Method, system and device for monitoring the operating state of a cable based on machine learning
By adaptively selecting features and optimizing the decision tree construction process in the random forest algorithm, the overfitting problem caused by an inappropriate number of features is solved, thereby improving the accuracy and efficiency of cable operation status monitoring.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- FUJIAN FUNENG NEW ENERGY CO LTD
- Filing Date
- 2025-07-01
- Publication Date
- 2026-06-09
AI Technical Summary
In existing random forest algorithms for cable operation status monitoring, too many or too few features can affect the performance of the decision tree, leading to overfitting or insufficient diversity, and reducing monitoring accuracy.
By adaptively selecting features, and utilizing the distribution balance, average difference, data similarity, and repetition of features, the decision tree construction process is optimized. The feature with the highest distribution balance is selected as the baseline feature, and iterative optimization is performed in combination with feature selection factors to construct a random forest algorithm model.
This improved the construction effect of decision trees, reduced overfitting, and enhanced the accuracy and efficiency of cable operation status monitoring.
Smart Images
Figure CN120763574B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of cable operation monitoring technology, specifically to a method, system, and equipment for monitoring cable operation status based on machine learning. Background Technology
[0002] In power systems, cables are the primary equipment for transmitting electrical energy, and their operational status is crucial to the safety and stability of the power system. With the widespread application of intelligent and digital technologies in power system operation, the requirements for analyzing cable operational status monitoring data are becoming increasingly stringent. Monitoring cable operational status through machine learning algorithms can accurately grasp the actual operating conditions of cables, thereby improving the stability and safety of the power system while reducing equipment maintenance costs and the probability of failure.
[0003] Random forest is a common algorithm in machine learning, belonging to the category of ensemble learning, and has broad application potential in cable operation status monitoring. By integrating multiple decision trees, random forest can effectively handle various complex problems in cable operation status monitoring, such as fault detection, condition prediction, anomaly detection, and feature selection. During the construction of a random forest, each decision tree considers the features of the data. However, if the number of features is too large, the similarity between decision trees increases, the overall variance rises, and overfitting may occur. Conversely, if the number of features is too small, the diversity of decision trees is high, but the performance of individual decision trees is weak, affecting the overall effect of the random forest model and thus the accuracy of cable operation status monitoring. Summary of the Invention
[0004] In light of the above, it is necessary to provide a machine learning-based cable operation status monitoring method, system, and equipment. Compared to traditional machine learning-based cable operation status monitoring methods, this approach improves the accuracy and efficiency of cable operation status monitoring by adaptively selecting features when constructing the decision tree.
[0005] In a first aspect, embodiments of this application provide a cable operation status monitoring method based on machine learning, the method comprising the following steps:
[0006] The original training dataset is composed of the values of all operating parameters of the cable at a preset number of historical moments, and the operating parameters are recorded as features.
[0007] When constructing any decision tree in the random forest algorithm, a subset of data is formed by randomly selecting a preset number of historical values from the original training dataset. For the subset of data, the distribution balance of each feature is obtained through the distribution of the values of each feature. Based on the distribution balance, a benchmark feature is selected from all features and added to the initialized empty benchmark feature set.
[0008] By using the average level of each feature, the values of each feature are divided into two categories, and the data differences between the two categories are compared to obtain the average difference degree of each feature. Each feature other than the benchmark feature is recorded as a candidate feature. By comparing the amount of data within each category of each benchmark feature and each candidate feature in the benchmark feature set, the corresponding class of each category of each benchmark feature in each candidate feature is obtained. By comparing the difference in the amount of data between each category of each benchmark feature and the corresponding class and the repetition at the time of data collection, the data similarity between each benchmark feature and each candidate feature is obtained. Combining the difference in the distribution balance and the difference in the average difference degree between each benchmark feature and each candidate feature, the feature repetition between each benchmark feature and each candidate feature is obtained. Combining the distribution balance of each candidate feature, the feature optimization factor of each candidate feature is obtained. Based on the feature optimization factor, a benchmark feature is selected again from the candidate features, the benchmark feature set is updated, and the process of selecting a benchmark feature by calculating the feature optimization factor is repeated. The difference between the benchmark feature set before and after the update is evaluated using the feature optimization factor, and the iteration termination condition is set.
[0009] The decision tree is constructed using the final set of baseline features, and the cable operation status is monitored using the completed random forest algorithm model.
