CN118035757A

CN118035757A - Electric drive assembly fault diagnosis method and device based on large language model

Info

Publication number: CN118035757A
Application number: CN202410014566.3A
Authority: CN
Inventors: 彭淦; 陈晓娇; 王晓旭; 齐嘉臣; 李凌舟; 方美娜; 杨亮
Original assignee: FAW Group Corp
Current assignee: FAW Group Corp
Priority date: 2024-01-04
Filing date: 2024-01-04
Publication date: 2024-05-14

Abstract

The application provides a large language model-based electric drive assembly fault diagnosis method and device, wherein the method comprises the following steps: inputting fault data into a distributed fault diagnosis model cluster, and performing fault diagnosis on the fault data by using each large language fault diagnosis model in the distributed fault diagnosis model cluster to obtain a plurality of fault cause result groups corresponding to the fault data; text similarity analysis is carried out on the segmented words in each fault cause result group so as to determine a target result group from a plurality of fault cause result groups; and constructing a fault diagnosis network map by utilizing the reasoning result text included in the target result group, and determining a target reason causing the fault of the electric drive assembly from the fault diagnosis network map. By the method and the device, the efficiency of fault diagnosis and reason exploration of the electric drive assembly is improved, and meanwhile, the accuracy of fault diagnosis is also improved.

Description

Electric drive assembly fault diagnosis method and device based on large language model

Technical Field

The application relates to the technical field of data processing, in particular to an electric drive assembly fault diagnosis method and device based on a large language model.

Background

The electric drive assembly is a core part of the new energy automobile, and the related state directly determines key indexes such as power performance, efficiency, safety, reliability and the like of the whole automobile. The diagnosis of the faulty electric drive assembly can help to find and solve related problems in time and avoid fault expansion. The accurate fault diagnosis can rapidly locate fault reasons and formulate an effective evading method and a maintenance strategy, so that the time, manpower and other resource costs are reduced.

At present, the traditional fault diagnosis method based on mechanism needs to write complex rules or logic judgment, and has high customization degree and limited scene coverage. This means that the rules are often not generic for different fault situations and cannot cope with some unaccounted fault situations. The fault diagnosis method based on the machine learning and deep learning model needs to limit strict working conditions, and performs a large amount of pretreatment and cleaning work on input data to eliminate the influence of noise and abnormal values. The algorithm complexity and the labor consumption index of the two are increased, the maintenance cost is greatly increased, and the fault association factors cannot be intuitively discovered. Therefore, how to increase the fault diagnosis speed of the electric drive assembly is a not-to-small technical problem.

Disclosure of Invention

Therefore, the application aims to provide the fault diagnosis method and the fault diagnosis device for the electric drive assembly based on the large language model, which solve the problems of low data characteristic utilization efficiency and poor model heuristic capability in the prior art, can mine relevant factors related to faults, improve the fault diagnosis and the cause exploration efficiency of the electric drive assembly, and improve the fault diagnosis accuracy.

In a first aspect, an embodiment of the present application provides a method for diagnosing a failure of an electric drive assembly based on a large language model, where the method for diagnosing a failure of an electric drive assembly includes:

Acquiring fault data of an electric drive assembly, inputting the fault data into a distributed fault diagnosis model cluster, and performing fault diagnosis on the fault data by using each large language fault diagnosis model in the distributed fault diagnosis model cluster to obtain a plurality of fault cause result groups corresponding to the fault data;

Extracting text features of each fault cause result group to obtain segmentation words included in each fault cause result group;

Text similarity analysis is carried out on the segmented words in each fault cause result group so as to determine a target result group from a plurality of fault cause result groups;

And constructing a fault diagnosis network map by utilizing the reasoning result text included in the target result group, and determining a target reason causing the fault of the electric drive assembly from the fault diagnosis network map.

Further, the distributed fault diagnosis model cluster is obtained through the following steps:

acquiring a pre-constructed training set corpus, and preprocessing data in the training set corpus to obtain an input corpus;

Generating a set of multi-sample labels using a plurality of different random seeds based on the input corpus;

Inputting each marking set into a large language fault diagnosis original model aiming at each marking set, and training the large language fault diagnosis original model to obtain the large language fault diagnosis model;

and obtaining the distributed fault diagnosis model cluster based on a plurality of large language fault diagnosis models.

Further, after the large language fault diagnosis model is obtained, the electrical drive assembly fault diagnosis method further includes:

And carrying out model evaluation on the large language fault diagnosis model based on a preset prompt word template and a preset test sample, when the evaluation result of the large language fault diagnosis model is judged not to meet the preset evaluation standard, adjusting model parameters of the large language fault diagnosis model, and returning to execute the step of inputting the mark set into the large language fault diagnosis original model to train the large language fault diagnosis original model until the evaluation result of the large language fault diagnosis model meets the preset evaluation standard.

