Engine tightening fault type classification method based on improved random forest

Abstract

The invention belongs to the technical field of industrial automation and machine learning application, and particularly relates to an engine tightening fault type classification method based on an improved random forest, which is characterized by comprising the following specific operation steps: 1) attaching a label to tightening curve data; 2) preprocessing data, removing noise points, and completing extraction of main curve segments; 3) feature extraction; and 4) training the improved random forest model to complete classification. Compared with the prior art, the method has the beneficial effects that the defect data can be classified based on the defect data classification algorithm of the improved random forest, the defect data of the new type can be effectively detected, the data of the existing defect type can be classified, the fault data of the new type can be effectively identified, and the fault detection accuracy is improved. The potential risk is reduced, and remarkable technical and economic application values are realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial automation and machine learning application, and in particular relates to an engine emergency fault type classification method based on improved random forest. Background Art

[0002] In the process of automobile engine production, the assembly quality of the engine directly affects the factory quality of the engine. The assembly quality of the engine directly affects the factory quality of the engine. In most of the engine assembly workshops of automobile companies in my country, it is not possible to achieve fully automated assembly by machines instead of people. Most of the workstations in the workshop still rely on workers to tighten the bolts with electric guns. Therefore, the quality of bolt tightening directly affects the quality of engine assembly, and the quality of bolt tightening must be effectively monitored. When tightening quality problems occur, timely analysis of the type of fault is of great guiding significance for targeted improvement of solutions. However, in the process of engine tightening, some new types of faults may occur, and most of the existing defect classification methods cannot identify new types of faults, resulting in classification errors.

[0003] Li Wei, Guo Jun, Wang Honglin and others proposed a fault classification method for transformers in a distribution network, which method includes: collecting target data of transformers in the distribution network, and establishing a first data matrix based on the target data; using a t-distributed random neighborhood embedding algorithm to extract key features of the first data matrix to obtain a second data matrix; using a classifier to train the second data matrix to obtain a fault classification model; collecting fault data of transformers in the distribution network; classifying the fault data through the fault classification model to obtain the fault category corresponding to the fault data.

[0004] Yin Hongshen proposed a bearing fault classification method and device, including: inputting the bearing fault data of the target bearing into a fault classification model, wherein the fault classification model is pre-trained based on source domain sample data, target domain sample data and a total classification loss function, and the total classification loss function is determined based on the source domain sample data and the target domain sample data; determining fault signals of different frequencies corresponding to the bearing fault data through a feature extraction module of the fault classification model; assigning preset weights to key features in fault signals of different frequencies to obtain a first feature through a feature processing module of the fault classification model; determining the fault classification corresponding to the target bearing according to the first feature through a feature classification module of the fault classification model.

[0005] Lin Xiangning, Wang Guangyuan, Ma Xiao and others proposed a distribution network defect classification method based on deep learning, which obtains multiple original fault waveform data groups; processes each original fault waveform data group separately to obtain target sample data; makes all target sample data into a data set, and divides the data set into a training set and a test set, constructs a deep learning network model, and uses the training set to train the deep learning network model to obtain the original fault classification model; uses the test set to tune the parameters of the original fault classification model to obtain an optimized fault classification model; obtains a real-time fault waveform data group, and processes the real-time fault waveform data group to obtain the fault data to be tested, uses the optimized fault classification model to perform real-time identification on the fault data to be tested, and obtains the real-time fault classification result.

[0006] Chen Haibo proposed a training method and device for a fault classification model based on a support vector machine. The method adopts a preset classification principle to classify the acquired fault sample data to obtain test sample data and training sample data. Based on a preset multiverse optimization algorithm and a preset second number of training parameters, the training sample data is iteratively trained for a first number of times to obtain a first number of SVM fault classification models, wherein the training parameters include a set of penalty coefficients and kernel widths. According to the test sample data, the first number of SVM fault classification models are tested to obtain a first number of fitnesses, wherein the fitnesses are determined based on the test values ​​of the test sample data and the corresponding true values. The SVM fault classification model corresponding to the minimum fitness among the first number of fitnesses is determined as the optimal SVM fault classification model.

[0007] Chen Jian, Cai Kunqi and others proposed a bearing fault classification method, which uses a bearing test device to obtain bearing vibration signal data under different fault states; decomposes the bearing vibration signal data into several rotational components; calculates the fuzzy entropy values ​​of the rotational components; uses the calculated fuzzy entropy values ​​to construct feature vectors to form a feature matrix; uses the feature matrix to train a bearing fault classifier model; and uses the trained bearing fault classifier model to classify the bearing fault vibration signals to be classified.

