Time domain AC equivalent DC resistance test and analysis method based on big data

Abstract

The invention relates to the technical field of power equipment detection and state evaluation, in particular to a time domain alternating current equivalent direct current resistance test and analysis method based on big data. According to the method, the equivalent direct current resistance is calculated through two independent paths of a linear prediction model and time domain signal iterative optimization, the time sequence characteristics of the equivalent resistance and the working condition correlation characteristics are fused, the high-dimensional characteristic matrix is constructed, the data internal law and the working condition influence mechanism are fully mined, and rich characteristic support is provided for the model; the geometric cascade forest model is good at basic feature classification, the gradient boosting tree and the isolated forest fusion model give consideration to health state evaluation and anomaly detection, the two complements each other to form dual research and judgment, the comprehensiveness and accuracy of state recognition are improved, and the accuracy of state recognition is improved through dual prediction result comparison and anomaly feedback list output. And data or model problems can be found in time, a direction is provided for subsequent optimization, and the stability and credibility of an analysis result are ensured.

Description

Technical Field

[0001] This invention relates to the field of power equipment testing and condition assessment technology, and in particular to a method for testing and analyzing time-domain AC equivalent DC resistance based on big data. Background Technology

[0002] The DC resistance of power equipment is a core parameter reflecting the health status of key components such as windings and joints. Its value changes are directly related to potential faults such as poor contact, winding aging, and insulation damage. Traditional DC resistance testing requires power outages, which not only affects the continuity of power supply but also has a long testing cycle and low efficiency, making it difficult to meet the needs of modern power systems for "condition-based maintenance".

[0003] Currently, most AC equivalent DC resistance testing methods are based on a single signal processing algorithm, which has problems such as weak anti-interference ability and insufficient testing accuracy. At the same time, the analysis of test data often relies on a single model and does not fully integrate equipment operating information and multi-dimensional features, resulting in limited accuracy of condition judgment and difficulty in effectively identifying potential faults. Therefore, a solution is proposed. Summary of the Invention

[0004] The purpose of this invention is to provide a method for testing and analyzing the equivalent DC resistance of AC power equipment based on big data. This method achieves high-precision testing of the equivalent DC resistance of power equipment and accurate identification of its health status through dual-path resistance calculation, multi-dimensional feature fusion, dual-model analysis and cross-validation.

[0005] The objective of this invention can be achieved through the following technical solution: a method for testing and analyzing the time-domain AC equivalent DC resistance based on big data, comprising the following steps: Step 1: Dual-path calculation of equivalent DC resistance: Collect sample data to build a prediction model and calculate the equivalent DC resistance ZR in real time; at the same time, denoise compensation is performed on the time-domain waveform data, the initial resistance is calculated in segments, and the final equivalent resistance RD is obtained through weighted iterative optimization. Step 2: Verify resistor validity: Calculate the deviation between ZR and RD. If the deviation is within the preset range, a valid signal is generated; otherwise, an invalid signal is generated and an alarm is triggered. Step 3: Constructing Features and Storing Data: After the effective signal is generated, the resistance, time, and operating condition data are associated and stored. At the same time, the resistance time series features and the operating condition associated features are fused to construct a high-dimensional feature matrix and basic feature vectors. Step 4: Dual-model assessment of equipment status: Input the basic feature vector into the geometric cascaded forest model to obtain the first prediction result; input the high-dimensional feature matrix into the gradient boosting tree + isolated forest fusion model, and combine with interactive analysis to obtain the second prediction result; Step 5: Verify and provide feedback: Compare the two prediction results. If they match, output the result; otherwise, generate a deviation signal and build an anomaly feedback list.

[0006] Preferably, the process for obtaining the equivalent DC resistance ZR is as follows: S1: Collect AC time-domain waveform data of n (n>5) sets of power equipment as sample data, and extract two core variables (predictor variables and reference variables) from the sample data. The predictor variables are standardized, and the final coefficient vector β is obtained by iterative processing of the standardized predictor variables and reference variables using the set least squares method. S2: Real-time acquisition of AC time-domain waveform data of m (m>0) groups of power equipment, and standardization processing of the AC time-domain waveform data of each group to obtain the prediction variable Xnew; S3: The equivalent DC resistance value of each group is calculated based on R = predicted variable Xnew × final coefficient vector β; S4: Calculate the average value of the equivalent DC resistance and use the result of the average calculation of the equivalent DC resistance as the final equivalent DC resistance ZR.