Further, the text similarity analysis is performed on the word segmentation in each fault cause result packet, so as to determine a target result packet from a plurality of fault cause result packets, including:

For each word in each fault cause result group, calculating word frequency and inverse text frequency of the word, and calculating text similarity of the word by using the word frequency and the inverse text frequency;

summing the text similarity of the plurality of segmented words to obtain the sum of the text similarity corresponding to the fault cause result grouping, and calculating the text vector corresponding to each segmented word by utilizing the sum of the text similarity, the text similarity of each segmented word and the average vector of each segmented word;

And screening the target result groups from a plurality of fault reason result groups by using the text vector of each word in each fault reason result group.

Further, the screening the target result packet from the plurality of fault cause result packets by using the text vector of each word in each fault cause result packet includes:

Aiming at any two segmented words in each fault cause result group, calculating a similarity evaluation value between the two segmented words by using the text vector of each segmented word in the two segmented words;

determining the number of similar texts in the fault cause result group based on the number of similar evaluation values larger than a preset evaluation threshold value in the plurality of similar evaluation values;

And when the number of the similar texts is larger than or equal to a preset threshold value, grouping the fault cause results as the target result group.

Further, the constructing a fault diagnosis network map by using the reasoning result text included in the target result packet includes:

determining an entity relation group contained in each reasoning result text based on a preset entity relation rule aiming at each reasoning result text in the target result group;

Aiming at each entity relation group, calculating a confidence coefficient corresponding to the entity relation group by utilizing the text similarity corresponding to each entity in the entity relation group and the identification sequence number of the large language fault diagnosis model for predicting the target result group;

Constructing a fault diagnosis network map corresponding to the target result group by utilizing a target entity relation group with a trust coefficient larger than a preset threshold value in a plurality of entity relation groups; the fault diagnosis network map comprises a plurality of nodes, and each node represents a fault source in the target entity relation group.

Further, the determining, from the fault diagnosis network map, a target cause causing the fault of the electric drive assembly includes:

calculating the degree centrality corresponding to each node in the fault diagnosis network map, and taking a fault source represented by the node with the highest degree centrality among a plurality of nodes as a target cause for causing the fault of the electric drive assembly.

In a second aspect, an embodiment of the present application further provides an electrical drive assembly fault diagnosis device based on a large language model, where the electrical drive assembly fault diagnosis device includes:

The reasoning result grouping determination module is used for acquiring fault data of the electric drive assembly, inputting the fault data into a distributed fault diagnosis model cluster, and performing fault diagnosis on the fault data by using each large language fault diagnosis model in the distributed fault diagnosis model cluster to obtain a plurality of fault cause result groupings corresponding to the fault data;

The word segmentation extraction module is used for extracting text characteristics of each fault cause result group to obtain words contained in each fault cause result group;

The target result grouping determining module is used for carrying out text similarity analysis on the word segmentation in each fault cause result grouping so as to determine a target result grouping from a plurality of fault cause result groupings;

And the target cause determining module is used for constructing a fault diagnosis network map by utilizing the reasoning result text included in the target result packet and determining a target cause causing the fault of the electric drive assembly from the fault diagnosis network map.

In a third aspect, an embodiment of the present application further provides an electronic device, including: the system comprises a processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor, the processor and the memory are communicated through the bus when the electronic device runs, and the machine-readable instructions are executed by the processor to execute the steps of the large language model-based electric drive assembly fault diagnosis method.

In a fourth aspect, embodiments of the present application also provide a computer readable storage medium having a computer program stored thereon, which when executed by a processor performs the steps of the above-described large language model-based electrical drive assembly fault diagnosis method.

The application utilizes a distributed large language model to carry out fault diagnosis of an electric drive assembly, screens out a target result group from a plurality of fault cause result groups, then constructs a graph network of the fault, namely a fault diagnosis map, based on an inference result text included in the target result group, analyzes the fault diagnosis map, and outputs a target cause related to the fault. The problems of low data characteristic utilization efficiency and poor model heuristic capability in the prior art are solved, related factors related to faults can be mined, the efficiency of fault diagnosis and cause exploration of the electric drive assembly is improved, and meanwhile, the accuracy of fault diagnosis is also improved.

In order to make the above objects, features and advantages of the present application more comprehensible, preferred embodiments accompanied with figures are described in detail below.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings that are needed in the embodiments will be briefly described below, it being understood that the following drawings only illustrate some embodiments of the present application and therefore should not be considered as limiting the scope, and other related drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.

FIG. 1 is a flow chart of a method for diagnosing faults of an electric drive assembly based on a large language model according to an embodiment of the present application;

FIG. 2 is a schematic diagram of an electrical drive assembly fault diagnosis device based on a large language model according to an embodiment of the present application;

FIG. 3 is a second schematic structural diagram of an electrical drive assembly fault diagnosis device based on a large language model according to an embodiment of the present application;

fig. 4 is a schematic structural diagram of an electronic device according to an embodiment of the present application.

