Failure mode data analysis method and system based on generative language models

Abstract

The application provides a failure mode data analysis method and system based on a generative language model, and relates to the field of data processing. The text data of a product can be split and classified by using a language model to obtain text information of each failure mode in multiple risk dimensions, so that the dependence on expert experience in the failure mode identification stage is avoided. Then, the text information is vectorized, the similarity between multiple failure modes is analyzed, similar failure modes to the failure mode to be predicted are screened out, a prediction score is obtained according to an initial score, the prediction score is corrected by using correction weights of multiple risk dimensions to obtain a corrected score, and the priority of each failure mode is obtained based on the corrected score. According to the priority, multiple failure modes are sorted. The sorting result weakens the dependence on expert experience and considers the reliability of multiple risk dimensions, so that the failure modes can be finely sorted, and the product research and development management has good guiding significance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a failure mode data analysis method and system based on a generative language model. BACKGROUND

[0002] Failure mode and effects analysis (FMEA) is a risk management method to ensure the safety and reliability of product development. It identifies potential failure modes, assesses risk factors (occurrence probability O, severity S, and detectability D), and calculates risk priority (RPN = O x S x D). It is widely used in high-risk fields such as aerospace and medicine.

[0003] Traditional FMEA methods rely on expert assessment and scoring of risk factors, then multiply the scores of multiple risk factors to get the risk factor priority.

[0004] The above-mentioned traditional FMEA method relies on expert experience and is subjective in evaluation, lacks objective data support, and the risk factor priority calculation method is single. Different risk factor scores may result in the same product, making it difficult to distinguish different risk scenarios and respond to differentiated needs, resulting in insufficient guidance of evaluation results for management practice. Subjective evaluation and simple product sorting lead to one-sided risk assessment, which cannot accurately guide management decisions. SUMMARY

[0005] The problem to be solved by the present application is that the existing failure mode analysis method relies on expert experience, and subjective evaluation and simple product sorting lead to one-sided risk assessment, resulting in insufficient guidance of evaluation results for management practice, and inability to accurately guide management decisions.

[0006] To solve the above problems, in a first aspect, the present application provides a failure mode data analysis method based on a generative language model, comprising:

[0007] Obtaining multiple failure modes of a product, initial scores of each failure mode in multiple risk dimensions, and text data of the product;

[0008] Splitting the text data into sentences according to punctuation marks, classifying multiple sentences according to the content of each sentence and the interpretation of each failure mode in multiple risk dimensions, to obtain text information corresponding to each failure mode in multiple risk dimensions;

[0009] Vectorizing the text information corresponding to each risk dimension using a generative language model to obtain vectors of each failure mode in multiple risk dimensions;

[0010] According to the vector of each failure mode in the plurality of risk dimensions, analyze the similarity between the plurality of failure modes;

[0011] According to the similarity and the initial score of the plurality of failure modes in the plurality of risk dimensions, obtain the predicted score of the plurality of failure modes in the plurality of risk dimensions;

[0012] According to the predicted score of each failure mode in the plurality of risk dimensions and the correction weight of the plurality of risk dimensions, obtain the correction score of the plurality of failure modes in the plurality of risk dimensions;

[0013] According to the correction score of the plurality of failure modes in the plurality of risk dimensions, obtain the priority of each failure mode;

[0014] According to the priority, sort the plurality of failure modes to guide the management decision of the product.

[0015] Optionally, the plurality of sentences are classified by using a generative language model.

[0016] Optionally, the analyzing the similarity between the plurality of failure modes according to the vector of each failure mode in the plurality of risk dimensions comprises:

[0017] Optionally, for any two failure modes, according to the vector of one failure mode in one risk dimension and the vector of another failure mode in the same risk dimension, the similarity between the two failure modes in the same risk dimension is obtained by using a cosine similarity algorithm;

[0018] The similarity between any two failure modes in the plurality of risk dimensions is obtained by cyclically analyzing all the failure modes.