[0007] Preferably, the process for obtaining the final equivalent resistance RD is as follows: The acquired AC time-domain waveform data is processed using existing wavelet threshold denoising and adaptive inductance compensation to obtain a purely resistive voltage signal containing only the voltage drop across the resistor. The processed pure resistive voltage and current signals are divided into N (N>10) segments according to the time series. The resistance value in each segment is initially calculated based on Ohm's law to obtain multi-group segmented resistance data. Extract the purely resistive voltage data URk and current data Ik of the k-th segment. For each segment k, calculate the initial resistance value Rk within that segment based on Ohm's law. The standard deviation of the initial resistance value Rk of each segment is used as the weighting coefficient σk. The initial weight wk is obtained by standardizing the weighting coefficient σk. Based on the initial weight wk, the initial equivalent resistance value RD is calculated using the weighted average method. The initial equivalent resistance value RD = ∑Rk × wk. Preferably, the initial weight wk is updated to obtain wk(b+1), and the equivalent resistance is updated to obtain RD(b+1). The convergence condition is: when |RD(b+1)-RD(b)|≤ a preset threshold, the iteration stops, and RD(b+1) is the final equivalent DC resistance value.

[0008] Preferably, the equivalent DC resistance ZR is compared and analyzed with the final equivalent resistance value RD(b+1). The difference between the equivalent DC resistance ZR and the final equivalent resistance value RD(b+1) is set as the equivalent DC resistance deviation. The equivalent DC resistance deviation is judged to obtain the result of valid signal or invalid signal. When a valid signal is generated, the calculated equivalent DC resistance ZR, acquisition time, and equipment operating condition data are associated and stored in the database, and a resistance-operating condition association library is constructed.

[0009] Preferably, the dual-model assessment of equipment status analysis process is as follows: Extract the feature parameters of power equipment from the database to form a feature vector; The feature vectors are input into the geometric cascaded forest algorithm model to obtain the class probability vectors of the Lth layer and ML geometric forests. The arithmetic mean of the collected ML probability vectors is taken according to the category dimension to obtain the final average probability vector. The probability values ​​of each category (including normal, potential risk and abnormal) in the final average probability vector are compared, and the category with the highest probability is selected as the current final prediction result.

[0010] Preferably, in addition to extracting the time-series features of the equivalent resistance value, operating condition correlation features are introduced to form a high-dimensional feature matrix; The high-dimensional feature matrix is ​​input into the gradient boosting tree + isolation forest fusion model to obtain the output real-time health index and real-time isolation degree. The real-time health index is compared and analyzed to obtain the state results of normal, potential risk and abnormal. Normal, potential risk and abnormal are collectively referred to as state level. A comparative analysis of real-time isolation is performed to obtain results of no potential faults and potential faults. The results of no potential faults and potential faults are collectively referred to as fault results. Preferably, interactive analysis is performed on the status level and fault results, and the final status judgment result is obtained based on the interactive analysis. The final status judgment result includes normal, potential risk and abnormal. The final state determination result is compared and analyzed with the final prediction result. If the final state determination result is consistent with the final prediction result, the final prediction result is output and displayed. If the final state determination result is inconsistent with the final prediction result, a deviation signal is generated and the deviation signal is responded to immediately. An anomaly feedback list is constructed based on the final state determination result, the final prediction result, the real-time health index, the real-time isolation degree, the fault result, the state level, and the probability value of each category.

[0011] The beneficial effects of this invention are as follows: This invention calculates the equivalent DC resistance through two independent paths: a linear prediction model and iterative optimization of the time-domain signal. Combined with a validity verification mechanism, it effectively eliminates the influence of abnormal data, ensuring the accuracy and reliability of the resistance value. Furthermore, it integrates the time-series characteristics of the equivalent resistance with the operating condition correlation characteristics to construct a high-dimensional feature matrix, fully exploring the inherent laws of the data and the operating condition influence mechanism, providing rich feature support for the model.