Detailed Description

For the purpose of making the objects, technical solutions and advantages of the embodiments of the present application more apparent, the technical solutions of the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application, and it is apparent that the described embodiments are only some embodiments of the present application, not all embodiments. The components of the embodiments of the present application generally described and illustrated in the figures herein may be arranged and designed in a wide variety of different configurations. Thus, the following detailed description of the embodiments of the application, as presented in the figures, is not intended to limit the scope of the application, as claimed, but is merely representative of selected embodiments of the application. Based on the embodiments of the present application, every other embodiment obtained by a person skilled in the art without making any inventive effort falls within the scope of protection of the present application.

First, an application scenario to which the present application is applicable will be described. The application can be applied to the technical field of data processing.

Meanwhile, through deep analysis of faults, the design of the electric drive system can be continuously optimized and improved, the service life and maintenance efficiency of the whole vehicle can be improved, and the safety and reliability of the whole vehicle can be improved. The electric drive assembly is generally composed of three parts, namely a motor, a motor control unit and a speed reducer. In recent years, with the advancement of new energy automobile technology, the electric drive assembly also shows higher integration level and complexity, such as an all-in-one integrated electric drive system, and other components except three parts, such as a PDU (power distribution unit), a DC-DC (direct current-direct current converter), an OBC (on-board charger) and the like, are deeply integrated, so that the failure rate of the all-in-one system is also improved, and the failure diagnosis faces more complex coupling judgment.

According to research, the traditional fault diagnosis method based on the mechanism needs to write complex rules or logic judgment, and has high customization degree and limited scene coverage. This means that the rules are often not generic for different fault situations and cannot cope with some unaccounted fault situations. In addition, this method relies on expert experience, and once expert knowledge is lacking, misdiagnosis is caused, and problems are difficult to solve. The fault diagnosis method based on the machine learning and deep learning model needs to limit strict working conditions, and performs a large amount of pretreatment and cleaning work on input data to eliminate the influence of noise and abnormal values. And with the complexity of the functional characteristics, the algorithm complexity and the labor consumption index of the two are increased, so that the maintenance cost is greatly increased, and the fault association factors cannot be intuitively discovered. Therefore, how to increase the fault diagnosis speed of the electric drive assembly is a not-to-small technical problem.

Based on the above, the embodiment of the application provides a large language model-based fault diagnosis method for an electric drive assembly, which improves the efficiency of fault diagnosis and cause exploration of the electric drive assembly and improves the accuracy of fault diagnosis.

Referring to fig. 1, fig. 1 is a flowchart of a fault diagnosis method for an electric drive assembly based on a large language model according to an embodiment of the present application. As shown in fig. 1, the fault diagnosis method for an electric drive assembly provided by the embodiment of the application includes:

S101, acquiring fault data of an electric drive assembly, inputting the fault data into a distributed fault diagnosis model cluster, and performing fault diagnosis on the fault data by using each large language fault diagnosis model in the distributed fault diagnosis model cluster to obtain a plurality of fault cause result groups corresponding to the fault data.

Here, the fault data may include a textual description of a fault data packet, a fault phenomenon, and a fault part location of the electric drive system assembly. Specifically, the fault data packet is a data segment with a certain time length, which is obtained by splicing, aligning, preprocessing and cleaning abnormal values of the returned data, and the fault sign position of the assembly. The fault phenomenon is a phenomenon generated when the electric drive assembly breaks down. The fault location is the location in the electric drive assembly where the fault occurs. The distributed fault diagnosis model cluster refers to a distributed diagnosis cluster composed of a plurality of large language fault diagnosis models. The large language fault diagnosis model is used for carrying out fault diagnosis on fault data. The fault cause result group comprises at least one fault diagnosis result.

For the above step S101, in the implementation, in the fault diagnosis stage, the fault data of the electric drive system assembly, such as the fault data packet, the fault phenomenon, and the text description of the fault part position, are input into the distributed fault diagnosis model cluster. In implementation, the feature distribution condition of the incoming fault data is calculated through the calculation cluster statistics, a fault moment data frame is obtained, the incoming text content is standardized based on a prompt word template, and all information is formed into an input distributed fault diagnosis model cluster. After the distributed fault diagnosis model cluster inputs the fault data, the distributed fault diagnosis model cluster comprises a plurality of large language fault diagnosis models, the fault data are input into each large language fault diagnosis model, fault diagnosis is carried out on the fault data, and a plurality of fault cause result groups corresponding to the fault data are obtained.

As an alternative embodiment, for the above step S101, the distributed fault diagnosis model cluster is obtained by the following steps:

Step 1011, obtaining a pre-built training set corpus, and preprocessing data in the training set corpus to obtain an input corpus.