[0008] Jiang Mingshun, Ma Mengyuan, Cao Hongyi and others proposed a method for classifying defects in the manufacturing of carbon fiber composite materials, including: using the ultrasonic detection signals of the artificial defect carbon fiber composite material samples obtained as the training set to train the defect classification model; performing an ultrasonic pre-scan on the carbon fiber composite material to be tested to determine the defect location; performing a secondary ultrasonic scan on the defect location to obtain the ultrasonic detection signal to be tested, using the trained defect classification model to classify the defect type.

[0009] In summary, the defect classification methods in the prior art first collect data and then extract features from the collected data to train the classification model. However, they all ignore the recognition of new types of data by the classification model, leading to potential risks. Because the background meaning of the curve is not taken into account, the accuracy of the original curve data is reduced. Summary of the invention

[0010] The purpose of the present invention is to provide an engine emergency fault type classification method based on improved random forest, overcome the shortcomings of the prior art, and improve the detection capability of new types of defects by introducing a credibility weighted mechanism. The defect data classification algorithm based on improved random forest can not only classify new types of data, but also effectively detect new types of defect data, and can not only classify existing defect type data, but also effectively identify new types of fault data, thereby improving the defect recognition rate and misjudgment rate.

[0011] To achieve the above object, the present invention is implemented through the following technical solutions:

[0012] An engine emergency fault type classification method based on improved random forest, the specific operation steps are as follows:

[0013] 1) Data labeling: When performing a tightening operation, the relationship between torque and angle shows how the torque changes as the bolt rotation angle increases. According to engineering standards or process requirements, rules for the tightening process are formulated, and unqualified curves are classified and labeled;

[0014] 2) Data preprocessing: first, smooth the curve, remove noise points, and complete the curve data cleaning; then segment and remove the data segment without torque to complete the extraction of the main curve segment;

[0015] 3) Feature extraction: extract features from the preprocessed data and perform normalization;

[0016] 4) Train the improved random forest model. First, expand the decision tree. Each decision tree classifies the test sample and obtains the corresponding credibility. The credibility is selected as the weight of each decision tree's vote. Then, the random forest performs weighted voting on the classification results of multiple decision trees. Then, the information entropy value of the voting is calculated. If it is greater than the set threshold obtained through training, the test sample is considered to be a new type of defect. Otherwise, the defect category with the highest credibility in the voting result is selected as the final result to complete the classification.

[0017] Furthermore, the curve data cleaning in step 2) is aimed at the noise interference in the original tightening curve, and a smoothing algorithm is used to remove the noise points, as shown in the following formula:

[0018]

[0019] The Gaussian filtering smoothing algorithm is used to eliminate invalid data segments and remove invalid intervals without torque.

[0020] Furthermore, the extraction of the main curve segments in step 2) is based on the background knowledge of the bolt tightening process, and only the key curve segments related to the defects are intercepted.

[0021] Furthermore, the data extraction feature in step 3) refers to extracting features related to the tightening quality from the key curve segment, including any one or any combination of two or more of the maximum torque, slope change rate, and curve shape factor features.

[0022] Furthermore, the normalization process in step 3) is to normalize the extracted feature values ​​to ensure that features of different dimensions have the same impact on the classification results.

[0023] Furthermore, the process steps of the improved random forest model in step 4) are as follows:

[0024] 1) Decision tree generation and credibility calculation: for each decision tree, the credibility of the classification result is calculated based on the credibility α of the classification result. The credibility of the decision tree is used as the voting weight of its classification result.

[0025]

[0026] 2) Weighted voting mechanism: In the forest model, the classification results of each decision tree are weighted according to their credibility, thus improving the classification accuracy of the overall model;

[0027] 3) Information entropy calculation: Calculate the information entropy H of the weighted voting results:

[0028]

[0029] Where: P_i is the voting weight of category k in the classification result;

[0030] 4) New defect discrimination mechanism, setting the information entropy threshold θ, automatically tuning through the training data set, if the information entropy H>θ, then the test sample is considered to belong to the new defect category.

[0031] Furthermore, in the calculation of the credibility using α, the decision tree attribute splitting node value is set to value, the feature attribute value corresponding to the test case is x, and p represents the reliability of the decision tree going to the right subtree. The calculation formula is:

[0032]

[0033] The value range of credibility is (0, 1). The sigmoid function is introduced to map the variable to between 0 and 1. The formula of the sigmoid function is:

[0034]

[0035] According to the attribute value, it is determined whether to go to the left subtree or the right subtree. α represents the credibility of going to the target subtree, and 1-α represents the credibility of going to the other subtree.

