Deep learning method and system fusing vmd noise reduction and attention dynamic evidence distribution

Abstract

The application belongs to the cross technical field of GIS high-voltage equipment online monitoring, multi-source sensing signal processing, deep learning and D-S evidence theory, and particularly relates to a deep learning method and system fusing VMD noise reduction and attention dynamic evidence distribution. The application comprises a GIS multi-parameter data preprocessing module, a VMD decomposition and modal attention noise reduction module, a GIS temperature-pressure coupling feature extraction module, a double-attention driven dynamic BPA distribution module, a D-S multi-source evidence fusion and GIS state decision output module and a training optimization module. A multi-constraint joint loss function is constructed, end-to-end joint iterative training is performed through a back propagation algorithm, full-link collaborative optimization is realized, a lightweight modal attention network (MLP+Softmax) is introduced after VMD decomposition, and a differentiated strategy is matched to realize fine-grained adaptive hierarchical noise reduction and solve the weak feature annihilation problem caused by unified noise reduction. A double-branch parallel network structure of a sample attention branch (Sigmoid activation) and a category attention branch (Softmax activation) is designed to realize data-driven dynamic evidence distribution and enhance the adaptability of BPA generation.

Description

Technical Field

[0001] This invention belongs to the technical fields of GIS high-voltage equipment online monitoring, multi-source sensor signal processing, deep learning, and DS evidence theory. Specifically, it involves a deep learning method and system that integrates VMD noise reduction and attention dynamic evidence allocation. Background Technology

[0002] GIS (Gas Insulated Switchgear) relies on the cavity Insulating gases achieve both insulation and arc extinguishing. The internal temperature and gas pressure of the cavity are core parameters reflecting the insulation status, sealing performance, and operating load of the equipment. However, due to factors such as the shielding of the GIS metal cavity, strong electromagnetic interference in the field, instantaneous pulse interference from switch opening and closing, environmental temperature and humidity drift, noise in the sensor acquisition lines, and the subtle, concealed characteristics of slow gas leakage, existing integrated online temperature and pressure monitoring solutions for GIS suffer from multiple technical bottlenecks, making it difficult to accurately and stably perceive the true operating status inside the cavity. Specific shortcomings are as follows:

[0003] (1) Independent training of modules cannot achieve global optimum.

[0004] In existing GIS multi-parameter monitoring algorithms, the VMD denoising module, feature extraction module, and DS evidence allocation and fusion module all adopt a step-by-step independent training and parameter tuning mode. Each module is independent, their parameters are not interconnected, and there is no end-to-end constraint mechanism. The algorithm can only optimize each module individually, failing to achieve backpropagation of decision-level errors to front-end modules and iterative parameter adjustment. This results in errors from the denoising stage being propagated step-by-step to feature extraction, evidence allocation, and the final decision-making stage, leading to severe error accumulation. End-to-end global collaborative optimization is impossible, and the overall model performance suffers from significant bottlenecks. Furthermore, the complex and redundant process of independently debugging multiple modules significantly increases the difficulty and engineering cost of parameter tuning for substation deployment.

[0005] (2) Multi-parameter sensor signals are mixed with noise, and traditional unified noise reduction mode is prone to obliterating early weak fault characteristics.

[0006] The time-series signals collected by the built-in temperature and pressure sensors of GIS are susceptible to interference from the complex operating conditions of substations, superimposed with multiple noises such as electromagnetic radiation noise, opening and closing operation pulse noise, ambient temperature periodic drift noise, and Gaussian noise from the sensing circuit. Weak fault characteristics are easily masked by noise. Existing traditional noise reduction techniques use a uniform noise reduction approach without differentiation. Mean filtering and wavelet global filtering can only smooth the signal as a whole, and cannot distinguish between the effective fault modes and pure noise modes contained in the signal. It is difficult to identify weak low-frequency fault characteristics that characterize early leakage and local overheating of GIS, such as slight pressure drops and slow temperature rises. Over-smoothing can easily cause early weak characteristics to be completely obliterated.

[0007] (3) The multi-parameter fusion adopts a fixed BPA allocation, which has poor adaptability to working conditions and cannot quantify and monitor uncertainty.

[0008] Current GIS monitoring methods mostly use independent threshold alarms for temperature and pressure as single parameters, making it difficult to distinguish between coupled operating conditions such as ambient temperature rise and contact overheating, and ambient air pressure fluctuation and cavity leakage pressure drop, resulting in significant false alarm problems. Existing multi-parameter fusion schemes rely on fixed mathematical mapping rules to generate the basic probability allocation value of DS evidence theory BPA, using static and fixed weight logic, which has several drawbacks: First, the fixed and uniform BPA allocation rules cannot adaptively adjust the weights according to the quality of input samples, feature clarity, and differences in operating conditions, resulting in poor adaptability to operating conditions; second, traditional static BPA is a deterministic output mode, which cannot quantify the uncertainty of model inference or characterize the credibility of samples. It lacks robustness in scenarios such as small leaks, small samples, outliers, sensor disconnection, and signal failure outside the signal domain, easily leading to missed alarms, false alarms, and misdiagnosis problems, and the interpretability of BPA generation is poor. Summary of the Invention

[0009] To address the shortcomings of the existing technologies, this invention provides a deep learning method and system that integrates VMD denoising and attention-based dynamic evidence allocation. The technical problems to be solved are the core challenges in the existing technologies, such as error accumulation caused by independent training of each module, inability to achieve global optimization, excessive smoothing of traditional unified denoising leading to the annihilation of early weak fault features, and poor adaptability to working conditions, inability to quantify monitoring uncertainty, and insufficient robustness in abnormal scenarios caused by fixed BPA weight allocation.

[0010] The technical solution adopted in this invention is as follows:

[0011] This invention provides a deep learning system integrating VMD denoising and attention-based dynamic evidence allocation, comprising a GIS multi-parameter data preprocessing module, a VMD decomposition and modal attention denoising module, a GIS temperature-pressure coupling feature extraction module, a dual-attention-driven dynamic BPA allocation module, a DS multi-source evidence fusion and GIS state decision output module, and a training optimization module. The GIS multi-parameter data preprocessing module, VMD decomposition and modal attention denoising module, GIS temperature-pressure coupling feature extraction module, dual-attention-driven dynamic BPA allocation module, and DS multi-source evidence fusion and GIS state decision output module are combined into a five-layer modular serial architecture, namely: GIS multi-parameter data preprocessing layer, VMD modal decomposition and modal attention adaptive denoising layer, GIS temperature-pressure coupling feature extraction layer, dual-attention-driven dynamic BPA allocation layer, and DS multi-source evidence fusion and GIS state decision output layer; the training optimization module is used for global training optimization.