Here, the corpus of the training set is derived from the information tuple of the failure working condition library of the electric drive assembly, the characteristic form and the logic document of the electric drive assembly, and the SQL code block of the mathematical statistics and Python. The tuple in the fault condition database of the electric drive assembly comprises a fault data packet, a fault moment data frame, data characteristic distribution, fault types and expert diagnosis fault reasons; and the electric drive system assembly returns the characteristic fields such as the sampling amount of each sensor, the calculated amount of the controller and the like to the cloud server in a certain frequency and time range. The fault data packet is a data segment with a certain time length, which is formed by splicing, aligning, preprocessing and cleaning abnormal values of the returned data, and the fault mark position bit of the assembly. The fault moment data frame is the data line at the moment when the jump state of the extraction zone bit is set. The data characteristic distribution is the distribution condition of each characteristic field of the electric drive system assembly in fault data fragments, namely the whole range, size and fluctuation trend, including the number of effective values, the number of different values, standard deviation, extremum, mean value and quartile. The fault category is the fault name which is corresponding to the fault code analysis uploaded in the fault range. Expert diagnosis is based on the textual description of the cause of the failure, the phenomenon associated with the failure and the location of the failed part by an expert group or engineer based on data and their own experience. The electric drive assembly characteristic form and the logic document comprise an analysis form, a field semantic matching description form and a system functional characteristic logic form corresponding to the electric drive assembly fault code. The analysis table of the fault codes of the electric drive assembly is used for analyzing the corresponding fault codes so as to distinguish fault types. The field semantic matching description table contains a corresponding description of each feature field in the data, a physical value range, transmission accuracy and semantic interpretation. The system functional characteristic logic form is a description of fault handling, control functional logic, technical standards and performance characteristics in the assembly. The mathematical statistics characteristic SQL statement and the Python code block are some preset data inquiry, range judgment, mechanism logic and machine learning codes are used for task planning and code generation of a heuristic model.

For the step 1011, when in implementation, a pre-built training set corpus is first obtained, and data in the training set corpus is preprocessed. Here, preprocessing includes text extraction of document classes, natural language paragraph filtering and deduplication. Removing abnormal values according to the document length, the ratio of symbols to words and the failure and mechanism association rate standard during filtering, and then filtering irrelevant corpus by using a linear correction filter; and during de-duplication, checking and processing the text fragments with fuzzy matching and character sequence repetition rate. Finally, generating training corpuses in batches in json mode and simultaneously standardizing to form an input corpus in a pre-training stage.

Step 1012, generating a set of multi-sample labels using a plurality of different random seeds based on the input corpus.

For the above step 1012, in implementation, the input corpus is input in a data stream, and multiple sample sets of markers are generated with different random seeds prior to training. Further, after the mark sets are generated, corpus construction and prompt word vocabulary correctness in each mark set are checked, and mark sets with poor quality, low standardization and large construction error deviation are deleted.

Step 1013, for each markup set, inputting the markup set into a large language fault diagnosis original model, and training the large language fault diagnosis original model to obtain the large language fault diagnosis model.

For the step 1013, in implementation, for each generated tag set, the tag set is sampled and sliced and then trained in the original large language fault diagnosis model to obtain the large language fault diagnosis model. Here, the training method for the large language model is described in detail in the prior art, and will not be described herein again.

As an alternative embodiment, for the step 1013, after obtaining the large language fault diagnosis model, the electrical drive assembly fault diagnosis method further includes:

For the above steps, in the specific implementation, after training in step 1013 to obtain a large language fault diagnosis model, model evaluation is also required for the trained model. Different preset prompt word templates and preset test sample cases are input into a large language fault diagnosis model for model evaluation. When it is determined that the evaluation result of the large language fault diagnosis model does not meet the preset evaluation standard, the model parameters of the large language fault diagnosis model are adjusted, and the step of inputting the set of marks into the large language fault diagnosis original model and training the large language fault diagnosis original model in step 1013 is performed, until the evaluation result of the large language fault diagnosis model meets the preset evaluation standard.

Step 1014, obtaining the distributed fault diagnosis model cluster based on a plurality of the large language fault diagnosis models.

For the above step 1014, when it is specifically implemented, after training to obtain a plurality of large language fault diagnosis models, each large language fault diagnosis model is deployed into a production application cluster to obtain a distributed fault diagnosis model cluster.

Thus, according to the steps 1011-1014, the input corpus of model training is generated by multi-language combination in the training stage, meanwhile, multiple large language models are generated by extracting different sample sets for training to form a distributed diagnosis cluster, multiple fault scenes can be covered and detected, multiple fault categories are processed and identified, the model has high interpretation, and the fault diagnosis time is short.

S102, extracting text features of each fault cause result group to obtain segmentation included in each fault cause result group.

For the above step S102, in implementation, text feature extraction is performed on each failure cause result packet to obtain the segmentation included in each failure cause result packet. In the specific implementation, in the training stage, a private library document set is constructed, word vectors are generated by word segmentation processing based on an electric drive assembly characteristic form and a logic document in an input corpus of the training stage, namely an analysis form, a field semantic matching description form and a system functional characteristic logic form corresponding to the electric drive assembly fault code, after stop words are removed, filtering and screening are carried out, only key electric drive assembly characteristics and professional words are reserved, parts of speech are marked, a custom dictionary is formed, so that the word segmentation requirement of text consistency judgment is better adapted, and the word segmentation accuracy is improved. In the operation stage, after receiving the fault cause result grouping, firstly loading a custom dictionary, marking and reserving host numbers of texts from clusters in the fault cause result grouping, and then performing word segmentation and preprocessing on the texts in the fault cause result grouping to obtain word segmentation included in the fault cause result grouping. Here, the word segmentation and preprocessing operations include removing stop words, punctuation marks, and other irrelevant characters, while reducing word variants by performing word drying or word shape reduction on a class of words.

