A method and system for classifying whole vehicle mass problems based on a mixture model

Abstract

The application discloses a kind of classification method and system of vehicle quality problem based on mixed model, according to the historical data of multiple vehicle detection bad data, multiple text classification models are trained, according to the accuracy of each text classification model in training set and test set respectively Set the weight of each text classification model;Vehicle detection bad data is input into multiple text classification models, obtain multiple prediction results, according to multiple prediction results and multiple weights obtain the category of the vehicle detection bad data.The application establishes multiple model integrated text classification method, can quickly and accurately determine the category of vehicle detection quality problem, application category can guide host factory to quickly analyze quality problem and find relevant countermeasures;At the same time, the integrated learning idea is used, let multiple text classification models vote to determine the final category, so that the classification result is more accurate.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of deep learning, and particularly relates to a classification method and system for vehicle quality problems based on a hybrid model. BACKGROUND

[0002] After the completion of vehicle equipment, it needs to be detected. The detection personnel detects the vehicle faults through artificial senses (ears, eyes, hands) and vehicle terminal equipment connected to the vehicle. For the detected vehicle faults, the classification needs to be found, and then the corresponding responsible department is found for rectification according to the classification. In order not to affect the normal detection of the production line, the classification of the vehicle fault by the detection personnel has the problems of long time consumption, low precision, and no coherence of the data before and after. Therefore, the vehicle detection field needs a fast and accurate method to quickly determine the classification of the vehicle fault.

[0003] The determination data of the vehicle quality problem can be regarded as text data with the classification as the label, so the determination of the classification can be converted into a text classification model. At present, there are few research literatures on the combination of the classification of the vehicle quality problem in the production process of the host factory and the natural language processing technology and the deep learning technology. Most methods in the field of text classification only consider single language and single model, and the classification accuracy of Chinese text depends on the segmentation accuracy and the word vector expression result. At the same time, the increase of the training algebra of the single model is easy to cause overfitting, and the accuracy will be reduced, and the result accuracy depends on the training quality. SUMMARY

[0004] In order to improve the classification of each vehicle fault in the vehicle detection process and improve the efficiency of vehicle detection, the application provides a classification method and system for vehicle quality problems based on a hybrid model.

[0005] A classification method for vehicle quality problems based on a hybrid model for achieving one of the purposes of the application, comprising the following steps:

[0006] S1, training a plurality of text classification models according to a plurality of historical data of vehicle detection bad data, and setting the weight of each text classification model according to the accuracy of each text classification model in the training set and the test set; the text classification model is used to predict the category of each vehicle detection bad data;

[0007] S2, inputting the vehicle detection bad data into the plurality of text classification models to obtain a plurality of prediction results, and obtaining the category of the vehicle detection bad data according to the plurality of prediction results and the plurality of weights. The category can be used for actual purposes such as quality traceability of vehicle quality problems.

[0008] Further, the step S1 further includes removing part of columns of the vehicle detection failure data by using information entropy, calculating the entropy value of each column, and removing the column with an entropy value lower than a set value. The calculation method includes:

[0009]

[0010] In the formula:

[0011] H(X): represents the entropy value of column X;

[0012] In the formula represents the probability of the column data X being x i ; x i is the i-th value of the column data X.

[0013] Further, in order to improve the accuracy of the model, the step S1 further includes removing the vehicle detection failure data by calculating the proportion of null values, calculating the proportion of null values in each column of data, and removing the column with a proportion lower than a set value.

[0014] Further, in order to improve the accuracy of the model, the step S1 further includes column removal based on information redundancy. If a column of data is a group of data of several other columns of data, the several other columns of data are deleted.

[0015] Further, in order to improve the accuracy of the model, the step S1 further includes unifying professional terms in the data, and unifying different descriptions of the same object to improve the prediction accuracy.

[0016] Further, in the step S1, each piece of data in the vehicle detection failure data is composed of multiple groups of data representing vehicle fault attributes. When training the text classification model, each group of data in each piece of vehicle detection failure data except the category is separated by a space and then spliced into description data. The category is used as label data. The description data and the label data are used as input parameters of the text classification model. The output parameter of the text classification model is the predicted category of each piece of vehicle detection failure data.

