 3 of the present invention also provides a computer program product, including computer instructions, which instruct computer equipment to perform the operation corresponding to the switchgear insulation degradation early warning method.

[0108] Preferably, the processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor, or the processor can be any conventional processor. The processor is the control center of the device, connecting various parts of the device through various interfaces and lines.

[0109] The memory mainly includes a program storage area and a data storage area. The program storage area can store the operating system, applications required for at least one function, etc., while the data storage area can store related data, etc. Furthermore, the memory can be a high-speed random access memory, or a non-volatile memory, such as a plug-in hard drive, a SmartMedia Card (SMC), a Secure Digital (SD) card, and a Flash Card, or other volatile solid-state storage devices.

[0110] It should be noted that the above-mentioned devices may include, but are not limited to, processors and memory, as will be understood by those skilled in the art.

[0111] For the working principle and process of the above embodiments, please refer to the description of Embodiment 1 of the present invention, which will not be repeated here.

[0112] Compared with existing technologies, this invention has the following significant advantages: By constructing a cross-modal fusion network based on Tucker high-order tensor decomposition, this invention organically fuses high-frequency transient magnetic vector signals and low-frequency characteristic gas concentration signals into a unified high-order tensor representation. Furthermore, it introduces an intermodal temporal attention mechanism to adaptively compensate for the physical lag effect of "discharge-gas generation," fundamentally solving the problem of misalignment of magnetic-gas heterogeneous features on the time scale. This overcomes the technical limitations of traditional simple splicing and fusion, deeply explores the intrinsic correlation between the two physical quantities, and improves the characterization ability of the fused features for insulation degradation by more than 40%, achieving a fault diagnosis accuracy of 98.5%. Building upon this foundation, this invention constructs a combined time-series prediction network cascaded with TCN and Mamba. Utilizing the large receptive field of TCN's causal dilated convolution, it accurately captures short-term abrupt changes in insulation state caused by partial discharge. Simultaneously, leveraging the long-range dependency preservation capability of the Mamba selective state-space model, it effectively overcomes the long-range information forgetting problem of traditional recurrent neural networks. Furthermore, by combining a hardware-friendly parallel scanning algorithm, the computational complexity is reduced from O(n²) to O(n), and the inference speed is improved by more than 5 times. This achieves high-precision multi-step advance prediction of insulation health index up to 7 days in the future, with the root mean square error controlled within 1.2%. In addition, this invention deeply integrates the physical mechanism of insulation pyrolysis kinetics with a data-driven model, possessing excellent resistance to environmental interference and load fluctuations. It is adaptable to switchgear of different models and service lives. The model supports lightweight edge deployment, requiring no modification to primary equipment, posing no electrical safety risks, and meeting the dual needs of pre-installation in new switchgear and intelligent transformation of existing switchgear. It has extremely high engineering promotion value and application prospects.

[0113] The above description is merely a preferred embodiment of the present invention and should not be construed as limiting the scope of the invention. Therefore, any equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A method for early warning of insulation degradation in switchgear, characterized in that, include: Step S1: Synchronously acquire transient magnetic vector signals and characteristic gas concentration signals in the switch cabinet, and obtain magnetic feature sequences and gas feature sequences after preprocessing; Step S2: Construct the magnetic feature sequence and the gas feature sequence into a high-order fusion tensor, extract cross-modal correlation features based on the tensor decomposition algorithm, and perform adaptive compensation and fusion of magnetic and gas features in the temporal dimension through the attention alignment mechanism to generate a cross-modal fusion feature sequence. Step S3: Construct a combined temporal prediction network. First, capture the local mutation features of the fused feature sequence through the first temporal feature extraction module, and then capture the long-range temporal dependency through the second temporal feature extraction module. Based on the captured features, output the insulation health index prediction sequence for multiple future time steps. Step S4: Based on the predicted sequence and the preset warning threshold, output a graded warning instruction.