[0010] In one embodiment, the process of obtaining the distribution balance is as follows:
[0011] Calculate the mean of all values for any feature, find the value with the smallest difference from the mean among all values of the feature, divide all values of the feature into two categories based on the statistical results, and calculate the average of all data in each category.
[0012] Calculate the difference in the average value and the difference in data volume between the two categories respectively;
[0013] The distribution balance of any feature is inversely proportional to the difference value and the difference in data volume, respectively.
[0014] In one embodiment, selecting a benchmark feature from all features and adding it to the benchmark feature set includes: taking the feature with the highest distribution balance as the benchmark feature, and the benchmark feature set is initially an empty set.
[0015] In one embodiment, the process of obtaining the average difference is as follows:
[0016] The maximum value among all possible values of any feature in the statistical data subset;
[0017] Calculate the ratio of the average value to the maximum value for each category of any feature, and denot it as the first ratio and the second ratio. The average difference of any feature is the difference between the first ratio and the second ratio.
[0018] In one embodiment, the process of obtaining the corresponding class is as follows:
[0019] For any baseline feature and any candidate feature, the class with the smallest difference in data volume among all classes of the candidate feature and all classes of the baseline feature is taken as the corresponding class of each class of the baseline feature in the candidate feature.
[0020] In one embodiment, the process of obtaining the data similarity is as follows:
[0021] The total number of data points for each class of any benchmark feature and its corresponding class for any candidate feature at the same acquisition time is counted.
[0022] Calculate the difference in data volume between each class of any benchmark feature and its corresponding class of any candidate feature;
[0023] Calculate the cumulative total value and the sum of the differences for all classes corresponding to any one of the baseline features;
[0024] Calculate the sum of the sum and a preset positive number, and use the ratio of the sum to the cumulative sum as the data similarity between any benchmark feature and any candidate feature.
[0025] In one embodiment, the process of obtaining the feature repeatability is as follows:
[0026] The differences in distribution balance and average difference between each baseline feature and each candidate feature are denoted as the first difference and the second difference, respectively. The product of the first difference and the second difference is calculated.
[0027] The feature repetition degree is directly proportional to the data similarity and inversely proportional to the product.
[0028] In one embodiment, obtaining the feature optimization factor of each candidate feature and selecting a benchmark feature from the candidate features includes:
[0029] Calculate the cumulative value of feature repetition between all baseline features and each candidate feature. The feature preference factor is directly proportional to the distribution balance of each candidate feature and inversely proportional to the cumulative value.
[0030] The candidate feature with the largest feature optimization factor is used as the benchmark feature for further selection.
[0031] Secondly, embodiments of this application also provide a cable operation status monitoring system based on machine learning, the system comprising:
[0032] The feature data acquisition module is used to form the original training dataset by taking the values of all operating parameters of the cable at a preset number of historical moments, and to record the operating parameters as features.
[0033] The feature selection module is used to randomly extract a preset number of historical values from the original training dataset to form a data subset when constructing any decision tree in the random forest algorithm; for the data subset, the distribution balance of each feature is obtained through the distribution of the values of each feature, and a benchmark feature is selected from all features based on the distribution balance and added to the initialized empty benchmark feature set.
[0034] By using the average level of each feature, the values of each feature are divided into two categories, and the data differences between the two categories are compared to obtain the average difference degree of each feature. Each feature other than the benchmark feature is recorded as a candidate feature. By comparing the amount of data within each category of each benchmark feature and each candidate feature in the benchmark feature set, the corresponding class of each category of each benchmark feature in each candidate feature is obtained. By comparing the difference in the amount of data between each category of each benchmark feature and the corresponding class and the repetition at the time of data collection, the data similarity between each benchmark feature and each candidate feature is obtained. Combining the difference in the distribution balance and the difference in the average difference degree between each benchmark feature and each candidate feature, the feature repetition between each benchmark feature and each candidate feature is obtained. Combining the distribution balance of each candidate feature, the feature optimization factor of each candidate feature is obtained. Based on the feature optimization factor, a benchmark feature is selected again from the candidate features, the benchmark feature set is updated, and the process of selecting a benchmark feature by calculating the feature optimization factor is repeated. The difference between the benchmark feature set before and after the update is evaluated using the feature optimization factor, and the iteration termination condition is set.
[0035] The operation status monitoring module is used to construct any decision tree through the final benchmark feature set, and monitor the cable operation status through the constructed random forest algorithm model.
