Lightweight waveform fault analysis method for wireless current sensor based on time domain feature and uncertainty dual-drive fusion decision tree

Abstract

The invention relates to the technical field of machine learning, in particular to a light-weight waveform fault analysis method for a wireless current sensor based on a time domain feature and uncertainty dual-drive fusion decision tree, and the method comprises the steps: collecting the real-time current data of a distribution line according to a fixed sampling rate through a wireless current sensor; a lightweight feature extraction method based on multi-dimensional time domain geometric features is adopted, and feature values capable of representing waveform forms and energy are calculated in real time from the current waveform time sequence; the feature vectors serve as input and are sent into an uncertainty double-drive fusion decision tree classification model, and the input feature vectors are mapped to predefined fault categories through hierarchical logic judgment rules; after training is completed, time domain feature vectors extracted in real time are input, and real-time identification and fault diagnosis are carried out on the operation state of the line; according to the invention, the problem that the current wireless current sensor cannot carry out real-time and accurate fault analysis on the power waveform at the terminal side due to the limitation of computing resources is solved.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a lightweight waveform fault analysis method for wireless current sensors based on a dual-drive fusion decision tree of time-domain features and uncertainty. Background Technology

[0002] With the rapid development of the power grid and the continuous expansion of the distribution network, the demand for line operation and maintenance and fault handling has surged. Live-line working has become a routine and crucial means to ensure the stable operation of the power grid. However, the current ability to perceive and diagnose the operating status of distribution lines is still lagging behind, especially the waveform fault analysis algorithms built into wireless current sensors widely deployed in the field, which face severe challenges.

[0003] Currently, most mainstream fault analysis methods rely on frequency domain analysis (such as Fast Fourier Transform, FFT) or complex deep learning models. While these methods offer high accuracy, they are computationally complex and resource-intensive. This leads to a significant contradiction: sensors collect massive amounts of waveform data but cannot perform real-time, intelligent analysis locally, often requiring the raw data to be transmitted back to the cloud or edge servers for processing. This not only incurs substantial communication overhead and latency but can also cause missed or false alarms for critical faults during network outages, preventing frontline maintenance personnel from receiving timely warnings and indirectly threatening their safety.

[0004] Therefore, researching a waveform fault analysis method that can accurately identify multiple faults while meeting the requirements of low power consumption and real-time performance has become an urgent need for intelligent operation and maintenance of power grids. An innovative, lightweight algorithm is urgently needed to enable truly intelligent wireless current sensors, allowing them to instantly detect line anomalies. Summary of the Invention

[0005] This invention provides a lightweight waveform fault analysis method for wireless current sensors based on a dual-drive fusion decision tree of time-domain features and uncertainties, which solves the problem that current wireless current sensors cannot perform real-time and accurate fault analysis of power waveforms at the terminal side due to limited computing resources.

[0006] To achieve the above objectives, the present invention employs the following technical solution:

[0007] A lightweight waveform fault analysis method for wireless current sensors based on a dual-drive fusion decision tree of time-domain features and uncertainties includes the following steps:

[0008] S1. Use a wireless current sensor to collect real-time current data of the power distribution line at a fixed sampling rate to obtain discrete current waveform time series;

[0009] S2. A lightweight feature extraction method based on multi-dimensional time-domain geometric features is adopted to calculate feature values ​​that can characterize the waveform shape and energy from the current waveform time series in real time. The set of feature values ​​includes the effective value of current, peak current, waveform factor, peak factor and zero crossing rate.

[0010] S3. After feature extraction is completed, the multi-dimensional temporal geometric feature values ​​are used as input and fed into the pre-trained uncertainty dual-drive fusion decision tree classification model. The uncertainty dual-drive fusion decision tree classification model maps the input feature vector to the predefined fault category through if-else hierarchical logic judgment rules.

[0011] S4. After training is completed, the real-time extracted time-domain feature vector is input to identify the line's operating status and diagnose faults in real time, and output the diagnosis results.

[0012] Furthermore, the diagnostic results include normal, overcurrent, undercurrent, suspected short circuit, and waveform distortion.