[0012] This invention also leverages the geometric cascaded forest model, which excels at basic feature classification, and the gradient boosting tree + isolated forest fusion model, which balances health status assessment and anomaly detection. The two complement each other to form a dual judgment, improving the comprehensiveness and accuracy of status identification. Furthermore, by comparing the dual prediction results and outputting anomaly feedback lists, data or model problems can be identified in a timely manner, providing direction for subsequent optimization and ensuring the stability and reliability of the analysis results. Attached Figure Description

[0013] The invention will now be further described with reference to the accompanying drawings; Figure 1 This is a reference diagram of the method of the present invention; Figure 2 This is a reference diagram for the ZR analysis of the equivalent DC resistance; Figure 3 This is a diagram illustrating the process of determining the deviation of equivalent DC resistance. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments; Example 1: Please refer to Figures 1 to 3 As shown, this invention is a method for testing and analyzing the time-domain AC equivalent DC resistance based on big data, including the following steps: Step 1: Dual-path calculation of equivalent DC resistance: Collect sample data to build a prediction model and calculate the equivalent DC resistance ZR in real time; at the same time, denoise compensation is performed on the time-domain waveform data, the initial resistance is calculated in segments, and the final equivalent resistance RD is obtained through weighted iterative optimization. Step 2: Verify resistor validity: Calculate the deviation between ZR and RD. If the deviation is within the preset range, a valid signal is generated; otherwise, an invalid signal is generated and an alarm is triggered. Step 3: Constructing Features and Storing Data: After the effective signal is generated, the resistance, time, and operating condition data are associated and stored. At the same time, the resistance time series features and the operating condition associated features are fused to construct a high-dimensional feature matrix and basic feature vectors. Step 4: Dual-model assessment of equipment status: Input the basic feature vector into the geometric cascaded forest model to obtain the first prediction result; input the high-dimensional feature matrix into the gradient boosting tree + isolated forest fusion model, and combine with interactive analysis to obtain the second prediction result; Step 5: Verify and provide feedback: Compare the two prediction results. If they match, output the result. If they do not match, generate a deviation signal and build an anomaly feedback list. The process of obtaining the equivalent DC resistance ZR in step one is as follows: S1: Collect AC time-domain waveform data from n (n>5) sets of power equipment as sample data, and extract two core variables from the sample data: Predicted variables (inputs): RMS AC voltage, RMS AC current, and voltage-current phase difference; Reference variable (benchmark): The actual value of DC resistance detected during historical power outages of the equipment; The predictor variables are standardized, and the final coefficient vector β is obtained by iterative processing of the standardized predictor variables and reference variables using the set least squares method. S2: Real-time acquisition of AC time-domain waveform data of m (m>0) groups of power equipment, and standardization processing of the AC time-domain waveform data of each group to obtain the prediction variable Xnew. The prediction variable Xnew represents a 1-row, 3-column feature vector, such as the standardized [AC voltage RMS value, AC current RMS value, voltage-current phase difference]; S3: The equivalent DC resistance value of each group is calculated based on R = predicted variable Xnew × final coefficient vector β; S4: Calculate the average value of the equivalent DC resistance and use the result of the average calculation of the equivalent DC resistance as the final equivalent DC resistance ZR; The process for obtaining the final equivalent resistance RD is as follows: The acquired AC time-domain waveform data is processed using existing wavelet threshold denoising and adaptive inductance compensation to obtain a purely resistive voltage signal containing only the voltage drop across the resistor. The processed pure resistive voltage and current signals are divided into N (N>10) segments according to the time series (e.g., each segment is 100ms long to ensure data stability). The resistance value is initially calculated based on Ohm's law within each segment to obtain multi-group segmented resistance data. The time interval of the k-th segment (k = 1, 2, ..., N) is [tk-1, tk], where t0 = 0, tN = T_total, and T_total is the total duration of a single data collection. Extract the purely resistive voltage data URk and current data Ik of the k-th segment; For each segment k, calculate the initial resistance value Rk within that segment based on Ohm's law; The standard deviation of the initial resistance value Rk of each segment is used as the weighting coefficient σk. The initial weight wk is obtained by standardizing the weighting coefficient σk. Based on the initial weight wk, the initial equivalent resistance value RD is calculated using the weighted average method. The initial equivalent resistance value RD = ∑Rk × wk. Weight update: wk(b+1)=wk(b)×[1-|Rk-RD(b)| / (max(Rk)-min(Rk))], where b is the iteration number. The smaller the deviation, the higher the weight increase ratio. Equivalent resistance update: RD(b+1)=[∑Rk×wk(b+1)] / ∑initial weight wk(b+1); Convergence condition: If |RD(b+1)-RD(b)|≤ preset threshold, then the iteration stops, and RD(b+1) is the final equivalent DC resistance value; The equivalent DC resistance ZR and the final equivalent resistance value RD(b+1) are compared and analyzed. The difference between the equivalent DC resistance ZR and the final equivalent resistance value RD(b+1) is set as the equivalent DC resistance deviation. The equivalent DC resistance deviation is judged. If the equivalent DC resistance deviation is within the preset range, a valid signal is generated. If the equivalent DC resistance deviation is not within the preset range, an invalid signal is generated. In response to the invalid signal, the preset warning text corresponding to the invalid signal is displayed immediately so as to promptly re-check the data preprocessing process and ensure the accuracy of the analysis results. When a valid signal is generated, the calculated equivalent DC resistance ZR, acquisition time, and equipment operating conditions (such as load rate, ambient temperature, and running time) are associated and stored in the database, and a resistance-operating condition association library is constructed.

