Intelligent diagnosis method for partial discharge of SF6 / N2 mixed gas under low temperature environment

Abstract

The present application belongs to the technical field of gas discharge diagnosis, and particularly relates to an intelligent diagnosis method for partial discharge of SF6 / N2 mixed gas under low-temperature environment. For the SF6 / N2 mixed gas, multi-source partial discharge signals and environmental physical quantities when the mixed gas occurs partial discharge are collected; for each sample in the source domain data set and the target domain data set, a compression factor deviation ΔZ is calculated; a double-flow adaptive feature fusion network is constructed, and the fusion weight is dynamically adjusted by ΔZ; an entropy weight adaptive loss function of the double-flow adaptive feature fusion network is set, and joint training is performed; the double-flow adaptive feature fusion network is pruned and fine-tuned to obtain a final optimized diagnosis model; multi-source partial discharge signals and environmental physical quantities when the mixed gas occurs partial discharge under actual working conditions are collected in real time, ΔZ is calculated, real-time data and ΔZ are input into the optimized diagnosis model, and a diagnosis result is output. The present application can accurately diagnose the partial discharge condition of the SF6 / N2 mixed gas.

Description

Technical Field

[0001] This invention belongs to the field of gas discharge diagnostic technology, specifically relating to an intelligent diagnostic method for partial discharge of SF6 / N2 mixed gas under low-temperature conditions. Background Technology

[0002] SF6 (sulfur hexafluoride) / N2 (nitrogen) mixture is one of the environmentally friendly alternatives to SF6. Due to its strong insulation and environmentally friendly properties, it has become a key insulating medium replacing SF6. However, partial discharge detection of SF6 / N2 mixtures faces numerous challenges in electrical equipment used in high-altitude or low-temperature applications. Traditional partial discharge diagnostic systems often struggle to adapt to low-temperature conditions, and their detection accuracy and reliability are affected by changes in gas parameters. The physical properties of gases change at low temperatures, causing differences in the characteristics of partial discharge signals compared to normal temperatures. Traditional diagnostic methods cannot effectively capture and analyze these changed signal characteristics, easily leading to misjudgments or missed detections, posing potential risks to the safe operation of equipment. Furthermore, because SF6 is no longer an ideal gas under low-temperature and high-pressure conditions, traditional gas law fails at low temperatures, making it difficult to accurately describe the microscopic particle states of SF6 / N2 mixtures.

[0003] Existing deep learning-based partial discharge diagnostic models (such as CNN and LSTM) are typically trained on ambient temperature datasets. The pulse waveform features they extract differ from those in low-temperature environments, and their detection accuracy and reliability are affected by changes in gas parameters. Conventional learning methods do not design adaptive mechanisms to address this inter-domain feature shift. Directly applying ambient temperature models to low-temperature environments results in significant errors in partial discharge detection, reducing diagnostic reliability. Summary of the Invention

[0004] In view of the shortcomings of the prior art, the purpose of this invention is to provide an intelligent diagnostic method for partial discharge of SF6 / N2 mixed gas in low-temperature environments, which can accurately diagnose the partial discharge situation of SF6 / N2 mixed gas and provide a reliable basis for the operation and maintenance of electrical equipment in low-temperature environments.

[0005] To achieve the above objectives, this invention provides an intelligent diagnostic method for partial discharge of SF6 / N2 mixed gas under low-temperature conditions, comprising the following steps:

[0006] S1. For SF6 / N2 mixed gas, collect multi-source partial discharge signals and environmental physical quantities when partial discharge occurs in the mixed gas, and construct source domain dataset and target domain dataset. The source domain dataset corresponds to normal temperature, and the target domain dataset corresponds to low temperature.

[0007] S2. For each sample in the source domain dataset and the target domain dataset, calculate the compressibility factor deviation ΔZ. ΔZ is a dimensionless physical quantity used to quantify the degree to which the mixed gas deviates from the ideal gas state.

[0008] S3. Construct a dual-stream adaptive feature fusion network, where the fusion weights are dynamically adjusted by ΔZ.

[0009] S4. Set the entropy weight adaptive loss function of the dual-stream adaptive feature fusion network, and input the source domain dataset, target domain dataset and the calculated ΔZ into the dual-stream adaptive feature fusion network for joint training.

[0010] S5. The temperature sensing channel pruning strategy is used to prune the dual-stream adaptive feature fusion network, and the pruned dual-stream adaptive feature fusion network is fine-tuned to obtain the final optimized diagnostic model.

[0011] S6. Real-time acquisition of multi-source partial discharge signals and environmental physical quantities when partial discharge occurs in mixed gas under actual working conditions, calculation of ΔZ, inputting real-time data and ΔZ into the optimized diagnostic model, and outputting diagnostic results.