[0013] Furthermore, step S2 includes the following steps:

[0014] S2.1. Perform data preprocessing on the collected real-time current data and apply moving average filtering;

[0015] S2.2 Core Feature Extraction, specifically including:

[0016] S.2.2.1 Calculate the effective value of the current and quantify the average power of the current;

[0017] S2.2.2 Calculate the peak and valley values ​​of the current and detect instantaneous spikes or drops in the waveform;

[0018] S2.2.3 Calculate the waveform factor that represents the shape difference between the waveform and the standard sine wave;

[0019] S2.2.4 Calculate the peak factor, which reflects the sharpness of the waveform;

[0020] S2.2.5 Calculate the zero-crossing rate of the current;

[0021] S2.2.6, Output feature vector F:

[0022] ;

[0023] in, It is a matrix; This is the effective value of the current, i.e., the root mean square value; This represents the peak current. Waveform factor; Peak factor; This refers to the zero-crossing rate.

[0024] Furthermore, step S3 specifically includes the following steps:

[0025] S3.1 Feature fuzzification, generating membership degrees: For each continuous / discrete feature ,definition Given a fuzzy subset, calculate the fuzzy membership degree of each sample feature value using a membership function. The formula is as follows:

[0026] ;

[0027] in, The precise value of the i-th sample on the j-th feature belongs to the k-th fuzzy subset of the j-th feature. Membership degree;

[0028] The precise value of the i-th sample on the j-th feature is the feature value. This is the lower bound of the fuzzy subset; when the feature value is less than or equal to this value, the membership degree is 0.

[0029] This is the upper bound for fuzzy subsets: when the eigenvalue is greater than this value, the membership degree is 0;

[0030] The peak value for a fuzzy subset: when the feature value equals this value, the membership degree is 1;

[0031] S3.2, Class Conditional Probability Initialization: Initialize the entire dataset. Calculate the category conditional probability for each category, which serves as the initial probability distribution for the root node, using the following formula:

[0032] ;

[0033] in, Given the entire dataset D, what is the conditional probability that a sample belongs to class c? For dataset China belongs to Number of samples in each category This is the total number of samples;

[0034] S3.3, Constructing a decision tree through dual-drive splitting, specifically including the following steps:

[0035] S3.3.1 Calculate the fuzzy entropy and probability entropy for the current node;

[0036] Fuzzy entropy:

[0037] ;

[0038] in, For fuzzy entropy; for nodes of One sample Given 1 feature, sum the membership degree of all samples on all fuzzy subsets × log (membership degree);

[0039] Probability entropy:

[0040] ;

[0041] in, For probability entropy; A collection of categories; For nodes The medium sample belongs to the category The conditional probability;

[0042] S3.3.2 Calculate the dual-drive split gain and select the optimal split: For each feature Each fuzzy subset , will node Split into two child nodes and Calculate the dual-drive split gain Choose the one with the highest gain. As a mode of splitting, among which Sample membership , This is the membership threshold. Sample membership ;

[0043] Dual-drive split gain formula:

[0044] ;

[0045] in:

[0046] ;

[0047] ;

[0048] in, For dual-drive split gain; For fuzzy split gain; For probability splitting gain, , child node , The number of samples, These are the weighting coefficients;

[0049] S3.3.3, Generate dual-drive child nodes and recursively split them;

[0050] Each child node generated after splitting ( It needs to store two types of information simultaneously:

[0051] Fuzzy information: The precise input feature value corresponding to the node is converted into the membership degree of the corresponding fuzzy subset. Based on the preconditions of all fuzzy rules, the activation intensity of each rule is calculated. The activation intensity of the rule is applied to the subsequent rules to generate the output fuzzy subset.

[0052] Probability information: The probability distribution of this node's category. ,in for Medium category The number of samples; Repeat steps 3.3.1 to 3.3.3 until the stopping condition is met, stop splitting and set the node as a dual-drive leaf node;

[0053] S.4, Dual-drive pruning optimization;

[0054] Calculate the error of dual-drive pruning The formula is as follows:

[0055] ;

[0056] in, This is for the error caused by dual-drive pruning; This is for fuzzy pruning error; This represents the probabilistic pruning error.