[0012] The GIS multi-parameter data preprocessing module connects to the temperature and pressure sensing probes of each cavity, and integrates abnormal pulse rejection, missing value interpolation and completion, temperature and pressure time sequence alignment, differential normalization, and sample caching unit to complete the front-end standardization of sensing data.

[0013] The VMD decomposition and modal attention denoising module incorporates a GIS-specific VMD parameter configuration unit, a lightweight modal attention network, a three-level differentiated denoising algorithm, and a temperature and pressure signal reconstruction unit to purify cavity sensing signals.

[0014] The GIS temperature-pressure coupling feature extraction module has a built-in 1D-CNN+GRU hybrid temporal backbone network, supports custom configuration of sampling step size and network depth, and extracts deep features of temperature-pressure coupling correlation.

[0015] The dual-attention-driven dynamic BPA allocation module includes a sample attention subunit, a GIS working condition category attention subunit, an original evidence value calculation unit, and a dynamic BPA generation unit, which outputs single-cavity body temperature and pressure evidence with uncertainty.

[0016] The DS multi-source evidence fusion and GIS status decision output module integrates a multi-point evidence iterative fusion unit, a Pignistic probability transformation unit, and a maximum belief real-time decision unit, and outputs real-time temperature and pressure values ​​of the cavity and equipment fault warning signals.

[0017] The training optimization module has a built-in GIS-specific joint loss calculation unit, backpropagation parameter update unit, and model saving / online loading unit, which are used for offline model training, parameter fixation, and on-site terminal deployment.

[0018] In the five-layer modular serial architecture adopted in this invention, the GIS multi-parameter data preprocessing layer serves as the entry point of the entire algorithm chain. It collects time-series signals output by the temperature and pressure sensors built into the GIS cavity, and completes outlier detection and removal, missing value completion, data dimension unification, unit normalization, and dataset partitioning operations. This completely avoids problems such as invalid data, dimension mismatch, and numerical deviation, laying a solid data foundation for subsequent signal decomposition and feature extraction.

[0019] The VMD mode decomposition and modal attention adaptive denoising layer, the core unit of data optimization in this invention, abandons the traditional VMD global uniform denoising mode. It decomposes the non-stationary time-series signals of GIS temperature and pressure into multiple independent frequency band modal components; then, a lightweight modal attention network autonomously learns the proportion of GIS leakage and overheating fault features in each mode, outputting quantized weights; based on the weights, it classifies the modal types (effective modes, mixed modes, pure noise modes), matches differentiated denoising strategies, and finally combines the attention weights to complete the weighted reconstruction of temperature and pressure signals. This module achieves full-process adaptive modification from modal decomposition to weight evaluation, hierarchical denoising, and signal restoration, filtering out electromagnetic and pulse noise from substations while maximizing the preservation of early, weak GIS fault features such as slight pressure drops and gradual temperature increases.

[0020] The GIS temperature-pressure coupling feature extraction layer sends the noise-reduced and purified cavity temperature and pressure time-series data into a deep learning backbone network adapted to the time-series signal. It extracts the bottom-level waveform features of temperature and pressure, the middle-level correlation features of temperature-pressure coupling, and the high-level abstract features of equipment operation status in layers. It outputs a high-dimensional feature vector with fixed dimensions and strong representation capabilities, which represents the temperature-pressure coordinated change law inside the GIS cavity. As the core carrier for subsequent evidence allocation and model inference, the feature quality directly determines the final inference accuracy.

[0021] The dual-attention driven dynamic BPA allocation layer overcomes the technical bottleneck of traditional fixed formula mapping for BPA generation, constructing a dual-branch parallel network structure of sample-level attention and GIS condition category-level attention. Based on temperature and pressure coupling characteristics, it learns the overall reliability weight of a single group of sensor samples and the matching weight between the sample and various GIS operating conditions. Combining the network classification score, it calculates the original evidence value, and then generates a single-category focal element BPA and a full-set focal element BPA representing monitoring uncertainty through normalization operations. This achieves "one set of evidence allocation rules for a single cavity and a single group of sensors," adaptively adapting to various samples such as normal, minor leakage, local overheating, compound faults, and sensor failure, mitigating multi-parameter evidence conflicts from the source and quantifying the degree of uncertainty in GIS monitoring.

[0022] The DS multi-source evidence fusion and GIS state decision output layer is designed for scenarios involving single-cavity single-measuring-point, multi-cavity distributed multi-measuring-point, and multi-evidence sources with dual temperature and pressure parameters. It employs a DS synthesis rule with conflict coefficient weighting correction to address the fusion failure issue of traditional DS rules under GIS boundary ambiguity and high-conflict temperature and pressure evidence. It iteratively fuses independent BPA evidence from multiple cavities' temperature and pressure parameters to obtain comprehensive global evidence. Then, based on real-time monitoring requirements, it selects either Pignistic probability transformation or the maximum belief criterion to complete the GIS condition determination, outputting the final classification, identification, and diagnostic results.

[0023] In addition to the five functional modules mentioned above, this invention adds a global training optimization system, designs a multi-constraint joint loss function for GIS multi-parameter monitoring, and uses the backpropagation algorithm to perform end-to-end joint iterative training on all trainable parameters in the link, breaking down the parameter barriers between modules, eliminating the error accumulation caused by step-by-step training, and achieving the optimal global performance of the GIS multi-parameter monitoring framework.

[0024] This invention also provides a deep learning method that integrates VMD denoising and attention-based dynamic evidence allocation, implemented based on the aforementioned deep learning system, and includes the following steps:

[0025] Step S1, raw data preprocessing;

[0026] Preferably, in step S1, the raw data is the internal temperature of each cavity of the GIS. For the one-dimensional time-series signals collected by pressure sensors, a dedicated preprocessing procedure has been developed to address the issues of data distortion and inconsistent dimensions in power field sensing data.

