Intelligent identification linkage protection method for 35kV and below equipment based on multi-source adaptive fusion and deep and shallow double-path network

Abstract

The invention discloses an intelligent identification linkage protection method for 35kV and below equipment based on multi-source adaptive fusion and a deep and shallow double-path network. The method comprises the following steps: acquiring multi-source data of target equipment, splitting the multi-source data into instantaneous and steady-state feature flows, extracting features by using a deep and shallow dual-path network, and performing dynamic self-adaptive fusion according to a real-time operation scene to obtain comprehensive features; inputting a time sequence formed by continuous comprehensive features into an LSTM network subjected to equipment exclusive gating adjustment, and outputting a combined recognition result of an equipment type and a working state and an initial confidence coefficient in combination with a hierarchical time sequence matching mechanism; the confidence coefficient is calibrated according to the characteristic fluctuation degree and the sample difficulty, and finally the protection strategy is adjusted in a hierarchical linkage mode according to the calibrated confidence coefficient. According to the method, the dynamic scene adaptability and the identification accuracy of a small number of devices are improved, the risk of misjudgment and missed judgment is effectively reduced through linkage of confidence coefficient calibration and a dynamic constant value, and the reliability of a protection system is enhanced.

Description

Technical Field

[0001] This invention relates to the field of power system protection and operation and maintenance technology, specifically to a method for intelligent identification and adaptive protection of 35kV and below power distribution equipment. Background Technology

[0002] In the field of power system protection, traditional relay protection methods, such as overcurrent, overvoltage, overload, underfrequency load shedding, and undervoltage load shedding, typically rely on preset static settings and switching on / off switches. These methods use basic electrical quantities such as current and voltage collected on-site as criteria and are set once during equipment commissioning, then run continuously thereafter. However, traditional static modes fail to fully consider the dynamic changes in the equipment's own state parameters (such as insulation resistance and dielectric loss angle) and system operating conditions (such as upstream and downstream circuit parameters and load levels) during long-term operation. Therefore, they are ill-suited to long-term operating scenarios such as equipment aging and load fluctuations, resulting in low fault identification accuracy for niche equipment with special parameters or indistinct operating characteristics, such as reactors.

[0003] To improve fault identification accuracy, several improvement schemes have been proposed by those skilled in the art. One approach is to use a fault diagnosis method based on convolutional neural networks, extracting features through single-path convolution operations for identification. Although this method improves accuracy to some extent, its network structure struggles to simultaneously meet the dual requirements of rapid fault removal and accurate fault identification. According to feedback from practical applications, approximately 15% to 20% of faults still cannot be effectively identified. Another approach attempts to introduce a fixed-weight feature contribution allocation mechanism to determine faults, aiming to strengthen the role of key features. However, this fixed-weight allocation method is poorly adaptable to dynamic operating scenarios such as drastic load fluctuations caused by motor start-up and shutdown, and power grid harmonic interference. Moreover, this method typically relies solely on the volatility of a single feature to correct the confidence level of the model output, making it prone to "overconfidence" misjudgments due to instantaneous feature mutations, or "insensitive" missed judgments due to slow steady-state feature changes, resulting in a high overall misjudgment rate. In addition, the settings in existing protection strategies are mostly static presets, which fail to be linked and dynamically adjusted with the real-time identification results and health status of the equipment. This makes it impossible for the protection system to adaptively follow the aging of the equipment and changes in load, thus limiting its long-term operational reliability. Summary of the Invention

[0004] This invention proposes an intelligent identification and linkage protection method for 35kV and below equipment based on multi-source adaptive fusion and deep / shallow dual-path networks. Its objectives are: 1. To effectively improve the accuracy of fault identification in power system protection, especially enhancing the identification capability for less common equipment such as reactors. 2. To improve the adaptability of the protection method to dynamic operating scenarios such as load fluctuations and harmonic interference, avoiding overconfidence-induced misjudgments and undersensitivity-induced missed judgments caused by fixed weight allocation and single feature correction. 3. To enable the threshold parameters to be dynamically adjusted based on the real-time identification results and health status of the equipment, adapting to long-term operating scenarios of equipment aging and load changes, thereby enhancing the overall reliability of the protection system.

[0005] The technical solution of this invention is as follows:

[0006] A method for intelligent identification and linkage protection of 35kV and below equipment based on multi-source adaptive fusion and deep and shallow dual-path network. Steps S1 to S5 are executed cyclically according to a preset sampling frequency. The comprehensive features of the target equipment at the current time are obtained in each execution cycle. The comprehensive features of multiple consecutive execution cycles constitute a time-series feature sequence, which is used for subsequent time-series identification and protection judgment.

[0007] Step S1: Collect multi-source data from the target device to obtain multi-dimensional features;

[0008] Step S2: Perform dimensionality optimization and feature splitting on the multidimensional features to obtain the instantaneous feature stream and the steady-state feature stream;

[0009] Step S3: Input the instantaneous feature stream and the steady-state feature stream into the deep and shallow dual-path network. Extract the instantaneous mutation features from the instantaneous feature stream through the shallow path and extract the steady-state global features from the steady-state feature stream through the deep path.

[0010] Step S4: Based on the real-time operating scenario of the target device, the instantaneous mutation features and steady-state global features are dynamically and adaptively fused to obtain the comprehensive features of the target device at the current moment.

[0011] Step S5: Input the temporal feature sequence composed of the comprehensive features of multiple consecutive time points into the LSTM network, enhance the temporal correlation through the hierarchical temporal matching mechanism, and output the combined recognition result of "device type-working status" and the corresponding initial confidence.

[0012] Step S6: Determine whether the initial confidence level is overconfident or undersensitive based on the volatility of the transient mutation features and the variance of the steady-state global features. Combine the sample difficulty coefficient to dynamically calculate the penalty coefficient and calibrate the initial confidence level to obtain the calibrated confidence level.

[0013] Step S7: Dynamically adjust the protection strategy of the target device based on the post-calibration confidence level.

[0014] As a further improvement to the intelligent identification and linkage protection method for 35kV and below equipment based on multi-source adaptive fusion and deep and shallow dual-path networks, the multi-source data collected in step S1 comprises a total of 18 dimensions:

[0015] (1) Electrical characteristics, 10 dimensions in total;

[0016] The static characteristics include the effective value of the three-phase current of the target equipment. , , Line voltage RMS value , , Power factor and zero-sequence current Dynamic characteristics include the maximum rate of change of three-phase current. and voltage fluctuation coefficient ;

[0017] Among them, voltage fluctuation coefficient The calculation formula is:

[0018]

[0019] In the above formula, and These represent the maximum and minimum line voltage values ​​for a single target device across multiple consecutive sampling periods. This is the rated line voltage of the target device;

[0020] (2) Test data, in four dimensions, including insulation resistance Dielectric loss tangent DC resistance and no-load loss ;

[0021] (3) Inspection data, in 4 dimensions, including equipment appearance status Operating noise Oil leakage status and instrument display status .

[0022] As a further improvement to the intelligent identification and linkage protection method for 35kV and below equipment based on multi-source adaptive fusion and deep and shallow dual-path networks, step S2 specifically includes:

[0023] Step S2.1: Based on the influence direction of the feature indicators on the equipment status, the 18-dimensional features obtained in step S1 are divided into extremely large indicators and extremely small indicators, and standardized respectively.

[0024] Step S2.2: Kernel principal component analysis is used to perform nonlinear dimensionality reduction on the standardized 18-dimensional features, retaining principal components with a cumulative contribution rate not less than a preset ratio to obtain KPCA fusion features; at the same time, the loading vectors of each principal component are analyzed to determine whether each principal component is a steady-state principal component.

[0025] Step S2.3: Based on the standardized feature indices and the principal components obtained from kernel principal component analysis, obtain the instantaneous feature flow. and steady-state characteristic flow ;

[0026] Instantaneous Feature Flow There are 7 dimensions in total, including the standardized ones. , , It also includes the initiation feature entropy And the three instantaneous principal components with the highest contribution rates;

[0027] Among them, the initiation feature entropy It is obtained during the equipment startup phase using the following calculation formula:

[0028]

[0029] In the above formula, This is the sampling sequence number during the startup phase. For the first The normalized probability of the current for each sampling period is derived from the mean of the effective values ​​of the three-phase currents in that period. The calculation shows that, ;

[0030] Steady-state characteristic flow There are 7 dimensions in total, including the standardized ones. , , , , , And the steady-state principal component with the highest contribution rate.