[0019] Optionally, the obtaining the predicted score of the plurality of failure modes in the plurality of risk dimensions according to the similarity and the initial score of the plurality of failure modes in the plurality of risk dimensions comprises:

[0020] Optionally, for a selected failure mode, the similarity corresponding to the selected failure mode in the current risk dimension is compared to determine the maximum similarity corresponding to the selected failure mode in the current risk dimension, wherein the similarity corresponding to the selected failure mode is the similarity between the selected failure mode and other failure modes;

[0021] According to the similarity corresponding to the selected failure mode in the current risk dimension and the maximum similarity corresponding to the selected failure mode in the current risk dimension, a plurality of similarity absolute differences corresponding to the selected failure mode in the current risk dimension are obtained, wherein the similarity absolute differences corresponding to all the failure modes in the current risk dimension constitute a set of similarity absolute differences;

[0022] each absolute difference in the set of absolute differences of the similarity corresponding to the current risk dimension is taken as a candidate threshold value, and a plurality of screening threshold values are determined according to the plurality of candidate threshold values and the maximum similarity corresponding to the failure mode to be predicted, wherein the plurality of candidate threshold values correspond to the plurality of screening threshold values one by one;

[0023] a predicted value of the failure mode to be predicted corresponding to each candidate threshold value is obtained according to the similarity between the failure mode to be predicted and other failure modes in the current risk dimension, each screening threshold value of the current risk dimension, and the initial score of the plurality of failure modes in the current risk dimension;

[0024] a predicted value of each failure mode in the current risk dimension corresponding to each candidate threshold value is determined through a loop analysis;

[0025] an average absolute error of all failure modes corresponding to each candidate threshold value in the current risk dimension is obtained according to the predicted value of all failure modes in the current risk dimension corresponding to each candidate threshold value and the initial score of all failure modes in the current risk dimension;

[0026] the average absolute errors of all failure modes corresponding to all candidate threshold values are compared, and a predicted value corresponding to the minimum average absolute error is taken as a predicted score of the plurality of failure modes in the current risk dimension;

[0027] a predicted score of the plurality of failure modes in each risk dimension is determined through a loop analysis, wherein the risk dimensions include occurrence probability, occurrence severity, detectability and management cost.

[0028] Optionally, the obtaining of the predicted value of the failure mode to be predicted corresponding to each candidate threshold value according to the similarity between the failure mode to be predicted and other failure modes in the current risk dimension, each screening threshold value of the current risk dimension and the initial score of the plurality of failure modes in the current risk dimension comprises:

[0029] similar failure modes corresponding to the failure mode to be predicted under each screening threshold value are determined according to the similarity between the failure mode to be predicted and other failure modes in the current risk dimension and each screening threshold value of the current risk dimension, wherein the other failure modes are other failure modes in the plurality of failure modes except the current failure mode;

[0030] a similarity weight corresponding to each similar failure mode is determined according to the similarity of the similar failure mode corresponding to each screening threshold value;

[0031] the predicted value of the failure mode to be predicted corresponding to each candidate threshold value is obtained according to the similarity weight and the initial score of each similar failure mode in the current risk dimension.

[0032] Optionally, the obtaining the corrected scores of the plurality of failure modes in the plurality of risk dimensions according to the predicted scores of the plurality of failure modes in the plurality of risk dimensions and the correction weights of the plurality of risk dimensions comprises:

[0033] normalizing the predicted scores of the plurality of failure modes in the plurality of risk dimensions to obtain standardized scores;

[0034] determining variabilities of the plurality of risk dimensions according to the standardized scores of the plurality of risk dimensions;

[0035] determining the correction weights of the plurality of risk dimensions according to the variabilities of the plurality of risk dimensions;

[0036] obtaining the corrected scores of the plurality of failure modes in the plurality of risk dimensions according to the correction weights and the standardized scores.

[0037] Optionally, the variability is:

[0038]

[0039] wherein G j represents the variability of the jth risk dimension, h nj represents the standardized score of the nth failure mode in the jth risk dimension, and q represents the total number of failure modes.

[0040] The correction weight is:

[0041]

[0042] wherein t j represents the correction weight of the jth risk dimension, G j represents the variability of the jth risk dimension.

[0043] The corrected score is:

[0044] U nj = t j h nj ,

[0045] wherein U nj represents the corrected score of the nth failure mode in the jth risk dimension, h nj represents the standardized score of the nth failure mode in the jth risk dimension.

[0046] Optionally, the obtaining the priority of each failure mode according to the corrected scores of the plurality of failure modes in the plurality of risk dimensions comprises:

[0047] determining the maximum corrected score and the minimum corrected score in each risk dimension according to the corrected scores of the plurality of failure modes in the plurality of risk dimensions;

[0048] According to the maximum correction score and the minimum correction score on each risk dimension and the correction score of each failure mode on multiple risk dimensions, the priority of each failure mode is determined.

