Intelligent extraction method and device of fault diagnosis information based on dual models

Abstract

The present invention discloses a method and device for intelligently extracting fault diagnosis information based on dual models, wherein the method comprises: based on a historical diagnosis database, extracting, standardizing, and cleaning historical fault diagnosis data in the historical diagnosis database to generate diagnosis data; marking the diagnosis data to filter out non-DTC diagnosis data; based on the non-DTC diagnosis data, sampling the non-DTC diagnosis data to generate sample diagnosis data; extracting the sample diagnosis data to generate benchmark diagnosis data; based on the benchmark diagnosis data, generating a training model learning corpus; based on the training model learning corpus, training and generating a first training model. The method for intelligently extracting fault diagnosis information based on dual models of the present invention can not only accurately and efficiently extract effective diagnosis information from complex diagnosis data, but also ensure the consistency of quality during the information extraction process.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent diagnosis of automobile faults, and in particular to a method and device for intelligent extraction of fault diagnosis information based on dual models. Background Art

[0002] Diagnostic data is an information carrier for recording fault diagnosis, repair, and maintenance that occurred during the actual driving of the vehicle, and contains a large number of different fault phenomena and corresponding fault cause information. Extracting information and analyzing the fault principle of a large amount of diagnostic data is an effective way to achieve intelligent fault diagnosis recommendation. However, the existing diagnostic information extraction method is inaccurate and inefficient, and cannot guarantee the quality consistency during the information extraction process.

[0003] The information disclosed in this background technology section is only intended to enhance the understanding of the overall background of the invention and should not be regarded as an acknowledgment or any form of suggestion that the information constitutes the prior art already known to a person skilled in the art. Summary of the invention

[0004] The purpose of the present invention is to provide a dual-model based fault diagnosis information intelligent extraction method and device, which can not only accurately and efficiently extract effective diagnostic information from complex diagnostic data, but also ensure quality consistency during the information extraction process.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a method for intelligently extracting fault diagnosis information based on a dual model, comprising:

[0006] Based on a historical diagnosis database, extracting, standardizing, and cleaning historical fault diagnosis data in the historical diagnosis database to generate diagnostic data;

[0007] Marking the diagnostic data to filter out non-DTC diagnostic data;

[0008] Based on the non-DTC diagnostic data, sampling the non-DTC diagnostic data to generate sample diagnostic data;

[0009] extracting the sample diagnostic data to generate benchmark diagnostic data;

[0010] Based on the benchmark diagnostic data, generating a training model learning corpus;

[0011] Based on the training model learning corpus, training and generating a first training model;

[0012] Based on the training model learning corpus, training and generating a second training model;

[0013] Based on the sample diagnostic data, effective diagnostic information in the sample diagnostic data is extracted using the first training model and the second training model respectively.

[0014] In one embodiment of the present invention, the extracting, standardizing, and cleaning the historical fault diagnosis data in the historical diagnosis database to generate the diagnosis data includes:

[0015] Based on the historical diagnosis database, extracting the diagnosis data in the historical diagnosis database;

[0016] Standardizing the names of the fault cause parts in the diagnostic data;

[0017] The fault phenomenon description content in the diagnostic data is identified, and the preset label type data is cleaned to generate the diagnostic data.

[0018] In one embodiment of the present invention, marking the diagnostic data to filter out non-DTC diagnostic data includes:

[0019] Marking the diagnostic data according to a preset DTC diagnostic data identification rule to generate DTC diagnostic data;

[0020] The DTC-type diagnostic data in the diagnostic data is removed, and the non-DTC-type diagnostic data is screened out.

[0021] In one embodiment of the present invention, sampling the non-DTC diagnostic data based on the non-DTC diagnostic data to generate sample diagnostic data includes:

[0022] Based on the non-DTC diagnostic data, the non-DTC diagnostic data is sampled according to a preset sampling rule to generate the sample diagnostic data.

[0023] In one embodiment of the present invention, extracting the sample diagnostic data to generate reference diagnostic data includes:

[0024] Extracting keyword information used to describe the fault phenomenon and the fault cause from the sample diagnostic data to obtain the fault phenomenon keywords and the fault cause keywords;

[0025] According to the fault phenomenon keywords and the fault cause keywords, respectively creating corresponding universal phenomenon words and universal cause words that are common across vehicle models;

[0026] The reference diagnostic data is generated based on the sample diagnostic data, the fault phenomenon keywords, the fault cause keywords, the common phenomenon words, and the common cause words.

[0027] In one embodiment of the present invention, the training and generating a first training model based on the training model learning corpus includes:

[0028] Install and call the information extraction model UIE, convert the training model learning corpus into a first data set in a data format required by the information extraction model UIE, and split the first data set into a first training data set and a first verification data set according to a preset ratio;

[0029] Loading a first pre-training module and loading the first training data set to perform model training to generate a first fine-tuned model;

[0030] The first fine-tuned model is loaded to perform result prediction and generate a first training model.