[0057] Furthermore, step S4 specifically includes the following steps:

[0058] S4.1, New Sample Blurring: Blurring the new sample Each eigenvalue The fuzzy membership degree of a new sample on the c-th fuzzy subset of the j-th feature is calculated using the membership function. ;

[0059] S4.2 Traversing the decision tree to reach a leaf node: Starting from the root node, determine the new sample based on the fuzzy rules of each internal node. By checking if the membership degree satisfies the rule, the process is traversed downwards until the unique dual-drive leaf node is reached. ;

[0060] S4.3 Obtain and output diagnostic results through dual-drive fusion prediction;

[0061] Calculate the average membership degree of the new sample across all fuzzy subsets of the traversal path. This fuzzy confidence level is then fused with the probability of the leaf nodes to obtain the final predicted probability. The formula is as follows:

[0062] ;

[0063] in, This is the final predicted probability; For the entire dataset Under the condition that the sample belongs to category The conditional probability; Dual-drive leaf node In the diagram, the class conditional probability that a sample belongs to class c is... This represents a dual-drive leaf node.

[0064] Compared with the prior art, the beneficial effects of the present invention are:

[0065] 1) Uncertainty Dual-Driven Fusion Decision Tree: It solves the dual uncertainty problems of "fuzzy boundaries" (such as unclear fault category boundaries) and "random distribution" (such as fluctuations in sample data distribution) in traditional decision trees. Compared with traditional single decision trees (only probability or only fuzzy logic), it is more suitable for complex data scenarios of power distribution line current waveforms.

[0066] 2) Probability output mechanism: Quantifiable prediction confidence (e.g., through the final prediction probability) This approach not only retains the logical interpretability and if-else hierarchical judgment of decision trees, but also improves the accuracy of fault identification, avoiding the limitation of traditional algorithms that "only output the category without confidence".

[0067] 3) Dual-drive pruning optimization: Simultaneously controlling fuzzy and probabilistic errors. By using dual-drive error pruning, overfitting of the model is effectively avoided. In uncertain scenarios such as power distribution line load fluctuations and noise interference, the fault diagnosis stability is significantly better than that of traditional pruning methods. Attached Figure Description

[0068] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0069] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings:

[0070] See Figure 1 This is a flowchart of the method of the present invention. The present invention provides a lightweight waveform fault analysis method for wireless current sensors based on a dual-drive fusion decision tree of time-domain features and uncertainties, comprising the following steps:

[0071] S1. Use a wireless current sensor to collect real-time current data of the power distribution line at a fixed sampling rate to obtain discrete current waveform time series.

[0072] S2. A lightweight feature extraction method based on multi-dimensional time-domain geometric features is adopted to calculate feature values ​​that can characterize the waveform shape and energy from the current waveform time series in real time. The feature values ​​include at least: current effective value, current peak value, waveform factor, peak factor and zero crossing rate.

[0073] S3. After feature extraction is completed, the multi-dimensional temporal geometric feature values ​​are used as input and fed into the pre-trained uncertain dual-drive fusion decision tree classification model. The uncertain dual-drive fusion decision tree classification model maps the input feature vector to the predefined fault category through if-else hierarchical logic judgment rules.

[0074] S4. After training is completed, the time-domain feature vector extracted in real time is input to identify the operating status of the line and diagnose faults in real time, and output the diagnosis results, which include normal, overcurrent, undercurrent, suspected short circuit and waveform distortion.

[0075] The lightweight time-domain feature extraction algorithm set includes several sub-algorithms such as current RMS value calculation algorithm and peak factor calculation algorithm. The algorithm set is a collective term for these sub-algorithms, and its essence is still an expansion of lightweight feature extraction methods. It does not rely on complex frequency domain transformations, but directly calculates a small number of highly representative statistical and geometric features from the waveform of the signal changing over time (i.e., the "time domain"). In wireless current sensors, it "translates" a continuous, seemingly incomprehensible waveform into a series of digital "languages" that can be easily understood by machine learning models. In IoT and edge computing scenarios, although traditional signal processing methods (such as Fourier transform and wavelet analysis) are powerful, their huge computational load and memory consumption make them unsuitable for real-time operation on miniature, battery-powered sensors. Directly transmitting massive amounts of raw data wirelessly to the cloud will lead to high latency and high power consumption.