[0036] Thirdly, embodiments of this application also provide a cable operation status monitoring device based on machine learning, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of any of the above-described machine learning-based cable operation status monitoring methods.
[0037] This application has at least the following beneficial effects:
[0038] This application randomly selects a subset of data and calculates the distribution balance based on the distribution of the values of each feature in the subset. The feature with the highest distribution balance is selected as the benchmark feature. This can effectively screen out features that make significant contributions to the classification effect, improve the construction effect of the decision tree, and avoid misclassifying data belonging to the same category as two categories.
[0039] Furthermore, considering the characteristics of faults that occur during cable operation, the redundancy between features is evaluated through data similarity, i.e., the possibility that features bring duplicate information in the classification of the decision tree; the average difference is calculated to reflect the difference between features in different categories; by combining the average difference, data similarity, and distribution balance, feature redundancy is obtained, which can measure the similarity of the data classification effect of the baseline feature and candidate features when constructing the decision tree, so as to avoid the duplicate benefits between features when selecting the baseline feature and improve the efficiency of the random forest algorithm model.
[0040] Furthermore, by optimizing the distribution balance through feature repetition, a feature optimization factor is obtained. Based on the feature optimization factor, the selection of features in the decision tree construction process is adaptively completed, which can improve the performance of a single decision tree, reduce the similarity between the constructed decision trees, and reduce the possibility of overfitting. This improves the accuracy and efficiency of the random forest algorithm for monitoring the cable operation status. Attached Figure Description
[0041] To more clearly illustrate the technical solutions and advantages in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0042] Figure 1 A flowchart illustrating the steps of a machine learning-based cable operation status monitoring method provided in one embodiment of this application;
[0043] Figure 2 A schematic diagram of the process for obtaining the feature optimization factors;
[0044] Figure 3 This is a schematic diagram illustrating the process of obtaining the baseline feature set. Detailed Implementation
[0045] In the description of the embodiments in this application, the words "exemplary," "or," and "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of the words "exemplary," "or," and "for example" is intended to present the relevant concepts in a specific manner.
[0046] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. It should be understood that, unless otherwise stated, " / " in this application means "or".
[0047] It should also be noted that the terms "first" and "second" in this application are used to distinguish similar objects, rather than to describe a specific order or sequence.
[0048] The following description, in conjunction with the accompanying drawings, details the specific solutions for the machine learning-based cable operation status monitoring method, system, and equipment provided in this application.
[0049] Please see Figure 1 The diagram illustrates a flowchart of a machine learning-based cable operation status monitoring method according to an embodiment of this application. The method includes the following steps:
[0050] Step 1: Compile the values of all operating parameters of the cable at a preset number of historical moments into an original training dataset, and denote the operating parameters as features.
[0051] When applying the random forest algorithm in cable operation status monitoring, multi-dimensional data needs to be collected to comprehensively reflect the cable's operating status. This application uses intelligent sensors to collect various operating parameter data of the cable. Intelligent sensors play a crucial role in cable operation status monitoring, providing a data foundation for analyzing cable operation status by collecting various operating parameter data in real time. Intelligent sensors can not only monitor traditional physical quantities such as temperature, current, voltage, and humidity, but also collect more complex indicators such as grounding current, insulation resistance, load, and aging indicators. All operating parameters of the cable at various historical moments are grouped into data points, and a preset number of data points are obtained to form the original training dataset. In this embodiment, the operating parameters include: temperature, current, voltage, humidity, grounding current, insulation resistance, load, and aging indicators. The aging indicator is partial discharge intensity, which is obtained through temperature sensors, current sensors, voltage sensors, humidity sensors, grounding current sensors, online insulation resistance monitors, smart meters, and partial discharge detectors. Each operating parameter is denoted as a feature.
[0052] In this embodiment, the preset quantity is 20000. The preset quantity is preset by humans and the implementer can set it according to the actual situation. This application does not impose any special restrictions.
[0053] Step 2: Adaptively complete the feature selection process in the decision tree construction process of the random forest algorithm.
[0054] The Random Forest algorithm extracts a subset of data with replacement from the original training dataset when constructing decision trees. Traditional Random Forest algorithms typically select a subset of features for splitting during decision tree construction. However, an excessively large number of features increases the similarity between the constructed decision trees, raising the overall variance and potentially leading to overfitting. Conversely, a small number of features results in high diversity of decision trees, but weak performance for individual trees. Therefore, this application analyzes the extracted data subset, adaptively selects features, and then performs node splitting based on these features to complete the construction of the decision tree. Here, features are the attributes used by the model for splitting. A subset of data is formed by randomly selecting M data points from the original training dataset.