[0027] Preferably, step S1 includes:

[0028] S1.1, Raw data collection and initial screening of invalid data;

[0029] Integrated temperature and pressure sensors are deployed inside each air chamber of the GIS to continuously collect real-time temperature data. Gas pressure time series data;

[0030] use The criteria and box plot anomaly detection algorithm are used to jointly identify abnormal values ​​of sudden pulses caused by opening and closing pulses and sensor short circuits; for missing values ​​caused by transmission interruption and acquisition failure, adjacent mean interpolation is used to fill in the missing values; consecutive failure samples are directly removed; and bad data is blocked from flowing into subsequent processes from the source.

[0031] S1.2, Multi-measurement point dimension alignment and format unification;

[0032] Standardize the sampling frequency and sampling sequence length of sensors in each GIS cavity; synchronize and align the timestamps of temperature and pressure signals in the same cavity to ensure a one-to-one correspondence between temperature and pressure parameters.

[0033] S1.3, Data normalization processing;

[0034] The temperature data range is narrow, so Z-Score standardization is used to convert it into a standard distribution with a mean of 0 and a variance of 1. The pressure data has a large amplitude range. Min-Max normalization is used to map it to the [0,1] interval to eliminate the interference of the difference in physical dimensions of temperature and pressure on network training, accelerate the model convergence speed, and improve training stability.

[0035] S1.4, Dataset partitioning and GIS condition labeling;

[0036] Following the common partitioning ratio for industrial algorithms, the preprocessed complete dataset is divided into a training set, a validation set, and a test set, with a typical allocation ratio of 7:2:1. Samples are uniformly labeled with four operating condition tags: ① Cavity in normal operation; ② ③ Minor leakage; ④ Localized overheating of the contact; ⑤ Combined leakage and overheating faults; ⑥ Simultaneously label samples outside the sensor failure domain to obtain a standardized labeled sample set. ,in This represents the total number of all samples.

[0037] Step S2: VMD decomposition and modal attention adaptive noise reduction;

[0038] This module is the core of data denoising and feature protection, solving the problems of mixed noise in GIS temperature and pressure signals and easy loss of early weak fault features. It is divided into five major stages: parameter initialization, mode decomposition, attention weight calculation, hierarchical denoising, and signal reconstruction.

[0039] Preferably, step S2 includes:

[0040] S2.1, Variational Mode Decomposition (VMD);

[0041] VMD is a non-recursive adaptive signal decomposition algorithm. Compared with wavelet decomposition and EMD decomposition, it has the advantages of anti-mode aliasing, high decomposition accuracy and clear frequency domain division. It can be used to decompose noise modes and fault characteristic modes in GIS temperature and pressure time series signals, and is suitable for non-stationary and slowly changing pressure leakage and temperature overheating signals.

[0042] S2.1.1, Definition and selection of core parameters;

[0043] Pre-configure three core parameters:

[0044] Number of mode decompositions : Represents the total number of intrinsic mode components ultimately decomposed from the signal, set according to the signal frequency band complexity. The temperature and pressure signals from a typical GIS cavity are taken as follows: Multi-interference outdoor substation scenarios ;

[0045] Secondary penalty factor Used to balance decomposition accuracy and noise robustness. The larger the value, the smoother the decomposition result and the stronger the noise resistance. The smaller the size, the more complete the detail is preserved; the larger the size should be appropriately increased in low signal-to-noise ratio scenarios. In high signal-to-noise ratio scenarios, appropriately reduce ;

[0046] Convergence tolerance : Controls the VMD iteration termination condition; the smaller the value, the higher the decomposition accuracy. In typical engineering scenarios, it is set to... This balances decomposition accuracy with computational efficiency.

[0047] S2.1.2, Signal mode decomposition operation;

[0048] For the samples obtained in step S1 Perform VMD iterative decomposition to break down the original signal into... The decomposition formula for each mode component with independent center frequencies and limited bandwidth is as follows:

[0049]

[0050] In the formula: Representing the First-order intrinsic modal components, Low-frequency mode carries the characteristics of long-term temperature and pressure changes and slow leakage in the cavity; medium and high-frequency mode carries the characteristics of short-term local overheating and sudden changes; high-frequency mode is mostly electromagnetic and pulse noise of substations.

[0051] S2.2, Modal attention weight calculation;

[0052] A lightweight modal attention network is constructed, with the main body being a shallow multilayer perceptron (MLP). The network has a simple structure and low computational load, making it suitable for the low computing power requirements of GIS online monitoring.

[0053] The network input is a single independent modal component. MLP is responsible for extracting single-modal frequency domain and time domain features and compressing dimensionality;

[0054] Access The activation function performs global normalization on the output of all modalities, ultimately obtaining the modal attention weights corresponding to each modality. Calculation formula:

[0055] ,

[0056] Weighted constraints and physical definitions: mandatory satisfaction ,and The weight values ​​directly represent the proportion of effective information in a mode. The closer the value is to 1, the more likely the mode contains effective features of GIS leakage and overheating. The closer a value is to 0, the more likely the mode is to consist entirely of noise and have no effective features. This weight is learned autonomously by the network without human intervention and possesses strong adaptability.

[0057] S2.3, Modal classification and differentiated noise reduction based on dual thresholds;

[0058] Set a two-level threshold: a high-weight threshold. Low-weight threshold And strictly meet The threshold can be adaptively adjusted using statistical analysis of the modal weight distribution in the training set. Based on the relationship between the two threshold levels, all modes are divided into three categories, and a specific noise reduction algorithm is matched for each mode based on the signal and noise type it carries.

[0059] High-weighted effective modes, This type of mode is the core feature carrier of the sample, containing early fault features such as slight pressure drop and slow temperature rise. Neighborhood mean smoothing and noise reduction are used for very light processing, filtering out only sporadic random noise, while preserving the time domain waveform, frequency domain peaks, and details of abrupt change points throughout; weak feature loss is eliminated, making it suitable for low-frequency main feature modes.