2. The method according to claim 1, characterized in that, Step S1 specifically includes: The transient magnetic vector signal and the characteristic gas concentration signal are acquired by deploying a three-axis tunneling magnetoresistive (TMR) magnetic sensing array and a multi-component electrochemical gas sensing module, respectively, and the two types of signals are synchronized in time. The transient magnetic vector signal is processed by wavelet threshold denoising method. The db4 wavelet basis is selected for 5-level decomposition. The high-frequency coefficients are processed by soft threshold function to reconstruct the denoised magnetic signal. The time domain, frequency domain and time-frequency domain features are extracted to construct the magnetic feature matrix. The characteristic gas concentration signal is processed by moving average filtering and outlier removal algorithm based on the Laida criterion. Missing data is filled by linear interpolation, and trend features, rate of change features and cumulative gas production features are extracted to construct a gas feature matrix. The magnetic feature matrix and the gas feature matrix are respectively subjected to Z-Score normalization to obtain the normalized magnetic feature sequence and gas feature sequence.

3. The method according to claim 1 or 2, characterized in that, In step S2, the higher-order fusion tensor is a bimodal feature tensor obtained by reshaping the magnetic feature sequence and the gas feature sequence into third-order tensors and then concatenating them; the tensor decomposition algorithm is the Tucker decomposition algorithm based on alternating least squares, and the bimodal feature tensor is decomposed into a core tensor and three factor matrices corresponding to the temporal dimension, feature dimension, and modal dimension, respectively. The sum of the products of the moduli is used, and the optimization objective is to minimize the F-norm of the fitted residual tensor.

4. The method according to claim 3, characterized in that, In step S2, the adaptive compensation and fusion of magnetic and gas features in the temporal dimension through the attention alignment mechanism specifically includes: The feature dimension factor matrix and time sequence dimension factor matrix obtained by decomposing the magnetic feature sequence and the gas feature sequence are projected onto a unified latent feature space to obtain magnetic latent features and gas latent features, respectively. A learnable modal alignment weight matrix is ​​introduced, and an attention mechanism is used to calculate the intermodal temporal attention matrix between the magnetic latent features and the gas latent features; Based on the temporal attention matrix, the gas latent features are temporally reweighted to obtain gas latent features that are temporally aligned with the magnetic latent features; The aligned magnetic latent features and gas latent features are concatenated and combined with the core tensor obtained from decomposition. The cross-modal fusion feature sequence is generated by reconstructing the feature sequence through learnable fusion weights and biases.

5. The method according to claim 1, characterized in that, In step S3, the first temporal feature extraction module is a multi-layer causal dilated convolutional network, which is composed of multiple stacked residual blocks. Each residual block contains two layers of causal convolution and dilated convolution. The dilation factor increases exponentially with the number of network layers and forms residual connections through identity mapping. The second temporal feature extraction module is a multi-layer selective state space model. The discretization step size and state parameter matrix of this model are dynamically generated by linear projection of the current input features to selectively remember and forget temporal information.

6. The method according to claim 1 or 5, characterized in that, In step S3, the training of the combined time series prediction network adopts a joint loss function, which is composed of a weighted sum of numerical accuracy loss and trend consistency loss. The numerical accuracy loss is the mean square error between the predicted insulation health index and the actual insulation health index, and the trend consistency loss is calculated based on the first-order difference sign difference between the predicted insulation health index and the actual insulation health index.

7. The method according to claim 1, characterized in that, In step S4, the preset early warning threshold is dynamically adjusted based on the type of insulation material, service life, load conditions, and ambient temperature and humidity of the switchgear using a fuzzy inference algorithm. The graded early warning instructions include at least a first-level early warning, a second-level early warning, and a third-level early warning. Each level of early warning corresponds to a different insulation health index range and a pyrolysis conversion rate range, as well as corresponding operation and maintenance strategies.

8. The method according to claim 1 or 7, characterized in that, The method also includes a field-engineered closed-loop implementation step: After model quantization and pruning, the trained high-order fusion tensor decomposition network and the combined temporal prediction network are lightweighted and deployed to the edge computing terminal to adapt to the computing power of the edge device. Regularly collect fault data confirmed by on-site inspections and supplementary experimental data from the laboratory, and use incremental learning methods to iteratively update the deployed model online.

9. A switchgear insulation degradation early warning device, characterized in that, include: One or more processors; Memory; One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the one or more processors, the one or more applications being configured to perform the switchgear insulation degradation early warning method as described in any one of claims 1 to 8.

10. A computer program product, characterized in that, Includes computer instructions that instruct computer equipment to perform the operation corresponding to the switchgear insulation degradation early warning as described in any one of claims 1 to 8.

Images