S103, performing text similarity analysis on the segmented words in each fault cause result group so as to determine a target result group from a plurality of fault cause result groups.

Here, text similarity analysis is a technique of judging the relationship between texts by calculating the similarity between them. The higher the similarity, the more similar the two texts are. Common text similarity analysis algorithms include cosine similarity, pearson correlation coefficient, jaccard similarity, and the like.

Here, according to the embodiment provided by the present application, TF-IDF method similarity analysis is used. TF-IDF (Term Frequency-Inverse Document Frequency) is a statistical method for performing keyword extraction to evaluate the importance of a word to one of the documents in a corpus or corpus. The importance of a word increases proportionally with the number of times it appears in the file, but at the same time decreases inversely with the frequency with which it appears in the corpus. TF (Term Frequency) refers to the Frequency with which a given word appears in text. IDF (Inverse Document Frequency, inverse text frequency) is a measure of the general importance of a word, the less text that contains a word, the greater the IDF.

For the above step S103, in implementation, text similarity analysis is performed on the word in each failure cause result group to determine a target result group from among the plurality of failure cause result groups.

As an optional embodiment, for the step S103, the text similarity analysis is performed on the word segmentation in each fault cause result packet to determine a target result packet from a plurality of fault cause result packets, where the method includes:

Step 1031, for each word segment in each failure cause result group, calculating word frequency and inverse text frequency of the word segment, and calculating text similarity of the word segment by using the word frequency and the inverse text frequency.

For the step 1031, in implementation, for each word segment in each failure cause result packet, the word frequency and the inverse text frequency of the word segment are calculated, and the text similarity of the word segment is calculated by using the word frequency and the inverse text frequency of the word segment.

Specifically, if the failure cause result group is T _i, there is (w ₁,w₂,…,w_n)∈T_i, the word frequency TF of the word segmentation is calculated by the following formula:

Where w _i is the ith word in the failure cause result packet T _i, N is the total number of words in the failure cause result packet, The number of occurrences of the word w _i in the failure cause result packet T _i.

The inverse text frequency of the segmentation word w _i is calculated by the following formula:

Where D is the total number of failure cause result packets, Total group number grouped for failure cause result containing the word w _i.

After the word frequency and the inverse text frequency of the word are determined, the word frequency is multiplied by the inverse text frequency to obtain the text similarity FT-IDF (w _i), namely TF-IDF (w _i)＝TF(w_i)×IDF(w_i) of the word w _i.

Step 1032, summing the text similarity of the plurality of segmented words to obtain a sum of the text similarity corresponding to the fault cause result group, and calculating the text vector corresponding to each segmented word by using the sum of the text similarity, the text similarity of each segmented word and the average vector of each segmented word.

For the above step 1032, when the method is specifically implemented, after obtaining the text similarity of each word segment in the fault cause result group, the text similarity of the plurality of word segments is summed to obtain the sum of the text similarities corresponding to the fault cause result group. Then, a text vector corresponding to each word segment is calculated by using the sum of the text similarity, the text similarity of each word segment and the average vector of each word segment. Here, the text vector representation method of each word is as follows: firstly, directly obtaining the average vector of each Word by using Word2Vec methodAnd then using the text similarity TF-IDF of each word segment to assign weight to each word segment, and calculating a weighted average vector. I.e., for each Word segment, its Word2Vec vector is multiplied by its TF-IDF value, and then weighted averaged. Specifically, the text vector is calculated by the following formula:

Wherein, sigma TF-IDF (w) is the sum of TF-IDF of all the words in the failure cause result group.

And 1033, screening the target result groups from the plurality of fault reason result groups by using the text vector of each word in each fault reason result group.

For the above step 1033, in implementation, the text vector of each word in each failure cause result packet is used to screen out the target result packet from the plurality of failure cause result packets.

As an optional embodiment, for step 1033, the selecting the target result packet from the plurality of failure cause result packets by using the text vector of each word in each failure cause result packet includes:

Step 10331, for any two tokens in each failure cause result group, calculates a similarity evaluation value between the two tokens using the text vector of each of the two tokens.

With respect to the above step 10331, in the specific implementation, for any two of the tokens in each failure cause result group, a similarity evaluation value between the two tokens is calculated using the text vector of each of the two tokens. Here, specifically, for the text vectors T _i and T _j of any two segmented words in the failure cause result group, cosine similarity is used to calculate the similarity evaluation value, that is, calculate the included angle of the text vectors, and the smaller the included angle is, the higher the similarity is. Specifically, a similarity evaluation value between two segmentation words is calculated by the following formula:

Wherein, The value of (2) is the similarity evaluation value.

Step 10332, determining the number of similar texts in the failure cause result packet based on the number of similar evaluation values greater than a preset evaluation threshold value in the plurality of similar evaluation values.