[0012] In a preferred embodiment of the present invention, in S1, a temperature-controlled GIS simulated gas chamber experimental platform is constructed. When partial discharge occurs in the mixed gas, multi-source partial discharge signals and environmental physical quantities are simultaneously collected by sensing devices: an HFCT sensor connected in series with the equipment grounding wire couples the partial discharge pulse current signal; a UHF sensor is deployed to sense the ultra-high frequency electromagnetic wave signal radiated by partial discharge; a SiPM optical signal sensor detects the weak optical pulse signal generated by partial discharge; simultaneously, a Pt100 platinum resistance temperature sensor collects the temperature T of the mixed gas in real time; and a piezoresistive absolute pressure sensor measures the pressure P of the mixed gas. t ;

[0013] The acquired partial discharge pulse current signal, ultra-high frequency electromagnetic wave signal, and weak light pulse signal are amplified by their respective preamplifiers and then converted into digital signals by a high-speed data acquisition card to form synchronous multi-dimensional time series data. The digital signals are processed including filtering and noise reduction, and temperature and pressure calibration and standardization. Finally, source domain datasets and target domain datasets covering multi-source partial discharge signals and corresponding environmental physical quantities are constructed under normal temperature and low temperature scenarios. The normal temperature is 25°C±5°C and the low temperature is -40°C to 0°C.

[0014] As a preferred embodiment of the present invention, in step S2, the compression factor deviation ΔZ is obtained as follows:

[0015] ;

[0016] In the formula, Zactual The actual compression factor is calculated as follows:

[0017] ;

[0018] In the formula, P mix V represents the corrected equivalent pressure of the gas mixture; m R represents the molar volume of the gas mixture; R represents the universal gas constant; T represents the temperature of the gas mixture.

[0019] As a preferred embodiment of the present invention, P mix The calculation method is as follows:

[0020] ;

[0021] In the formula, , These represent the second-order and third-order mixed virial coefficients related to T, respectively, which are calculated according to the mixing rule and the temperature interpolation table. The gas correction factor, which is related to temperature T and used to compensate for the nonlinear pressure drop of SF6 molecules at low temperatures, is defined as follows:

[0022] ;

[0023] In the formula, T c This indicates the critical temperature of SF6; The cluster polymerization strength coefficient is obtained by fitting PVT experimental data of SF6 gas at low temperature. The pressure sensitivity coefficient is obtained through experimental fitting; P t P0 is the pressure of the gas mixture; P0 is the standard atmospheric pressure; e is the natural constant.

[0024] As a preferred embodiment of the present invention, in S3, the dual-stream adaptive feature fusion network includes:

[0025] The signal input layer normalizes and segments the multi-source partial discharge signal, and then uses it as a multi-channel input. The multi-channel includes a time-series feature extraction branch and a frequency domain feature extraction branch.

[0026] The temporal feature extraction branch employs a gated convolutional network. One-dimensional causal convolution is used to ensure temporal order, and gated linear units are used to learn the local impulse features of multi-source partial discharge signals along the time axis, outputting the temporal feature F. time ;

[0027] The frequency domain feature extraction branch employs a multi-scale dilated convolutional network. By setting one-dimensional dilated convolutional layers with different dilation rates in parallel, it captures different frequency components and features in multi-source partial discharge signals and outputs frequency domain features F. freq ;

[0028] Adaptive feature fusion gating module, receives F time F freq The features corresponding to ΔZ for the current sample are aggregated along the time dimension using global average pooling to obtain the global descriptor f for the temporal feature extraction branch. time The global descriptor f of the frequency domain feature extraction branch freq ;

[0029] f time and f freq By concatenating along the channel dimension, we obtain the concatenated feature f. concat The ΔZ corresponding to the current sample is copied and expanded to become a vector ΔZ that matches the batch size. vec , then f concat With ΔZ vec Concatenate again, and then concatenate the vectors. Input a lightweight subnetwork;

[0030] The subnetwork is followed by a Sigmoid activation function, which outputs a scalar fusion weight. Ultimately, the fusion features ;

[0031] F fused After several fully connected layers, a Softmax classifier maps the high-dimensional features to the probability distribution of each local emission type. The class with the highest output probability is the diagnostic result of the dual-stream adaptive feature fusion network.

[0032] As a preferred embodiment of the present invention, the subnetwork consists of two fully connected layers, with ReLU activation used in the middle, and the last fully connected layer is a single neuron fully connected layer.

[0033] As a preferred embodiment of the present invention, in S4, the entropy-weighted adaptive loss function is a dynamic loss function, expressed as:

[0034] ;

[0035] In the formula, L CE L represents the cross-entropy classification loss; MMD This represents the loss due to the maximum mean difference. The entropy weight adjustment function related to ΔZ is expressed as:

[0036] ;

[0037] In the formula, Indicates the basic migration weight; Z0 represents the gain coefficient; Z0 represents the normalized reference value; ln represents the logarithm to the base e.

[0038] During the training iteration, each batch of samples in the target domain dataset carries its calculated ΔZ.