[0031] As a further improvement to the intelligent identification and linkage protection method for 35kV and below equipment based on multi-source adaptive fusion and deep and shallow dual-path networks, the feature extraction process in step S3 is as follows:

[0032] Step S3.1: Calculate the instantaneous feature flow obtained in step S2. Input the shallow path of the deep-shallow dual-path network; the shallow path consists of 3 lightweight convolutional layers, and the final output transient mutation feature is denoted as... ;

[0033] Step S3.2: Calculate the steady-state characteristic flow obtained in step S2. Input the deep path of the shallow-deep dual-path network; the deep path first undergoes preliminary processing through two depthwise separable convolutions, and the output is denoted as... ; then After deep feature extraction using 3 layers of pre-activated residual blocks, the output is denoted as... Finally, a feature filter is used to... Weighting is performed to obtain steady-state global features .

[0034] As a further improvement to the intelligent identification and linkage protection method for 35kV and below equipment based on multi-source adaptive fusion and deep and shallow dual-path networks, the formula for calculating the importance weight of steady-state features by the feature filter is as follows:

[0035]

[0036] In the above formula, It is the sigmoid activation function. and The parameters learned during training, the weight vector Dimensions and Consistent, express The The element, i.e., the th element Importance weights of steady-state features;

[0037] Weighted steady-state global characteristics Calculated via element-wise product: ,in This indicates element-wise multiplication.

[0038] As a further improvement to the intelligent identification and linkage protection method for 35kV and below equipment based on multi-source adaptive fusion and deep and shallow dual-path networks, step S4 specifically includes:

[0039] Step S4.1: Calculate the load density of the real-time running scenario parameters. and degree of interference And determine the current scene type;

[0040] Load density The calculation formula is:

[0041]

[0042] The value range of is [0, 2];

[0043] Interference level The calculation formula is:

[0044]

[0045] The value range is [0, 1];

[0046] According to load density and degree of interference Determine the current scene type;

[0047] Step S4.2: Based on the scene template library Dynamic calculation of instantaneous feature fusion weights ;

[0048] The calculation formula is:

[0049]

[0050] In the above formula, This represents the calculation of mutual information entropy. The current 7-dimensional transient mutation features With the current scene type In the scene template library The corresponding 7-dimensional instantaneous template Mutual information entropy, Represents the current 7-dimensional steady-state global characteristics With the current scene type In the scene template library The corresponding 7-dimensional steady-state template Mutual information entropy;

[0051] Step S4.3: Fuse weights based on instantaneous features Transient mutation characteristics and steady-state global characteristics Perform feature fusion to generate comprehensive features ;

[0052] The fusion method is as follows:

[0053]

[0054] In the above formula, This represents a 1×1 convolution operation, used to unify feature dimensions; For the residual term, the calculation formula is: ;

[0055] The scene template library The method of obtaining it is:

[0056] For each scenario, at least 1000 sets of 7-dimensional instantaneous feature streams and 7-dimensional steady-state feature streams from the target devices are collected. Then, using fixed weights, the corresponding instantaneous and steady-state feature streams of each set are weighted and fused into 7-dimensional pseudo-mixed features. A 1×1 convolutional layer is then used to upscale the pseudo-mixed features to 128 dimensions, resulting in 128-dimensional pseudo-fusion features. Finally, K-means clustering is performed on the 128-dimensional pseudo-fusion features, 7-dimensional instantaneous feature streams, and 7-dimensional steady-state feature streams for each scenario, with clustering parameter K=1, to obtain the corresponding 128-dimensional fusion template, 7-dimensional instantaneous template, and 7-dimensional steady-state template for each scenario, forming a scenario template library. .

[0057] As a further improvement to the intelligent identification and linkage protection method for 35kV and below equipment based on multi-source adaptive fusion and deep and shallow dual-path networks, step S5 specifically includes:

[0058] Step S5.1: Construct a time-series feature sequence;

[0059] Continuous acquisition target device 128-dimensional composite features of the frame , constitutes a time-series feature sequence The superscript indicates the frame number;

[0060] Step S5.2: Use the device type classifier to predict the device type of the current target device;

[0061] Before inputting the time sequence into the LSTM, the synthesized features of any frame are first processed. Input a lightweight device type classifier, which is a fully connected network with two hidden layers, and finally output the probability distribution of the target device type through a Softmax layer. ,Pick The type corresponding to the maximum value is used as the device type prediction result;

[0062] Step S5.3: Transform the time-series feature sequence Input the data into the LSTM network and perform device-specific gating adjustments based on the predicted device type;

[0063] The main structure of the LSTM network is uniform, but its internal control parameters are dynamically adjusted based on the device type prediction results obtained in step S5.2. The specific adjustment method is as follows:

[0064] (1) If the prediction is that it is an electric motor, then strengthen the forget gate and introduce a load adaptation factor. , The enhanced forget gate corresponds to the load density at a given time. The calculation formula is:

[0065]

[0066] (2) If the problem is predicted to be a transformer, then strengthen the input gate and introduce an interference adaptation factor. , To mitigate interference, the enhanced input gate The calculation formula is:

[0067]

[0068] In the above formula, The hidden state of the LSTM at time t-1. Let be the input features at time t. , , , The parameters learned during training, For batch normalization, It is the sigmoid activation function;

[0069] (3) For capacitors and reactors, universal gating parameters are used;

[0070] Step S5.4: In the LSTM network processing, a hierarchical timing matching mechanism is introduced;

[0071] Specifically, the length is The temporal feature sequence is divided into multiple time window levels such as short-term, medium-term, and long-term. Within each level, an attention mechanism is used to calculate the cosine similarity between frames, and only strong correlations with similarity not lower than a preset relationship threshold are retained, thereby reducing cross-level interference.

[0072] Step S5.5: Obtain the combined recognition result and initial confidence level based on the LSTM network;

[0073] The final output layer of the LSTM network is a Softmax layer, whose output is a probability distribution of the "device type-operating state" combination category; this probability distribution contains The probability value with the largest value in the probability distribution is taken as the initial confidence level. The corresponding "device type - working status" combination is the combination identification result of the current target device.

[0074] As a further improvement to the intelligent identification and linkage protection method for 35kV and below equipment based on multi-source adaptive fusion and deep and shallow dual-path networks, step S6 specifically includes:

[0075] Step S6.1: Determining overconfidence versus undersensitivity;

[0076] Set up a sliding window containing 5 consecutive frames of data, and calculate the 7-dimensional transient mutation features within the window. volatility and 7-dimensional steady-state global features variance ;

[0077] Instantaneous characteristic volatility The calculation formula is:

[0078]

[0079] In the above formula, , , Representing the sliding window respectively The maximum, minimum, and mean values;

[0080] The rules for judging overconfidence and lack of sensitivity are as follows: If Greater than the first threshold and initial confidence level If it exceeds the second threshold, it is judged as "overconfidence in transient features"; if Less than the third threshold and If the value is less than the fourth threshold, it is judged as "insufficiently sensitive steady-state characteristics";

[0081] Step S6.2: Based on the comprehensive features and scene template library at the current moment Calculate the sample difficulty coefficient :

[0082]

[0083] In the above formula, The 128-dimensional comprehensive features at the current moment, For the current scene type In the scene template library The corresponding 128-dimensional fusion template in the middle, This represents the calculation of mutual information entropy;

[0084] Step S6.3: Based on the judgment results of overconfidence and undersensitivity and the sample difficulty coefficient Calculate the penalty coefficient And calibrate the initial confidence level;

[0085] Penalty coefficient The calculation rules are as follows:

[0086] If it is a case of "overconfidence due to transient features":

[0087]

[0088] If the situation is "insufficiently sensitive steady-state characteristics":

[0089]

[0090] If it is a normal situation:

[0091]

[0092] In the above formula, The overall feature volatility is calculated as follows: a sliding window containing 5 consecutive frames of data is set up. The standard deviation of each dimension of the KPCA fusion features obtained by kernel principal component analysis in step S2 is calculated for each dimension in the sliding window. The mean of the standard deviations of all dimensions is then taken as the overall feature volatility. ;

[0093] Obtain the penalty coefficient Then, the initial confidence level Perform calibration:

[0094] .

[0095] As a further improvement to the intelligent identification and linkage protection method for 35kV and below equipment based on multi-source adaptive fusion and deep and shallow dual-path networks, step S7 includes the following measures:

[0096] (1) If the working status in the combined identification result is "normal operation", then the protection setting value used to trigger the protection measures of the target device shall not be adjusted;

[0097] (2) If the working state in the combined identification result is a fault state, then based on the post-calibration confidence level Different ranges trigger a tiered linkage strategy:

[0098] (2.1) When In this case, a dynamic setpoint is used as the protection setpoint for the target device to trigger protection measures; the dynamic setpoint is calculated by multiplying the baseline setpoint, the device health index, and the scenario coefficient.