[0055] In this embodiment, the value of M is 4000. The value of M is preset by the user and can be limited by the implementer according to the actual situation. This application does not impose any special restrictions.
[0056] First, the parameters of the random forest algorithm are initialized, with the number of decision trees set to 200, the maximum depth of a single tree set to 20, the minimum number of samples required for node splitting set to 20, and the minimum number of samples required for leaf nodes set to 5.
[0057] Step 2.1: For the data subset, obtain the distribution balance of each feature by analyzing the distribution of the values of each feature.
[0058] Taking the construction process of any decision tree in the random forest algorithm as an example, the corresponding extracted data subset is W. Since the data subset is randomly extracted from the original training dataset, the data subsets of different decision trees are usually different, and the optimal features for node splitting are also usually different. This application first obtains the distribution balance of each feature by analyzing the distribution of the values of each feature in the data subset, specifically:
[0059] Taking feature A as an example, calculate the mean of all values of feature A in the data subset W, count the value of feature A in the data subset W that has the smallest difference from the mean, and record it as the most recent value. Divide all values of feature A in the data subset W into two classes based on the most recent value, calculate the average value of all data in each class, and obtain the distribution balance of feature A by the difference between the average values and the difference in data volume between the two classes.
[0060] In this embodiment, when obtaining the most recent value, the difference between the value of feature A and the mean is the absolute value of the difference.
[0061] In this embodiment, the expression for the distribution balance of feature A is:
[0062] In the formula, The distributional uniformity of feature A is represented; exp() represents an exponential function with the natural constant as the base. , These represent the amount of data in the first and second classes of feature A, respectively. , Let A and B represent the average values of all data in the first and second classes of feature A, respectively. This indicates an absolute value operation. The average value in the first category is greater than the average value in the second category.
[0063] In another embodiment, the expression for the distribution balance of feature A is:
[0064] In the formula, The distribution uniformity of feature A is represented by norm(); norm() represents the normalization function. , These represent the amount of data in the first and second classes of feature A, respectively. , Let A and B represent the average values of all data in the first and second classes of feature A, respectively. This indicates the absolute value operation. The normalization function is the Min-Max normalization function.
[0065] It should be noted that: the greater the difference in the amount of data in the two categories obtained by segmentation, and the greater the difference in the values of the data in the two categories, the more dispersed the values of feature A in the data subset, and the smaller the calculated distribution balance. The greater the distribution balance, the more regular the distribution of the values of feature A, and the more likely the data points in the feature subset belong to the same category under feature A. In this case, if feature A is selected to segment the data points in the data subset, it is more likely that data of the same category will be mistakenly divided into two categories. The smaller the distribution balance, the more irregular the values of feature A in the data subset, and the more likely the data points belong to multiple categories under feature A. In this case, selecting feature A to split the nodes will have a better effect.
[0066] Step 2.2: Based on the distribution balance, select a benchmark feature from all features and add it to the initialized empty benchmark feature set; divide the values of each feature into two categories by the average level of each feature and compare the data differences between the two categories to obtain the average difference of each feature; denote each feature other than the benchmark feature as a candidate feature; by comparing the amount of data within each category of each benchmark feature and each candidate feature in the benchmark feature set, obtain the corresponding class of each category of each benchmark feature in each candidate feature; by comparing the difference in the amount of data between each category of each benchmark feature and the corresponding class and the repetition at the time of data collection, obtain the data similarity between each benchmark feature and each candidate feature; and by combining the difference in distribution balance and the difference in average difference between each benchmark feature and each candidate feature, obtain the feature repetition between each benchmark feature and each candidate feature.
[0067] Following the calculation method for the distribution balance of feature A, the distribution balance of each feature can be calculated. However, directly using the distribution balance for feature selection and subsequent decision tree construction presents problems. When a cable malfunctions, its operating parameters often fluctuate to varying degrees. For example, partial discharge during cable operation can lead to instantaneous current pulses, temperature increases, and decreased insulation resistance. In cable fault detection, different types of smart sensors, such as current sensors, voltage sensors, and temperature sensors, may have overlapping responses to the same fault phenomenon. Directly using the distribution balance of each feature for feature selection and decision tree construction may result in redundant information. This is because the instantaneous fluctuations caused by faults, such as partial discharge, short circuits, or overloads, can cause similar fluctuation patterns in the data collected by multiple smart sensors, leading to "duplication effects" in feature construction. In this situation, when using "current" or "temperature" features to divide the decision tree, the distribution of data points in the leaf nodes may be quite similar. That is, the benefits brought by the two types of features to the decision tree construction are duplicated. In this case, only one type of feature needs to be used when constructing the decision tree.