[0039] As a preferred embodiment of the present invention, in S5, the temperature-sensing channel pruning strategy for the dual-stream adaptive feature fusion network specifically involves removing redundant channels with low diagnostic contribution and insensitivity to temperature changes from the convolutional layers of the dual-stream adaptive feature fusion network. During removal, a comprehensive score based on channel importance is used, and a temperature sensitivity bias term is introduced.

[0040] ;

[0041] In the formula, S c The importance score of channel c is indicated by W. c This represents the channel weight of channel c; Indicates the balance coefficient; This represents the temperature sensitivity bias term for channel c;

[0042] As a temperature sensitivity, forward propagation is performed on a pruned calibration dataset containing samples at different temperatures to calculate the L2 norm of the gradient of the average value of the output feature map of channel c with respect to the input variable T, where T is the temperature of the mixed gas.

[0043] For each convolutional layer, calculate the overall importance score of all channels, retain the top k% of channels with the highest scores, and remove the remaining channels and their corresponding connections and weights.

[0044] The pruned dual-stream adaptive feature fusion network is iteratively fine-tuned to obtain the final optimized diagnostic model.

[0045] As a preferred embodiment of the present invention, in step S6, the partial discharge type corresponding to the category with the highest output probability in the real-time data and ΔZ input optimization diagnostic model is taken as the diagnostic result.

[0046] The beneficial effects of this invention are:

[0047] This invention addresses the intelligent diagnosis of partial discharge in SF6 / N2 mixed gases. By introducing an improved state equation with a low-temperature correction factor and embedding a deep learning diagnostic process, the resulting compressibility factor deviation ΔZ index provides interpretable physical guidance, effectively avoiding the risk of failure of purely data-driven models under extreme conditions.

[0048] This invention achieves refined adaptive compensation for low-temperature characteristic drift through dual ΔZ adjustment. On the simulated low-temperature partial discharge test set, the diagnostic accuracy of this invention is higher than that of the method of directly applying the room temperature model to the low-temperature data.

[0049] This invention introduces temperature-sensing pruning, which not only reduces the number of parameters and speeds up inference in the final deployed model, but also retains the ability to extract key features that are sensitive to temperature changes, thereby enhancing the model's generalization ability and stability in complex low-temperature environments. Attached Figure Description

[0050] Figure 1 This is a schematic diagram illustrating the steps of the method of the present invention;

[0051] Figure 2 This is a schematic diagram of the overall process of Embodiment 1 of the present invention;

[0052] Figure 3 This is a comparison chart of accuracy rates in Embodiment 1 of the present invention. Detailed Implementation

[0053] The embodiments of the present invention will be further described below with reference to the accompanying drawings:

[0054] Example 1: As Figure 1 As shown, the intelligent diagnostic method for partial discharge of SF6 / N2 mixed gas under low-temperature conditions includes the following steps:

[0055] S1. For SF6 / N2 mixed gas, collect multi-source partial discharge signals and environmental physical quantities when partial discharge occurs in the mixed gas, and construct source domain dataset and target domain dataset. The source domain dataset corresponds to normal temperature, and the target domain dataset corresponds to low temperature.

[0056] S2. For each sample in the source domain dataset and the target domain dataset, calculate the compressibility factor deviation ΔZ. ΔZ is a dimensionless physical quantity used to quantify the degree to which the mixed gas deviates from the ideal gas state.

[0057] S3. Construct a dual-stream adaptive feature fusion network, where the fusion weights are dynamically adjusted by ΔZ.

[0058] S4. Set the entropy weight adaptive loss function of the dual-stream adaptive feature fusion network, and input the source domain dataset, target domain dataset and the calculated ΔZ into the dual-stream adaptive feature fusion network for joint training.

[0059] S5. The temperature sensing channel pruning strategy is used to prune the dual-stream adaptive feature fusion network, and the pruned dual-stream adaptive feature fusion network is fine-tuned to obtain the final optimized diagnostic model.

[0060] S6. Real-time acquisition of multi-source partial discharge signals and environmental physical quantities when partial discharge occurs in mixed gas under actual working conditions, calculation of ΔZ, inputting real-time data and ΔZ into the optimized diagnostic model, and outputting diagnostic results.