With respect to step 10332, in implementation, the number of similar texts in the failure cause result packet is determined based on the number of similar evaluation values greater than the preset evaluation threshold value among the plurality of similar evaluation values. Here, specifically, if the similarity value of the text vectors T _i and T _j is denoted as v _i,j, and the similarity value v _i,j is specified to be greater than the preset evaluation threshold α, the similarity Wen Benshu is incremented by 1.

And step 10333, when the number of the similar texts is larger than or equal to a preset threshold value, grouping the fault cause results as the target results.

Here, the preset threshold may be half the number of all the failure cause result packets, and the present application is not particularly limited.

For the above step 10333, in a specific implementation, when it is determined that the number of similar texts is greater than or equal to the preset threshold, the failure cause result group is taken as the target result group.

Here, according to the steps 1031 to 1033, the results of the fault cause returned by the distributed large language model cluster are collected in groups, and text feature extraction and TF-IDF similarity analysis are performed to make a plurality of text consistency diagnoses. Outputting most consistent answers when the returned answers reach most consistent, otherwise, packaging all the text answers and delivering the text answers to an expert system for further manual diagnosis and judgment.

S104, constructing a fault diagnosis network map by utilizing the reasoning result text included in the target result group, and determining a target reason causing the fault of the electric drive assembly from the fault diagnosis network map.

For the step S104, in the implementation, for the target result group, a fault diagnosis network map of the fault is constructed by using the reasoning result text included in the target result group, and then the fault characteristics of the correlation influence are analyzed and searched by using a centrality algorithm for the map, so as to output the target cause of the fault of the electric drive assembly. In this way, the fault diagnosis network map is automatically constructed by utilizing the reasoning result of the cluster, the influence degree of the fault related factors can be analyzed by analyzing the relation between the nodes and the edges in the network map, the reliability is high, deep fault association mining is provided, and the understanding of the relation and the cause of the fault and the common characteristics and modes of the fault are facilitated.

As an optional embodiment, for the step S104, the constructing a fault diagnosis network map by using the reasoning result text included in the target result packet includes:

Step 1041, for each inference result text in the target result packet, determining an entity relationship group contained in the inference result text based on a preset entity relationship rule.

Here, as an example, the entity relationship rule may include "rule 1, X is caused by Y", "rule two, the reason for X is Y", and the like, which is not particularly limited to the present application. And forming an entity relation group by using the fault source and the association reason, wherein X and Y are the entities in the entity relation group.

For the step 1041, in the implementation, for each inference result text in the target result group, a structured text is extracted, and a relationship extraction is performed based on a preset entity relationship rule, so as to identify an entity relationship group included in the inference result text.

Here, after determining the entity relationship group included in the inference result text, a relationship extraction table is formed corresponding to each inference result text, where the relationship extraction table includes a fault source, an association reason, a relationship identifier, and an association coefficient. The relationship identification is the identification serial number of the large language fault diagnosis model of the reasoning result source, and the association coefficient corresponds to the TF-IDF value corresponding to the entity.

Step 1042, for each entity relationship group, calculating the confidence coefficient corresponding to the entity relationship group by using the text similarity corresponding to each entity in the entity relationship group and the identification sequence number of the large language fault diagnosis model for predicting the target result group.

Confidence coefficient calculation is performed for each entity relationship group in step 1041. In the implementation, for each entity relation group, calculating a confidence coefficient corresponding to the entity relation group by utilizing the text similarity corresponding to each entity in the entity relation group and the identification sequence number of the large language fault diagnosis model for predicting the target result group. Specifically, the confidence coefficient corresponding to the entity relation group is calculated through the following steps:

Where A and B represent entities in the entity-relationship group. f _A,B represents the confidence coefficient between A and B, D _i represents the relation identification of the reasoning result under the large language fault diagnosis model with the identification number of i, and lambda _A,B is the association coefficient.

Step 1043, constructing a fault diagnosis network map corresponding to the target result packet by using a plurality of target entity relationship groups with confidence coefficients greater than a preset threshold in the entity relationship groups.

For the above step 1043, in the implementation, a confidence coefficient corresponding to each entity relationship group is determined, and the entity relationship group with the confidence coefficient greater than the preset threshold value in the plurality of entity relationship groups is used as the target entity relationship group. And constructing a fault diagnosis network map corresponding to the target result group by utilizing the target entity relation group. Here, the fault diagnosis network map includes a plurality of nodes, each of which characterizes a fault source in the target entity relationship group.

For the above step S104, in implementation, the determining, from the fault diagnosis network map, the target cause causing the fault of the electric drive assembly includes:

For the steps, when the method is specifically implemented, fault characteristics of relevant influences are analyzed and searched for by using a centrality algorithm on the graph. Specifically, after the fault diagnosis network map is constructed, calculating the degree centrality corresponding to each node in the fault diagnosis network map. Here, the node's ingress degree is used as the degree-centering degree, with respect to node O _i Wherein the method comprises the steps ofIs the weight on the vertex O _i and other vertex edges. Among the plurality of nodes, the node with the highest degree center represents that the fault source of the type has the greatest influence on the whole fault map and is the most important, and the node is the true cause of the fault of the current electric drive assembly possibly caused directly or indirectly. Therefore, a fault source represented by the node with the highest degree of center among the plurality of nodes is used as a target cause for causing the fault of the electric drive assembly.