[0068] In this application, the feature with the highest distribution balance is taken as the benchmark feature, and the benchmark feature is added to the benchmark feature set, which is initially an empty set. All features other than the benchmark feature are denoted as candidate features. Based on the above analysis, the feature redundancy between each benchmark feature and each candidate feature is obtained by comparing the distribution of values between the benchmark features and candidate features in the benchmark feature set. The specific process is as follows:
[0069] (1) By using the average level of each feature, the values of each feature are divided into two categories and the data differences between the two categories are compared to obtain the average difference of each feature.
[0070] The expression for the average dissimilarity of each feature is:
[0071] ; This represents the average degree of difference for feature B; , These represent the average values of all data in the first and second classes of feature B, respectively. This represents the maximum value among all possible values of feature B in the data subset. This indicates the absolute value operation. In this case, the average of all data in the first category is greater than the average of all data in the second category.
[0072] It should be noted that the average dissimilarity reflects the difference between features in different categories; the greater the average dissimilarity, the greater the degree of difference between features in different categories.
[0073] (2) By comparing the amount of data in each class between each benchmark feature and each candidate feature in the benchmark feature set, the corresponding class of each class of each benchmark feature in each candidate feature is obtained. By comparing the difference in the amount of data between each class of each benchmark feature and the corresponding class and the repetition of the data collection time, the data similarity between each benchmark feature and each candidate feature is obtained.
[0074] Taking baseline feature D and candidate feature E as an example, the class with the smallest difference in data volume between all classes of candidate feature E and all classes of baseline feature D is taken as the corresponding class of each class of baseline feature D in candidate feature E.
[0075] In this embodiment, the difference in data volume refers to the absolute value of the difference between data volumes.
[0076] By comparing the data volume differences between each class of the baseline feature D and the corresponding class, as well as the repetition of data collection times, the data similarity between the baseline feature D and the candidate feature E is obtained, expressed as:
[0077] In the formula, This represents the data similarity between baseline feature D and candidate feature E; it also represents the total number of data points for each class of baseline feature D and its corresponding class in candidate feature E at the same data collection time. This represents the cumulative value of the total number corresponding to all classes of the baseline feature D; This represents the amount of data in the i-th class of the baseline feature D; This represents the amount of data in the corresponding class of the baseline feature D in the candidate feature E; This indicates the operation of taking the absolute value; α represents a preset positive number used to avoid the denominator being 0. The value of α is preset by the user. In order to avoid affecting the calculation results of data similarity, α should be a very small positive number. In this embodiment, the value of α is 0.01.
[0078] It should be noted that data similarity reflects the redundancy between features; the greater the data similarity, the more similar the baseline feature D and the candidate feature E are in terms of data distribution, and the more likely the baseline feature D and the candidate feature E are to bring duplicate information in the classification of the decision tree.
[0079] (3) By combining the data similarity between each benchmark feature and each candidate feature, and the differences in distribution balance and average difference between each benchmark feature and each candidate feature, the feature repetition between each benchmark feature and each candidate feature is obtained.
[0080] ; This indicates the feature overlap between the baseline feature D and the candidate feature E; This represents the data similarity between baseline feature D and candidate feature E; , These represent the distribution balance of the baseline feature D and the candidate feature E, respectively. , These represent the average differences between the baseline feature D and the candidate feature E, respectively. ε represents the absolute value operation; ε represents a preset positive number used to avoid the denominator being 0. The value of ε is preset by the user. In order to avoid affecting the calculation result of feature repeatability, ε should be a very small positive number. In this embodiment, the value of ε is 0.01.
[0081] It should be noted that: the greater the feature repetition, the more similar the data classification results of the baseline feature D and the candidate feature E when constructing the decision tree, and only one of them should be retained; the smaller the feature repetition, the greater the difference in the data classification results of the baseline feature D and the candidate feature E when constructing the decision tree, and both should be retained.