[0061] In S1, a temperature-controlled GIS simulated gas chamber experimental platform was built. When partial discharge occurred in the mixed gas, multi-source partial discharge signals and environmental physical quantities were simultaneously collected through sensing devices: an HFCT sensor connected in series with the equipment grounding wire was used to couple the partial discharge pulse current signal; a UHF sensor was deployed to sense the ultra-high frequency electromagnetic wave signal radiated by partial discharge (high-frequency alternating electromagnetic waves between 300kHz and 300GHz); a SiPM optical signal sensor was used to detect the weak light pulse signal generated by partial discharge (a light radiation signal with extremely weak intensity and extremely short duration generated by gas molecules due to ionization / excitation during partial discharge); simultaneously, the temperature T of the mixed gas was collected in real time using a Pt100 platinum resistance temperature sensor; and the pressure P of the mixed gas was measured using a piezoresistive absolute pressure sensor. t ;

[0062] The temperature-controlled GIS simulated gas chamber experimental platform can be built based on existing technologies: the core temperature-controlled sealed gas chamber can use a stainless steel high-pressure cavity with a commercial insulation layer, combined with a semiconductor cooling / liquid nitrogen circulation temperature control system (existing technology has achieved ±1℃ accuracy adjustment from -40℃ to room temperature); the gas system can integrate a mature SF6 / N2 mixed gas ratio device, a high-pressure charging / stabilizing module, and a leak detection component; defect simulation uses replaceable artificial defect electrodes (such as tips, air gaps, and floating potential electrodes, all of which are standard designs for partial discharge experiments); the signal acquisition part can directly use commercial HFCT (pulse current), UHF (ultra-high frequency), and SiPM (low light) sensors, combined with a standardized preamplifier and a high-speed data acquisition card (DAQ), and complete signal processing through existing data preprocessing algorithms (normalization / standardization, segmentation, noise reduction); the interface adaptation between modules and the closed-loop control of temperature control and the gas system are all supported by mature industrial control solutions, and only system integration and parameter debugging are required according to experimental needs.

[0063] The acquired partial discharge pulse current signal, ultra-high frequency electromagnetic wave signal, and weak light pulse signal are amplified by their respective preamplifiers and then converted into digital signals by a high-speed data acquisition card to form synchronous multi-dimensional time series data. The digital signals are processed including filtering and noise reduction, and temperature and pressure calibration and standardization. Finally, source domain datasets and target domain datasets covering multi-source partial discharge signals and corresponding environmental physical quantities are constructed under normal temperature and low temperature scenarios. The normal temperature is 25°C±5°C and the low temperature is -40°C to 0°C.

[0064] In S2, the compression factor deviation ΔZ is obtained as follows:

[0065] ;

[0066] In the formula, Z actual The actual compression factor is calculated as follows:

[0067] ;

[0068] In the formula, P mix V represents the corrected equivalent pressure of the gas mixture; m R represents the molar volume of the gas mixture; R represents the universal gas constant; T represents the temperature of the gas mixture.

[0069] Z actual Z is a direct indicator of the non-ideality of a gas. When the gas is in an ideal state, Z... actual The value of Z is equal to 1; if the gas is a non-ideal gas (such as an SF6 / N2 mixture at low temperature), Z actual It will deviate from 1. The greater the deviation, the stronger the non-ideality of the gas.

[0070] ΔZ is a key physical sensing signal. At normal temperature and pressure, ΔZ is close to 0. As the temperature decreases, the pressure increases, or the proportion of SF6 increases, the ΔZ value increases, directly and continuously reflecting the degree of physical deviation between the current operating environment and the model's standard training environment (usually a near-ideal gas state).

[0071] P mix The calculation method (improved equation of state for mixed gases) is as follows:

[0072] ;

[0073] In the formula, , These represent the second-order and third-order mixed virial coefficients related to T, respectively, calculated according to the mixing rule (such as the Kay rule) and the temperature interpolation table (one column is the predetermined standard temperature point, and the other column is the pure virial coefficient of SF6 or N2 at the corresponding temperature, which is a known value obtained through experimental or theoretical calculation). The gas correction factor, which is related to temperature T and used to compensate for the nonlinear pressure drop of SF6 molecules at low temperatures, is defined as follows:

[0074] ;

[0075] In the formula, T c This indicates the critical temperature of SF6; The cluster polymerization strength coefficient is a dimensionless empirical parameter related to the intermolecular forces of SF6. Its value is usually between 0.05 and 0.3, and it is obtained by fitting PVT experimental data of SF6 gas at low temperature. The pressure sensitivity coefficient, used to adjust the degree of influence of pressure on the correction factor, is obtained through experimental fitting and is set to 1.2; P tP0 is the pressure of the gas mixture; P0 is the standard atmospheric pressure (101.325 kPa); e is the natural constant.

[0076] The formula is used when T is lower than T. c Activated by time, it corrects calculation biases in the standard virial equation. When T is greater than T0. c When, the exp term ( The value approaches 0. , It degenerates into the regular virial equation. When T is less than Tc, the exp term takes effect. Its function is to measure the pressure P at low temperatures. t This leads to a corrected virtual pressure that is closer to the pressure value that an ideal gas should exhibit, thus correcting the problem of low pressure measurements.

[0077] In S3, the dual-stream adaptive feature fusion network includes:

[0078] The signal input layer normalizes and segments the multi-source partial discharge signal, and then uses it as a multi-channel input. The multi-channel includes a time-series feature extraction branch and a frequency domain feature extraction branch.