[0060] Medium-weighted mixed modes, This type of mode belongs to the hybrid mode of features and noise, containing some effective detailed features as well as impulse noise and Gaussian noise. Median filtering is used to suppress the interference of opening and closing pulses, and wavelet thresholding is superimposed to filter out electromagnetic Gaussian noise, so as to retain the weak effective features within the mode to the maximum extent while reducing noise.

[0061] Low-weighted pure noise mode, No GIS features are present. Modal components are directly set to zero and removed; noise transmission to the backend link is completely cut off, reducing data interference at the source.

[0062] S2.4, Signal weighted reconstruction and boundary verification;

[0063] After performing hierarchical noise reduction, the processed modal components of each order are obtained. Combine modal attention weights By performing weighted summation, the denoised pure temperature and pressure signal can be reconstructed. Reconstructing the formula:

[0064] ,

[0065] After reconstruction, a boundary verification step is added: the amplitude range and temporal continuity of the reconstructed temperature and pressure signal are detected to avoid data distortion during noise reduction and reconstruction, and to output GIS temperature and pressure time series samples with high signal-to-noise ratio and prominent fault characteristics.

[0066] This step fully realizes the closed-loop process of "modal differentiation → weight adaptation → hierarchical noise reduction → signal reconstruction → boundary verification", comprehensively solving the technical defects of traditional VMD noise reduction.

[0067] Step S3, Deep learning feature extraction;

[0068] Preferably, step S3 includes:

[0069] S3.1, Refined Network Selection;

[0070] To address the temporal coupling characteristics of temperature and pressure in GIS, the following approach is adopted. Hybrid backbone network, hierarchical mining of temperature and pressure correlation features:

[0071] 1D-CNN extracts short-term local abrupt change features of temperature and pressure signals, including instantaneous overheating and sudden pressure changes;

[0072] GRU captures long-term, slow-changing characteristics, including long-term slow leakage and continuous load temperature rise, adapting to long-term online monitoring time-series data of GIS.

[0073] S3.2, Hierarchical Feature Extraction;

[0074] Deep learning networks consist of three layers of feature extraction logic: shallow, medium, and high.

[0075] Shallow networks are used to extract temperature waveforms and pressure fluctuations from the underlying temporal texture.

[0076] The middle-layer network integrates the coupled correlation characteristics of temperature and pressure at the same time;

[0077] High-level networks condense and distinguish global abstract features of normal, leakage, overheating, and complex faults.

[0078] S3.3, Feature Output;

[0079] After the network completes multi-level feature extraction, it outputs the first... Temperature-pressure coupling feature vector with fixed sample dimensions Calculation formula:

[0080] .

[0081] Feature vector It condenses all the effective information on the coordinated changes in temperature and pressure inside the GIS cavity, and is the only input for subsequent dual attention branch operations and dynamic BPA generation. The discriminability and completeness of the features directly determine the rationality of evidence allocation.

[0082] Step S4, attention mechanism-driven dynamic evidence allocation;

[0083] Traditional methods map network scores to BPA using a fixed formula, which fails to adapt to sample differences. This invention designs a dual-branch parallel structure of sample attention and category attention to dynamically generate temperature and pressure evidence allocation values ​​with uncertainty, adapting to various GIS operating conditions.

[0084] Preferably, step S4 includes:

[0085] S4.1, dual attention branch design;

[0086] Based on temperature-pressure coupling feature vector Two attention branches are designed to operate synchronously without interfering with each other, representing sample quality and class matching degree respectively:

[0087] (1) Sample attention branch : Quantify the overall quality, data reliability, and feature completeness of a single-group cavity temperature and pressure sensor sample.

[0088] ,

[0089] use Activation functions naturally output weights Constraints The range, with its numerical value, directly represents the reliability of the sample. The larger the value, the stronger the sample reliability.

[0090] (2) Category Attention Branch : Quantify the matching similarity between the sample and five types of GIS operating conditions (normal / minor leakage / local overheating / combined fault / sensor anomaly); Characterizing the first The nth sample pair Matching degree of class features:

[0091] ,

[0092] use Activation function, forced constraint This achieves global normalization of category weights; The larger the value, the more similar the current temperature and pressure characteristics are to the first... The higher the GIS-like working condition matching degree, This represents the total number of preset operating condition categories.

[0093] S4.2, Calculation of original dynamic evidence value;

[0094] Fusion Sample Credibility Weights Category matching weight The fully connected layer classification score, based on three dimensions of information, is used to calculate the first... The sample corresponds to the first Original evidence value for similar working conditions :

[0095] ,

[0096] in The first output of the fully connected layer The original score for the similar working conditions; this evidence value integrates sample quality, working condition matching degree, and network discrimination tendency. Compared with a single score mapping, it is more in line with the true characteristic distribution of the sample and reduces the probability of conflict of multi-parameter evidence of temperature and pressure from the source.

[0097] S4.3, Generation of Dynamic Basic Probability Assignment Value (BPA);

[0098] Based on the DS evidence theory framework, BPA is divided into single-class focal element BPA and whole-set focal element BPA. Single-class focal element BPA is assigned to known target conditions; whole-set focal element BPA is assigned to the entire identification framework. This represents the uncertainty in GIS monitoring; normalization is performed based on the original evidence values, and numerical stability is added to avoid anomalies where the denominator is zero.

[0099] Single-category coke element BPA, valid evidence for various operating conditions: ,

[0100] Complete set of focal element BPA, evidence of uncertainty: .

[0101] BPA variation patterns under different operating conditions:

[0102] Normal cavity samples: The sum of evidence values ​​for various working conditions is large, the proportion of single-category BPA is high, and the proportion of overall uncertainty BPA is low;

[0103] Minor leaks / critical overheating fuzzy / boundary samples: Evidence values ​​for various operating conditions are close, BPA differences for multiple operating conditions are small, uncertainty BPA is moderately increased, and the model output indicates potential issues.

[0104] Sensor disconnection / outside anomaly samples: The sum of evidence values ​​for various operating conditions is extremely small, and the overall uncertainty (BPA) is significantly increased. The system actively determines sensor failure and eliminates false alarms.

[0105] This design perfectly compensates for the shortcomings of traditional static BPA in expressing uncertainty, while relying on dynamic weights to adaptively adjust the evidence distribution and effectively suppress evidence conflicts.