[0082] Step 2.3: Obtain the feature optimization factor for each candidate feature by considering the feature repetition between each benchmark feature and each candidate feature, as well as the distribution balance of each candidate feature. Based on the feature optimization factor, select another benchmark feature from the candidate features, update the benchmark feature set, repeat the process of selecting benchmark features by calculating the feature optimization factor, evaluate the difference between the benchmark feature set before and after the update using the feature optimization factor, and set the iteration termination condition.
[0083] The feature optimization factor for each candidate feature is obtained by considering the feature redundancy between each baseline feature and each candidate feature, as well as the distribution balance of each candidate feature. The expression is as follows:
[0084] In the formula, The feature preference factor represents the candidate feature E; The distribution balance of candidate feature E is represented; U represents the baseline feature set. This indicates the feature overlap between the baseline feature D and the candidate feature E.
[0085] It should be noted that the feature optimization factor is obtained by optimizing the distribution balance through feature repetition. The larger the feature optimization factor of candidate feature E, the better the classification effect of the data points in the data subset under candidate feature E, and the better the construction effect of the decision tree. Conversely, the smaller the feature optimization factor, the worse the classification effect of the data points in the data subset under candidate feature E, and the worse the construction effect of the decision tree. The process of obtaining the feature optimization factor is illustrated in the diagram below. Figure 2 As shown.
[0086] The candidate feature with the largest feature optimization factor is selected as the benchmark feature and added to the benchmark feature set. The benchmark feature set is updated, and the feature optimization factor of the benchmark feature is recorded. Step 2.2 is repeated to recalculate the feature optimization factor of each candidate feature, and the candidate feature with the largest feature optimization factor is added to the benchmark feature set as the new benchmark feature. The maximum value among the recalculated feature optimization factors is counted, and the minimum value among the recorded feature optimization factors is counted. When the difference between the normalized value of the minimum value and the normalized value of the maximum value is less than a preset threshold, step 2.2 is stopped. The schematic diagram of the benchmark feature set acquisition process is shown below. Figure 3 As shown.
[0087] In this embodiment, the Min-Max normalization function is used to obtain the normalized value of the maximum value and the normalized value of the minimum value.
[0088] In this embodiment, the preset threshold value is 0.2. The preset threshold value is preset by a person and can be set by the implementer. This application does not impose any special restrictions.
[0089] Step 3: Construct any decision tree using the final set of baseline features, and monitor the cable operation status using the completed random forest algorithm model.
[0090] The decision tree nodes are split based on the final obtained baseline feature set until the decision tree no longer meets the splitting conditions or reaches its maximum depth, thus completing the construction of the random forest algorithm model. The construction of the random forest algorithm model is a well-known technique and will not be elaborated upon in this application. The completed random forest algorithm model is used to monitor the cable's operating status. All cable operating parameter data at the current moment are used as input to the random forest algorithm, and the output is the cable's operating status, such as "overheating," "partial discharge," or "aging."
[0091] Based on the same inventive concept as the above method, embodiments of this application also provide a cable operation status monitoring system based on machine learning, including:
[0092] The feature data acquisition module is used to form the original training dataset by taking the values of all operating parameters of the cable at a preset number of historical moments, and to record the operating parameters as features.
[0093] The feature selection module is used to randomly extract a preset number of historical values from the original training dataset to form a data subset when constructing any decision tree in the random forest algorithm; for the data subset, the distribution balance of each feature is obtained through the distribution of the values of each feature, and a benchmark feature is selected from all features based on the distribution balance and added to the initialized empty benchmark feature set.
[0094] By using the average level of each feature, the values of each feature are divided into two categories, and the data differences between the two categories are compared to obtain the average difference degree of each feature. Each feature other than the benchmark feature is recorded as a candidate feature. By comparing the amount of data within each category of each benchmark feature and each candidate feature in the benchmark feature set, the corresponding class of each category of each benchmark feature in each candidate feature is obtained. By comparing the difference in the amount of data between each category of each benchmark feature and the corresponding class and the repetition at the time of data collection, the data similarity between each benchmark feature and each candidate feature is obtained. Combining the difference in the distribution balance and the difference in the average difference degree between each benchmark feature and each candidate feature, the feature repetition between each benchmark feature and each candidate feature is obtained. Combining the distribution balance of each candidate feature, the feature optimization factor of each candidate feature is obtained. Based on the feature optimization factor, a benchmark feature is selected again from the candidate features, the benchmark feature set is updated, and the process of selecting a benchmark feature by calculating the feature optimization factor is repeated. The difference between the benchmark feature set before and after the update is evaluated using the feature optimization factor, and the iteration termination condition is set.