[0049] Optionally, the priority is:

[0050]

[0051] wherein, l n represents the priority of the nth failure mode, a represents the maximum or minimum value category of the correction score on the risk dimension, U nj represents the correction score of the nth failure mode on the jth risk dimension, represents the maximum correction score on the jth risk dimension, represents the minimum correction score on the jth risk dimension.

[0052] In a second aspect, the present application further provides a failure mode data analysis system based on a generative language model, comprising:

[0053] a data acquisition module, configured to acquire multiple failure modes of a product, initial scores of each failure mode on multiple risk dimensions, and text data of the product;

[0054] an information classification module, configured to split the text data into sentences according to punctuation marks, classify the multiple sentences according to the content of each sentence and the interpretation of each failure mode on multiple risk dimensions, and obtain text information corresponding to each failure mode on multiple risk dimensions;

[0055] an information vectorization module, configured to vectorize the text information corresponding to each risk dimension by using a generative language model, and obtain vectors of each failure mode on multiple risk dimensions;

[0056] a similarity analysis module, configured to analyze the similarity between multiple failure modes according to the vectors of each failure mode on multiple risk dimensions;

[0057] a score prediction module, configured to obtain predicted scores of multiple failure modes on multiple risk dimensions according to the similarity and the initial scores of multiple failure modes on multiple risk dimensions;

[0058] a score correction module, configured to obtain correction scores of multiple failure modes on multiple risk dimensions according to the predicted scores of each failure mode on multiple risk dimensions and correction weights of multiple risk dimensions;

[0059] a priority analysis module, configured to obtain the priority of each failure mode according to the correction scores of multiple failure modes on multiple risk dimensions.

[0060] A failure mode ranking module ranks the plurality of failure modes according to the priority to guide the management decision of the product.

[0061] The application provides a failure mode data analysis method and system based on a generative language model.

[0062] The obtained text data of the product can be split by using the generative language model and the like, and the sentences in the text data are classified in combination with the interpretation of each failure mode in a plurality of risk dimensions, so that the text information corresponding to each failure mode in the plurality of risk dimensions is obtained, and the dependence on expert experience in the failure mode identification stage is avoided; then the text information corresponding to each risk dimension is vectorized by using the generative language model, the similarity between the plurality of failure modes is analyzed according to the vectors of each failure mode in the plurality of risk dimensions, the failure modes similar to the to-be-predicted failure mode are screened out according to the similarity, and the prediction scores of the plurality of failure modes in the plurality of risk dimensions are obtained one by one by using the initial scores of the plurality of failure modes in the plurality of risk dimensions; the reliability and importance degree of the plurality of risk dimensions are further considered, the prediction scores are corrected by using the correction weights of the plurality of risk dimensions, and the corrected scores of the plurality of failure modes in the plurality of risk dimensions are obtained; the priority of each failure mode is obtained based on the corrected scores, and the plurality of failure modes are ranked according to the priority, the ranking result weakens the dependence on expert experience, the reliability of the plurality of risk dimensions is considered, even if the final priority is the same, the ranking can be further refined according to the reliability of the risk dimension and the score of the corresponding dimension, the failure modes can be finely ranked, and the product development, management and testing and the like have good guiding significance. BRIEF DESCRIPTION OF DRAWINGS

[0063] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description only constitute some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0064] Figure 1 A flowchart of a failure mode data analysis method based on a generative language model provided by an embodiment of the present application;

[0065] Figure 2 An execution flow diagram of a failure mode data analysis method based on a generative language model provided by an embodiment of the present application;

[0066] Figure 3A structural schematic diagram of a failure mode data analysis system based on a generative language model is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0067] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application are described clearly and completely. Obviously, the described embodiments are some embodiments of the present application, but not all embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0068] In order to better understand the above technical solutions, the above technical solutions will be described in detail below in combination with the accompanying drawings and specific embodiments.

[0069] Generative language model: a language model whose core function is to generate new, coherent text content based on input text content. It learns the patterns, structures and semantics of language by learning given text data, and can generate natural language text according to given prompts or context. Through training or adjustment, it can be used for text classification tasks; for example, a generative model using the Transformer architecture can be used for text classification by adding a classifier to the output layer of the model. It can also extract vector representations of text through its internal embedding layer.