[0079] The temporal feature extraction branch employs a gated convolutional network. One-dimensional causal convolution is used to ensure temporal order, and gated linear units are used to learn the local impulse features of multi-source partial discharge signals along the time axis, outputting the temporal feature F. time ;

[0080] The frequency domain feature extraction branch employs a multi-scale dilated convolutional network. By setting one-dimensional dilated convolutional layers with different dilation rates (e.g., 1, 2, 4) in parallel, the receptive field is expanded without increasing the parameters. Branches with smaller dilation rates focus on details and high-frequency components, while branches with larger dilation rates have wider receptive fields, enabling them to capture broader contextual information and low-frequency trends. This allows for the capture of different frequency components and features in multi-source partial discharge signals, outputting the frequency domain feature F. freq ;

[0081] Adaptive feature fusion gating module, receives F time F freq The features corresponding to ΔZ for the current sample are aggregated along the time dimension using global average pooling to obtain the global descriptor f for the temporal feature extraction branch. time The global descriptor f of the frequency domain feature extraction branch freq ;

[0082] f time and f freq By concatenating along the channel dimension, we obtain the concatenated feature f. concatThe ΔZ corresponding to the current sample is copied and expanded to become a vector ΔZ that matches the batch size. vec , then f concat With ΔZ vec Concatenate again, and then concatenate the vectors. Input a lightweight subnetwork;

[0083] The subnetwork is followed by a Sigmoid activation function, which outputs a scalar fusion weight. A β close to 1 indicates that the network considers the temporal features more reliable under the current ΔZ condition; a β close to 0 indicates that the frequency domain features are more reliable, and the final fused features... This means that for samples with a large ΔZ (significant low-temperature effect), the network may automatically adjust β, focusing more on extracting relatively stable features affected by the medium from the frequency domain flow.

[0084] F fused After several (two) fully connected layers, a Softmax classifier maps the high-dimensional features to the probability distribution of each local discharge type. The class with the highest output probability is the diagnostic result of the dual-stream adaptive feature fusion network.

[0085] The subnetwork consists of two fully connected layers, with ReLU activation in between, and the last fully connected layer is a single neuron.

[0086] In S4, the entropy-weighted adaptive loss function is a dynamic loss function, expressed as:

[0087] ;

[0088] In the formula, L CE L represents the cross-entropy classification loss; MMD This represents the loss due to the maximum mean difference. The entropy weight adjustment function related to ΔZ is expressed as:

[0089] ;

[0090] In the formula, This represents the base migration weight, set to 0.5; Z0 represents the gain coefficient, set to 2.0; Z0 represents the normalized reference value, set to 0.1; ln represents the logarithm to the base e.

[0091] During training iterations, each batch of samples in the target domain dataset carries its calculated ΔZ. When encountering a sample with a large ΔZ, Approaching 1, Increase, so that L MMDWhen the weight of the loss term increases, the gas environment of the sample differs greatly from that of the source domain, and the domain adaptation strength needs to be enhanced. Conversely, for samples with small ΔZ, the domain adaptation strength should be appropriately reduced to avoid excessive distortion of the learned useful features.

[0092] After training, structured pruning is performed to obtain a model that is more focused on key features of low temperature.

[0093] In S5, a temperature-sensing channel pruning strategy is used to prune the dual-stream adaptive feature fusion network. Specifically, in the convolutional layers of the dual-stream adaptive feature fusion network, redundant channels with low diagnostic contribution and insensitivity to temperature changes are removed. During the removal process, a comprehensive score of channel importance is used, and a temperature sensitivity bias term is introduced.

[0094] ;

[0095] In the formula, S c The importance score of channel c is indicated by W. c This represents the channel weight of channel c; This represents the balance coefficient, which is set to 0.7. This represents the temperature sensitivity bias term for channel c;

[0096] As a temperature sensitivity, forward propagation is performed on a pruned calibration dataset containing samples at different temperatures to calculate the L2 norm of the gradient of the average value of the output feature map of channel c with respect to the input variable T, where T is the temperature of the mixed gas. A high value indicates that the activation intensity of channel c changes drastically with temperature, and it is very likely to capture key discrimination information related to low temperature.

[0097] For each convolutional layer, calculate the overall importance score of all channels, retain the top k% (30%~70%) channels with the highest scores, and remove the remaining channels and their corresponding connections and weights;

[0098] The pruned dual-stream adaptive feature fusion network was fine-tuned with a few iterations to obtain the final optimized diagnostic model.

[0099] In S6, the partial discharge type corresponding to the category with the highest output probability in the diagnostic model, based on real-time data and ΔZ input, is used as the diagnostic result, including:

[0100] Tip discharge: Partial discharge caused by sharp / burr defects in metal parts inside the equipment;

[0101] Surface discharge: Partial discharge that occurs on the surface of solid insulation (such as the surface of an insulator);

[0102] Internal air gap discharge: Partial discharge generated in tiny air gaps inside insulating materials;

[0103] Floating potential discharge: Partial discharge caused by a floating potential formed by poor contact of metal parts (such as shielding covers) in the equipment.