[0031] In one embodiment of the present invention, the training and generating a second training model based on the training model learning corpus includes:

[0032] Install and call the knowledge enhancement model PaddleNLP ERNIE3.0, convert the training model learning corpus into a second data set in the data format required by the knowledge enhancement model PaddleNLP ERNIE3.0, and split the second data set into a second training data set and a second verification data set according to a preset ratio;

[0033] Loading a second pre-training module and loading the second training data set to perform model training to generate a second fine-tuned model;

[0034] The second fine-tuned model is loaded to perform result prediction and generate a second training model.

[0035] In a second aspect, the present invention provides a dual-model fault diagnosis information intelligent extraction device, comprising: a first generation module, a second generation module, a third generation module, a fourth generation module, a fifth generation module, a first training module, a second training module, a sixth generation module and a seventh generation module. The first generation module is used to extract, standardize and clean the historical fault diagnosis data in the historical diagnosis database based on the historical diagnosis database to generate diagnosis data; the second generation module is used to mark the diagnosis data and filter out non-DTC diagnosis data; the third generation module is used to sample the non-DTC diagnosis data based on the non-DTC diagnosis data to generate sample diagnosis data; the fourth generation module is used to extract the sample diagnosis data to generate reference diagnosis data; the fifth generation module is used to generate a training model learning corpus based on the reference diagnosis data; the first training module is used to train and generate a first training model based on the training model learning corpus; the second training module is used to train and generate a second training model based on the training model learning corpus; the sixth generation module is used to input the sample diagnosis data into the first training model to generate first effective diagnosis information; and the seventh generation module is used to input the first effective diagnosis information into the second training model to generate second effective diagnosis information.

[0036] In one embodiment of the present invention, the first generation module includes: a first extraction unit, a processing unit, and a first generation unit. The first extraction unit is used to extract the diagnostic data in the historical diagnostic database based on the historical diagnostic database; the processing unit is used to standardize the names of the fault cause parts in the diagnostic data; and the first generation unit is used to identify the description content of the fault phenomenon in the diagnostic data, clean the preset label type data, and generate the diagnostic data.

[0037] In one embodiment of the present invention, the second generating module includes: a second generating unit and a third generating unit. The second generating unit is used to mark the diagnostic data according to a preset DTC diagnostic data identification rule to generate DTC diagnostic data; and the third generating unit is used to remove the DTC diagnostic data from the diagnostic data to filter out the non-DTC diagnostic data.

[0038] Compared with the prior art, the dual-model based intelligent extraction method and device for fault diagnosis information according to the present invention can not only accurately and efficiently extract effective diagnostic information from complex diagnostic data, but also ensure quality consistency during the information extraction process. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a flow chart of a method for intelligent extraction of fault diagnosis information based on dual models in Embodiment 1 of the present invention;

[0040] Figure 2 is a flowchart of generating diagnostic data in Embodiment 1 of the present invention;

[0041] Figure 3 is a flowchart for filtering out non-DTC diagnostic data in the first embodiment of the present invention;

[0042] Figure 4 is a flowchart of generating benchmark diagnostic data in Embodiment 1 of the present invention;

[0043] Figure 5 is a flowchart of training to generate a first training model in Embodiment 1 of the present invention;

[0044] Figure 6 is a flow chart of generating a second training model through training in Embodiment 1 of the present invention;

[0045] Figure 7 It is a structural diagram of a dual-model-based intelligent extraction device for fault diagnosis information in Embodiment 2 of the present invention. DETAILED DESCRIPTION

[0046] The embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the embodiments of the present invention, rather than to limit the embodiments of the present invention. It is also necessary to explain that, for ease of description, only parts related to the embodiments of the present invention are shown in the accompanying drawings, rather than all structures.

[0047] To facilitate understanding, the main implementation concepts of the embodiments of the present invention are first briefly described.

[0048] Diagnostic data is an information carrier that records the fault diagnosis, repair, and maintenance that occurred during the actual driving of the vehicle. It contains a large number of different fault phenomena and the corresponding fault cause information. Extracting information and analyzing the fault principle of a large amount of diagnostic data is an effective way to achieve intelligent fault diagnosis recommendations. However, there are some challenges in the process of extracting diagnostic information:

[0049] 1. Complexity of information extraction: Diagnostic data will record a large amount of data that is useless for analyzing the fault principle, and maintenance personnel will be very arbitrary in the language description when recording diagnostic information, which will cause great trouble in identifying and accurately extracting effective diagnostic information;

[0050] 2. Information consistency requirements: When extracting effective information from diagnostic data, different personnel have different abilities to identify effective information. It is difficult to ensure the consistency of information extraction by relying solely on manual batch processing.