[0076] The lightweight temporal feature extraction algorithm solves the core problem: extracting valuable information from the raw data to the maximum extent and discarding redundant data, laying the foundation for subsequent intelligent judgment. The specific steps S2 include the following steps:

[0077] S2.1. Perform data preprocessing on the collected real-time current data and apply moving average filtering. The purpose is to eliminate high-frequency random noise in the original sampled signal, smooth the waveform, provide a cleaner and more stable signal source for subsequent feature extraction, and keep the computation lightweight.

[0078] S2.1.1 Define a window of fixed length L (e.g., L=5, i.e., 5 sampling points), where L is an odd number, usually 3 or 5;

[0079] S2.1.2 Starting from the beginning of the sequence, place the window on the first to Lth data points;

[0080] S2.1.3 Calculate the arithmetic mean of all data within the window, and output this mean as the new value of the window's center point, using the following formula:

[0081] ;

[0082] in, This is the original current sample value; This is the smoothed current value; This is the length of the filtering window; This is the index for the current time. This is the index variable in the summation loop.

[0083] S2.2 Core Feature Extraction: The purpose is to extract key digital features that can represent the shape and energy of the smoothed waveform.

[0084] S2.2.1 Calculating the effective value of the current and quantifying the average power (heating effect) of the current are core indicators for judging energy-related faults such as overcurrent and undercurrent. The calculation formula is as follows:

[0085] ;

[0086] in, The effective value of the current, also known as the root mean square value, is the core indicator for measuring the ability of current to do work (the amount of energy). This represents the total number of sampling points within a calculation window. To represent the first in the smoothed sequence The value of each point;

[0087] S2.2.2 Calculate the peak and valley values ​​of the current to detect instantaneous spikes or drops in the waveform, which is used to diagnose transient faults such as lightning surges, switching impulses, and short circuits; the calculation formula is as follows;

[0088] ;

[0089] ;

[0090] in, This is the peak current; Valley current; The entire smoothed current data sequence;

[0091] S2.2.3 Calculate the waveform factor, which characterizes the difference in shape between the waveform and the standard sine wave. This value changes significantly when harmonic distortion or waveform clipping occurs. The calculation formula is as follows:

[0092] ;

[0093] ;

[0094] in, The absolute value is the average of the sums of the absolute values ​​of all sample points. Waveform factor; This is the effective value of the current;

[0095] S2.2.4 Calculate the peak factor. The peak factor reflects the sharpness of the waveform. An abnormally high peak factor is a typical characteristic of sharp pulses such as arc faults and partial discharges. The calculation formula is as follows:

[0096] ;

[0097] in, Peak factor; This is calculated by dividing the peak value by the effective value.

[0098] S2.2.5 Calculate the zero-crossing rate of the current. In stable power frequency AC, the zero-crossing rate should be basically constant. Abnormal changes in it indicate frequency fluctuations or the presence of a large number of harmonics. The calculation formula is as follows:

[0099] ;

[0100] in, Zero-crossing rate; This is an indicator function; the function value is 1 if the condition inside the parentheses is true, and 0 otherwise. This is the condition for the indicator function; This is the second point to the [number]. Sum all zero-crossing events (using 1 as the digit) at each point to obtain the total number of zero-crossing events;

[0101] S2.2.6, Final Output:

[0102] ;

[0103] For feature vectors; It is a matrix;

[0104] The entire algorithm suite successfully transforms complex waveform analysis problems into lightweight computational tasks suitable for running on resource-constrained wireless current sensors through a series of basic yet efficient operations; furthermore, the final output is the perfect input for subsequent uncertain dual-drive fusion decision tree for fast and accurate fault diagnosis.

[0105] In step S3, the uncertainty dual-drive fusion decision tree DU-FDT is adopted. The core is to embed fuzzy processing and probabilistic modeling dual-drive modules in the entire decision tree process, from data preprocessing to prediction inference, to solve fuzzy uncertainty (such as semantic boundary fuzziness) and random uncertainty (such as sample distribution fluctuation).