[0104] Based on the above, the overall flowchart of this embodiment is as follows: Figure 2 As shown.

[0105] An application example of the method in this embodiment is as follows:

[0106] First, in a temperature-controlled GIS simulation chamber (a sealed experimental cavity used to simulate the actual operating environment of gas-insulated switchgear GIS, which can be precisely controlled in temperature, and is the core of the temperature-controlled GIS simulation chamber experimental platform), a typical metal tip defect (simulating tip discharge) is set up. The UHF, HFCT and optical signals of partial discharge are collected by the data acquisition system. At the same time, the temperature sensor and the pressure sensor measure the temperature and pressure of the gas environment, respectively, to provide key parameters for subsequent processing.

[0107] At room temperature (25°C ± 2°C), different voltage levels were applied to the defects to collect stable partial discharge signals, resulting in approximately 500 effective discharge pulse sequences and a total of 500 room temperature samples. Each sample included synchronized HFCT, UHF, and SiPM waveform data (10 μs in length, i.e., 2500 sampling points) and corresponding temperature and pressure readings. The ambient temperature control chamber was set to -30°C. After the gas temperature inside the chamber stabilized, the discharge experiment was repeated, yielding approximately 70 effective samples.

[0108] The collected signal data underwent preliminary processing. Pulse current signals (partial discharge pulse current signals), UHF signals (ultra-high frequency electromagnetic wave signals), and optical signals (weak optical pulse signals) were filtered to remove noise interference and extract effective partial discharge characteristic signals. Simultaneously, data collected from temperature and pressure sensors were calibrated and standardized to ensure data consistency and accuracy.

[0109] For each sample in the dataset (whether from the source or target domain), a low-temperature correction factor is activated based on the collected temperature and pressure data. Calculate P mix ,in V m Calculated based on the mixture ratio, molar mass, and measured density. , The values ​​were obtained by consulting the virial coefficient tables for SF6 and N2 and interpolating at 243.15K according to the Kay mixing rule.

[0110] set up , (Determined by fitting previous PVT experimental data of SF6 / N2 mixed gas from -40°C to 0°C), P0 = 0.101325 MPa, calculated as follows: .

[0111] Using P mix Calculate the actual compressibility factor Z of the gas mixture in this sample. actual Calculate the compression factor deviation ΔZ of the sample, and use this ΔZ value as a key physical feature label of the sample, storing it together with the original waveform data.

[0112] A low-temperature adaptive intelligent diagnostic network (dual-stream adaptive feature fusion network) is constructed. After normalizing the three-channel waveforms (HFCT, UHF, SiPM) of each sample, a multi-channel input is formed. , Let F represent a B×3×2500 dimensional real space, where B is the batch size. The temporal feature extraction branch consists of four layers of alternating one-dimensional causal convolutions (kernel size 5) and GLU gated units (gated linear units), ultimately outputting temporal features F. time The frequency domain branch contains three sets of parallel dilated convolutional layers with dilation rates of 1, 2, and 4, followed by 1×1 convolutional fusion to output the frequency domain feature F. freq ;

[0113] The adaptive feature fusion gating module receives F time F freq The ΔZ corresponding to the current sample has two fully connected subnetworks (32 and 16 neurons respectively, using ReLU activation), which ultimately output a scalar fusion weight β. The fusion feature F... fused After passing through two fully connected layers (128-dimensional and 64-dimensional), the Softmax layer with 3 neurons outputs the probabilities of three types of discharges (corona discharge, surface discharge, and suspension discharge), among which corona discharge is a type of tip discharge.

[0114] The entropy-weighted adaptive loss function was used for training, with an initial learning rate of 0.001 and a batch size of 32. During training, , Z0 = 0.1. During training iterations, for each target domain sample in a batch, its pre-computed ΔZ is read. After training, the performance of the initial model is evaluated on an independent validation set. Forward propagation is performed using a calibration dataset containing multiple temperature points from -35°C to 25°C. For each convolutional channel in the network (two-stream adaptive feature fusion network), the gradient of its output with respect to the input temperature variable T is calculated, and its temperature sensitivity is obtained by taking the L2 norm. For each convolutional layer, the overall score S for each channel is calculated according to the formula. cThe global pruning rate is set to 30%, meaning that the lowest 30% of channels in each layer are removed. The pruned lightweight network is then fine-tuned on the training set with a low learning rate (0.0001) to restore its accuracy to the level before pruning.