[0051] 3. The workload of data processing is huge: In the automotive aftermarket, the amount of diagnostic data recorded each year is over one million, and the amount of data processing is extremely large. It is difficult to meet the actual data processing needs if it is analyzed and processed entirely by manual methods.

[0052] The inventors discovered the technical defects as described in the previous background technology, and then hoped to find a way to not only accurately and efficiently extract effective diagnostic information from complex diagnostic data, but also ensure quality consistency during the information extraction process.

[0053] In order to solve the technical problems in the prior art that the diagnostic information extraction methods are inaccurate, inefficient, and cannot ensure quality consistency during the information extraction process, the inventors of the present invention, through creative work, have obtained the present invention's method and device for intelligent extraction of fault diagnostic information based on dual models. Example

[0054] Figure 1 FIG. 1 is a flow chart of a method for intelligently extracting fault diagnosis information based on dual models in Embodiment 1 of the present invention. Figure 1 As shown, a dual-model fault diagnosis information intelligent extraction method provided in Embodiment 1 includes:

[0055] Step S100, based on a historical diagnosis database, extracting, standardizing, and cleaning historical fault diagnosis data in the historical diagnosis database to generate diagnosis data;

[0056] Specifically, the dual-model-based fault diagnosis information intelligent extraction method of the present embodiment is applied to a dual-model-based fault diagnosis information intelligent extraction system, in which an original diagnostic data storage module is created to store the original fault diagnosis data actually occurring in the market. A template is preset in the original diagnostic data storage module, and the original diagnostic data storage module automatically extracts the information columns such as the maintenance type, brand, vehicle model, fault phenomenon description, fault cause part number, and fault cause part name in the imported original fault diagnosis data according to the preset template. Among them, since the diagnostic data involves vehicle models with different configurations, there are differences in the names of the same parts in the loading lists of different vehicle models. In order to facilitate the unified management of the name information of each part, the different names of the same parts are unified to create a common common name by analyzing the parts atlas, and the corresponding relationship between the brand name and the common name of the loading part is saved. The fault cause part names in the fault cause part name information column of the above-extracted diagnostic data are uniformly replaced with the corresponding fault cause part common name, for example: ECM has different brand names in different vehicle models, such as: ECM, engine electronic control unit, engine ECM controller, engine ECU, etc. An automatic labeling function is preset in the original diagnostic data storage module. By identifying the description of the fault phenomenon in the diagnostic data, the diagnostic data containing the preset keywords "quality rectification" and "passive upgrade" are automatically labeled with the label "quality rectification". The type data of the maintenance type is ordinary maintenance or quality rectification, and the diagnostic data marked as "quality rectification" are cleaned (i.e. deleted), and the cleaned diagnostic data is output.

[0057] Based on the above analysis, it can be seen that the dual-model fault diagnosis information intelligent extraction method of this embodiment is based on the historical fault diagnosis data in the historical diagnosis database. The historical fault diagnosis data includes a large number of different fault phenomenon situations of a large number of different automobiles and the corresponding fault cause information. This embodiment extracts, standardizes and cleans the information in the historical diagnosis database in a targeted manner, facilitates subsequent identification and marking, and avoids the waste of manpower due to processing invalid data to a great extent.

[0058] Step S200, marking the diagnostic data to filter out non-DTC diagnostic data;

[0059] Specifically, the DTC-type diagnostic data recognition rules are preset in the original diagnostic data storage module. When the words and letters such as "current, read, fault code, report code, instrument display, computer check, OBD, FC, SPN, FMI, DTC" are identified in the fault phenomenon description, the original diagnostic data storage module automatically marks the diagnostic data as "DTC-type diagnostic data". The recognition range of DTC-type diagnostic data keywords can be maintained according to the actual data characteristics. For example: the fault phenomenon description of the diagnostic data is: engine failure, the instrument displays SPN:91; FMI:2, and it is judged to be an accelerator pedal failure. If the fault phenomenon description contains the preset "instrument display", "SPN", and "FMI" keywords, it is marked as DTC-type diagnostic data, and the unmarked data are all non-DTC-type diagnostic data.

[0060] Based on the above analysis, it can be seen that by presetting the DTC diagnostic data keyword recognition rules in the original diagnostic data storage module, DTC diagnostic data can be quickly obtained and identified, thereby filtering out non-DTC diagnostic data, which is convenient for the subsequent separate extraction of non-DTC fault information.