[0036] Furthermore, the credibility reflects the degree of confidence of the decision tree in its decision results, so in subsequent voting, the voting weight of those with higher credibility is greater than that of those with lower credibility.

[0037] Furthermore, the setting of the information entropy threshold is obtained through training, and the threshold obtained through automatic training of the training set is selected to distinguish new types of defects from predefined defect types.

[0038] Furthermore, the decision tree has multiple layers of nodes, and has different credibility for going to the left subtree and the right subtree, and the credibility is multiplied as a weight.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] 1) The defect data classification algorithm based on the improved random forest can classify defect data and effectively detect new types of defect data. It can not only classify data of existing defect types, but also effectively identify new types of fault data, thus reducing potential risks;

[0041] 2) The model takes into account the background of the tightening process, intercepts the main curve segments for analysis, eliminates noise interference, and solves the influence of different curve lengths on the classifier by extracting features from the main curve segments. Compared with the traditional defect classification method, the algorithm first smoothes and denoises the data, removes invalid data segments, extracts key features, and then uses the improved random forest model for classification, which improves the accuracy of judging the original curve data;

[0042] 3) Confidence is introduced into the decision tree. When each decision tree makes a classification result for the test sample, it can also obtain the corresponding credibility. The credibility reflects the degree of confidence of the decision tree in its decision result. Through the credibility weighting of the decision tree and the information entropy threshold judgment, the manual setting of the threshold size when detecting the curve is avoided. This method can more accurately classify and detect unknown defects. Through the information entropy threshold judgment, the algorithm can identify undefined defect types, solving the problem that traditional methods cannot detect new defects. In the classification process, the credibility of each decision tree is used for weighted voting, thereby improving the classification accuracy and robustness of the model;

[0043] 4) Experimental results show that in the automotive manufacturing industry and other industrial fields, the algorithm performs superiorly in terms of defect recognition rate and false positive rate, significantly improves the monitoring effect of engine assembly quality, and provides intelligent support for production quality control. It has important application value in the automotive manufacturing field and can be widely used in quality control and fault warning in the assembly process, which helps to improve the reliability and factory quality of engine assembly, and has significant technical and economic application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a classification flow chart of an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The technical solution of the present invention will be clearly and completely described below in conjunction with specific embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments.

[0046] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the specific embodiments required to be used in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the specific embodiments described below are some implementation methods of the present invention. For ordinary technicians in this field, other specific embodiments can be obtained based on these specific embodiments without paying creative work.

[0047] The components of the embodiments of the present invention generally described and shown in the specific embodiments herein can be arranged and designed in countless different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the specific embodiments is not intended to limit the scope of the invention claimed, but only represents selected embodiments of the present invention.

[0048] See Figure 1 , is a classification flow chart of an embodiment of an engine emergency fault type classification method based on improved random forest of the present invention, and the specific operation steps are as follows:

[0049] 1) Data labeling: When performing a tightening operation, the relationship between torque and angle shows how the torque changes as the bolt rotation angle increases. According to engineering standards or process requirements, rules for the tightening process are formulated, and unqualified curves are classified and labeled, such as "slipping", "thread deformation", "slip jamming" and other labels;

[0050] 2) Data preprocessing: firstly, smooth the curve, remove noise points, and complete the curve data cleaning; then segment and remove the data segment without torque to complete the extraction of the main curve segment; curve data cleaning is aimed at the noise interference in the original tightening curve, and the noise points are removed by using a smoothing algorithm, as shown in the following formula 1):

[0051]

[0052] Formula 1) uses Gaussian filtering smoothing algorithm to eliminate invalid data segments and remove invalid intervals without torque.

[0053] The extraction of main curve segments is based on the background knowledge of the bolt tightening process, and only the key curve segments related to defects are intercepted.

[0054] 3) Feature extraction: extract features from the preprocessed data and perform normalization processing; data extraction features refer to extracting features related to tightening quality from key curve segments, including any one or any combination of two or more of the maximum torque, slope change rate, and curve shape factor features. Normalization processing is to normalize the extracted feature values ​​to ensure that features of different dimensions have the same impact on the classification results.

[0055] 4) Train the improved random forest model. First, expand the decision tree. Each decision tree classifies the test sample and obtains the corresponding credibility. The credibility is selected as the weight of each decision tree's vote. Then, the random forest performs weighted voting on the classification results of multiple decision trees. Then, the information entropy value of the voting is calculated. If it is greater than the set threshold obtained through training, the test sample is considered to be a new type of defect. Otherwise, the defect category with the highest credibility in the voting result is selected as the final result to complete the classification.