[0017] Further, in the step S1, the text classification model includes a Fasttext model. When constructing the Fasttext model, a prefix “_label_” is added before the label data as an input parameter of the Fasttext model.

[0018] Further, the general word vector training is 100-200-300 latitude. In the present application, it is set to 300 latitude, in order to speed up the training speed, by calculating the variance of each latitude, it is found that the variance of the first 50 latitudes can well represent all data information. Therefore, in the step S1, the text classification model includes a Word2Vec-SVM model, when constructing and training the Word2Vec-SVM model, the word vector dimension trained by the Word2Vec algorithm is set to 300 dimensions, and then the SVM model is trained after dimension reduction to 50 dimensions by the PCA algorithm.

[0019] Further, in the step S1, the text classification model includes a Word2Vec-SVM model and a Fasttext model, and when training the Word2Vec-SVM model and the Fasttext model, the GridSearch algorithm is used to find the optimal solution in the hyperparameter space.

[0020] Further, in the step S1, the text classification model includes a DistilBERT model, and when fine-tuning the pre-trained model DistilBERT, the batch_size is set to 8 or 16 to prevent the machine memory from being insufficient when running the model.

[0021] Further, in the step S1, the text classification model includes a DistilBERT model, and when establishing the DistilBERT model, a prefix [CLS] and a suffix [SEP] are added to each label data as input parameters of the DistilBERT model.

[0022] Further, in the step S1, the method for setting multiple weights of each text classification model includes:

[0023]

[0024] In the formula:

[0025] ω: represents the weight of the text classification model, wherein ω1<ω2<ω3;

[0026] N1, N2: set value, N1<N2;

[0027] accuracy test : represents the accuracy of the text classification model in the training set;

[0028] accuracy train : represents the accuracy of the text classification model in the test set.

[0029] Further, the method of obtaining the category of the vehicle detection failure data in step S2 comprises: obtaining the vote number of each category according to the category output by each text classification model; and the category with the highest vote number is the final category of the vehicle detection failure data.

[0030] Further, the method of calculating the vote number comprises:

[0031]

[0032] wherein:

[0033] V label is the vote number of the category of the vehicle detection failure data as label;

[0034] P i-label represents whether the category output by the i-th model is label, and P i-label = 1 when the category output by the i-th model is label, and P = 0 otherwise.

[0035] w i represents the weight of the i-th model.

[0036] Further, since the prediction ability and processing manner of the models are different for different languages, and the data expressed in multiple languages can enrich the data set for model training and improve the prediction accuracy of the model; and since the data set of the vehicle detection failure data in the application is recorded in the form of entity multiple groups (‘vehicle model’, ‘primary parts’, ‘secondary parts’, ‘defects’...), each group is manually recorded, and the word segmentation accuracy is high; when automatic translation is performed, only a single word needs to be translated, without considering the context semantics; therefore, the translation accuracy is high when language conversion is performed, and the application is suitable for multi-language conversion. In addition, multiple languages greatly enrich the data set, and the prediction ability of different languages is different, which can make up for the deficiency of single language prediction. Therefore, in step S1, the multiple vehicle detection failure data are translated into multiple languages, and the multiple text classification models are trained.

[0037] A vehicle quality problem classification system based on a hybrid model for achieving the second purpose of the application comprises a text classification model training module and a classification module.

[0038] The text classification model training module is used for training multiple text classification models according to historical data of multiple vehicle detection failure data, and setting the weight of each text classification model according to the accuracy of each text classification model in the training set and the test set; and the text classification model is used for predicting the category of each vehicle detection failure data.

[0039] The classification module is used for inputting the whole vehicle detection poor data into multiple text classification models to obtain multiple prediction results, and obtaining the category of the whole vehicle detection poor data according to the multiple prediction results and the multiple weights.