The application is based on the information tuple pre-training model of multi-language combination, fully utilizes the characteristics of historical fault data, and has wide model identification coverage and strong interpretation. In the reasoning stage, the large language model firstly makes task planning for the auxiliary computing cluster to further explore data according to the input information tuple reasoning, and then combines context correlation reasoning, so that the degree of multiplexing of the developed mechanism and the artificial intelligent model is high, and the method is compatible with possible future expansion. The method has the advantages that the inference information is collected on the basis of the distributed large language model clusters to conduct text consistency inference, the possible interference and abnormal influence of single model training and inference processes can be reduced, the adaptability of the models in different scenes is enhanced, the models are helped to better cope with noise and abnormal conditions, and the accuracy of diagnosis results is improved. The consistency judgment can also be used for forming an index basis for evaluating the model effect, identifying the model with abnormal performance and guiding the adjustment direction of the model. The fault diagnosis network map is constructed based on the cluster reasoning result, so that relevant factors of fault association can be deeply mined, the influence degree among faults is intuitively displayed, the fault occurrence mechanism and the propagation rule are clear, and the fault diagnosis speed is increased.

As an optional embodiment, the electrical drive assembly fault diagnosis method provided by the application corresponds to an electrical drive assembly fault diagnosis system based on a large language model, and the system is divided into four layers of a data storage layer, a model layer, a calculation layer and a decision layer according to the type and the function distinction of processed data. The data storage layer is a data source input, storage and management layer and contains a data source to be processed (fault data packet group), a text information source (text description corresponding to the phenomenon of faults, the position of fault parts and the like), an algorithm and model storage library, which is used for storing and managing the versions of diagnostic models and large language models for calculation, and an electric drive assembly fault working condition library contains information tuples after fault marking is finished and filed. At the model application layer, the input information tuples are distributed to all hosts of the large language model cluster, the distributed cluster is responsible for connecting the hosts in a offline mode or responding to overtime before inputting information, and the hosts which do not react and execute tasks for more than preset time are removed when the reasoning results are collected to form groups. Meanwhile, the distributed cluster is responsible for retrying repairing unresponsive hosts when no task is arranged, checking the model version corresponding conditions of the hosts before task execution and reporting error corresponding information in time. At the computing application layer, the computing cluster is responsible for fetching task execution from the message queue and returning the result after execution. The calculation content comprises data feature distribution calculation and fault frame extraction before input information tuples are formed, data features/data source exploration and mining tasks of each large language model plan, feature extraction and similarity calculation of texts after reasoning results are received. And in the decision layer, after receiving the similarity calculation result, judging most of the consistency, and judging the grouping which does not meet the consistency by submitting the grouping to an expert diagnosis system. And constructing a fault diagnosis map for the group meeting the consistency standard, analyzing the fault relevance, then aggregating the result, and submitting the diagnosis report and the reasoning result to a data storage layer for marking and archiving.

Referring to fig. 2 and 3, fig. 2 is a schematic structural diagram of an electric drive assembly fault diagnosis device based on a large language model according to an embodiment of the present application, and fig. 3 is a second schematic structural diagram of an electric drive assembly fault diagnosis device based on a large language model according to an embodiment of the present application. As shown in fig. 2, the electric drive assembly fault diagnosis apparatus 200 includes:

The reasoning result grouping determination module 201 is configured to obtain fault data of an electric drive assembly, input the fault data into a distributed fault diagnosis model cluster, and perform fault diagnosis on the fault data by using each large language fault diagnosis model in the distributed fault diagnosis model cluster to obtain a plurality of fault cause result groupings corresponding to the fault data;

The word segmentation extraction module 202 is configured to perform text feature extraction on each fault cause result packet to obtain a word segment included in each fault cause result packet;

A target result grouping determining module 203, configured to perform text similarity analysis on the word segmentation in each fault cause result grouping, so as to determine a target result grouping from a plurality of fault cause result groupings;

the target cause determining module 204 is configured to construct a fault diagnosis network map by using the inference result text included in the target result packet, and determine a target cause causing the fault of the electric drive assembly from the fault diagnosis network map.