[0056] The process steps of the improved random forest model in step 4) are as follows:

[0057] 1) Decision tree generation and credibility calculation: for each decision tree, the credibility of the classification result is calculated based on the credibility α of the classification result. The credibility of the decision tree is used as the voting weight of its classification result (see formula 2);

[0058]

[0059] 2) Weighted voting mechanism: In the forest model, the classification results of each decision tree are weighted according to their credibility, thus improving the classification accuracy of the overall model;

[0060] 3) Information entropy calculation: Calculate the information entropy H of the weighted voting results:

[0061]

[0062] Where: P_i is the voting weight of category k in the classification result;

[0063] 4) New defect discrimination mechanism, set the information entropy threshold θ, automatically tune through the training data set, if the information entropy H>θ, then the test sample is considered to belong to the new defect category. The information entropy threshold is set through training, and the threshold automatically trained by the training set is used to distinguish new types of defects from predefined defect types.

[0064] In the calculation of credibility using α, let the decision tree attribute splitting node value be value, the feature attribute value corresponding to the test case be x, and p represents the reliability of the decision tree going to the right subtree. The calculation formula is shown in Formula 4) as follows:

[0065]

[0066] The value range of credibility is (0, 1). The sigmoid function is introduced to map the variable to between 0 and 1. The formula of the sigmoid function is:

[0067]

[0068] According to the attribute value, it is determined whether to go to the left subtree or the right subtree. α represents the credibility of going to the target subtree, and 1-α represents the credibility of going to the other subtree.

[0069] Credibility reflects the degree of confidence of the decision tree in its decision results, so in subsequent voting, the vote weight of the one with higher credibility is greater than that of the one with lower credibility. Decision trees generally have multiple layers of nodes, and there are different credibility for going to the left subtree and the right subtree, and the credibility is multiplied as the weight.

[0070] It can be clearly seen from Table 1 that Algorithm 1 (improved random forest algorithm) further improves the accuracy and reliability of classification by introducing a threshold training algorithm (Algorithm 2). In the standard random forest training process, although the generalization ability of the model is improved through diversified decision trees and feature selection, the classification accuracy of Algorithm 1 may still be affected by certain categories of data. For further optimization, Algorithm 2 calculates the confidence and information entropy of each sample to find the most reliable division point in the classification process, and uses this division point as the threshold to improve the accuracy of the classifier.

[0071] The core function of Algorithm 2 is to determine an optimal threshold value, which is obtained through information gain and the confidence of the classification result, to help Algorithm 1 determine the classification standard, thereby effectively reducing misclassification and improving the ability to recognize new types of data. The threshold training algorithm (Algorithm 2) is used to optimize the improved random forest algorithm (Algorithm 1) to improve classification accuracy, especially when facing complex data sets or new types of data, to improve the recognition ability and decision accuracy of the classifier. Specifically, Algorithm 2 provides a key decision criterion (threshold value) for Algorithm 1, which can help the improved random forest model make more accurate and reliable judgments when classifying.

[0072] Table 1

[0073]

[0074]

[0075] The present invention improves the detection capability of new types of defects by introducing a credibility weighting mechanism. The defect data classification algorithm based on the improved random forest can not only classify new types of data, but also effectively detect new types of defect data. It can not only classify data of existing defect types, but also effectively identify new types of fault data, thereby significantly improving the defect recognition rate and the misjudgment rate.

[0076] The data set used in the embodiment of the present invention is introduced as follows: The embodiment of the present invention collects 500 pieces of engine tightening fault data, which are divided into 8 categories. In the training set, we use the data of 6 known categories, and the test set contains data of all categories. The classification performance of the three models is examined by cross-validation, and the classification indicators use accuracy and F1 score. The results of the comparison method are shown in Table 2.

[0077] F1 is the abbreviation of F1-Score, which is mainly used to evaluate the performance of classification models. It is the harmonic mean of precision and recall. The definition of F1 score is

[0078]

[0079] In formula 6):

[0080] Precision indicates how many of the samples predicted as positive are actually positive samples.

[0081]

[0082] In formula 7), TP is the true positive and FP is the false positive.

[0083] Recall indicates how many of the actual positive samples are correctly predicted as positive samples:

[0084]

[0085] In formula 8), FN is the false negative example (False Negatives).