[0061] Step S300, based on the non-DTC diagnostic data, sampling the non-DTC diagnostic data to generate sample diagnostic data;

[0062] Specifically, a data sampling module is created in the system, and sampling rules are preset in the data sampling module. A small sample is taken from the cleaned non-DTC diagnostic data, and the sample is ensured to cover as much diagnostic data of all fault cause parts as possible. The non-DTC diagnostic big data in the fault phenomenon is sampled. When sampling, the fault phenomenon description + fault cause part information in the diagnostic data is used as the sampling basis, and the sample size is set to N:

[0063] In the non-DTC diagnostic data, all fault cause item numbers with a corresponding diagnostic data frequency of less than 5 times are extracted;

[0064] If the fault cause number is higher than 5 times, sampling is performed based on the proportion of the corresponding diagnostic data frequency in the sum of all non-DTC diagnostic big data frequencies;

[0065] For diagnostic data with a common name frequency of fault cause parts higher than 5, random sampling is performed by moving forward 5 years from the current year and setting weights for each year. If the number of entries for a certain year with set weights is insufficient, all entries will be selected, and the remaining sampling quota will be rolled over to other years in proportion to the above weights;

[0066] Set the sampling sample number threshold (the number of single samples is controlled within 40,000). The threshold can be adjusted according to actual conditions.

[0067] Based on the above analysis, it can be seen that by setting sampling rules, non-DTC diagnostic data is sampled and information is extracted, and the sampling process covers all the diagnostic data of the fault cause parts that have appeared, which is conducive to cutting the workload of data processing in steps and summarizing and improving the subsequent information extraction rules for different types of diagnostic data.

[0068] Step S400, extracting the sample diagnostic data to generate benchmark diagnostic data;

[0069] Specifically, a fault diagnosis information extraction module 708 is created in the system, and the keyword information used to describe the fault phenomenon and the fault cause information in the fault description of the first batch of sample data after sampling is manually extracted; because the keywords for describing the same fault phenomenon and cause are inconsistent in different models, in order to accurately analyze the association logic between different fault phenomena and the technical logic of each fault phenomenon and the fault cause, corresponding universal phenomenon words and universal cause words that are common to all models are created according to the phenomenon keywords and cause keywords; the manually extracted and frozen valid data are stored as the benchmark data, for example: the fault phenomenon description in the fault diagnosis data is: the right side window lifter cannot be raised or lowered, and it is found that the window lifter motor is not working after inspection. The extractable phenomenon keywords are: right side glass, cannot be raised or lowered. The phenomenon universal word can be created as: the right door glass cannot be raised or lowered. The extractable cause keywords are: right side glass, window lifter motor, not working. The cause universal word can be created as: the right door window lifter motor is damaged.

[0070] Step S500, generating a training model learning corpus based on the benchmark diagnostic data;

[0071] Specifically, a learning corpus production module is created in the system to produce learning corpus for training the model. The keyword tagging model standardizes the format of the sample data, automatically adds the "Description:" field before the sentence of the fault phenomenon description content of each sample data, and adds the "The common name of the fault cause is xxx" field at the end of the sentence. The information of the fault phenomenon description after format standardization + the manually extracted (phenomenon + cause) keywords + the manually created (phenomenon + cause) common words is used as a complete learning corpus. Among them, "xxx" is the information automatically extracted by the keyword tagging model from the fault cause name information column of the sample data. For example: The fault phenomenon description in the sample data is: The right glass lifter cannot be raised or lowered. After inspection, it is found that the glass lifter motor is not working. Among them, the content of the fault cause name information column is: Glass lifter motor. The sample data format is standardized as: Description: The right glass lifter cannot be raised or lowered. After inspection, it is found that the glass lifter motor is not working. The fault cause is the glass lifter motor.

[0072] Specifically, the learning corpus marking rules are preset in the learning corpus production module to mark the phenomenon and cause keywords respectively. The general marking rules are as follows:

[0073] The keyword tagging model identifies punctuation marks other than ","," / ,"()," and ")" in the fault phenomenon description content and automatically divides the content into several segments;

[0074] When marking, each keyword is only allowed to be marked once;

[0075] In the same extracted attribute (i.e. phenomenon tag or cause tag), characters in the same position are not allowed to be repeated;

[0076] When marking, mark the keywords one by one according to their length, from long to short.

[0077] When multiple phenomena or multiple cause keywords appear in the same learning corpus, they must be merged and duplicated, and the keywords in the complete sentence must be searched and marked starting from the end of the sentence and moving toward the beginning of the sentence.

[0078] Step S600, training and generating a first training model based on the training model learning corpus;

[0079] Specifically, the information extraction model UIE is trained and the information extraction model PaddleNLPUIE is installed in the python environment. The learning corpus in step S500 is converted into the data format required by the PaddleNLP UIE dataset. The dataset is split into a training set and a validation set according to a 9:1 ratio. The training model performs deep learning on the description content and tag information of the fault phenomenon, and intelligently recommends (phenomenon + cause) keywords.