[0040] Beneficial effects:

[0041] The method for text classification of mixed languages and multiple models is established, the category of the whole vehicle detection quality problem can be quickly and accurately determined, the quality problem can be quickly analyzed and the related countermeasures can be found according to the category, meanwhile, the integrated learning idea is adopted, multiple text classification models are used for voting to determine the final category, and the classification result is more accurate. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 is a flowchart of the method embodiment of the present application;

[0043] Figure 2 is a model training flowchart in the method embodiment of the present application;

[0044] Figure 3 is a model prediction flowchart in the method embodiment of the present application. DETAILED DESCRIPTION

[0045] The following detailed description is used to explain the technical solutions of the claims of the present application, so that those skilled in the art can understand the claims. The protection scope of the present application is not limited to the following specific implementation structure. The technical solutions of the claims of the present application which are different from the following specific embodiments and contain the technical solutions of the claims of the present application are also within the protection scope of the present application.

[0046] As Figure 1 shown is a flowchart of the method embodiment of the present application, specifically including:

[0047] S1, collect the historical data of whole vehicle quality problem determination in the WQIS system, including 25-dimensional data such as factory, vehicle type, vehicle body number, poor project, repair department, etc. The data sample is as follows:

[0048] Factory Model Bad data …… Category Occurrence time Factory 1 XR-V Roof rear deformation …… FQ 2020-12-2610:36 Factory 1 XR-V 2D out of frame injury …… FQ 2020-12-2610:25 Factory 2 XR-V Roof front deformation …… AF 2020-12-2610:37 Factory 3 XR-V Roof rear deformation …… FQ 2020-12-2610:36

[0049] Table 1

[0050] The original text data is preprocessed by column elimination and professional term unification, specifically as follows:

[0051] ① Column elimination based on information entropy: the information entropy of each column is calculated based on the following formula:

[0052]

[0053] In the formula The probability of this column of data X is x i .

[0054] By comparing the information entropy of each column with the set minimum threshold value to eliminate the column.

[0055] In this embodiment, the value_counts function of pandas library is used to quickly calculate the information entropy of each column; taking the first column "factory" as an example, the information entropy of this column is calculated: the column data X is divided into 'one factory', 'two factory', 'three factory' three categories; the probability of one factory is The probability of two factories is The probability of three factories is The entropy of this column is 1.54, and the calculation process is as follows, the threshold value set in this case is 0.3, and the column is retained.

[0056] P = [23318 / 76521, 33926 / 76521, 19927 / 76521]

[0057] -tf.reduce_sum(P*tf.math.log(P) / tf.math.1og(2.)).numpy()

[0058] P = 1.5481832.

[0059] After calculation, it is found that the "inspection department" and "repair method" columns can be eliminated; at the same time, it is found that the information entropy values of "vehicle body number" and "occurrence time" columns are very large, but considering that these two columns have no actual meaning, they are eliminated.

[0060] ② Column elimination based on null value proportion: calculate the null value proportion of each column, set the threshold value to 80%, and eliminate the "responsibility system", "displacement", and "part" columns;

[0061] ③ Column elimination based on information redundancy: it is observed that the column "defect item" is a combination of columns "primary part", "secondary part", "associated part", and "defect phenomenon", so the column "defect item" is eliminated.

[0062] ④ Professional term unification: in the column "responsibility system", the description of the same object is not unified, for example: "AF1", "AF2", "AF3" all represent "AF", which is unified as "AF".

[0063] ⑤ Merge multiple columns: after processing, the column with the label "label" is merged into the "text" column with spaces.

[0064] S2, language conversion is performed on the text data after preprocessing. The present application is only converted into 4 languages, which can be extended to more. Call the python non-standard library translate, create the Translator class, set the initial parameters from_lang="Chinese", to_lang="Jananese", "English", "French", "Korean", and expand the Chinese text data into Chinese, Japanese, English, French, and Korean five languages; then process the five language data according to the different classification models, wherein for the model Word2Vec-SVM, the five languages are respectively processed by word segmentation; for the model Fasttext, the "__label__" label needs to be added; for the model DistilBert, the '[CLS]' and '[SEP]' symbols need to be added.