[0086] Table 2

[0087]

[0088] In Table 2, "Support Vector Machine" and "Random Forest" are used as comparison models to obtain the accuracy and F1 values ​​of these two classic machine learning models and the embodiments of the present invention. The improved random forest model of the present invention has obvious advantages in classification tasks and provides the best classification performance, especially in terms of F1 score, indicating that it is more effective in dealing with classification problems. This method is superior to the traditional random forest model because it improves the model's ability to detect new types of defects, especially when faced with unseen defect data, and can effectively improve the accuracy of classification. The improved random forest algorithm, combined with the credibility weighting mechanism, significantly improves the defect recognition rate and reduces misjudgment. The model shows higher accuracy in detecting new types of fault data, thereby improving the reliability of the overall system.

[0089] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An engine fault type classification method based on improved random forest, characterized in that: The specific steps are as follows: 1) Data labeling: When performing a tightening operation, the relationship between torque and angle shows how the torque changes as the bolt rotation angle increases. According to engineering standards or process requirements, rules for the tightening process are formulated, and unqualified curves are classified and labeled; 2) Data preprocessing: first, smooth the curve, remove noise points, and complete the curve data cleaning; then segment and remove the data segment without torque to complete the extraction of the main curve segment; 3) Feature extraction: extract features from the preprocessed data and perform normalization; 4) Train the improved random forest model. First, expand the decision tree. Each decision tree classifies the test sample and obtains the corresponding credibility. The credibility is selected as the weight of each decision tree's vote. Then, the random forest performs weighted voting on the classification results of multiple decision trees. Then, the information entropy value of the voting is calculated. If it is greater than the set threshold obtained through training, the test sample is considered to be a new type of defect. Otherwise, the defect category with the highest credibility in the voting result is selected as the final result to complete the classification.

2. The engine fault type classification method based on improved random forest according to claim 1 is characterized in that: The curve data cleaning in step 2) is aimed at the noise interference in the original tightening curve, and a smoothing algorithm is used to remove the noise points, as shown in the following formula: The Gaussian filtering smoothing algorithm is used to eliminate invalid data segments and remove invalid intervals without torque.

3. The engine fault type classification method based on improved random forest according to claim 1 is characterized in that: The extraction of the main curve segments in step 2) is based on the background knowledge of the bolt tightening process, and only the key curve segments related to the defects are intercepted.

4. The engine fault type classification method based on improved random forest according to claim 1 is characterized in that: The data extraction feature in step 3) refers to extracting features related to tightening quality from key curve segments, including any one or any combination of two or more of the maximum torque, slope change rate, and curve shape factor features.

5. The engine fault type classification method based on improved random forest according to claim 1 is characterized in that: The normalization process in step 3) is to normalize the extracted feature values ​​to ensure that features of different dimensions have the same impact on the classification results.

6. The engine fault type classification method based on improved random forest according to claim 1 is characterized in that: The process steps of the improved random forest model in step 4) are as follows: 1) Decision tree generation and credibility calculation: for each decision tree, the credibility of the classification result is calculated based on the credibility α of the classification result. The credibility of the decision tree is used as the voting weight of its classification result. 2) Weighted voting mechanism: In the forest model, the classification results of each decision tree are weighted according to their credibility, thus improving the classification accuracy of the overall model; 3) Information entropy calculation: Calculate the information entropy H of the weighted voting results: Where: P_i is the voting weight of category k in the classification result; 4) New defect discrimination mechanism, setting the information entropy threshold θ, automatically tuning through the training data set, if the information entropy H>θ, then the test sample is considered to belong to the new defect category.

7. The engine fault type classification method based on improved random forest according to claim 6 is characterized in that: In the calculation of the credibility using α, the decision tree attribute splitting node value is set to value, the feature attribute value corresponding to the test case is x, and p represents the reliability of the decision tree going to the right subtree. The calculation formula is: The value range of credibility is (0, 1). The sigmoid function is introduced to map the variable to between 0 and 1. The formula of the sigmoid function is: According to the attribute value, it is determined whether to go to the left subtree or the right subtree. α represents the credibility of going to the target subtree, and 1-α represents the credibility of going to the other subtree.

8. The engine fault type classification method based on improved random forest according to claim 6 is characterized in that: The credibility reflects the degree of confidence of the decision tree in its decision results, so in subsequent voting, the voting weight of those with higher credibility is greater than that of those with lower credibility.

9. The engine fault type classification method based on improved random forest according to claim 6, characterized in that: The setting of the information entropy threshold is obtained through training, and the threshold obtained by automatic training of the training set is selected to distinguish new types of defects from predefined defect types.

10. The engine fault type classification method based on improved random forest according to claim 7, characterized in that: The decision tree has multiple layers of nodes, and has different credibility for going to the left subtree and the right subtree, and the credibility is multiplied as a weight.

Images