[0099] in:

[0100] The baseline setting is the rated protection threshold set at the factory when the equipment leaves the factory;

[0101] Equipment Health Index The calculation formula is:

[0102]

[0103] In the above formula, and These are the time-attenuation weighted fusion values ​​of insulation resistance and dielectric loss tangent from the experimental data in the multi-source data;

[0104] Scene coefficient The calculation formula is:

[0105]

[0106] In the above formula, Equipment load density;

[0107] The operating logic in this case is as follows: For overcurrent and overvoltage faults, if the corresponding electrical quantity exceeds the dynamic set value for no less than 2ms, the circuit breaker will trip immediately; for overload and insulation faults, if the corresponding electrical quantity exceeds the dynamic set value for no less than 5 seconds, the circuit breaker will trip.

[0108] (2.2) When At this time, a default setpoint is used as the protection setpoint for triggering protection measures. The default setpoint is a threshold preset by on-site power personnel and is higher than the benchmark setpoint. At the same time, inspection data verification is introduced.

[0109] The specific measures are as follows:

[0110] First, the default settings are loaded; then, the quantified inspection data is checked. If the inspection data indicates a fault, an audible and visual alarm is triggered; if the inspection data has not been updated for more than 24 hours, it automatically switches to the general settings, which are preset industry-standard protection thresholds.

[0111] The operating logic in this case is as follows: For overcurrent and overvoltage faults, if the corresponding electrical quantity exceeds the default setting value by no less than 5ms, the circuit breaker will trip immediately; for overload and insulation faults, if the corresponding electrical quantity exceeds the default setting value by no less than 10 seconds, the circuit breaker will trip immediately.

[0112] (2.3) When If this occurs, the protection setting that triggers the protection measures will be locked to the general setting, and an emergency alarm will be triggered.

[0113] If the corresponding electrical quantity exceeds the general setting value, the circuit breaker will trip within 30 seconds after remote or on-site confirmation by maintenance personnel, and will automatically trip after 30 seconds.

[0114] As a further improvement to the intelligent identification and linkage protection method for 35kV and below equipment based on multi-source adaptive fusion and deep / shallow dual-path networks, the training process of the deep / shallow dual-path network is supervised by a multi-task joint loss function. The calculation formula is:

[0115]

[0116] In the above formula, For classifying losses, For feature importance loss, To calibrate the loss for confidence level; , , This refers to the task weighting coefficient;

[0117] (1) Classification loss The multi-class cross-entropy loss, which combines equipment type and equipment status, is expressed as follows:

[0118]

[0119] In the above formula, The number of samples in a single training session; This represents the total number of combined categories for equipment type and equipment status. For the sample Belongs to the category of combination The tags use one-hot encoding; The predicted samples output by the final Softmax layer of the LSTM network Belongs to the category of combination The original probability;

[0120] (2) Feature importance loss The expression is:

[0121]

[0122] In the above formula, steady-state characteristic flow The number of dimensions; For sample-based The first obtained during the calculation of steady-state global characteristics Importance weights of steady-state features; For the first The steady-state characteristics are based on preset target weights according to the physical properties of the equipment;

[0123] (3) Confidence calibration loss The expression is:

[0124]

[0125] In the above formula, Number of bins for confidence level; For the first Number of samples in each bin; For the first The average raw confidence level of the samples within each bin; For the first Actual identification accuracy of samples within each bin.

[0126] Compared with the prior art, the present invention has the following beneficial effects:

[0127] 1. This invention integrates electrical features, test data, and inspection data to form a multi-source feature system, and uses a deep and shallow dual-path network to extract instantaneous change features and steady-state global features respectively. This can more comprehensively characterize the operating status of equipment, especially improving the identification accuracy of equipment with inconspicuous parameter features such as reactors, and solving the problem of low identification rate of traditional protection methods when facing niche equipment.

[0128] 2. This invention designs an adaptive feature fusion mechanism based on mutual information entropy and real-time scenario parameters for dynamic operating scenarios such as load fluctuations and harmonic interference. By calculating load density and interference level, the fusion weight of instantaneous and steady-state features is dynamically adjusted, thereby improving the adaptability and robustness of the protection system to complex operating environments.

[0129] 3. This invention further introduces a hierarchical time-series matching mechanism and a device-specific gating adjustment strategy in the time-series recognition stage. The parameters of the forget gate or input gate inside the LSTM are dynamically adjusted according to the predicted device type, which enhances the pertinence and accuracy of time-series modeling. At the same time, by dividing short, medium and long-term time windows and retaining strongly correlated frames, cross-level interference is effectively reduced, which reduces the computational load while ensuring recognition accuracy.

[0130] 4. This invention uses a confidence calibration mechanism to determine overconfidence and undersensitivity using a sliding window. It dynamically calculates the penalty coefficient by combining instantaneous feature volatility, steady-state feature variance, and sample difficulty coefficient to correct the initial confidence of the model output. This allows the deviation between the calibrated confidence and the actual recognition accuracy to be controlled at a low level, providing a more reliable basis for subsequent protection decisions.

[0131] 5. This invention realizes the linkage between protection settings and real-time equipment status and operating scenarios. It executes protection strategies based on the confidence level after calibration, and adopts protection settings that are dynamically adjusted by the equipment health index and scenario coefficient under high confidence fault conditions. This enables the protection action to adaptively follow equipment aging and load changes, overcoming the limitation of traditional static settings that cannot adapt to changes in operating status over a long period of time. Attached Figure Description

[0132] Figure 1 This is a flowchart illustrating the method of the present invention. Detailed Implementation

[0133] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0134] like Figure 1As shown, a method for intelligent identification and linkage protection of 35kV and below equipment based on multi-source adaptive fusion and deep / shallow dual-path networks is presented. Its core is to construct an intelligent closed-loop system for each target device that can dynamically adapt to its operating scenario and health status and link with protection devices. Specifically, this method includes:

[0135] Steps S1 to S5 are executed cyclically according to a preset sampling frequency. In each execution cycle, the comprehensive features of the target device at the current moment are obtained. The comprehensive features of multiple consecutive execution cycles constitute a time-series feature sequence, which is used for subsequent time-series identification and protection judgment.

[0136] Step S1: Collect multi-source data from the target device to obtain multi-dimensional features.

[0137] This step aims to construct a comprehensive feature system for the target equipment, covering electrical operation, periodic testing, and on-site inspections. The collected multi-source data comprises 18 dimensions, specifically including:

[0138] (1) Electrical characteristics, 10 dimensions in total.

[0139] The static characteristics include the effective value of the three-phase current of the target equipment. , , Line voltage RMS value , , Power factor and zero-sequence current Dynamic characteristics include the maximum rate of change of three-phase current. (Calculate the relative current change for each value and take the maximum value) and voltage fluctuation coefficient. .

[0140] Among them, voltage fluctuation coefficient The calculation formula is:

[0141]

[0142] In the above formula, and These represent the maximum and minimum line voltage values ​​for a single target device over 10 consecutive sampling periods. This is the rated line voltage of the target device.

[0143] (2) Experimental data, totaling 4 dimensions.

[0144] Test data are regularly updated parameters reflecting the health status of the target equipment, including insulation resistance. Dielectric loss tangent DC resistance and no-load loss .

[0145] To integrate historical data, a time-decrease weighted method is used to fuse each type of parameter. Insulation resistance is used as an example. For example, its fusion value is:

[0146]

[0147] In the above formula, The data is from the past year. For testing data from one to two years ago, weighting coefficients of 0.7 and 0.3 were set based on the timeliness of the data. Other test data parameters... , , The fusion method is similar. Later in the text, insulation resistance... Dielectric loss tangent DC resistance and no-load loss All represent the values ​​after fusion.

[0148] (3) Inspection data, totaling 4 dimensions.

[0149] The inspection data is the result of quantifying the qualitative parameters collected on-site during operation and maintenance, and is synchronized once a day.

[0150] The quantification rules are as follows: Equipment appearance status (No cracks = 1, minor cracks = 0.6, severe cracks = 0.2); Operating sound (No abnormality = 1, slight abnormal noise = 0.5, severe abnormal noise = 0); Oil leakage status (No leakage = 1, slight leakage = 0.4, severe leakage = 0); Instrument display status (Normal = 1, Abnormal = 0).