[0065] The preprocessed multi-column data samples are as follows:

[0066]

[0067] Table 2

[0068] The corresponding Korean version text data samples are as follows (other language text data is similar):

[0069]

[0070] Table 3

[0071] Taking the Chinese text data set as an example, the corresponding model Fasttext, the "category" is prefixed with "_label_", such as "_label_FQ" and "_label_AF";

[0072] Taking the Chinese text data set as an example, the corresponding model DistilBERT, the text data is added with the prefix and suffix '[CLS]' and '[SEP]' symbols by the tokenizer function, and the sample is as follows:

[0073] "[CLS] a factory xrv interior ceiling rear deformation offline tkl hf7 crystal Yao white 2021 not repaired 1 diligence level high flying vq [SEP]";

[0074] Among them: '[CLS]' represents classification, which is a classification model;'[SEP]'represents separate, which represents the end of a sentence.

[0075] S3, three text classification models based on Fasttext algorithm, Word2Vec-SVM algorithm and DistilBERT algorithm are established. Specifically, it includes:

[0076] ①Take 80% of the original data as the training set, and the remaining 20% of the data as the test set. During sampling, due to the imbalance of the data set, the RandomOverSampler method is used according to the label tag;

[0077] ②During the training of the Word2Vec-SVM model, the word vector dimension trained by the Word2Vec algorithm is set to 300 dimensions, and the SVM model is trained after dimension reduction to 50 dimensions by the PCA algorithm;

[0078] ③When training the Word2Vec-SVM model and the Fasttext model, the GridSearch algorithm is used to find the optimal solution in the hyperparameter space. When fine-tuning the pre-trained model DistilBERT, in order to prevent insufficient memory, set batch_size=8 or 16;

[0079] ④After training the three models, input the 20% test set into the model to evaluate the model. The evaluation rules and weight setting method are as follows:

[0080] Evaluation rules:

[0081] 1. The test set accuracy is significantly higher than the training set accuracy, which exceeds the set threshold value, and the model training effect is good. The model should be assigned a higher weight; 2. The test set accuracy is comparable to the training set accuracy, within the set threshold value, and the model is general, and is assigned a general weight; 3. The test set accuracy is significantly lower than the training set accuracy, which exceeds the set threshold value, and the model is considered to be overfitting, and the model should be assigned a lower weight.

[0082] The calculation method of the weight ω in this embodiment is as follows:

[0083] When ω=0.5

[0084] When ω=1

[0085] When ω=1.5

[0086] In the formula:

[0087] accuracy train is the accuracy of the model in the training set;

[0088] accuracy test is the accuracy of the model in the test set;

[0089] accuracy=number of correctly predicted samples / total number of samples.

[0090] Supplementary explanation using Chinese text training of the Word2Vec-SVM model as an example;

[0091] The hyperparameter space for training the SVM classifier using GridSearch is as follows:

[0092]

[0093] Table 4

[0094] Accuracy of the Word2Vec-SVM model on Chinese text data training set train =89.98%, test set accuracy test =90.56%, according to the weight calculation formula, set w=1. The weight setting method for other models is similar, such as... Figure 2 As shown.

[0095] S4. Fifteen text classification models were obtained through step S3. A single piece of vehicle inspection defect data was input and processed according to steps S1-S2. Then, the data was input into each of the 15 text classification models from step S3, resulting in 15 classification results. Based on the classification results and the model weights calculated in step S3, a final classification was determined through voting. In this embodiment, the final classification corresponds one-to-one with the responsible department for the fault, and it can be used for quality backtracking of the problem. Figure 3 As shown. The formula for calculating the number of votes for each tag is:

[0096]

[0097] In the formula: V label ω represents the number of votes predicted as the label. i P represents the weights of the i-th model. i-label This represents the probability that model i predicts the label. For example, in this case, the weights of each model are calculated based on the evaluation rules and the accuracy of the training and test sets, as shown in the table below:

[0098] Chinese Japanese English French Korean Word2Vec-SVM 1.5 1 1.5 1 1 Fasttext 1.5 1 1.5 1 1.5 DistilBERT 1 1 1 0.5 1

[0099] Table 5

[0100] Input a defect record "XRV interior roof rear deformation", and the output results for each model are shown in the table below:

[0101] Chinese Japanese English French Korean Word2Vec-SVM FQ FQ FQ FQ AF Fasttext FQ FQ FQ FQ FQ DistilBERT FQ FQ FQ AF FQ

[0102] Table 6

[0103] Then we have: V FQ =15.5,V AF =1.5,V other= 0. Therefore, the classification of the piece of vehicle detection bad data is FQ.