[0080] For example: Data set format: Taking phenomenon keyword extraction as an example

[0081] {

[0082] "content":"Description: The customer called to report that the vehicle was not generating electricity. Our maintenance personnel found that the generator was damaged, which caused this failure and required the engine to be replaced. The cause of the failure was the generator",

[0083] "result_list":[

[0084] {

[0085] "start":9,

[0086] "end":14,

[0087] "text":"The vehicle does not generate electricity"

[0088] }

[0089] ],

[0090] "prompt":"gzxx"

[0091] }

[0092] Use uie-base as the pre-trained model to fine-tune the model and save the fine-tuned model to the custom model_dir directory.

[0093] Load the pre-trained model:

[0094] model=UIE.from_pretrained("uie-base")

[0095] optimizer=paddle.optimizer.AdamW(learning_rate=learning_rate,parameters=model.parameters())

[0096] criterion = paddle.nn.BCELoss()

[0097] metric=SpanEvaluator()

[0098] Use model to load the training dataset for model training.

[0099] Load the fine-tuned model parameters to make predictions and save the prediction results.

[0100] Load the fine-tuned model:

[0101] my_ie=Taskflow("information_extraction",schema=schema,task_path=model_dir)

[0102] Use my_ie to make predictions:

[0103] results = uie_predict(my_ie,datas), where my_ie represents the trained model, datas represents the data set to be predicted, and results represents the result data after the data set is predicted.

[0104] Step S700, training and generating a second training model based on the training model learning corpus;

[0105] Specifically, the pre-trained large-scale knowledge enhancement model PaddleNLP ERNIE3.0 is trained. The pre-trained large-scale knowledge enhancement model PaddleNLP ERNIE3.0 is installed in the python environment, and the learning corpus in step 4.2.2 is converted into the data format required by the PaddleNLPERNIE3.0 dataset. The dataset is split into a training set and a validation set according to a 9:1 ratio. The training model deeply learns the relationship between artificially created keywords and common words, and intelligently recommends common words (phenomenon + cause) based on the (phenomenon + cause) keywords.

[0106] For example: Data set format: Taking phenomenon keywords-general words as an example

[0107] Text label

[0108] Battery power main switch; ablation power main switch cannot be turned off

[0109] Seat belt; Seat belt stuck and not returning to its original position

[0110] Use roberta-wwm-ext-large as the pre-trained model to fine-tune the model and save the fine-tuned model to the custom . / model_lib / model_best directory.

[0111] Load the pre-trained model:

[0112] MODEL_NAME = 'roberta-wwm-ext-large'

[0113] model=ppnlp.transformers.RobertaForSequenceClassification.from_pretrained(MODEL_NAME, num_classes=classnum)

[0114] tokenizer=ppnlp.transformers.RobertaTokenizer.from_pretrained(MODEL_NAME)

[0115] Load the fine-tuned model parameters to make predictions and save the prediction results.

[0116] Loading model parameters

[0117] my_model=ppnlp.transformers.RobertaForSequenceClassification.from_pretrained('roberta-wwm-ext-large',num_classes=classnum)

[0118] Use my_model to make predictions and get the results

[0119] results = uie_predict(my_model,datas), where my_model represents the trained model, datas represents the data set to be predicted, and results represents the result data after the data set is predicted.

[0120] Step S800, inputting the sample diagnostic data into the first training model to generate first valid diagnostic information;

[0121] Step S900, input the first valid diagnostic information into the second training model to generate the second valid diagnostic information; specifically, the non-DTC diagnostic data information is intelligently recommended, and the historical non-DTC diagnostic data is sampled in batches according to the sample sampling rules in step S300; the diagnostic data is extracted with (phenomenon + cause) keywords according to the fine-tuned information extraction model PaddleNLP UIE, that is, A=UIE_predict(D): where A represents the (phenomenon + cause) keywords extracted from the diagnostic data, and D represents the diagnostic data; the (phenomenon + cause) keywords in the diagnostic data (i.e., the first valid diagnostic information) are recommended with the fine-tuned pre-trained large-scale knowledge enhancement model PaddleNLP ERNIE3.0, that is, B=ERNIE_predict(A): D represents the (phenomenon + cause) common words, and A represents the (phenomenon + cause) keywords in the diagnostic data. Using the dual model method, the information extraction of each sample data is completed in batches.

[0122] Based on the above analysis, it can be seen that the use of the above method can not only accurately and efficiently extract effective diagnostic information from complex diagnostic data, but also ensure the consistency of quality during the information extraction process.

[0123] After the dual model outputs the first effective diagnostic information and the second effective diagnostic information, it is necessary to compare the first effective diagnostic information and the second effective diagnostic information with the manually extracted benchmark diagnostic data, so as to calibrate the first training model and the second training model. Specifically:

[0124] Training model correction: Use manually extracted (phenomenon + cause) keywords and common (phenomenon + cause) words created based on the keywords as comparison benchmarks, set a prediction result consistency threshold (the consistency threshold is set to 99%), and when the consistency between the dual model prediction results and the manual benchmark results is lower than the set threshold, perform manual benchmark data analysis, adjustment, and continuously correct the dual training model based on the comparison results until the consistency between the dual training model prediction results and the manual benchmark results reaches the set threshold standard.