[0106] Step S5, DS evidence fusion and decision output;

[0107] This module is designed for GIS scenarios with distributed measurement points in multiple cavities and multiple evidence sources with dual parameters of temperature and pressure in the same cavity. It adopts a DS synthesis rule with a conflict coefficient weighting correction to replace the traditional DS rule, which solves the problem that the fusion results of the traditional rule under high conflict evidence violate common sense, and completes evidence fusion and monitoring decision-making.

[0108] Preferably, step S5 includes:

[0109] S5.1, Iterative fusion of multi-source temperature and pressure evidence;

[0110] The dynamic BPA output from each cavity and each temperature / pressure sensor is treated as an independent source of evidence. The DS synthesis rule is used to iteratively fuse the evidence to solve the problem of fusion distortion under high conflict of temperature and pressure evidence from multiple measurement points, and output a global comprehensive BPA.

[0111] S5.2, dual decision criteria adapted to power monitoring scenarios;

[0112] (1) Probability Transformation: Converts the BPA probability distribution into a standard probability form, outputting the confidence probabilities of each of the five operating conditions. It is suitable for non-real-time analysis scenarios such as refined fault tracing and early leakage trend analysis in substations.

[0113] (2) Maximum belief criterion: The working condition with the largest BPA value is directly selected as the final reasoning result. The operation logic is simple, the amount of calculation is small, and the latency is low. After code optimization, it can meet the millisecond-level response and is suitable for GIS on-site real-time online early warning and local sound and light alarm strong real-time scenarios.

[0114] S5.3, Output the results;

[0115] Synchronous output: ① Current real-time internal temperature of the GIS cavity, ① Pressure values; ② Equipment operating condition judgment results (normal / minor leakage / local overheating / complex fault / sensor abnormality); ③ Quantitative indicators of monitoring uncertainty; ④ Early warning signals of abnormal fault levels, providing maintenance personnel with complete information on the cavity status.

[0116] Step S6: Iterative training of the model.

[0117] Preferably, step S6 includes:

[0118] To address the problem of independent module training and progressive error accumulation in traditional methods, this invention constructs a multi-constraint joint loss function, which incorporates all trainable modules across the entire training chain into a unified training framework. Combined with a dedicated training strategy, optimizer, and early stopping mechanism, it achieves end-to-end global optimization through model training.

[0119] The multi-constraint joint loss function includes three types of losses:

[0120] Temperature and pressure condition classification loss: Cross-entropy loss is used to constrain the accuracy of the feature extraction network in distinguishing the five types of conditions in GIS;

[0121] Modal attention regularization loss: constrains the distribution of modal attention weights to ensure that fault feature modes and noise modes are clearly distinguishable;

[0122] Multi-parameter evidence conflict penalty loss: Constrains the degree of conflict between multiple temperature and pressure BPAs to guide the model to generate low-conflict monitoring evidence;

[0123] The total loss is a weighted sum of the three types of losses to balance the optimization objectives of each module.

[0124] Preferably, in step S6, the model training method is as follows: using the backpropagation algorithm combined with the gradient descent optimizer, updating all parameters of the VMD mode decomposition and modal attention adaptive noise reduction layer, the GIS temperature and pressure coupling feature extraction layer, and the dual attention-driven dynamic BPA allocation layer layer by layer from back to front, with the validation set early stopping mechanism, iterating until the model monitoring accuracy converges and then solidifying the model, which is then directly deployed on the GIS online monitoring edge terminal;

[0125] The gradient descent optimizer includes Adam or SGD.

[0126] The overall operational logic of this invention can be summarized as follows: GIS cavity temperature and pressure sensor raw data acquisition → multi-parameter standardization preprocessing → VMD decomposition + modal attention adaptive noise reduction and purification → deep learning network mining of deep temperature and pressure coupling features → dynamic calculation of weights using dual attention branches and generation of adaptive BPA → DS rules completing multi-cavity / temperature and pressure multi-parameter evidence fusion → decision criteria outputting GIS internal temperature and pressure parameters and equipment operating status early warnings → joint loss function driving iterative optimization of all modules. The entire chain is progressive and interconnected, achieving comprehensive performance optimization from the data source to the decision-making end.

[0127] The beneficial effects obtained by adopting the above technical solution are as follows:

[0128] (1) This invention constructs a multi-constraint joint loss function (including temperature and pressure condition classification loss, modal attention regularization loss, and multi-parameter evidence conflict penalty loss), and performs end-to-end joint iterative training on the VMD modal decomposition and modal attention adaptive denoising layer, GIS temperature and pressure coupling feature extraction layer, and dual attention-driven dynamic BPA allocation layer through backpropagation algorithm; it breaks down the parameter barriers between VMD denoising module, feature extraction module, dynamic BPA module, and evidence fusion module, so that the error of the decision layer can be backpropagated to guide the parameter adjustment of the front-end denoising and feature extraction modules, realizes full-link global collaborative optimization, improves the overall performance of the model, and simplifies the engineering debugging process and reduces the difficulty of on-site deployment.

[0129] (2) This invention introduces a lightweight modal attention network (MLP+Softmax) after VMD decomposition. It autonomously learns the effective information proportion weight of each modal component and divides the modality into high-weight effective mode, medium-weight mixed mode and low-weight pure noise mode according to dual thresholds. It matches three differentiated strategies respectively: neighborhood mean smoothing, median filtering + wavelet threshold joint noise reduction and direct zeroing elimination. Finally, the signal is reconstructed by weighting. It realizes fine-grained adaptive hierarchical noise reduction. While filtering out electromagnetic and pulse noise in substations, it retains the early weak fault features of GIS such as pressure drop and temperature rise to the maximum extent, effectively solving the problem of weak feature annihilation caused by uniform noise reduction.