[0115] The multi-source partial discharge signals and environmental physical quantities (environmental parameters) collected on-site are input into the optimized diagnostic network (optimized diagnostic model), which outputs the type of partial discharge. Taking tip discharge as an example, the accuracy of the method in this embodiment and the conventional method for identifying low-temperature partial discharge are compared as follows: Figure 3 As shown, the accuracy of the conventional method is about 80%, while the accuracy of the method in this embodiment (low temperature adaptive method) is significantly higher (close to 95%), which intuitively demonstrates that the method in this embodiment is superior to the conventional method.

[0116] Example 2: The difference between this example and Example 1 is that, after calculating the compression factor deviation ΔZ in S2, a physical feature vector V is further constructed. phy In subsequent steps, use V phy Replace ΔZ.

[0117] V phy The vector encapsulates the physical essence of the changes in gas insulation properties (effective ionization coefficient) under current low-temperature environments. The electron drift velocity (ve) allows the dual-flow adaptive feature fusion network to directly obtain the gas microphysical properties altered by cryogenic temperatures. phy The construction method is as follows:

[0118] Based on the temperature T and pressure P of the mixed gas t and actual compression factor Calculate the gas molecule number density N:

[0119] ;

[0120] In the formula, kB is the Boltzmann constant;

[0121] Based on N and the preset electric field strength E / N of the mixed gas (E represents the electric field strength), the electric field strength of the mixed gas is calculated using empirical formulas. and ve (which can be calculated using existing technology);

[0122] Constructing physical feature vectors .

[0123] V phy It provides multi-dimensional physical information that combines macroscopic deviations and microscopic processes, which can help the dual-stream adaptive feature fusion network to more accurately match the signal feature patterns in low-temperature scenarios and improve the model's adaptability to low-temperature environments.

[0124] Example 3: Intelligent diagnostic device for partial discharge of SF6 / N2 mixed gas under low temperature environment, comprising:

[0125] One or more processors;

[0126] Memory, used to store one or more computer programs;

[0127] When one or more programs are executed by one or more processors, the one or more processors perform the method in Embodiment 1 or Embodiment 2.

[0128] Example 4: A computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, cause the processor to perform the method in Example 1 or Example 2.

Claims

1. A method for intelligent diagnosis of partial discharge in SF6 / N2 mixed gas under low-temperature conditions, characterized in that... Includes the following steps: S1. For SF6 / N2 mixed gas, collect multi-source partial discharge signals and environmental physical quantities when partial discharge occurs in the mixed gas, and construct source domain dataset and target domain dataset. The source domain dataset corresponds to normal temperature, and the target domain dataset corresponds to low temperature. S2. For each sample in the source domain dataset and the target domain dataset, calculate the compressibility factor deviation ΔZ. ΔZ is a dimensionless physical quantity used to quantify the degree to which the mixed gas deviates from the ideal gas state. S3. Construct a dual-stream adaptive feature fusion network, where the fusion weights are dynamically adjusted by ΔZ. S4. Set the entropy weight adaptive loss function of the dual-stream adaptive feature fusion network, and input the source domain dataset, target domain dataset and the calculated ΔZ into the dual-stream adaptive feature fusion network for joint training. S5. The temperature sensing channel pruning strategy is used to prune the dual-stream adaptive feature fusion network, and the pruned dual-stream adaptive feature fusion network is fine-tuned to obtain the final optimized diagnostic model. S6. Real-time acquisition of multi-source partial discharge signals and environmental physical quantities when partial discharge occurs in mixed gas under actual working conditions, calculation of ΔZ, input of real-time data and ΔZ into the optimized diagnostic model, and output of diagnostic results; In S1, a temperature-controlled GIS simulated gas chamber experimental platform was built. When partial discharge occurred in the mixed gas, multi-source partial discharge signals and environmental physical quantities were simultaneously collected by sensing devices: an HFCT sensor connected in series with the equipment grounding wire coupled the partial discharge pulse current signal; a UHF sensor was deployed to sense the ultra-high frequency electromagnetic wave signal radiated by partial discharge; a SiPM optical signal sensor was used to detect the weak optical pulse signal generated by partial discharge; simultaneously, a Pt100 platinum resistance temperature sensor was used to collect the temperature T of the mixed gas in real time; and a piezoresistive absolute pressure sensor was used to measure the pressure P of the mixed gas. t ; In S2, the compression factor deviation ΔZ is obtained as follows: ； In the formula, Z actual The actual compression factor is calculated as follows: ； In the formula, P mix V represents the corrected equivalent pressure of the gas mixture; m R represents the molar volume of the gas mixture; R represents the universal gas constant; T represents the temperature of the gas mixture. P mix The calculation method is as follows: ； In the formula, , These represent the second-order and third-order mixed virial coefficients related to T, respectively, which are calculated according to the mixing rule and the temperature interpolation table. The gas correction factor, which is related to temperature T and used to compensate for the nonlinear pressure drop of SF6 molecules at low temperatures, is defined as follows: ； In the formula, T c This indicates the critical temperature of SF6; The cluster polymerization strength coefficient is obtained by fitting PVT experimental data of SF6 gas at low temperature; The pressure sensitivity coefficient is obtained through experimental fitting; P t P0 is the pressure of the gas mixture; P0 is the standard atmospheric pressure. e is a natural constant; In S3, the dual-stream adaptive feature fusion network includes: The signal input layer normalizes and segments the multi-source partial discharge signal, and then uses it as a multi-channel input. The multi-channel includes a time-series feature extraction branch and a frequency domain feature extraction branch. The temporal feature extraction branch employs a gated convolutional network. One-dimensional causal convolution is used to ensure temporal order, and gated linear units are used to learn the local impulse features of multi-source partial discharge signals on the time axis, outputting the temporal feature F. time ; The frequency domain feature extraction branch employs a multi-scale dilated convolutional network. By setting one-dimensional dilated convolutional layers with different dilation rates in parallel, it captures different frequency components and features in multi-source partial discharge signals and outputs frequency domain features F. freq ; Adaptive feature fusion gating module, receives F time F freq The ΔZ corresponding to the current sample is aggregated along the time dimension using global average pooling to obtain the global descriptor f of the temporal feature extraction branch. time The global descriptor f of the frequency domain feature extraction branch freq ; f time and f freq By concatenating along the channel dimension, we obtain the concatenated feature f. concat The ΔZ corresponding to the current sample is copied and expanded to become a vector ΔZ that matches the batch size. vec , then f concat With ΔZ vec Concatenate again, and then concatenate the vectors. Input a lightweight subnetwork; The subnetwork is followed by a Sigmoid activation function, which outputs a scalar fusion weight. Ultimately, the fusion features ; F fused After several fully connected layers, a Softmax classifier maps the high-dimensional features to the probability distribution of each local discharge type. The class with the highest output probability is the diagnostic result of the dual-stream adaptive feature fusion network.