[0106] Specifically, the steps include the following:

[0107] S3.1 Feature fuzzification, generating membership degrees: For each continuous / discrete feature ,definition Given a fuzzy subset, calculate the fuzzy membership degree of each sample feature value using a membership function. The commonly used triangle membership function (simple to calculate and with smooth boundary transitions) has the following formula:

[0108] ;

[0109] This function will use eigenvalues Membership degrees mapped to [0, 1] in fuzzy subsets peak The membership degree is 1 (completely belong) at the minimum value. and maximum value Outside of these regions, the membership degree is 0 (not belonging at all), and the intermediate region is a linear transition.

[0110] in, The precise value of the i-th sample on the j-th feature belongs to the k-th fuzzy subset of the j-th feature. Membership degree;

[0111] The precise value of the i-th sample on the j-th feature is the feature value. This is the lower bound of the fuzzy subset; when the feature value is less than or equal to this value, the membership degree is 0.

[0112] This is the upper bound for fuzzy subsets: when the eigenvalue is greater than this value, the membership degree is 0;

[0113] The peak value of the fuzzy subset: when the feature value is equal to this value, the membership degree is 1.

[0114] S3.2, Class Conditional Probability Initialization: Initialize the entire dataset. Calculate the category conditional probability for each category, which serves as the initial probability distribution for the root node, using the following formula:

[0115] ;

[0116] in, Given the entire dataset D, what is the conditional probability that a sample belongs to class c? For dataset China belongs to Number of samples in each category It represents the total number of samples; this formula quantifies random uncertainty, that is, whether a randomly selected sample belongs to a certain category. The probability of.

[0117] S3.3, Constructing a decision tree through dual-drive splitting, specifically including the following steps:

[0118] S3.3.1 For the current node, calculate the fuzzy entropy and probability entropy respectively; the former measures fuzzy uncertainty, and the latter measures random uncertainty.

[0119] Fuzzy entropy:

[0120] ;

[0121] in, For fuzzy entropy; for nodes of One sample The entropy is calculated by summing the membership degrees of all samples on all fuzzy subsets × log(membership degree); the smaller the entropy value, the clearer the sample's belonging to the fuzzy subset (low fuzzy uncertainty).

[0122] Probability entropy:

[0123] ;

[0124] in, For probability entropy; A collection of categories; For nodes The medium sample belongs to the category The conditional probability;

[0125] S3.3.2 Calculate the dual-drive split gain and select the optimal split: For each feature Each fuzzy subset , will node Split into two child nodes (sample membership degree) , (This is the membership threshold, usually taken as 0.5) and (sample membership degree) ), calculate the dual-drive split gain Choose the one with the highest gain. As a means of division;

[0126] Dual-drive split gain formula:

[0127] ;

[0128] in:

[0129] ;

[0130] ;

[0131] in, For dual-drive split gain; For fuzzy split gain; For probability splitting gain, , child node , The number of samples, The larger the split gain, the greater the reduction in the total uncertainty of the nodes after the split, indicating a better splitting effect.

[0132] S3.3.3, Generate dual-drive child nodes and recursively split them;

[0133] Each child node generated after splitting ( It needs to store two types of information simultaneously:

[0134] Fuzzy information: The feature corresponding to this node - fuzzy subset splitting rule, specifically including the process:

[0135] 1) Convert precise input feature values ​​into membership degrees of corresponding fuzzy subsets;

[0136] 2) Rule evaluation: Calculate the activation strength of each rule based on the preconditions of all fuzzy rules;

[0137] 3) Reasoning: Apply the activation strength of the rule to the subsequent rules to generate a fuzzy subset of the output;

[0138] 4) Defuzzification: Aggregate the fuzzy subsets of output generated by all activated rules and convert them back to a precise output value through an algorithm for actual control;

[0139] Probability information: The probability distribution of this node's category. ,in for Medium category The number of samples; Repeat steps 3.3.1 to 3.3.3 until the stopping condition is met, such as the number of node samples ≤ minimum threshold and split gain ≤ 0, then stop splitting and set the node as a dual-drive leaf node.