[0095] The operation status monitoring module is used to construct any decision tree through the final benchmark feature set, and monitor the cable operation status through the constructed random forest algorithm model.
[0096] Based on the same inventive concept as the above methods, this application also provides a cable operation status monitoring device based on machine learning, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of any one of the above-described methods for monitoring cable operation status based on machine learning.
[0097] In summary, this application, by randomly sampling a subset of data and calculating the distribution balance based on the distribution of the values of each feature in the subset, selects the feature with the highest distribution balance as the benchmark feature. This effectively filters out features that significantly contribute to the classification effect, improves the construction effect of the decision tree, and avoids misclassifying data belonging to the same category as two categories.
[0098] Furthermore, considering the characteristics of faults that occur during cable operation, the redundancy between features is evaluated through data similarity, i.e., the possibility that features bring duplicate information in the classification of the decision tree; the average difference is calculated to reflect the difference between features in different categories; by combining the average difference, data similarity, and distribution balance, feature redundancy is obtained, which can measure the similarity of the data classification effect of the baseline feature and candidate features when constructing the decision tree, so as to avoid the duplicate benefits between features when selecting the baseline feature and improve the efficiency of the random forest algorithm model.
[0099] Furthermore, by optimizing the distribution balance through feature repetition, a feature optimization factor is obtained. Based on the feature optimization factor, the selection of features in the decision tree construction process is adaptively completed, which can improve the performance of a single decision tree, reduce the similarity between the constructed decision trees, and reduce the possibility of overfitting. This improves the accuracy and efficiency of the random forest algorithm for monitoring the cable operation status.
[0100] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. In the descriptions corresponding to the flowcharts and block diagrams in the accompanying drawings, the operations or steps corresponding to different blocks may also occur in a different order than disclosed in the description, and sometimes there is no specific order between different operations or steps. For example, two consecutive operations or steps may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. Each block in a block diagram and / or flowchart, and combinations of blocks in a block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.
[0101] It will be apparent to those skilled in the art that this application is not limited to the details of the exemplary embodiments described above, and that this application can be implemented in other specific forms without departing from its essential characteristics. Therefore, the embodiments described above should be considered exemplary and non-limiting in all respects.
Claims
1. A method for monitoring cable operating status based on machine learning, characterized in that, The method includes the following steps: The original training dataset is composed of the values of all operating parameters of the cable at a preset number of historical moments, and the operating parameters are recorded as features. When constructing any decision tree in the random forest algorithm, a subset of data is formed by randomly selecting a preset number of historical values from the original training dataset. For the subset of data, the distribution balance of each feature is obtained through the distribution of the values of each feature. Based on the distribution balance, a benchmark feature is selected from all features and added to the initialized empty benchmark feature set. By using the average level of each feature, the values of each feature are divided into two categories, and the data differences between the two categories are compared to obtain the average difference degree of each feature. Each feature other than the benchmark feature is recorded as a candidate feature. By comparing the amount of data within each category of each benchmark feature and each candidate feature in the benchmark feature set, the corresponding class of each category of each benchmark feature in each candidate feature is obtained. By comparing the difference in the amount of data between each category of each benchmark feature and the corresponding class and the repetition at the time of data collection, the data similarity between each benchmark feature and each candidate feature is obtained. Combining the difference in the distribution balance and the difference in the average difference degree between each benchmark feature and each candidate feature, the feature repetition between each benchmark feature and each candidate feature is obtained. Combining the distribution balance of each candidate feature, the feature optimization factor of each candidate feature is obtained. Based on the feature optimization factor, a benchmark feature is selected again from the candidate features, the benchmark feature set is updated, and the process of selecting a benchmark feature by calculating the feature optimization factor is repeated. The difference between the benchmark feature set before and after the update is evaluated using the feature optimization factor, and the iteration termination condition is set. The decision tree is constructed using the final set of baseline features, and the cable operation status is monitored using the completed random forest algorithm model.
2. The cable operation status monitoring method based on machine learning as described in claim 1, characterized in that, The process of obtaining the distribution balance is as follows: Calculate the mean of all values for any feature, find the value with the smallest difference from the mean among all values of the feature, divide all values of the feature into two categories based on the statistical results, and calculate the average of all data in each category. Calculate the difference in the average value and the difference in data volume between the two categories respectively; The distribution balance of any feature is inversely proportional to the difference value and the difference in data volume, respectively.