Further, as shown in fig. 3, the electrical drive assembly fault diagnosis apparatus 200 further includes a model training module 205, where the model training module 205 is configured to obtain the distributed fault diagnosis model cluster by:

Further, after obtaining the large language fault diagnosis model, the model training module 205 is further configured to:

Further, when the target result packet determining module 203 is configured to perform text similarity analysis on the word in each fault cause result packet to determine a target result packet from a plurality of fault cause result packets, the target result packet determining module 203 is further configured to:

Further, when the target result packet determining module 203 is configured to screen the target result packet from the plurality of failure cause result packets by using the text vector of each word in each failure cause result packet, the target result packet determining module 203 is further configured to:

Further, the target cause determining module 204 is further configured to, when configured to construct a fault diagnosis network map using the inference result text included in the target result packet, the target cause determining module 204 is further configured to:

Further, when the target cause determining module 204 is configured to determine a target cause causing the fault of the electric drive assembly from the fault diagnosis network map, the target cause determining module 204 is further configured to:

Referring to fig. 4, fig. 4 is a schematic structural diagram of an electronic device according to an embodiment of the application. As shown in fig. 4, the electronic device 400 includes a processor 410, a memory 420, and a bus 430.

The memory 420 stores machine-readable instructions executable by the processor 410, when the electronic device 400 is running, the processor 410 communicates with the memory 420 through the bus 430, and when the machine-readable instructions are executed by the processor 410, the steps of the electrical drive assembly fault diagnosis method based on the large language model in the method embodiment shown in fig. 1 can be executed, and specific implementation can be referred to method embodiments, which are not repeated herein.

The embodiment of the present application further provides a computer readable storage medium, where a computer program is stored on the computer readable storage medium, and when the computer program is executed by a processor, the steps of the method for diagnosing faults of an electric drive assembly based on a large language model in the method embodiment shown in fig. 1 can be executed, and a specific implementation manner can refer to the method embodiment and is not repeated herein.

It will be clear to those skilled in the art that, for convenience and brevity of description, specific working procedures of the above-described systems, apparatuses and units may refer to corresponding procedures in the foregoing method embodiments, and are not repeated herein.

In the several embodiments provided by the present application, it should be understood that the disclosed systems, devices, and methods may be implemented in other manners. The above-described apparatus embodiments are merely illustrative, for example, the division of the units is merely a logical function division, and there may be other manners of division in actual implementation, and for example, multiple units or components may be combined or integrated into another system, or some features may be omitted, or not performed. Alternatively, the coupling or direct coupling or communication connection shown or discussed with each other may be through some communication interface, device or unit indirect coupling or communication connection, which may be in electrical, mechanical or other form.

The units described as separate units may or may not be physically separate, and units shown as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

In addition, each functional unit in the embodiments of the present application may be integrated in one processing unit, or each unit may exist alone physically, or two or more units may be integrated in one unit.

The functions, if implemented in the form of software functional units and sold or used as a stand-alone product, may be stored in a non-volatile computer readable storage medium executable by a processor. Based on this understanding, the technical solution of the present application may be embodied essentially or in a part contributing to the prior art or in a part of the technical solution, in the form of a software product stored in a storage medium, comprising several instructions for causing a computer device (which may be a personal computer, a server, a network device, etc.) to perform all or part of the steps of the method according to the embodiments of the present application. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (Random Access Memory, RAM), a magnetic disk, or an optical disk, or other various media capable of storing program codes.

Finally, it should be noted that: the above examples are only specific embodiments of the present application, and are not intended to limit the scope of the present application, but it should be understood by those skilled in the art that the present application is not limited thereto, and that the present application is described in detail with reference to the foregoing examples: any person skilled in the art may modify or easily conceive of the technical solution described in the foregoing embodiments, or perform equivalent substitution of some of the technical features, while remaining within the technical scope of the present disclosure; such modifications, changes or substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present application, and are intended to be included in the scope of the present application. Therefore, the protection scope of the application is subject to the protection scope of the claims.

Claims

1. The electric drive assembly fault diagnosis method based on the large language model is characterized by comprising the following steps of:

2. The electrical drive assembly fault diagnosis method of claim 1, wherein the distributed fault diagnosis model cluster is obtained by:

3. The electrical drive assembly fault diagnosis method according to claim 2, wherein after obtaining the large language fault diagnosis model, the electrical drive assembly fault diagnosis method further comprises:

4. The method of claim 1, wherein the performing text similarity analysis on the word segment in each failure cause result group to determine a target result group from a plurality of failure cause result groups comprises:

5. The method of claim 4, wherein the screening the target result group from the plurality of failure cause result groups using the text vector of each word in each failure cause result group comprises:

6. The method of claim 4, wherein constructing a fault diagnosis network map using the inference result text included in the target result packet comprises:

7. The method of claim 6, wherein determining a target cause of the failure of the electric drive assembly from the failure diagnosis network map comprises:

8. An electric drive assembly fault diagnosis device based on a large language model, which is characterized by comprising:

9. An electronic device, comprising: a processor, a memory and a bus, the memory storing machine readable instructions executable by the processor, the processor and the memory in communication via the bus when the electronic device is running, the machine readable instructions when executed by the processor performing the steps of the large language model based electro-drive assembly fault diagnosis method of any one of claims 1 to 7.

10. A computer readable storage medium, characterized in that the computer readable storage medium has stored thereon a computer program which, when executed by a processor, performs the steps of the large language model based electro-drive assembly fault diagnosis method according to any one of claims 1 to 7.