[0125] In this embodiment, step S100 includes:

[0126] Step S101, based on the historical diagnosis database, extracting the diagnosis data in the historical diagnosis database;

[0127] Step S102, standardizing the names of the fault cause parts in the diagnosis data;

[0128] Step S103, identifying the fault phenomenon description content in the diagnostic data, and cleaning the preset tag type data to generate the diagnostic data.

[0129] In this embodiment, step S200 includes:

[0130] Step S201, marking the diagnostic data according to a preset DTC diagnostic data identification rule to generate DTC diagnostic data;

[0131] Step S202 , removing the DTC-type diagnostic data from the diagnostic data, and filtering out the non-DTC-type diagnostic data.

[0132] In this embodiment, step S300 includes:

[0133] Based on the non-DTC diagnostic data, the non-DTC diagnostic data is sampled according to a preset sampling rule to generate the sample diagnostic data.

[0134] In this embodiment, step S400 includes:

[0135] Step S401, extracting keyword information used to describe fault phenomena and fault causes in the sample diagnostic data to obtain fault phenomenon keywords and fault cause keywords;

[0136] Step S402, creating corresponding universal phenomenon words and universal cause words that are common across vehicle types, respectively, according to the fault phenomenon keywords and the fault cause keywords;

[0137] Step S403: Generate the reference diagnostic data based on the sample diagnostic data, the fault phenomenon keywords, the fault cause keywords, the common phenomenon words and the common cause words.

[0138] In this embodiment, step S600 includes:

[0139] Step S601, installing and calling an information extraction model, converting the training model learning corpus into a first data set in a data format required by the information extraction model, and splitting the first data set into a first training data set and a first verification data set according to a preset ratio;

[0140] Step S602, loading a first pre-training module, and loading the first training data set to perform model training, to generate a first fine-tuned model;

[0141] Step S603: load the first fine-tuned model to perform result prediction and generate a first training model.

[0142] In this embodiment, step S700 includes:

[0143] Step S701, installing and calling the knowledge enhancement model, converting the training model learning corpus into a second data set in a data format required by the knowledge enhancement model, and splitting the second data set into a second training data set and a second verification data set according to a preset ratio;

[0144] Step S702, loading a second pre-training module, and loading the second training data set to perform model training, to generate a second fine-tuned model;

[0145] Step S703: load the second fine-tuned model to perform result prediction and generate a second training model. Example

[0146] Figure 7 FIG. 1 is a schematic diagram of a structure of a dual-model based fault diagnosis information intelligent extraction device in Embodiment 2 of the present invention. Figure 7As shown, embodiment 2 provides an intelligent extraction device for dual-model fault diagnosis information, including: a first generation module 701, a second generation module 702, a third generation module 703, a fourth generation module 704, a fifth generation module 705, a first training module 706, a second training module 707, a sixth generation module 708 and a seventh generation module 709. The first generation module 701 is used to extract, standardize and clean the historical fault diagnosis data in the historical diagnostic database to generate diagnostic data based on the historical diagnostic database; the second generation module 702 is used to mark the diagnostic data and filter out non-DTC diagnostic data; the third generation module 703 is used to sample the non-DTC diagnostic data based on the non-DTC diagnostic data to generate sample diagnostic data; the fourth generation module 704 is used to extract the sample diagnostic data to generate benchmark diagnostic data; the fifth generation module 705 is used to generate a training model learning corpus based on the benchmark diagnostic data; the first training module 706 is used to train and generate a first training model based on the training model learning corpus; the second training module 707 is used to train and generate a second training model based on the training model learning corpus; the sixth generation module 708 is used to input the sample diagnostic data into the first training model to generate first valid diagnostic information; and the seventh generation module 709 is used to input the first valid diagnostic information into the second training model to generate second valid diagnostic information.

[0147] In this embodiment, the first generation module 701 includes: a first extraction unit, a processing unit, and a first generation unit. The first extraction unit is used to extract the diagnostic data in the historical diagnostic database based on the historical diagnostic database; the processing unit is used to standardize the names of the fault cause parts in the diagnostic data; and the first generation unit is used to identify the description content of the fault phenomenon in the diagnostic data, clean the preset label type data, and generate the diagnostic data.

[0148] In this embodiment, the second generation module 702 includes: a second generation unit and a third generation unit. The second generation unit is used to mark the diagnostic data according to a preset DTC diagnostic data identification rule to generate DTC diagnostic data; and the third generation unit is used to remove the DTC diagnostic data from the diagnostic data and filter out the non-DTC diagnostic data.