[0130] (3) This invention designs a dual-branch parallel network structure with sample attention branch (Sigmoid activation) and category attention branch (Softmax activation). Sample attention independently quantifies the overall credibility of the current input sample, and category attention dynamically calculates the matching degree between the sample and various working conditions. After the two are fused with the classification score of the fully connected layer, they are normalized to generate single-class focal element BPA and full-set focal element BPA. It realizes data-driven dynamic evidence allocation, which can adaptively adjust the evidence weight according to the sample quality and working condition matching degree. At the same time, it explicitly quantifies the uncertainty of GIS monitoring, enhances the interpretability of BPA generation and robustness to sensor disconnection / extra-domain abnormal samples, and improves adaptability. Attached Figure Description

[0131] Figure 1 This is a flowchart of the deep learning method that integrates VMD noise reduction and attention-based dynamic evidence allocation according to the present invention. Detailed Implementation

[0132] The technical solution of the present invention will now be described more clearly and completely with reference to the accompanying drawings.

[0133] This embodiment provides a deep learning system that integrates VMD noise reduction and attention-based dynamic evidence allocation. Six modules operate collaboratively, supporting deployment on multiple platforms including embedded devices, servers, and industrial control computers. The detailed system composition is as follows:

[0134] GIS multi-parameter data preprocessing module: Connects to temperature and pressure sensing probes in various cavities, integrates abnormal pulse removal, missing value interpolation and completion, temperature and pressure time sequence alignment, differential normalization, and sample caching unit to complete the front-end standardization of sensor data;

[0135] VMD decomposition and modal attention denoising module: Built-in GIS-specific VMD parameter configuration unit, lightweight modal attention network, three-level differentiated denoising algorithm, and temperature and pressure signal reconstruction unit to achieve purification of cavity sensing signals;

[0136] GIS Temperature-Pressure Coupling Feature Extraction Module: Built-in 1D-CNN+GRU hybrid temporal backbone network, supports custom configuration of sampling step size and network depth, and extracts deep features of temperature-pressure coupling correlation;

[0137] Dual attention-driven dynamic BPA allocation module: includes sample attention sub-unit, GIS working condition category attention sub-unit, original evidence value calculation unit, and dynamic BPA generation unit, outputting single-cavity body temperature and pressure evidence with uncertainty;

[0138] DS Multi-Source Evidence Fusion and GIS Status Decision Output Module: Integrates a multi-point evidence iterative fusion unit, a pignistic probability transformation unit, and a maximum belief real-time decision unit, outputting real-time temperature and pressure values ​​of the cavity and equipment fault warning signals;

[0139] Training optimization module: It has a built-in GIS-specific joint loss calculation unit, backpropagation parameter update unit, and model saving / online loading unit, which is responsible for offline model training, parameter fixation and on-site terminal deployment.

[0140] The modules adopt a pipeline-style data interaction, supporting both offline training and online real-time GIS monitoring modes, and are suitable for scenarios such as online monitoring and transformation of newly built GIS stations and sensor upgrades of existing GIS equipment.

[0141] This embodiment also provides a deep learning method that integrates VMD denoising and attention-based dynamic evidence allocation, such as... Figure 1 As shown, this embodiment focuses on the interior of the GIS segmented cavity in a substation. Taking online temperature and pressure monitoring as an application scenario, this paper provides a comprehensive and detailed explanation of the complete implementation process of the present invention:

[0142] ①GIS cavity temperature and pressure sensor acquisition and multi-parameter preprocessing

[0143] An integrated temperature and pressure sensor probe was installed inside each of the three independent air chambers of the GIS system to continuously collect data from the interior of the chambers. Gas temperature and absolute pressure time-series data; using The criteria filter out instantaneous pulse anomalies during switch opening and closing; short-term disconnection segments in the sensors are repaired using third-order mean interpolation; the sampling frequency of the three cavity sections is unified to 1Hz, and the temperature and pressure timestamps of the same cavity are synchronized and aligned; temperature data is standardized using Z-Score, and pressure data is normalized using Min-Max; the dataset is divided into four categories according to a 7:2:1 ratio, and labeled with four operating conditions: normal operation, ... Minor leakage, localized overheating of contacts, and combined leakage and overheating faults were identified. Simultaneously, samples from outside the sensor disconnection domain were recorded to obtain a standardized temperature and pressure time series sample set.

[0144] ②VMD mode decomposition and adaptive modal attention denoising of GIS temperature and pressure signals

[0145] Set the number of VMD decomposition modes Punishment factor Convergence tolerance The preprocessed temperature and pressure signals were decomposed using VMD to obtain 8 independent frequency band modal components. These components were then fed into a lightweight MLP modal attention network to calculate the weights of each mode. Based on an adaptive threshold, the modes were classified into three categories: low-frequency high-weight modes, which carry the characteristics of long-term slight pressure drop and slow temperature rise in the cavity, resulting in leakage / overheating, were only lightly mean-smoothed; mid-frequency mixed modes were filtered using median filtering and wavelet denoising to remove pulse electromagnetic interference; and high-frequency low-weight pure noise modes were directly zeroed out. Finally, the pure temperature and pressure time sequence signal was reconstructed based on the modal weights to completely eliminate the masking of minor fault characteristics by substation noise.

[0146] ③ Extracting GIS temperature and pressure coupling features using a 1D-CNN+GRU network

[0147] The noise-reconstructed temperature and pressure time series samples are divided into frames according to 300 time steps (300 time series sampling points) to construct the network input; 1D-CNN extracts short-term temperature and pressure change features, and GRU mines the long-term slow leakage and continuous temperature rise time series dependencies; the shallow layer of the network extracts basic waveform features, the middle layer fuses the temperature and pressure coupling correlation features at the same time, and the high layer outputs a unified dimension coupling feature vector that distinguishes four types of working conditions, accurately representing the temperature and pressure coordinated change law inside the GIS cavity.

[0148] ④ Dual Attention Branch Dynamic GIS Generation with Adaptive BPA

[0149] The sample attention branch calculates the credibility weight of single-cavity sensor samples based on residual noise and waveform integrity in temperature and pressure signals. The operating condition attention branch matches coupled features with four types of GIS standard operating condition features and outputs the matching similarity for each category. The original evidence value is calculated by fusing the dual weights and network classification scores, and normalized to generate the single-focal element BPA for the four operating conditions and the full set BPA representing the monitoring uncertainty. For samples with slow leakage at the cavity boundary, the uncertainty BPA is automatically increased, and a "Monitoring doubtful, retest recommended" prompt is output. For samples outside the sensor disconnection domain, the full set uncertainty value increases significantly, indicating sensor failure and avoiding false alarms.