3. The cable operation status monitoring method based on machine learning as described in claim 1, characterized in that, The step of selecting a benchmark feature from all features and adding it to the benchmark feature set includes: taking the feature with the highest distribution balance as the benchmark feature, and the benchmark feature set is initially an empty set.
4. The cable operation status monitoring method based on machine learning as described in claim 2, characterized in that, The process for obtaining the average difference is as follows: The maximum value among all possible values of any feature in the statistical data subset; Calculate the ratio of the average value to the maximum value for each category of any feature, and denot it as the first ratio and the second ratio. The average difference of any feature is the difference between the first ratio and the second ratio.
5. The cable operation status monitoring method based on machine learning as described in claim 1, characterized in that, The process of obtaining the corresponding class is as follows: For any baseline feature and any candidate feature, the class with the smallest difference in data volume among all classes of the candidate feature and all classes of the baseline feature is taken as the corresponding class of each class of the baseline feature in the candidate feature.
6. The cable operation status monitoring method based on machine learning as described in claim 5, characterized in that, The process of obtaining the data similarity is as follows: The total number of data points for each class of any benchmark feature and its corresponding class for any candidate feature at the same acquisition time is counted. Calculate the difference in data volume between each class of any benchmark feature and its corresponding class of any candidate feature; Calculate the cumulative total value and the sum of the differences for all classes corresponding to any one of the baseline features; Calculate the sum of the sum and a preset positive number, and use the ratio of the sum to the cumulative sum as the data similarity between any benchmark feature and any candidate feature.
7. The cable operation status monitoring method based on machine learning as described in claim 1, characterized in that, The process of obtaining the feature repeatability is as follows: The differences in distribution balance and average difference between each baseline feature and each candidate feature are denoted as the first difference and the second difference, respectively. The product of the first difference and the second difference is calculated. The feature repetition degree is directly proportional to the data similarity and inversely proportional to the product.
8. The cable operation status monitoring method based on machine learning as described in claim 7, characterized in that, The process of obtaining feature optimization factors for each candidate feature and selecting a benchmark feature from the candidate features includes: Calculate the cumulative value of feature repetition between all baseline features and each candidate feature. The feature preference factor is directly proportional to the distribution balance of each candidate feature and inversely proportional to the cumulative value. The candidate feature with the largest feature optimization factor is used as the benchmark feature for further selection.
9. A cable operation status monitoring system based on machine learning, employing the cable operation status monitoring method based on machine learning as described in claim 1, characterized in that, The system includes: The feature data acquisition module is used to form the original training dataset by taking the values of all operating parameters of the cable at a preset number of historical moments, and to record the operating parameters as features. The feature selection module is used to randomly extract a preset number of historical values from the original training dataset to form a data subset when constructing any decision tree in the random forest algorithm; for the data subset, the distribution balance of each feature is obtained through the distribution of the values of each feature, and a benchmark feature is selected from all features based on the distribution balance and added to the initialized empty benchmark feature set. By using the average level of each feature, the values of each feature are divided into two categories, and the data differences between the two categories are compared to obtain the average difference degree of each feature. Each feature other than the benchmark feature is recorded as a candidate feature. By comparing the amount of data within each category of each benchmark feature and each candidate feature in the benchmark feature set, the corresponding class of each category of each benchmark feature in each candidate feature is obtained. By comparing the difference in the amount of data between each category of each benchmark feature and the corresponding class and the repetition at the time of data collection, the data similarity between each benchmark feature and each candidate feature is obtained. Combining the difference in the distribution balance and the difference in the average difference degree between each benchmark feature and each candidate feature, the feature repetition between each benchmark feature and each candidate feature is obtained. Combining the distribution balance of each candidate feature, the feature optimization factor of each candidate feature is obtained. Based on the feature optimization factor, a benchmark feature is selected again from the candidate features, the benchmark feature set is updated, and the process of selecting a benchmark feature by calculating the feature optimization factor is repeated. The difference between the benchmark feature set before and after the update is evaluated using the feature optimization factor, and the iteration termination condition is set. The operation status monitoring module is used to construct any decision tree through the final benchmark feature set, and monitor the cable operation status through the constructed random forest algorithm model.
10. A machine learning-based cable operation status monitoring device, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the machine learning-based cable operation status monitoring method as described in any one of claims 1-8.