[0016] Example 2: Analysis of the first prediction result obtained from inputting the basic feature vector into the geometric cascade forest model in step four, including: Extract feature parameters of power equipment from the database (such as the slope of the linear trend of resistance and the mean resistance value) to form a feature vector; The feature vector is input into the geometric cascaded forest algorithm model to obtain the class probability vectors of the Lth layer ML geometric forests. The arithmetic mean of the collected ML probability vectors is taken according to the category dimension to obtain the final average probability vector. The probability values ​​of each category (including normal, potential risk and abnormal) in the final average probability vector are compared, and the category with the highest probability is selected as the current final prediction result. For example: if the probability of the "normal" category (0.895) is the highest, then the corresponding power equipment status is determined to be "normal"; The training process for the geometric cascade forest algorithm model is as follows: Initialization: Use the original feature vector as the output C of layer 0. 0 (xi) = xi, where xi is a single sample in the training set, containing d original features, and both i and d are greater than zero; Cascaded growth: For each layer L (1≤L≤Lmax), ML training subsets are generated by Bootstrap sampling, and a forest containing T geometric decision trees is trained for each subset; the training set is input into the current layer forest to generate class probability vectors, which are concatenated with the enhancement features of the previous layer to form a new enhancement feature vector; the performance is evaluated through the validation set, and if the performance no longer improves, training is stopped, and the final number of layers L is determined; Training complete: L-layer cascaded model obtained; The analysis process for the second prediction result in step four is as follows: In addition to extracting the time-series features (mean, variance, trend slope, number of abrupt changes) of the equivalent resistance value, additional operating condition-related features (such as the correlation coefficient between load rate and resistance value, and the influence coefficient of ambient temperature on resistance value) are introduced to form a high-dimensional feature matrix, thereby improving the ability of subsequent models to make judgments. The high-dimensional feature matrix is ​​input into the gradient boosting tree + isolation forest fusion model to obtain the output real-time health index and real-time isolation degree. The real-time health index is compared and analyzed, and the preset health index ranges JKmax and JKmin are retrieved. If the real-time health index > JKmax, it is judged as normal; if JKmin ≤ real-time health index ≤ JKmax, it is judged as potential risk; if the real-time health index < JKmin, it is judged as abnormal. Normal, potential risk, and abnormal are collectively referred to as status levels; The real-time isolation degree is compared and analyzed. If the real-time isolation degree is less than the preset real-time isolation degree threshold, it is determined that there is no potential fault. If the real-time isolation degree is greater than or equal to the preset real-time isolation degree threshold, it is determined that there is no potential fault. The absence of potential fault and potential fault are collectively referred to as fault results. Interactive analysis is performed on the status level and fault results. The final status judgment result is obtained based on the interactive analysis. The final status judgment result includes normal, potential risk and abnormal. Verify and provide feedback: Compare and analyze the final state determination result with the final prediction result. If the final state determination result is consistent with the final prediction result, output the final prediction result. If the final state determination result is inconsistent with the final prediction result, generate a deviation signal and respond immediately to the deviation signal. Based on the final state determination result, the final prediction result, the real-time health index, the real-time isolation degree, the fault result, the state level, and the probability value of each category, construct an anomaly feedback list and output the anomaly feedback list for display. This allows for targeted data processing based on the anomaly feedback list to improve the accuracy of the result analysis. The interactive analysis process is as follows: If the status level is normal and the fault result indicates no potential fault, then the final result is determined to be normal. If the status level is normal and the fault result is a potential fault, then the final result is determined to be a potential risk. If the status level is potential risk and the failure result is no potential failure, then the final result is determined to be potential risk. If the status level is potential risk and the failure result is a potential failure, then the final result is determined to be abnormal. If the status level is abnormal and the fault result is no potential fault, then the final result is determined to be abnormal. If the status level is abnormal and the fault result is a potential fault, then the final result is determined to be abnormal. For example, the training process of Gradient Boosting Tree (XGBoost) + Isolation Forest is as follows: Gradient Boosting Tree (XGBoost) Training: Training set construction: Select the feature matrix of historical health equipment (including equivalent resistance time series features and operating condition correlation features), and label the health status (normal: 1, potential risk: 0.5, abnormal: 0). Model parameter settings: learning rate 0.1, tree depth 6, number of estimators 100, objective function uses mean squared error (MSE), and parameters are optimized through 5-fold cross-validation; Output: Current health index of the device (0-100 points, obtained by linear mapping of the predicted value output by the model).