[0149] In this embodiment, the third generating module 703 includes: a fourth generating unit, which is used to sample the non-DTC diagnostic data based on the non-DTC diagnostic data by a preset sampling rule to generate the sample diagnostic data.

[0150] In this embodiment, the fourth generation module 704 includes: a second extraction unit, a creation unit, and a fifth generation unit. The second extraction unit is used to extract the keyword information used to describe the fault phenomenon and the fault cause in the sample diagnostic data to obtain the fault phenomenon keyword and the fault cause keyword; the creation unit is used to create corresponding universal phenomenon words and universal cause words that are common across vehicle models according to the fault phenomenon keyword and the fault cause keyword; and the fifth generation unit is used to generate the benchmark diagnostic data based on the sample diagnostic data, the fault phenomenon keyword, the fault cause keyword, the universal phenomenon word, and the universal cause word.

[0151] In this embodiment, the first training module 706 includes: a first splitting unit, a sixth generating unit, and a seventh generating unit. The first splitting unit is used to install and call the information extraction model, convert the training model learning corpus into a first data set in a data format required by the information extraction model, and split the first data set into a first training data set and a first verification data set according to a preset ratio; the sixth generating unit is used to load the first pre-training module, load the first training data set for model training, and generate a first fine-tuned model; and the seventh generating unit is used to load the first fine-tuned model for result prediction and generate a first training model.

[0152] In this embodiment, the second training module 707 includes: a second splitting unit, an eighth generating unit and a ninth generating unit. The second splitting unit is used to install and call the knowledge enhancement model, convert the training model learning corpus into a second data set in a data format required by the knowledge enhancement model, and split the second data set into a second training data set and a second verification data set according to a preset ratio; the eighth generating unit is used to load the second pre-training module, load the second training data set for model training, and generate a second fine-tuned model; and the ninth generating unit is used to load the second fine-tuned model for result prediction and generate a second training model.

[0153] The various variations and specific examples of the dual-model based intelligent extraction method for fault diagnosis information provided in Example 1 are also applicable to the dual-model based intelligent extraction device for fault diagnosis information provided in this example. Through the above detailed description of a dual-model based intelligent extraction method for fault diagnosis information, a person skilled in the art can clearly understand the implementation method of the dual-model based intelligent extraction device for fault diagnosis information in this example. Therefore, for the sake of brevity of the specification, it will not be described in detail here.

[0154] In summary, the dual-model based intelligent fault diagnosis information extraction method and device of the present invention can not only accurately and efficiently extract effective diagnostic information from complex diagnostic data, but also ensure quality consistency during the information extraction process.

[0155] Note that the above are only preferred embodiments of the present invention and the technical principles used. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and that various obvious changes, readjustments and substitutions can be made by those skilled in the art without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in more detail through the above embodiments, the present invention is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of the present invention, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. A dual-model fault diagnosis information intelligent extraction method, characterized in that: include: Based on a historical diagnosis database, extracting, standardizing, and cleaning historical fault diagnosis data in the historical diagnosis database to generate diagnostic data; Marking the diagnostic data to filter out non-DTC diagnostic data; Based on the non-DTC diagnostic data, sampling the non-DTC diagnostic data to generate sample diagnostic data; extracting the sample diagnostic data to generate benchmark diagnostic data; Based on the benchmark diagnostic data, generating a training model learning corpus; Based on the training model learning corpus, training and generating a first training model; Based on the training model learning corpus, training and generating a second training model; Inputting the sample diagnostic data into the first training model to generate first valid diagnostic information; Inputting the first valid diagnostic information into the second training model to generate second valid diagnostic information; The step of extracting sample diagnostic data to generate reference diagnostic data includes: Extracting keyword information used to describe the fault phenomenon and the fault cause from the sample diagnostic data to obtain the fault phenomenon keywords and the fault cause keywords; According to the fault phenomenon keywords and the fault cause keywords, respectively creating corresponding universal phenomenon words and universal cause words that are common across vehicle models; generating the reference diagnostic data based on the sample diagnostic data, the fault phenomenon keywords, the fault cause keywords, the common phenomenon words and the common cause words; The training and generating a first training model based on the training model learning corpus includes: Install and call the information extraction model UIE, convert the training model learning corpus into a first data set in a data format required by the information extraction model UIE, and split the first data set into a first training data set and a first verification data set according to a preset ratio; Loading a first pre-training module and loading the first training data set to perform model training to generate a first fine-tuned model; Loading the first fine-tuned model to perform result prediction and generate a first training model; The training and generating a second training model based on the training model learning corpus includes: Install and call the knowledge enhancement model PaddleNLP ERNIE3.0, convert the training model learning corpus into a second data set in the data format required by the knowledge enhancement model PaddleNLP ERNIE3.0, and split the second data set into a second training data set and a second verification data set according to a preset ratio; Loading a second pre-training module and loading the second training data set to perform model training to generate a second fine-tuned model; The second fine-tuned model is loaded to perform result prediction and generate a second training model.