[0150] ⑤ Multi-cavity, multi-parameter evidence fusion and GIS status, temperature, and pressure parameter output

[0151] Each of the three GIS cavity segments outputs independent dynamic BPA evidence. The three sets of multi-source evidence are iteratively fused using the DS evidence synthesis rule to eliminate the limitations of single-cavity sensing and evidence conflicts. The maximum belief criterion is selected to adapt to the real-time online monitoring requirements, and the real-time internal data of each cavity segment is output synchronously. Temperature and pressure values, as well as equipment operating status determination results (normal / minor leakage / local overheating / compound fault / sensor abnormality), are used to output graded early warning signals and upload them to the substation backend under abnormal operating conditions.

[0152] ⑥ Verification of experimental results

[0153] Comparative experiments were conducted using multiple datasets from actual measurements of GIS cavities in substations, categorized as normal, slightly leaking, experiencing localized overheating, and suffering from sensor failure. The control group employed a traditional fixed-parameter VMD noise reduction + fixed-formula BPA fusion scheme. Experimental results showed that:

[0154] 1) This invention improves the overall accuracy of GIS working condition diagnosis. ,small Early detection of leaks reduces false alarm rate ;

[0155] 2) The average conflict coefficient of multi-parameter evidence for temperature and pressure decreased by 24%, and the number of false alarms for temperature-pressure coupled fuzzy conditions was significantly reduced;

[0156] 3) It can accurately quantify and monitor uncertainties, effectively distinguish between actual equipment faults and environmental or sensor interference anomalies, and is suitable for online monitoring engineering scenarios of complex electromagnetic GIS cavities in substations.

Claims

1. A deep learning system integrating VMD denoising and attention-based dynamic evidence allocation, characterized in that, The system includes a GIS multi-parameter data preprocessing module, a VMD decomposition and modal attention denoising module, a GIS temperature-pressure coupling feature extraction module, a dual-attention-driven dynamic BPA allocation module, a DS multi-source evidence fusion and GIS state decision output module, and a training optimization module. These modules are combined into a five-layer modular, serial architecture: a GIS multi-parameter data preprocessing layer, a VMD modal decomposition and modal attention adaptive denoising layer, a GIS temperature-pressure coupling feature extraction layer, a dual-attention-driven dynamic BPA allocation layer, and a DS multi-source evidence fusion and GIS state decision output layer. The training optimization module is used for global training optimization. The GIS multi-parameter data preprocessing module connects to the temperature and pressure sensing probes of each cavity, and integrates abnormal pulse rejection, missing value interpolation and completion, temperature and pressure time sequence alignment, differential normalization, and sample caching unit to complete the front-end standardization of sensing data. The VMD decomposition and modal attention denoising module incorporates a GIS-specific VMD parameter configuration unit, a lightweight modal attention network, a three-level differentiated denoising algorithm, and a temperature and pressure signal reconstruction unit to purify cavity sensing signals. The GIS temperature-pressure coupling feature extraction module has a built-in 1D-CNN+GRU hybrid temporal backbone network, supports custom configuration of sampling step size and network depth, and extracts deep features of temperature-pressure coupling correlation. The dual-attention-driven dynamic BPA allocation module includes a sample attention subunit, a GIS working condition category attention subunit, an original evidence value calculation unit, and a dynamic BPA generation unit, which outputs single-cavity body temperature and pressure evidence with uncertainty. The DS multi-source evidence fusion and GIS status decision output module integrates a multi-point evidence iterative fusion unit, a Pignistic probability transformation unit, and a maximum belief real-time decision unit, and outputs real-time temperature and pressure values ​​of the cavity and equipment fault warning signals. The training optimization module has a built-in GIS-specific joint loss calculation unit, backpropagation parameter update unit, and model saving / online loading unit, which are used for offline model training, parameter fixation, and on-site terminal deployment.

2. A deep learning method integrating VMD denoising and attention-based dynamic evidence allocation, implemented based on the deep learning system of claim 1, characterized in that, Includes the following steps: Step S1, raw data preprocessing; Step S2: VMD decomposition and modal attention adaptive noise reduction; Step S3, Deep learning feature extraction; Step S4, attention mechanism-driven dynamic evidence allocation; Step S5, DS evidence fusion and decision output; Step S6: Iterative training of the model.

3. The deep learning method according to claim 2, characterized in that, In step S1, the raw data is the internal temperature of each cavity of the GIS. One-dimensional time-series signal acquired by a pressure sensor.

4. The deep learning method according to claim 3, characterized in that, Step S1 includes: S1.1, Raw data collection and initial screening of invalid data; Based on the built-in temperature of each cavity of the GIS, Pressure sensor batch acquisition of real-time temperature, Gas pressure time series data; use The criteria and box plot anomaly detection algorithm are used to jointly identify abnormal values ​​of sudden pulses caused by opening and closing pulses and sensor short circuits; for missing values ​​caused by transmission interruption and acquisition failure, adjacent mean interpolation is used to fill in the missing values; and consecutive failure samples are directly removed. S1.2, Dimensional alignment and formatting consistency; Standardize the sampling frequency and sampling sequence length of sensors in various GIS cavity; synchronize and align the timestamps of temperature and pressure signals within the same cavity; S1.3, Data normalization processing; S1.4, Dataset partitioning and sample labeling; The preprocessed complete dataset was divided into a training set, a validation set, and a test set, with a ratio of 7:2:1; the samples were uniformly labeled with four operating condition tags: ① normal cavity operation; ② Minor leakage; ③ Localized overheating of the contact; ④ Combined leakage and overheating faults; Simultaneously label samples outside the sensor failure domain to obtain a standardized labeled sample set. ,in This represents the total number of all samples.