[0017] Isolation Forest Training: Training set construction: Select abnormal feature data of historical faulty equipment (fault types such as poor contact and winding aging) and normal feature data of healthy equipment to construct a mixed training set; Model parameter settings: 100 trees, 256 subsamples, and parameters optimized using grid search; Output: The isolation degree of the test sample (0-1, isolation degree > 0.7 (threshold) is judged as a potential fault).

[0018] Fusion rules: Set a "dual-condition trigger" mechanism. When the health index is <80 points and the isolation degree is >0.7, trigger a fault warning. In summary, this invention calculates the equivalent DC resistance through two independent paths: a linear prediction model and iterative optimization of the time-domain signal. Combined with a validity verification mechanism, it effectively eliminates the influence of abnormal data, ensuring the accuracy and reliability of the resistance value. Furthermore, it integrates the time-series characteristics of the equivalent resistance with the operating condition correlation characteristics to construct a high-dimensional feature matrix, fully exploring the inherent patterns of the data and the operating condition influence mechanism, providing rich feature support for the model. The geometric cascaded forest model excels at basic feature classification, while the gradient boosting tree + isolated forest fusion model takes into account both health status assessment and anomaly detection. The two complement each other to form a dual judgment, improving the comprehensiveness and accuracy of status identification. Moreover, by comparing the dual prediction results and outputting an anomaly feedback list, data or model problems can be identified in a timely manner, providing direction for subsequent optimization and ensuring the stability and credibility of the analysis results.

[0019] The threshold is set for comparative analysis of results to determine whether they are good or bad. The value of the threshold is determined by a combination of large-scale model analysis of sample data and human experience. It can also be adjusted appropriately based on seasonal or common-sense influencing factors.

[0020] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for testing and analyzing the time-domain AC equivalent DC resistance based on big data, characterized in that, Includes the following steps: Step 1: Dual-path calculation of equivalent DC resistance: Collect sample data to build a prediction model and calculate the equivalent DC resistance ZR in real time; at the same time, denoise compensation is performed on the time-domain waveform data, the initial resistance is calculated in segments, and the final equivalent resistance RD is obtained through weighted iterative optimization. Step 2: Verify resistor validity: Calculate the deviation between ZR and RD. If the deviation is within the preset range, a valid signal is generated; otherwise, an invalid signal is generated and an alarm is triggered. Step 3: Constructing Features and Storing Data: After the effective signal is generated, the resistance, time, and operating condition data are associated and stored. At the same time, the resistance time series features and the operating condition associated features are fused to construct a high-dimensional feature matrix and basic feature vectors. Step 4: Dual-model assessment of equipment status: Input the basic feature vector into the geometric cascaded forest model to obtain the first prediction result; input the high-dimensional feature matrix into the gradient boosting tree + isolated forest fusion model, and combine with interactive analysis to obtain the second prediction result; Step 5: Verify and provide feedback: Compare the two prediction results. If they match, output the result; otherwise, generate a deviation signal and build an anomaly feedback list.

2. The method for testing and analyzing time-domain AC equivalent DC resistance based on big data according to claim 1, characterized in that, The process for obtaining the equivalent DC resistance ZR is as follows: S1: Collect AC time-domain waveform data of n (n>5) sets of power equipment as sample data, and extract two core variables (predictor variables and reference variables) from the sample data. The predictor variables are standardized, and the final coefficient vector β is obtained by iterative processing of the standardized predictor variables and reference variables using the set least squares method. S2: Real-time acquisition of AC time-domain waveform data of m (m>0) groups of power equipment, and standardization processing of the AC time-domain waveform data of each group to obtain the prediction variable Xnew; S3: The equivalent DC resistance value of each group is calculated based on R = predicted variable Xnew × final coefficient vector β; S4: Calculate the average value of the equivalent DC resistance and use the result of the average calculation of the equivalent DC resistance as the final equivalent DC resistance ZR.