[0140] S.4. Dual-drive pruning optimization avoids overfitting and controls two types of uncertainty errors simultaneously. The goal is to remove redundant branches in the tree and ensure that the pruned tree has the minimum "fuzzy + probability" combined error on the test set.

[0141] Calculate the error of dual-drive pruning The formula is as follows:

[0142] ;

[0143] in, This is for the error caused by dual-drive pruning; This is for fuzzy pruning error; This represents the probabilistic pruning error.

[0144] Step S4 employs dual-drive predictive inference, inputting a new sample and outputting the fused class probability. The goal is to reach the leaf node along the splitting path of the tree for the new sample, fusing the fuzzy membership degree of the path and the probability of the leaf node, and outputting the final class prediction probability. Specifically, it includes the following steps:

[0145] S4.1, New Sample Blurring: Blurring the new sample Each eigenvalue The fuzzy membership degree of a new sample on the c-th fuzzy subset of the j-th feature is calculated using the membership function. ;

[0146] S4.2 Traversing the decision tree to reach a leaf node: Starting from the root node, determine the new sample based on the feature-fuzzy subset rule of each internal node. By checking if the membership degree satisfies the rule, the process is traversed downwards until the unique dual-drive leaf node is reached. ;

[0147] S4.3 Obtain and output diagnostic results through dual-drive fusion prediction;

[0148] Calculate the average membership degree of the new sample across all fuzzy subsets of the traversal path. This fuzzy confidence level is then fused with the probability of the leaf nodes to obtain the final predicted probability. The formula is as follows:

[0149] ;

[0150] in, This is the final predicted probability; For the entire dataset Under the condition that the sample belongs to category The conditional probability; Dual-drive leaf node In the diagram, the class conditional probability that a sample belongs to class c is... This represents a dual-drive leaf node.

[0151] The above embodiments are implemented based on the technical solution of the present invention, providing detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the above embodiments. Unless otherwise specified, the methods used in the above embodiments are conventional methods.

Claims

1. A lightweight waveform fault analysis method for wireless current sensors based on a dual-drive fusion decision tree of time-domain characteristics and uncertainties, characterized in that, Includes the following steps: S1. Use a wireless current sensor to collect real-time current data of the power distribution line at a fixed sampling rate to obtain discrete current waveform time series; S2. A lightweight feature extraction method based on multi-dimensional time-domain geometric features is adopted to calculate feature values ​​that can characterize the waveform shape and energy from the current waveform time series in real time. The set of feature values ​​includes the effective value of current, peak current, waveform factor, peak factor and zero crossing rate. S3. After feature extraction is completed, the multi-dimensional temporal geometric feature values ​​are used as input and fed into the pre-trained uncertainty dual-drive fusion decision tree classification model. The uncertainty dual-drive fusion decision tree classification model maps the input feature vector to the predefined fault category through if-else hierarchical logic judgment rules. S4. After training is completed, the real-time extracted time-domain feature vector is input to identify the line's operating status and diagnose faults in real time, and output the diagnosis results.

2. The lightweight waveform fault analysis method for wireless current sensors based on a dual-drive fusion decision tree of time-domain features and uncertainties as described in claim 1, characterized in that, The diagnostic results include normal, overcurrent, undercurrent, suspected short circuit, and waveform distortion.

3. The lightweight waveform fault analysis method for wireless current sensors based on a dual-drive fusion decision tree of time-domain features and uncertainties as described in claim 1, characterized in that, Step S2 includes the following steps: S2.

1. Perform data preprocessing on the collected real-time current data and apply moving average filtering; S2.2 Core Feature Extraction, specifically including: S.2.2.1 Calculate the effective value of the current and quantify the average power of the current; S2.2.2 Calculate the peak and valley values ​​of the current and detect instantaneous spikes or drops in the waveform; S2.2.3 Calculate the waveform factor that represents the shape difference between the waveform and the standard sine wave; S2.2.4 Calculate the peak factor, which reflects the sharpness of the waveform; S2.2.5 Calculate the zero-crossing rate of the current; S2.2.6, Output feature vector F: ； in, It is a matrix; This is the effective value of the current, i.e., the root mean square value; This represents the peak current. Waveform factor; Peak factor; This refers to the zero-crossing rate.