2. The intelligent diagnostic method for partial discharge of SF6 / N2 mixed gas under low-temperature environment according to claim 1, characterized in that, In S1, the acquired partial discharge pulse current signal, ultra-high frequency electromagnetic wave signal, and weak light pulse signal are amplified by their respective preamplifiers and then converted into digital signals by a high-speed data acquisition card to form synchronous multi-dimensional time series data. The digital signals are processed including filtering and noise reduction, and temperature and pressure calibration and standardization. Finally, source domain datasets and target domain datasets covering multi-source partial discharge signals and corresponding environmental physical quantities are constructed under normal temperature and low temperature scenarios. The normal temperature is 25°C±5°C and the low temperature is -40°C to 0°C.

3. The intelligent diagnostic method for partial discharge of SF6 / N2 mixed gas under low-temperature environment according to claim 1, characterized in that, The subnetwork consists of two fully connected layers, with ReLU activation in between, and the last fully connected layer is a single neuron.

4. The intelligent diagnostic method for partial discharge of SF6 / N2 mixed gas under low-temperature environment according to claim 1, characterized in that, In S4, the entropy-weighted adaptive loss function is a dynamic loss function, expressed as: ； In the formula, L CE L represents the cross-entropy classification loss; MMD This represents the loss due to the maximum mean difference. The entropy weight adjustment function related to ΔZ is expressed as: ； In the formula, Indicates the basic migration weight; Z0 represents the gain coefficient; Z0 represents the normalized reference value; ln represents the logarithm to the base e. During the training iteration, each batch of samples in the target domain dataset carries its calculated ΔZ.

5. The intelligent diagnostic method for partial discharge of SF6 / N2 mixed gas under low-temperature environment according to claim 1, characterized in that, In S5, the temperature-sensing channel pruning strategy for the dual-stream adaptive feature fusion network specifically involves removing redundant channels with low diagnostic contribution and insensitivity to temperature changes from the convolutional layers of the network. During this removal, a comprehensive score based on channel importance is used, and a temperature sensitivity bias term is introduced. ； In the formula, S c The importance score of channel c is indicated by W. c This represents the channel weight of channel c; Indicates the balance coefficient; This represents the temperature sensitivity bias term for channel c; As a temperature sensitivity, forward propagation is performed on a pruned calibration dataset containing samples at different temperatures to calculate the L2 norm of the gradient of the average value of the output feature map of channel c with respect to the input variable T, where T is the temperature of the mixed gas. For each convolutional layer, calculate the overall importance score of all channels, retain the top k% of channels with the highest scores, and remove the remaining channels and their corresponding connections and weights. The pruned dual-stream adaptive feature fusion network is iteratively fine-tuned to obtain the final optimized diagnostic model.

6. The intelligent diagnostic method for partial discharge of SF6 / N2 mixed gas under low-temperature environment according to claim 1, characterized in that, In S6, the partial discharge type corresponding to the category with the highest output probability in the real-time data and ΔZ input optimization diagnostic model is used as the diagnostic result.

Images