[0151] It should be noted that the inspection data may be subject to human subjectivity and incomplete collection (such as failing to distinguish between small transformers and reactors). The inspection data provides a reference for the on-site condition, while the subsequent model prediction provides multi-dimensional quantitative verification to ensure the accuracy of equipment type identification.

[0152] Step S2: Perform dimensionality optimization and feature splitting on the multidimensional features to obtain the instantaneous feature flow and the steady-state feature flow.

[0153] First, the 18-dimensional multidimensional features collected in step S1 are standardized and dimensionality reduced. Then, the features are split according to their temporal characteristics. This includes the following sub-steps:

[0154] Step S2.1: Based on the direction of influence of the feature indicators on the equipment status, the 18-dimensional features obtained in step S1 are divided into extremely large indicators and extremely small indicators, and standardized respectively.

[0155] A higher value for a very large indicator indicates a better equipment condition, including... , , , , , , , , , , , , , Its standardized formula is:

[0156] ;

[0157] In the above formula, These are the original eigenvalues. and These are the maximum and minimum values ​​of the feature in the training set, respectively. This is the standardized result.

[0158] The smaller the value of the miniature indicator, the better the equipment condition, including... , , , Its standardized formula is:

[0159] .

[0160] Step S2.2: Kernel Principal Component Analysis (KPCA) is used to perform nonlinear dimensionality reduction on the standardized 18-dimensional features, retaining principal components with a cumulative contribution rate of not less than 85%, resulting in 14-dimensional KPCA fused features. Simultaneously, the load vectors of each principal component, i.e., their nonlinear correlation weights with the original 18-dimensional features, are analyzed. Based on the load vectors, principal components dominated by rapidly changing features (such as current change rate, voltage fluctuation coefficient, operating sound, etc.) are identified as instantaneous principal components capable of capturing rapid abrupt changes; principal components dominated by slowly changing features (such as insulation resistance, power factor, appearance condition, test parameters, etc.) are identified as steady-state principal components capable of capturing long-term stable characteristics.

[0161] This step uses the radial basis function (RBF) kernel function, kernel parameters... The value is set to 0.8 based on the characteristic distribution of the power distribution equipment. This process reduces the computational load of the original feature space by approximately 30%.

[0162] Step S2.3: Based on the standardized feature indices and the principal components obtained from kernel principal component analysis, obtain the instantaneous feature flow. and steady-state characteristic flow .

[0163] Instantaneous Feature Flow There are 7 dimensions in total, including the standardized ones. , , , Start Feature Entropy And the three instantaneous principal components with the highest contribution rates.

[0164] Among them, the initiation feature entropy The acquisition method is as follows: During the device startup phase (0-3 seconds after power-on), additionally calculate the feature entropy of the startup phase. As an auxiliary feature, its calculation formula is:

[0165]

[0166] In the above formula, This is the sampling sequence number during the startup phase. For the first The normalized probability of the current for each sampling period is derived from the mean of the effective values ​​of the three-phase currents in that period. The calculation shows that, The sampling frequency for electrical characteristics is 20Hz.

[0167] Steady-state characteristic flow There are 7 dimensions in total, including the standardized ones. , , , , , And the steady-state principal component with the highest contribution rate.

[0168] The core instantaneous feature changes at a frequency greater than 1 Hz, while the core steady-state feature changes at a frequency less than 0.1 Hz, thereby ensuring that the feature stream can effectively separate the device's abrupt changes and steady-state information.

[0169] Step S3: Input the instantaneous feature stream and the steady-state feature stream into the deep and shallow dual-path network. Extract instantaneous mutation features from the instantaneous feature stream through the shallow path and extract steady-state global features from the steady-state feature stream through the deep path.

[0170] The feature extraction process is as follows:

[0171] Step S3.1: Calculate the 7-dimensional instantaneous feature flow obtained in step S2. Input the shallow path of the deep-shallow dual-path network. The shallow path consists of three lightweight convolutional layers with kernel sizes of 1×1, 3×1, and 1×1, and kernel numbers of 16, 32, and 32, respectively. Each convolutional layer is followed by batch normalization. The final output is the transient abrupt change feature after three layers of processing, denoted as . .

[0172] Step S3.2: Calculate the 7-dimensional steady-state feature flow obtained in step S2.3. Input the deep path of the shallow-deep dual-path network. The deep path first undergoes preliminary processing through two depthwise separable convolution layers (kernel size 5×1), and the output is denoted as... ; then After deep feature extraction using 3 layers of pre-activated residual blocks, the output is denoted as... Finally, a feature filter (based on an attention mechanism) is used to filter the data. Weighting is applied.

[0173] Furthermore, the feature filter calculates the importance weights of steady-state features using the following formula:

[0174]

[0175] In the above formula, It is the sigmoid activation function. and The parameters learned during training, , express The The element, i.e., the th element The importance weight of the steady-state feature is defined, with a value range of [0,1].

[0176] Weighted steady-state global characteristics Calculated via element-wise product: ,in This indicates element-wise multiplication. It still has 7-dimensional features, but redundant features have been suppressed and core steady-state features have been strengthened.

[0177] Step S4: Based on the real-time operating scenario of the target device, the instantaneous mutation features and steady-state global features are dynamically and adaptively fused to obtain the comprehensive features of the target device at the current moment.

[0178] The core of this step is to dynamically calculate the fusion weights based on the target device's current load and interference levels, and to use a pre-built scene template library to guide the fusion process, ultimately generating 128-dimensional comprehensive features. Specifically, it includes the following sub-steps:

[0179] Step S4.1: Calculate the load density of the real-time running scenario parameters. and degree of interference And determine the current scene type.

[0180] Load density The calculation formula is:

[0181]

[0182] The value range is [0, 2].

[0183] Interference level The calculation formula is:

[0184]

[0185] The value range is [0, 1].

[0186] According to load density and degree of interference Determine the current scene type:

[0187] (1) According to load density Load partitioning scenarios based on size: For low load, For medium load, For high load.

[0188] (2) According to the degree of interference Size division interference scenarios: For low interference, For interference, This is a high-interference signal.

[0189] (3) A total of 9 scenario types are constructed by combining 3 load conditions and 3 interference conditions. The current scenario type is determined based on the load conditions and interference conditions.

[0190] Step S4.2: Based on the scene template library Dynamic calculation of instantaneous feature fusion weights .

[0191] The calculation formula is:

[0192]

[0193] In the above formula, This represents the calculation of mutual information entropy. The current 7-dimensional transient mutation features With the current scene type In the scene template library The corresponding 7-dimensional instantaneous template Mutual information entropy, Represents the current 7-dimensional steady-state global characteristics With the current scene type In the scene template library The corresponding 7-dimensional steady-state template Mutual information entropy.

[0194] Mutual information entropy The range is [0,1], and it is calculated using conventional formulas in this field, which will not be elaborated here.

[0195] . The larger, The more it leans towards 1, the more important the instantaneous features are in high-load scenarios; The larger, The smaller, The more it leans towards 1, the more important the instantaneous features are in high-interference scenarios.

[0196] Furthermore, the scene template library The information is obtained and solidified in advance, and the acquisition method is as follows:

[0197] There are nine scenario types: low load low disturbance, low load medium disturbance, low load high disturbance, medium load low disturbance, medium load medium disturbance, medium load high disturbance, high load low disturbance, high load medium disturbance, and high load high disturbance. Six core scenarios are selected from all scenario types (i.e., three extremely rare scenarios are removed; more or even all scenarios can be retained depending on the situation; this example uses six scenarios): high load low disturbance, high load medium disturbance, high load high disturbance, medium load high disturbance, medium load medium disturbance, and low load low disturbance. For each core scenario, at least 1000 sets of 7-dimensional instantaneous feature streams and 7-dimensional steady-state feature streams of the target devices are collected. Then, fixed weights (such as...) are used... The instantaneous and steady-state feature streams of each pair are weighted and fused into a 7-dimensional pseudo-fusion feature. This pseudo-fusion feature is then upscaled to 128 dimensions using a 1×1 convolutional layer, resulting in a 128-dimensional pseudo-fusion feature. Finally, K-means clustering (K=1) is performed on the 128-dimensional pseudo-fusion feature, the 7-dimensional instantaneous feature stream, and the 7-dimensional steady-state feature stream for each core scene, yielding a 128-dimensional fusion template, a 7-dimensional instantaneous template, and a 7-dimensional steady-state template for each core scene. These 18 templates across 3 categories for 6 core scenes constitute the scene template library. .