[0104] It should be understood that the size of the serial number of each step in the above embodiment does not mean the order of execution, and the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0105] The embodiments of the present application also provide a mixed model vehicle quality problem classification system, comprising a text classification model training module and a classification module;

[0106] The text classification model training module is used for training a plurality of text classification models according to historical data of a plurality of pieces of vehicle detection bad data, and setting a weight of each text classification model according to the accuracy of each text classification model in the training set and the test set; the text classification model is used for predicting the category of each piece of vehicle detection bad data;

[0107] The classification module is used for inputting the vehicle detection bad data into the plurality of text classification models to obtain a plurality of prediction results, and obtaining the category of the vehicle detection bad data according to the plurality of prediction results and the plurality of weights.

Claims

1. A method for classifying whole vehicle mass problems based on a mixture model, characterized by, The method comprises the following steps: S1, training a plurality of text classification models according to historical data of a plurality of vehicle detection bad data, setting a weight of each text classification model according to the accuracy of each text classification model in the training set and the test set respectively; The text classification model is used for predicting the category of each vehicle detection bad data; S2, inputting the vehicle detection bad data into the plurality of text classification models to obtain a plurality of prediction results, and obtaining the category of the vehicle detection bad data according to the plurality of prediction results and the plurality of weights; In S1, the method for setting the plurality of weights of each text classification model comprises: ； In the formula: : represents weights of a text classification model, where ​​​ N1, N2: set value, N1 < N2; : indicates the accuracy of the text classification model on the training set; : represents the accuracy of the text classification model on the test set; In S1, the plurality of vehicle detection bad data is translated into a plurality of languages to train the plurality of text classification models.

2. The hybrid model based classification method of vehicle quality problems as claimed in claim 1, wherein, In the step S2, the method for obtaining the category of the vehicle detection bad data comprises: obtaining the vote number of each category according to the category output by each text classification model; and the category with the highest vote number is the final category of the vehicle detection bad data.

3. The classification method for vehicle quality problems based on a hybrid model as described in claim 2, characterized in that, The calculation method of the vote number is: ； In the formula: : the number of tickets of the category of the bad data of the whole vehicle detection for the line as label; : indicates whether the class of the i-th model output is label, 1 if label, otherwise 0; = 1, otherwise 0; : represents the weight of the i-th model.

4. The hybrid model based classification method of vehicle quality problems as claimed in claim 1, wherein, In the step S1, the text classification model comprises a Word2Vec-SVM model, when constructing and training the Word2Vec-SVM model, the word vector dimension trained by the Word2Vec algorithm is set to 300 dimensions, and the SVM model is trained after dimension reduction to 50 dimensions by the PCA algorithm.

5. The hybrid model based classification method of vehicle quality problems as claimed in claim 1, wherein, The text classification model comprises a Word2Vec-SVM model, when training the Word2Vec-SVM model, the GridSearch algorithm is used to find the optimal solution in the hyperparameter space.

6. The hybrid model based classification method of vehicle quality problems as claimed in claim 1, wherein, The text classification model comprises a Fasttext model, when training the Fasttext model, the GridSearch algorithm is used to find the optimal solution in the hyperparameter space.

7. The hybrid model based classification method of vehicle quality problems as claimed in claim 1, wherein, The text classification model comprises a DistilBERT model, when fine-tuning the pre-trained model DistilBERT, the batch_size is set to 8 or 16.

8. A hybrid model based classification system for vehicle mass problems using the method of claim 1, wherein, The text classification model training module and the quality problem classification module; The text classification model training module is used for training a plurality of text classification models according to historical data of a plurality of vehicle detection bad data, and setting a weight of each text classification model according to the accuracy of each text classification model in the training set and the test set respectively; The text classification model is used for predicting the category of each vehicle detection bad data; The quality problem classification module is used for inputting the vehicle detection bad data into the plurality of text classification models to obtain a plurality of prediction results, and obtaining the category of the vehicle detection bad data according to the plurality of prediction results and the plurality of weights.

Images