5. The deep learning method according to claim 4, characterized in that, Step S2 includes: S2.1, Variational Mode Decomposition (VMD); S2.1.1, Definition and selection of core parameters; Three core parameters are pre-configured, including: Number of mode decompositions : Represents the total number of intrinsic mode components that are finally decomposed from the signal; Secondary penalty factor Used to balance decomposition accuracy and noise robustness; Convergence tolerance : Control the termination condition of VMD iteration; S2.1.2, Signal mode decomposition operation; For the samples obtained in step S1 Perform VMD iterative decomposition to break down the original signal into... The decomposition formula for each mode component with independent center frequencies and limited bandwidth is as follows: ； In the formula: Representing the First-order modal components, ; S2.2, Modal attention weight calculation; A lightweight modal attention network was constructed, with the main body of the network being a shallow multilayer perceptron (MLP). The network input is a single independent modal component. MLP is responsible for extracting single-modal frequency domain and time domain features and compressing dimensionality; Access The activation function performs global normalization on the output of all modalities, ultimately obtaining the modal attention weights corresponding to each modality. Calculation formula: ； Forced satisfaction ,and ; S2.3, Modal classification and differentiated noise reduction based on dual thresholds; Set a two-level threshold: a high-weight threshold. Low-weight threshold And strictly meet ;in accordance with Based on the relationship between the two threshold levels, all modes are divided into three categories: High-weighted effective modes, The neighborhood mean smoothing noise reduction method is used to filter out sporadic random noise and retain the detailed features of time-domain waveform, frequency-domain peaks, and abrupt change points throughout the process. Medium-weighted mixed modes, Median filtering is used to suppress opening and closing pulse interference, and wavelet threshold denoising is superimposed to filter out electromagnetic Gaussian noise; Low-weighted pure noise mode, Directly set the modal components to zero and discard them; S2.4, Signal weighted reconstruction and boundary verification; After step S2.3 completes the hierarchical noise reduction, the processed modal components of each order are obtained. Combine modal attention weights By performing weighted summation, the denoised pure temperature and pressure signal can be reconstructed. Reconstructing the formula: ； After reconstruction, a boundary verification step is added: the amplitude range and timing continuity of the reconstructed temperature and pressure signal are checked to ensure the integrity of the output signal.

6. The deep learning method according to claim 5, characterized in that, Step S3 includes: S3.1, Refined Network Selection; A hybrid backbone network of 1D-CNN and GRU is used to hierarchically mine temperature and pressure correlation features: 1D-CNN extracts short-term local abrupt change features of temperature and pressure signals, including instantaneous overheating and sudden pressure changes; GRU captures long-term, slow-changing characteristics, including long-term slow leakage and continuous load temperature rise. S3.2, Hierarchical Feature Extraction; Feature extraction is divided into three layers: shallow, middle, and high. Shallow networks are used to extract temperature waveforms and pressure fluctuations from the underlying temporal texture. The middle-layer network integrates the coupled correlation characteristics of temperature and pressure at the same time; High-level networks condense and distinguish global abstract features of normal, leakage, overheating, and complex faults; S3.3, Feature Output; After completing multi-level feature extraction in step S3.2, output the first... Temperature-pressure coupling feature vector with fixed sample dimensions Calculation formula: 。 7. The deep learning method according to claim 6, characterized in that, Step S4 includes: S4.1, dual attention branch design; Based on temperature-pressure coupling feature vector Two attention branches are designed to represent sample quality and working condition matching degree, respectively: (1) Sample attention branch : Quantify the overall quality, data reliability, and feature completeness of a single set of cavity temperature and pressure sensor samples; ； use The activation function outputs weights. Constraints interval; (2) Category Attention Branch The similarity between quantitative samples and five types of GIS operating conditions is measured. The five types of GIS operating conditions are: normal, minor leakage, local overheating, compound fault, and sensor abnormality. Characterizing the first The nth sample pair Matching degree of class features: ； use Activation function, forced constraint ; S4.2, Calculation of original dynamic evidence value; Fusion Sample Credibility Weights Category matching weight The classification score of the fully connected layer is calculated. The sample corresponds to the first Original evidence value for similar working conditions : ； in The first output of the fully connected layer Original score for the type of working condition; S4.3, Generation of Dynamic Basic Probability Assignment Value (BPA); The BPA is divided into single-class focal element BPA and full-set focal element BPA. Single-class focal element BPA is assigned to known target conditions; full-set focal element BPA is assigned to the entire identification frame set. This represents the uncertainty in GIS monitoring; normalization is performed based on the original evidence values: Single-category coke element BPA, valid evidence for various operating conditions: ; Complete set of focal element BPA, evidence of uncertainty: .

8. The deep learning method according to claim 7, characterized in that, Step S5 includes: S5.1, Iterative fusion of multi-source temperature and pressure evidence; The dynamic BPA output of each cavity and each temperature / pressure sensor is treated as an independent source of evidence. The DS synthesis rule is used to iteratively fuse the evidence to obtain the fused global comprehensive BPA. S5.2, dual decision criteria adapted to power monitoring scenarios; (1) Probability Transformation: The BPA probability distribution is converted into a standard probability form, and the confidence probabilities of the five types of GIS operating conditions are output for substation background refined fault tracing and early leakage trend analysis scenarios. (2) Maximum belief criterion: The working condition with the largest BPA value is directly selected as the final reasoning result, which is used for real-time online early warning and local sound and light alarm scenarios in GIS field; S5.3, Output the results; Synchronous output: ① Current real-time internal temperature of the GIS cavity, ① Pressure value; ② Equipment operating condition judgment result; ③ Quantitative indicators of monitoring uncertainty; ④ Early warning signal of abnormal fault level.

9. The deep learning method according to claim 8, characterized in that, Step S6 includes: Construct a multi-constraint joint loss function, which includes three types of losses: Temperature and pressure condition classification loss: Cross-entropy loss is used to constrain the accuracy of the feature extraction network in distinguishing the five types of GIS conditions; Modal attention regularization loss: constrains the distribution of modal attention weights; Multi-parameter evidence conflict penalty loss: constraining the degree of conflict between multiple temperature and pressure BPAs; The total loss is the weighted sum of the three types of losses.

10. The deep learning method according to claim 9, characterized in that, In step S6, the model training method is as follows: using the backpropagation algorithm combined with the gradient descent optimizer, updating all parameters of the VMD mode decomposition and modal attention adaptive noise reduction layer, GIS temperature and pressure coupling feature extraction layer, and dual attention driven dynamic BPA allocation layer from back to front, iterating repeatedly until the model performance converges on the validation set, thus completing model training and solidification. The gradient descent optimizer includes Adam or SGD.

Images