[0198] Step S4.3: Fuse weights based on instantaneous features Transient mutation characteristics and steady-state global characteristics Perform feature fusion to generate comprehensive features .

[0199] The fusion method is as follows:

[0200]

[0201] In the above formula, This represents a 1×1 convolution operation, used to unify feature dimensions; For the residual term, the calculation formula is: This is used to prevent information loss.

[0202] Final output The 128-dimensional design takes into account the computing power and memory limitations of existing protection devices while retaining complete key feature information.

[0203] Step S5: Input the temporal feature sequence composed of the comprehensive features of multiple consecutive time points into the LSTM network, enhance the temporal correlation through the hierarchical temporal matching mechanism, and output the combined recognition result of "device type-working status" and the corresponding initial confidence level.

[0204] This step first uses a lightweight device type classifier to perform preliminary device type determination based on the comprehensive features, adapting it to the LSTM gating parameters. Then, it uses an LSTM network for temporal modeling and joint classification. Specifically, it includes the following sub-steps:

[0205] Step S5.1: Construct a time-series feature sequence.

[0206] Continuously collect data from the target device at a sampling frequency of 20Hz. Frames (e.g.) 128-dimensional comprehensive features (corresponding to 0.75 seconds) , constitutes a time-series feature sequence The superscript indicates the frame number.

[0207] Step S5.2: Use the device type classifier to predict the device type of the current target device.

[0208] Before inputting the time sequence into the LSTM, the 128-dimensional comprehensive features of the first frame (or any frame) are first processed. Input a lightweight device type classifier. This classifier is a fully connected network with two hidden layers, and finally outputs the probability distribution of the target device belonging to one of the four device types: motor, transformer, capacitor, and reactor through a Softmax layer. .Pick The type corresponding to the maximum value in the middle is used as the device type prediction result. This result is only used for subsequent adjustment of the LSTM gating parameters and does not participate in the final joint classification of "device type-working status", thus avoiding the accumulation of errors from two independent classifications.

[0209] Specifically, the classifier is a feedforward neural network based on a fully connected layer, and its input is the 128-dimensional comprehensive feature representing the current operating state of the target device, which is output from step S4. The network structure is as follows:

[0210] (1) The input layer receives 128-dimensional integrated features .

[0211] (2) The first hidden layer is a fully connected layer, mapping the 128-dimensional input to 64 dimensions. The forward computation logic of this layer is as follows:

[0212]

[0213] In the above formula, It is The weight matrix, It is a 64-dimensional bias vector. This is the activation function.

[0214] (3) The second hidden layer is a fully connected layer, which further compresses the 64-dimensional features to 32 dimensions. The forward computation logic of this layer is as follows:

[0215]

[0216] In the above formula, It is The weight matrix, It is a 32-dimensional bias vector.

[0217] (4) The output layer is a fully connected layer followed by a Softmax function, which maps the 32-dimensional features to 4 dimensions, corresponding to four target device types: motor, transformer, capacitor, and reactor. The forward computation logic of this layer is as follows:

[0218]

[0219] In the above formula, It is The weight matrix, It is a 4-dimensional bias vector. The function transforms the output into a probability distribution. It is a 4-dimensional vector, where each element represents the probability that the target device belongs to the corresponding type, and the sum of all elements is 1.

[0220] This classifier is trained separately during the training phase before model deployment. Training uses a large number of samples labeled with device type (the labels only specify device type, not operating status) for supervised learning using the cross-entropy loss function. Typical training parameters are: batch size 32, 50 iterations, and a learning rate of [missing information]. .

[0221] During the real-time operation phase, the current 128 dimensions will be... Inputting this classifier yields the probability distributions of the four device classes. .Pick The device type corresponding to the maximum value is used as the preliminary judgment result. For example, if If so, the target device type is determined to be "electric motor".

[0222] It is important to emphasize that the device type determination result here is not the final device identification result output, nor does it participate in the loss calculation of the subsequent "device type-operating status" joint classification. Its sole purpose is to provide the subsequent LSTM network with the basis for dynamically adjusting its internal forget gate or input gate parameters (i.e., the device-specific gating adjustment in step S5.3). This design avoids logical contradictions caused by unknown device type during LSTM inference, and also prevents the accumulation of errors that may result from performing two independent classifications (one for type and one for status). The final "device type-operating status" joint determination is entirely completed by the LSTM network.

[0223] Step S5.3: Transform the time-series feature sequence Input the LSTM network and perform device-specific gating adjustments based on the predicted device type.

[0224] The main structure of the LSTM network is uniform, but its internal control parameters are dynamically adjusted based on the device type prediction results obtained in step S5.2. The specific adjustment method is as follows:

[0225] (1) If the problem is predicted to be an electric motor, then the forgetting gate is strengthened. A load adaptation factor is introduced. , The enhanced forget gate corresponds to the load density at a given time. The calculation formula is:

[0226]

[0227] (2) If the problem is predicted to be a transformer, then strengthen the input gate. Introduce an interference adaptation factor. , To mitigate interference, the enhanced input gate The calculation formula is:

[0228]

[0229] In the above formula, The hidden state of the LSTM at time t-1. Let be the input features at time t. , , , The parameters learned during training, For batch normalization, It is the sigmoid activation function.

[0230] (3) For capacitors and reactors, use common gating parameters.

[0231] Step S5.4: In the LSTM network processing, a hierarchical timing matching mechanism is introduced.

[0232] A hierarchical timing matching mechanism is employed to enhance timing correlation. Specifically, a length of [missing information] is used to [missing information]. The temporal feature sequence is divided into multiple time window levels, such as short-term, medium-term, and long-term (e.g., the first 3 frames are short-term, the middle 5 frames are medium-term, and the last 7 frames are long-term). Within each level, an attention mechanism is used to calculate the cosine similarity between frames, retaining only strong correlations with similarity of not less than 0.8, thereby reducing cross-level interference and reducing the computational load by about 25%.

[0233] Step S5.5: Obtain the combined recognition result and initial confidence level based on the LSTM network.

[0234] The final output layer of an LSTM network is a Softmax layer, whose output is a probability distribution of all possible "device type-operation state" combinations. This probability distribution contains... A value (as mentioned above, Take the probability value with the largest value in the probability distribution as the initial confidence level. The corresponding "device type - working status" combination is the combination identification result of the current target device.

[0235] For example, if the probability corresponding to "motor-overcurrent" is 0.85 and is the maximum value, then the combined identification result is "motor-overcurrent", with an initial confidence level of... .

[0236] Step S6: Determine whether the initial confidence level is overconfident or undersensitive based on the volatility of the transient mutation features and the variance of the steady-state global features. Combine the sample difficulty coefficient to dynamically calculate the penalty coefficient and calibrate the initial confidence level to obtain the calibrated confidence level.

[0237] This step aims to address the confidence distortion issues caused by "overconfidence in transient features" and "lack of sensitivity in steady-state features." Specifically, it includes the following sub-steps:

[0238] Step S6.1: Determine whether you are overconfident or insensitive.

[0239] Set up a sliding window containing 5 consecutive frames of data. Calculate the 7-dimensional transient change features within the window. volatility and 7-dimensional steady-state global features variance .

[0240] Instantaneous characteristic volatility The calculation formula is:

[0241]

[0242] In the above formula, , , Representing the sliding window respectively The maximum, minimum and mean values.

[0243] The rules for judging overconfidence and lack of sensitivity are as follows: If And initial confidence level If so, it is judged as "overconfidence in transient features"; if and If so, it is judged as "insufficiently sensitive steady-state characteristics".

[0244] The above thresholds of 0.3, 0.9, 0.1, and 0.7 can be adjusted according to the actual situation of the project.

[0245] Step S6.2: Based on the comprehensive features and scene template library at the current moment Calculate the sample difficulty coefficient .

[0246] Sample difficulty level This is used to measure the degree of matching between the features of the current sample and the scene template; the higher the matching degree, the simpler the sample. The calculation formula is:

[0247]

[0248] In the above formula, The 128-dimensional comprehensive features at the current moment, For the current scene type In the scene template library The corresponding 128-dimensional fusion template in the middle, This represents the calculation of mutual information entropy. , A larger value indicates a simpler sample.

[0249] Step S6.3: Based on the judgment results of overconfidence and undersensitivity and the sample difficulty coefficient Calculate the penalty coefficient And calibrate the initial confidence level.