3. The method for testing and analyzing time-domain AC equivalent DC resistance based on big data according to claim 1, characterized in that, The process for obtaining the final equivalent resistance RD is as follows: The acquired AC time-domain waveform data is processed using existing wavelet threshold denoising and adaptive inductance compensation to obtain a purely resistive voltage signal containing only the voltage drop across the resistor. The processed pure resistive voltage and current signals are divided into N (N>10) segments according to the time series. The resistance value in each segment is initially calculated based on Ohm's law to obtain multi-group segmented resistance data. Extract the purely resistive voltage data URk and current data Ik of the k-th segment. For each segment k, calculate the initial resistance value Rk within that segment based on Ohm's law. The standard deviation of the initial resistance value Rk of each segment is used as the weighting coefficient σk. The initial weight wk is obtained by standardizing the weighting coefficient σk. Based on the initial weight wk, the initial equivalent resistance value RD is calculated using the weighted average method. The initial equivalent resistance value RD = ∑Rk × wk.

4. The method for testing and analyzing time-domain AC equivalent DC resistance based on big data according to claim 3, characterized in that, The initial weight wk is updated to obtain wk(b+1), and the equivalent resistance is updated to obtain RD(b+1). The convergence condition is: when |RD(b+1)-RD(b)|≤ preset threshold, the iteration stops and RD(b+1) is the final equivalent DC resistance value.

5. The method for testing and analyzing time-domain AC equivalent DC resistance based on big data according to claim 4, characterized in that, The equivalent DC resistance ZR and the final equivalent resistance value RD(b+1) are compared and analyzed. The difference between the equivalent DC resistance ZR and the final equivalent resistance value RD(b+1) is set as the equivalent DC resistance deviation. The equivalent DC resistance deviation is judged to obtain the result of valid signal or invalid signal. When a valid signal is generated, the calculated equivalent DC resistance ZR, acquisition time, and equipment operating condition data are associated and stored in the database, and a resistance-operating condition association library is constructed.

6. The method for testing and analyzing time-domain AC equivalent DC resistance based on big data according to claim 1, characterized in that, The dual-model analysis process for determining equipment status is as follows: Extract the feature parameters of power equipment from the database to form a feature vector; The feature vectors are input into the geometric cascaded forest algorithm model to obtain the class probability vectors of the Lth layer and ML geometric forests. The arithmetic mean of the collected ML probability vectors is taken according to the category dimension to obtain the final average probability vector. The probability values ​​of each category (including normal, potential risk and abnormal) in the final average probability vector are compared, and the category with the highest probability is selected as the current final prediction result.

7. The method for testing and analyzing time-domain AC equivalent DC resistance based on big data according to claim 6, characterized in that, In addition to extracting the time-series features of the equivalent resistance value, additional operating condition correlation features are introduced to form a high-dimensional feature matrix. The high-dimensional feature matrix is ​​input into the gradient boosting tree + isolation forest fusion model to obtain the output real-time health index and real-time isolation degree. The real-time health index is compared and analyzed to obtain the state results of normal, potential risk and abnormal. Normal, potential risk and abnormal are collectively referred to as state level. A comparative analysis of real-time isolation degrees is performed to obtain results of no potential faults and potential faults. The results of no potential faults and potential faults are collectively referred to as fault results.

8. The method for testing and analyzing time-domain AC equivalent DC resistance based on big data according to claim 7, characterized in that, Interactive analysis is performed on the status level and fault results. The final status judgment result is obtained based on the interactive analysis. The final status judgment result includes normal, potential risk and abnormal. The final state determination result is compared and analyzed with the final prediction result. If the final state determination result is consistent with the final prediction result, the final prediction result is output and displayed. If the final state determination result is inconsistent with the final prediction result, a deviation signal is generated and the deviation signal is responded to immediately. An anomaly feedback list is constructed based on the final state determination result, the final prediction result, the real-time health index, the real-time isolation degree, the fault result, the state level, and the probability value of each category.

Images