2. The dual-model based fault diagnosis information intelligent extraction method according to claim 1, characterized in that: The extracting, standardizing, and cleaning the historical fault diagnosis data in the historical diagnosis database to generate the diagnosis data includes: Based on the historical diagnosis database, extracting the diagnosis data in the historical diagnosis database; Standardizing the names of the fault cause parts in the diagnostic data; The fault phenomenon description content in the diagnostic data is identified, and the preset label type data is cleaned to generate the diagnostic data.

3. The dual-model based fault diagnosis information intelligent extraction method according to claim 1, characterized in that: The marking of the diagnostic data to filter out the non-DTC diagnostic data includes: Marking the diagnostic data according to a preset DTC diagnostic data identification rule to generate DTC diagnostic data; The DTC-type diagnostic data in the diagnostic data is removed, and the non-DTC-type diagnostic data is screened out.

4. The method for intelligent extraction of fault diagnosis information based on dual models according to claim 1, characterized in that: The sampling of the non-DTC diagnostic data based on the non-DTC diagnostic data to generate sample diagnostic data includes: Based on the non-DTC diagnostic data, the non-DTC diagnostic data is sampled according to a preset sampling rule to generate the sample diagnostic data.

5. A dual-model fault diagnosis information intelligent extraction device, characterized in that: include: A first generating module is used to extract, standardize and clean the historical fault diagnosis data in the historical diagnosis database to generate diagnostic data based on the historical diagnosis database; A second generating module is used to mark the diagnostic data and filter out non-DTC diagnostic data; A third generating module, configured to sample the non-DTC diagnostic data based on the non-DTC diagnostic data by using a preset sampling rule to generate sample diagnostic data; A fourth generating module, used for extracting the sample diagnostic data to generate reference diagnostic data; A fifth generation module, used to generate a training model learning corpus based on the benchmark diagnostic data; A first training module, used for learning corpus based on the training model, and training to generate a first training model; A second training module is used to learn the corpus based on the training model and train to generate a second training model; a sixth generating module, configured to input the sample diagnostic data into the first training model to generate first valid diagnostic information; as well as a seventh generating module, configured to input the first valid diagnostic information into the second training model to generate second valid diagnostic information; The fourth generation module includes a second extraction unit, a creation unit and a fifth generation unit; the second extraction unit is used to extract keyword information for describing the fault phenomenon and the fault cause in the sample diagnostic data to obtain the fault phenomenon keyword and the fault cause keyword; the creation unit is used to create corresponding universal phenomenon words and universal cause words that are common to all vehicle models according to the fault phenomenon keyword and the fault cause keyword; the fifth generation unit is used to generate the benchmark diagnostic data based on the sample diagnostic data, the fault phenomenon keyword, the fault cause keyword, the universal phenomenon word and the universal cause word; Among them, the first training module includes a first splitting unit, a sixth generating unit and a seventh generating unit; the first splitting unit is used to install and call the information extraction model UIE, convert the training model learning corpus into a first data set in a data format required by the information extraction model UIE, and split the first data set into a first training data set and a first verification data set according to a preset ratio; the sixth generating unit is used to load the first pre-training module, and load the first training data set for model training to generate a first fine-tuned model; the seventh generating unit is used to load the first fine-tuned model for result prediction to generate a first training model; Among them, the second training module includes a second splitting unit, an eighth generation unit and a ninth generation unit; the second splitting unit is used to install and call the knowledge enhancement model PaddleNLP ERNIE3.0, convert the training model learning corpus into a second data set in the data format required by the knowledge enhancement model PaddleNLP ERNIE3.0, and split the second data set into a second training data set and a second verification data set according to a preset ratio; the eighth generation unit is used to load the second pre-training module, and load the second training data set for model training to generate a second fine-tuned model; and the ninth generation unit is used to load the second fine-tuned model for result prediction to generate a second training model.

6. The dual-model based fault diagnosis information intelligent extraction device according to claim 5, characterized in that: The first generation module comprises: A first extraction unit, configured to extract diagnostic data from the historical diagnostic database based on the historical diagnostic database; a processing unit, configured to perform standardization processing on the names of the fault cause parts in the diagnosis data; and The first generating unit is used to identify the fault phenomenon description content in the diagnostic data, clean the preset label type data, and generate the diagnostic data.

7. The dual-model based fault diagnosis information intelligent extraction device according to claim 5, characterized in that: The second generation module comprises: A second generating unit, configured to mark the diagnostic data according to a preset DTC diagnostic data identification rule, and generate DTC diagnostic data; and The third generating unit is used to remove the DTC-type diagnostic data from the diagnostic data and filter out the non-DTC-type diagnostic data.

Images