[0250] Penalty coefficient The calculation rules are as follows:

[0251] If it is a case of "overconfidence due to transient features":

[0252]

[0253] If the situation is "insufficiently sensitive steady-state characteristics":

[0254]

[0255] If it is a normal situation:

[0256]

[0257] In the above formula, The overall feature volatility is calculated as follows: a sliding window containing 5 consecutive frames of data is set up. The standard deviation of each dimension of the 14-dimensional KPCA fusion feature obtained by kernel principal component analysis in step S2 is calculated for each dimension in the sliding window. The mean of the standard deviations of all dimensions is then taken as the overall feature volatility. .

[0258] Obtain the penalty coefficient Then, the initial confidence level Perform calibration:

[0259]

[0260] Confidence level after calibration The deviation from the actual accuracy rate should not exceed 3%, which is the core decision-making basis for triggering subsequent linkage protection.

[0261] Step S7: Dynamically adjust the protection strategy of the target device based on the post-calibration confidence level.

[0262] This step is based on post-calibration confidence. Based on the combined identification results of the target device (device type and operating status), a hierarchical linkage protection strategy is implemented, and point-to-point coordination with associated protection devices is achieved. Specifically, this includes the following measures:

[0263] (1) If the working status in the combined recognition result is "normal operation", then regardless of Regardless of the value, the protection settings used to trigger protection measures on the target equipment will not be adjusted to avoid malfunctions under normal conditions. Only differentiated monitoring will be performed, including:

[0264] (1.1) Monitor electrical parameters in real time and upload status data every 5 seconds;

[0265] (1.2) If If so, only local records are generated, and no alarms are triggered;

[0266] (1.3) If If this occurs, a "low confidence level is normal" alert will be triggered and pushed to the operations and maintenance backend (not an emergency alarm), prompting manual verification of whether the inspection data or test data has not been updated in a timely manner.

[0267] (2) If the operating status in the combined identification result is a fault status (including overcurrent, overvoltage, overload, and insulation fault), then according to Different ranges trigger a tiered linkage strategy:

[0268] (2.1) When In this case, a dynamic setpoint is used as the protection setpoint for the target device to trigger protection measures. The dynamic setpoint is calculated by multiplying the baseline setpoint, the device health index, and the scenario coefficient.

[0269] in:

[0270] The reference setting is the rated protection threshold set at the factory when the equipment leaves the factory. For example, the reference setting for overcurrent of a motor is 8 times the rated current, and the reference setting for overcurrent of a transformer is 1.2 times the rated current.

[0271] Equipment Health Index The calculation formula is:

[0272]

[0273] In the above formula, and These are the time-decrease-weighted fusion values ​​of insulation resistance and dielectric loss tangent from the test data, respectively. The equipment health index is normalized to the [0,1] interval, with a larger value indicating a healthier equipment.

[0274] Scene coefficient The calculation formula is:

[0275]

[0276] In the above formula, To account for the load density, its maximum value is limited to no more than 1.5 to avoid the coefficient being too large and causing the set value to become uncontrollable.

[0277] The action logic in this case is as follows: For overcurrent and overvoltage faults, if the corresponding electrical quantity exceeds the dynamic set value for no less than 2ms, the circuit breaker will trip immediately (to prevent transient interference); for overload and insulation faults, if the corresponding electrical quantity exceeds the dynamic set value for no less than 5 seconds, the circuit breaker will trip (allowing short-term overload). Simultaneously, the dynamic set value calculation process, equipment health index, and scenario parameters are uploaded to the backend for easy traceability.

[0278] (2.2) When At this time, a default setpoint is used as the protection setpoint for triggering protection measures. The default setpoint is a threshold preset by on-site power personnel and is slightly higher than the benchmark setpoint. At the same time, inspection data verification is introduced.

[0279] The specific measures are as follows:

[0280] First, load the default settings. Then, check the quantified inspection data. If the inspection data indicates a fault, such as an oil leak, then... or operating sound or the appearance of the equipment , an audible and visual alarm will be triggered. If the patrol data has not been updated for more than 24 hours, it will automatically switch to the general fixed value to avoid invalid verification. The general fixed value is a preset unified protection threshold that is commonly used in the industry.

[0281] The action logic in this case is as follows: for overcurrent and overvoltage faults, if the corresponding electrical quantity exceeds the default fixed value for no less than 5 ms, a trip will be immediately triggered; for overload and insulation faults, if the corresponding electrical quantity exceeds the default fixed value for no less than 10 seconds, a trip will be immediately triggered to reduce the risk of misoperation.

[0282] (2.3) When it is the case, the protection fixed value that triggers the protection measure will be locked as the general fixed value, and an emergency alarm will be triggered.

[0283] If the corresponding electrical quantity exceeds the general fixed value, the circuit breaker needs to be tripped after being remotely or on-site confirmed by the maintenance personnel within 30 seconds, and it will automatically trip after 30 seconds. At the same time, all characteristic data and classification results will be uploaded to the background to prompt manual verification of the device status.

[0284] To achieve reliable linkage with the existing protection system, this method also includes a multi-device coordination mechanism. This mechanism defines how the intelligent protection device running this invention method acts as the main controller and coordinates with the associated protection devices of the target equipment. Its core is point-to-point instruction and information synchronization to adapt to the distribution scenarios of 35 kV and below.

[0285] Step C1: Determine the main controller.

[0286] The main controller refers to the core unit responsible for making the final decision and executing the linkage protection strategy of the target equipment. In this method, the main controller is determined by one of the following two methods:

[0287] (1) Default binding method: For a single target equipment, the core protection device installed on it and executing steps S1 to S7 of this invention is defaulted as the main controller of the target equipment. This method is applicable to the vast majority of distribution scenarios of 35 kV and below, aiming to avoid complex selection.

[0288] (2) Competition determination method (alternative): In specific complex scenarios such as multi-device linkage of a busbar, if it is necessary to select one from multiple target equipment and use its protection device as the leading controller of the coordination strategy, an auxiliary determination logic will be enabled. This determination is based on the status of each target equipment itself. Calculate the decision priority score of each candidate target equipment, and its corresponding core protection device will be determined as the leading controller. The calculation formula for the decision priority score is:

[0289]

[0290] In the above formula, is the equipment health index of the connected target equipment, This represents the real-time load density of the connected target devices. This formula comprehensively considers the health status of the devices themselves and their load conditions. The better the health status of the target devices and the higher the load, the higher the weight given to their corresponding protection decisions in the coordination, and their protection devices thus become the dominant controllers.

[0291] Step C2: Determine the collaborating objects.

[0292] The main controller determined in step C1 only coordinates with the incoming-side protection devices that are electrically directly associated with the single target device. This coordination relationship is singular and point-to-point.

[0293] Step C3: Perform information synchronization.

[0294] After completing the identification and confidence calibration of the target equipment, the main controller will synchronize the following key information to the incoming line protection device in real time:

[0295] (1) The identified combination of "equipment type - working status";

[0296] (2) Post-calibration confidence ;

[0297] (3) According to and the protection settings determined by the hierarchical linkage strategy (such as dynamic settings or default settings).

[0298] The synchronization process is delayed by no more than 5ms to ensure that the incoming line protection device can obtain the accurate status of the target equipment in a timely manner.

[0299] Step C4: Execute the protection setting linkage command.

[0300] Based on real-time identification results of the target equipment and analysis of the operating scenario, the main controller can issue temporary setting adjustment commands to the incoming line protection device to achieve protection coordination. Typical linkage scenarios are as follows:

[0301] When the main controller determines that the target device is in the "motor-start" state, and the current operating scenario is high load, low interference (i.e., load density) Interference level When the overcurrent protection setting is activated, the main controller will immediately send a command to the incoming line protection device, requesting it to temporarily increase its overcurrent protection setting by 10%. The duration of this setting increase command is 3 seconds to cover the typical motor startup process.

[0302] The purpose of this linkage operation is to enable the incoming line protection device to withstand a large starting current during motor startup, and to avoid false tripping due to the starting current exceeding its normal specified value. After 3 seconds, the incoming line protection setting value will automatically return to its original value.

[0303] Before deployment, the deep and shallow dual-path network is trained. The training of this network aims to optimize its joint identification ability of device type and state, the importance weight allocation of steady-state features, and the calibration of model output confidence. The training process is supervised using a multi-task joint loss function. Specifically, the loss function is constructed... Its expression is:

[0304]

[0305] In the above formula, For classifying losses, For feature importance loss, To calibrate the loss for confidence level; , , The task weighting coefficient is set according to the requirements of the power distribution system project. Because the importance of features directly affects the quality of feature extraction, they are given slightly higher weights. Next.