4. The lightweight waveform fault analysis method for wireless current sensors based on a dual-drive fusion decision tree of time-domain features and uncertainties as described in claim 3, characterized in that, Step S3 specifically includes the following steps: S3.1 Feature fuzzification, generating membership degrees: For each continuous / discrete feature ,definition Given a fuzzy subset, calculate the fuzzy membership degree of each sample feature value using a membership function. The formula is as follows: ； in, The precise value of the i-th sample on the j-th feature belongs to the k-th fuzzy subset of the j-th feature. Membership degree; The precise value of the i-th sample on the j-th feature is the feature value. This is the lower bound of the fuzzy subset; when the feature value is less than or equal to this value, the membership degree is 0. This is the upper bound for fuzzy subsets: when the eigenvalue is greater than this value, the membership degree is 0; The peak value for a fuzzy subset: when the feature value equals this value, the membership degree is 1; S3.2, Class Conditional Probability Initialization: Initialize the entire dataset. Calculate the category conditional probability for each category, which serves as the initial probability distribution for the root node, using the following formula: ； in, Given the entire dataset D, what is the conditional probability that a sample belongs to class c? For dataset China belongs to Number of samples in each category This is the total number of samples; S3.3, Constructing a decision tree through dual-drive splitting, specifically including the following steps: S3.3.1 Calculate the fuzzy entropy and probability entropy for the current node; Fuzzy entropy: ； in, For fuzzy entropy; for nodes of One sample Given 1 feature, sum the membership degree of all samples on all fuzzy subsets × log (membership degree); Probability entropy: ； in, For probability entropy; A collection of categories; For nodes The medium sample belongs to the category The conditional probability; S3.3.2 Calculate the dual-drive split gain and select the optimal split: For each feature Each fuzzy subset , will node Split into two child nodes and Calculate the dual-drive split gain Choose the one with the highest gain. As a mode of splitting, among which Sample membership , This is the membership threshold. Sample membership ; Dual-drive split gain formula: ； in: ； ； in, For dual-drive split gain; For fuzzy split gain; For probability splitting gain, , child node , The number of samples, These are the weighting coefficients; S3.3.3, Generate dual-drive child nodes and recursively split them; Each child node generated after splitting ( It needs to store two types of information simultaneously: Fuzzy information: The precise input feature value corresponding to the node is converted into the membership degree of the corresponding fuzzy subset. Based on the preconditions of all fuzzy rules, the activation intensity of each rule is calculated. The activation intensity of the rule is applied to the subsequent rules to generate the output fuzzy subset. Probability information: The probability distribution of this node's category. ,in for Medium category The number of samples; Repeat steps 3.3.1 to 3.3.3 until the stopping condition is met, stop splitting and set the node as a dual-drive leaf node; S.4, Dual-drive pruning optimization; Calculate the error of dual-drive pruning The formula is as follows: ； in, This is for the error caused by dual-drive pruning; This is for fuzzy pruning error; This represents the probabilistic pruning error.

5. The lightweight waveform fault analysis method for wireless current sensors based on a dual-drive fusion decision tree of time-domain features and uncertainties as described in claim 4, characterized in that, Step S4 specifically includes the following steps: S4.1, New Sample Blurring: Blurring the new sample Each eigenvalue The fuzzy membership degree of a new sample on the c-th fuzzy subset of the j-th feature is calculated using the membership function. ; S4.2 Traversing the decision tree to reach a leaf node: Starting from the root node, determine the new sample based on the fuzzy rules of each internal node. By checking if the membership degree satisfies the rule, the process is traversed downwards until the unique dual-drive leaf node is reached. ; S4.3 Obtain and output diagnostic results through dual-drive fusion prediction; Calculate the average membership degree of the new sample across all fuzzy subsets of the traversal path. This fuzzy confidence level is then fused with the probability of the leaf nodes to obtain the final predicted probability. The formula is as follows: ； in, This is the final predicted probability; For the entire dataset Under the condition that the sample belongs to category The conditional probability; Dual-drive leaf node In the diagram, the class conditional probability that a sample belongs to class c is... This represents a dual-drive leaf node.

Images