[0306] (1) Classification loss The multi-class cross-entropy loss, which combines equipment type and equipment status, is expressed as follows:

[0307]

[0308] In the above formula, The number of samples in a single training session; This represents the total number of combined categories for equipment type and equipment status. For example, if equipment types include four categories: motors, transformers, capacitors, and reactors, and equipment status includes five categories: normal operation, overcurrent, overvoltage, overload, and insulation fault, then... ; For the sample Belongs to the category of combination The tags use one-hot encoding; The predicted samples output by the final Softmax layer of the LSTM network Belongs to the category of combination The original probability.

[0309] (2) Feature importance loss The feature-weighted mean squared error loss is expressed as follows:

[0310]

[0311] In the above formula, steady-state characteristic flow The number of dimensions, i.e., 7; For sample-based Calculated Importance weights of steady-state features; For the first Steady-state characteristics are based on preset target weights for the physical characteristics of the equipment. These weights are derived from the physical characteristics of the equipment and the preset operating scenarios, such as in transformer insulation assessment. The target weight is 0.8 in the steady-state evaluation of the electric motor. The target weight is 0.7.

[0312] (3) Confidence calibration loss The expected calibration error (ECE) loss is expressed as follows:

[0313]

[0314] In the above formula, The number of bins for the confidence level is typically 10 to 15 (e.g., 10 bins evenly divided into [0, 0.1), [0.1, 0.2)...[0.9, 1.0]). For the first Number of samples in each bin; For the first The average raw confidence level (i.e., uncalibrated) of samples within each bin. (mean) For the first Actual identification accuracy of samples within each bin (number of correctly predicted samples within the bin / total number of samples in the bin).

[0315] It should be noted that, as will be apparent to those skilled in the art, the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics thereof. The scope of the present invention is defined by the claims rather than the foregoing description.

Claims

1. A method for intelligent identification and linkage protection of 35kV and below equipment based on multi-source adaptive fusion and deep / shallow dual-path networks, characterized in that: Steps S1 to S5 are executed cyclically according to a preset sampling frequency. In each execution cycle, the comprehensive features of the target device at the current moment are obtained. The comprehensive features of multiple consecutive execution cycles constitute a time-series feature sequence, which is used for subsequent time-series identification and protection judgment. Step S1: Collect multi-source data from the target device to obtain multi-dimensional features; Step S2: Perform dimensionality optimization and feature splitting on the multidimensional features to obtain the instantaneous feature stream and the steady-state feature stream; Step S3: Input the instantaneous feature stream and the steady-state feature stream into the deep and shallow dual-path network. Extract the instantaneous mutation features from the instantaneous feature stream through the shallow path and extract the steady-state global features from the steady-state feature stream through the deep path. Step S4: Based on the real-time operating scenario of the target device, the instantaneous mutation features and steady-state global features are dynamically and adaptively fused to obtain the comprehensive features of the target device at the current moment. Step S5: Input the temporal feature sequence composed of the comprehensive features of multiple consecutive time points into the LSTM network, enhance the temporal correlation through the hierarchical temporal matching mechanism, and output the combined recognition result of "device type-working status" and the corresponding initial confidence. Step S6: Determine whether the initial confidence level is overconfident or undersensitive based on the volatility of the transient mutation features and the variance of the steady-state global features. Combine the sample difficulty coefficient to dynamically calculate the penalty coefficient and calibrate the initial confidence level to obtain the calibrated confidence level. Step S7: Dynamically adjust the protection strategy of the target device based on the post-calibration confidence level.

2. The intelligent identification and linkage protection method for 35kV and below equipment based on multi-source adaptive fusion and deep / shallow dual-path network as described in claim 1, characterized in that, The multi-source data collected in step S1 comprises a total of 18 dimensions: (1) Electrical characteristics, 10 dimensions in total; in Static characteristics include the effective value of the three-phase current of the target device. , , Line voltage RMS value , , Power factor and zero-sequence current Dynamic characteristics include the maximum rate of change of three-phase current. and voltage fluctuation coefficient ; Among them, voltage fluctuation coefficient The calculation formula is: ； In the above formula, and These represent the maximum and minimum line voltage values ​​for a single target device across multiple consecutive sampling periods. This is the rated line voltage of the target device; (2) Test data, in four dimensions, including insulation resistance Dielectric loss tangent DC resistance and no-load loss ; (3) Inspection data, in 4 dimensions, including equipment appearance status Operating noise Oil leakage status and instrument display status .

3. The intelligent identification and linkage protection method for 35kV and below equipment based on multi-source adaptive fusion and deep / shallow dual-path network as described in claim 2, characterized in that, Step S2 specifically includes: Step S2.1: Based on the influence direction of the feature indicators on the equipment status, the 18-dimensional features obtained in step S1 are divided into extremely large indicators and extremely small indicators, and standardized respectively. Step S2.2: Kernel principal component analysis is used to perform nonlinear dimensionality reduction on the standardized 18-dimensional features, retaining principal components with a cumulative contribution rate not less than a preset ratio to obtain KPCA fusion features; at the same time, the loading vectors of each principal component are analyzed to determine whether each principal component is a steady-state principal component. Step S2.3: Based on the standardized feature indices and the principal components obtained from kernel principal component analysis, obtain the instantaneous feature flow. and steady-state characteristic flow ; Instantaneous Feature Flow There are 7 dimensions in total, including the standardized ones. , , It also includes the initiation feature entropy And the three instantaneous principal components with the highest contribution rates; Among them, the initiation feature entropy It is obtained during the equipment startup phase using the following calculation formula: ； In the above formula, This is the sampling sequence number during the startup phase. For the first The normalized probability of the current for each sampling period is derived from the mean of the effective values ​​of the three-phase currents in that period. The calculation shows that, ; Steady-state characteristic flow There are 7 dimensions in total, including the standardized ones. , , , , , And the steady-state principal component with the highest contribution rate.

4. The intelligent identification and linkage protection method for 35kV and below equipment based on multi-source adaptive fusion and deep / shallow dual-path network as described in claim 1, characterized in that, The feature extraction process in step S3 is as follows: Step S3.1: Calculate the instantaneous feature flow obtained in step S2. Input the shallow path of the deep-shallow dual-path network; the shallow path consists of 3 lightweight convolutional layers, and the final output transient mutation feature is denoted as... ; Step S3.2: Calculate the steady-state characteristic flow obtained in step S2. Input the deep path of the shallow-deep dual-path network; the deep path first undergoes preliminary processing through two depthwise separable convolutions, and the output is denoted as... Subsequently After deep feature extraction using 3 layers of pre-activated residual blocks, the output is denoted as... Finally, a feature filter is used to... Weighted summation yields steady-state global features .

5. The intelligent identification and linkage protection method for 35kV and below equipment based on multi-source adaptive fusion and deep / shallow dual-path network as described in claim 4, characterized in that, The formula for calculating the importance weights of steady-state features by the feature filter is: ； In the above formula, It is the sigmoid activation function. and The parameters learned during training, the weight vector Dimensions and Consistent express The The element, i.e., the th element Importance weights of steady-state features; Weighted steady-state global characteristics Calculated via element-wise product: ,in This indicates element-wise multiplication.

6. The intelligent identification and linkage protection method for 35kV and below equipment based on multi-source adaptive fusion and deep / shallow dual-path network as described in claim 3, characterized in that, Step S4 specifically includes: Step S4.1: Calculate the load density of the real-time running scenario parameters. and degree of interference And determine the current scene type; Load density The calculation formula is: ； The value range of is [0, 2]; Interference level The calculation formula is: ； The value range is [0, 1]; According to load density and degree of interference Determine the current scene type; Step S4.2: Based on the scene template library Dynamic calculation of instantaneous feature fusion weights ; The calculation formula is: ； In the above formula, This represents the calculation of mutual information entropy. The current 7-dimensional transient mutation features With the current scene type In the scene template library The corresponding 7-dimensional instantaneous template Mutual information entropy, Represents the current 7-dimensional steady-state global characteristics With the current scene type In the scene template library The corresponding 7-dimensional steady-state template Mutual information entropy; Step S4.3: Fuse weights based on instantaneous features Transient mutation characteristics and steady-state global features Perform feature fusion to generate comprehensive features ; The fusion method is as follows: ； In the above formula, This represents a 1×1 convolution operation, used to unify feature dimensions; For the residual term, the calculation formula is: ; The scene template library The method of obtaining it is: For each scenario, at least 1000 sets of 7-dimensional instantaneous feature streams and 7-dimensional steady-state feature streams from the target devices are collected. Then, using fixed weights, the corresponding instantaneous and steady-state feature streams of each set are weighted and fused into 7-dimensional pseudo-mixed features. A 1×1 convolutional layer is then used to upscale the pseudo-mixed features to 128 dimensions, resulting in 128-dimensional pseudo-fusion features. Finally, K-means clustering is performed on the 128-dimensional pseudo-fusion features, 7-dimensional instantaneous feature streams, and 7-dimensional steady-state feature streams for each scenario, with clustering parameter K=1, to obtain the corresponding 128-dimensional fusion template, 7-dimensional instantaneous template, and 7-dimensional steady-state template for each scenario, forming a scenario template library. .

7. The intelligent identification and linkage protection method for 35kV and below equipment based on multi-source adaptive fusion and deep / shallow dual-path network as described in claim 1, characterized in that, Step S5 specifically includes: Step S5.1: Construct a temporal feature sequence; Continuous acquisition target device 128-dimensional composite features of the frame , constitutes a time-series feature sequence The superscript indicates the frame number; Step S5.2: Use the device type classifier to predict the device type of the current target device; Before inputting the time sequence into the LSTM, the synthesized features of any frame are first processed. Input a lightweight device type classifier, which is a fully connected network with two hidden layers, and finally output the probability distribution of the target device type through a Softmax layer. ,Pick The type corresponding to the maximum value is used as the device type prediction result; Step S5.3: Transform the time-series feature sequence Input the data into the LSTM network and perform device-specific gating adjustments based on the predicted device type; The main structure of the LSTM network is uniform, but its internal control parameters are dynamically adjusted based on the device type prediction results obtained in step S5.

2. The specific adjustment method is as follows: (1) If the prediction is that it is an electric motor, then strengthen the forget gate and introduce a load adaptation factor. , The enhanced forget gate corresponds to the load density at a given time. The calculation formula is: ； (2) If the problem is predicted to be a transformer, then strengthen the input gate and introduce an interference adaptation factor. , To mitigate interference, the enhanced input gate The calculation formula is: ； In the above formula, The hidden state of the LSTM at time t-1. Let be the input features at time t. , , , The parameters learned during training, For batch normalization, It is the sigmoid activation function; (3) For capacitors and reactors, universal gating parameters are used; Step S5.4: In the LSTM network processing, a hierarchical timing matching mechanism is introduced; Specifically, the length is The temporal feature sequence is divided into multiple time window levels such as short-term, medium-term, and long-term. Within each level, an attention mechanism is used to calculate the cosine similarity between frames, and only strong correlations with similarity not lower than a preset relationship threshold are retained, thereby reducing cross-level interference. Step S5.5: Obtain the combined recognition result and initial confidence level based on the LSTM network; The final output layer of the LSTM network is a Softmax layer, whose output is a probability distribution of the "device type-operating state" combination category; this probability distribution contains The probability value with the largest value in the probability distribution is taken as the initial confidence level. The corresponding "device type - working status" combination is the combination identification result of the current target device.

8. The intelligent identification and linkage protection method for 35kV and below equipment based on multi-source adaptive fusion and deep / shallow dual-path network as described in claim 6, characterized in that, Step S6 specifically includes: Step S6.1: Determining overconfidence versus undersensitivity; Set up a sliding window containing 5 consecutive frames of data, and calculate the 7-dimensional transient mutation features within the window. volatility and 7-dimensional steady-state global features variance ; Instantaneous characteristic volatility The calculation formula is: ； In the above formula, , , Representing the sliding window respectively The maximum, minimum, and mean values; The rules for judging overconfidence and lack of sensitivity are as follows: If Greater than the first threshold and initial confidence level If it exceeds the second threshold, it is judged as "overconfidence in transient features"; if Less than the third threshold and If the value is less than the fourth threshold, it is judged as "insufficiently sensitive steady-state characteristics"; Step S6.2: Based on the comprehensive features and scene template library at the current moment Calculate the sample difficulty coefficient : ； In the above formula, The 128-dimensional comprehensive features at the current moment, For the current scene type In the scene template library The corresponding 128-dimensional fusion template in the middle, This represents the calculation of mutual information entropy; Step S6.3: Based on the judgment results of overconfidence and undersensitivity and the sample difficulty coefficient Calculate the penalty coefficient And calibrate the initial confidence level; Penalty coefficient The calculation rules are as follows: If it is a case of "overconfidence due to transient features": ； If the situation is "insufficiently sensitive steady-state characteristics": ； If it is a normal situation: ； In the above formula, The overall feature volatility is calculated as follows: a sliding window containing 5 consecutive frames of data is set up. The standard deviation of each dimension of the KPCA fusion features obtained by kernel principal component analysis in step S2 is calculated for each dimension in the sliding window. The mean of the standard deviations of all dimensions is then taken as the overall feature volatility. ; Obtain the penalty coefficient Then, the initial confidence level Perform calibration: 。 9. The intelligent identification and linkage protection method for 35kV and below equipment based on multi-source adaptive fusion and deep / shallow dual-path network as described in claim 1, characterized in that, Step S7 includes the following measures: (1) If the working status in the combined identification result is "normal operation", then the protection setting value used to trigger the protection measures of the target device shall not be adjusted; (2) If the working state in the combined identification result is a fault state, then based on the post-calibration confidence level Different ranges trigger a tiered linkage strategy: (2.1) When In this case, a dynamic setpoint is used as the protection setpoint for the target device to trigger protection measures; the dynamic setpoint is calculated by multiplying the baseline setpoint, the device health index, and the scenario coefficient. in: The baseline setting is the rated protection threshold set at the factory when the equipment leaves the factory; Equipment Health Index The calculation formula is: ； In the above formula, and These are the time-attenuation weighted fusion values ​​of insulation resistance and dielectric loss tangent from the experimental data in the multi-source data; Scene coefficient The calculation formula is: ； In the above formula, Equipment load density; The operating logic in this case is as follows: For overcurrent and overvoltage faults, if the corresponding electrical quantity exceeds the dynamic set value for no less than 2ms, the circuit breaker will trip immediately; for overload and insulation faults, if the corresponding electrical quantity exceeds the dynamic set value for no less than 5 seconds, the circuit breaker will trip. (2.2) When At this time, a default setpoint is used as the protection setpoint for triggering protection measures. The default setpoint is a threshold preset by on-site power personnel and is higher than the benchmark setpoint. At the same time, inspection data verification is introduced. The specific measures are as follows: First, the default settings are loaded; then, the quantified inspection data is checked. If the inspection data indicates a fault, an audible and visual alarm is triggered; if the inspection data has not been updated for more than 24 hours, it automatically switches to the general settings, which are preset industry-standard protection thresholds. The operating logic in this case is as follows: For overcurrent and overvoltage faults, if the corresponding electrical quantity exceeds the default setting value by no less than 5ms, the circuit breaker will trip immediately; for overload and insulation faults, if the corresponding electrical quantity exceeds the default setting value by no less than 10 seconds, the circuit breaker will trip immediately. (2.3) When If this occurs, the protection setting that triggers the protection measures will be locked to the general setting, and an emergency alarm will be triggered. If the corresponding electrical quantity exceeds the general setting value, the circuit breaker will trip within 30 seconds after remote or on-site confirmation by maintenance personnel, and will automatically trip after 30 seconds.

10. The intelligent identification and linkage protection method for 35kV and below equipment based on multi-source adaptive fusion and deep / shallow dual-path network as described in claim 1, characterized in that, The training process of the deep and shallow dual-path network is supervised by a multi-task joint loss function. The loss function used is... The calculation formula is: ； In the above formula, For classifying losses, For feature importance loss, Loss is calibrated for confidence level; , , This refers to the task weighting coefficient; (1) Classification loss The multi-class cross-entropy loss, which combines equipment type and equipment status, is expressed as follows: ； In the above formula, The number of samples in a single training session; This represents the total number of combined categories for equipment type and equipment status. For the sample Belongs to the category of combination The tags use one-hot encoding; The predicted samples output by the final Softmax layer of the LSTM network Belongs to the category of combination The original probability; (2) Feature importance loss The expression is: ； In the above formula, steady-state characteristic flow The number of dimensions; For sample-based The first obtained during the calculation of steady-state global characteristics Importance weights of steady-state features; For the first The steady-state characteristics are based on preset target weights according to the physical properties of the equipment; (3) Confidence calibration loss The expression is: ； In the above formula, Number of bins for confidence level; For the first Number of samples in each bin; For the first The average raw confidence level of the samples within each bin; For the first Actual identification accuracy of samples within each bin.

Images