[0070] As shown in Figure 1 A failure mode data analysis method based on a generative language model is provided in an embodiment of the present application, which comprises:

[0071] S1: obtaining a plurality of failure modes of a product, an initial score of each failure mode in a plurality of risk dimensions, and text data of the product.

[0072] Specifically, data collection mainly collects data from various sources such as expert opinions, historical records, and technical documents related to the development of the analyzed product. Then the data is processed and organized into a structured format to form a list of typical potential failure modes in the development process of a certain product, which includes descriptions of each failure mode (i.e. the interpretation of failure modes in multiple risk dimensions), which is prepared by the project-related department based on historical experience and management focus, and each failure mode (FM) is initially scored by experts as an initial score to distinguish from subsequent predicted scores. Text data includes text data related to the development process of the product (similar products, existing batches, etc.), such as various expert opinions, meeting records, technical documents, etc. The risk dimensions of each failure mode include occurrence probability, severity of occurrence, detectability, and management cost.

[0073] S2: split the text data into sentences according to punctuation marks, and classify the multiple sentences according to the content of each sentence and the interpretation of each failure mode on multiple risk dimensions, to obtain text information corresponding to each failure mode on multiple risk dimensions. A generative language model can be used to split and classify the text data.

[0074] S3: vectorize the text information corresponding to each risk dimension using a generative language model to obtain a vector for each failure mode on multiple risk dimensions.

[0075] S4: analyze the similarity between multiple failure modes according to the vector of each failure mode on multiple risk dimensions. The similarity between two text vectors can be analyzed by calculating the cosine similarity, Euclidean distance, Manhattan distance, etc. between the two vectors.

[0076] S5: obtain the predicted score of multiple failure modes on multiple risk dimensions according to the similarity and the initial score of multiple failure modes on multiple risk dimensions.

[0077] Specifically, analyzing the text similarity of multiple failure modes on multiple risk dimensions can determine that the text descriptions of similar failure modes are similar. If the descriptions of two failure modes on a certain risk dimension are similar or similar, it means that the two failure modes are related or even mutually triggered, so the evaluation scores of the two failure modes on that risk dimension should be similar or similar. According to this characteristic, the similarity of the text of multiple failure modes on multiple risk dimensions can be analyzed first, and then the initial scores of similar failure modes are used to predict the score of the failure mode to be predicted, rather than directly using the initial score of the failure mode to be predicted for evaluation and sorting.

[0078] S6: obtain the corrected score of multiple failure modes on multiple risk dimensions according to the predicted score of each failure mode on multiple risk dimensions and the correction weight of multiple risk dimensions.

[0079] Specifically, although the predicted scores of multiple failure modes have been calculated, there is no correlation between the scores of these failure modes on each risk dimension, that is, there is no priority between each risk dimension. According to the importance of each risk dimension or the reliability of each risk dimension, the correction weight corresponding to multiple risk dimensions can be assigned to correct the predicted score, so that the scores of the failure modes on multiple risk dimensions are biased, and the corrected score has more guiding significance.

[0080] S7: obtain the priority of each failure mode according to the corrected score of multiple failure modes on multiple risk dimensions.

[0081] S8: The plurality of failure modes are ranked according to the priority to guide the management decision of the product.

[0082] In the embodiment, the text data of the product obtained (for example, the text data in the product development process, the data in the product production process, the data in the test process, etc.) can be split by using a generative language model or the like, and the text information corresponding to each failure mode in the plurality of risk dimensions can be obtained by classifying the sentences in the text data in combination with the interpretation of each failure mode in the plurality of risk dimensions, so as to avoid the dependence on the expert experience in the failure mode identification stage. Then, the text information corresponding to each risk dimension is vectorized by using the generative language model, the similarity between the plurality of failure modes is analyzed according to the vectors of each failure mode in the plurality of risk dimensions, the failure modes similar to the failure mode to be predicted can be screened out according to the similarity, and the prediction scores of the plurality of failure modes in the plurality of risk dimensions are obtained one by one by using the initial scores of the plurality of failure modes in the plurality of risk dimensions. Further, the prediction scores are corrected by using the correction weights of the plurality of risk dimensions according to the reliabilities and importance degrees of the plurality of risk dimensions, so as to obtain the corrected scores of the plurality of failure modes in the plurality of risk dimensions. Based on the corrected scores, the priority of each failure mode is obtained, and the plurality of failure modes are ranked according to the priority. The ranking result weakens the dependence on the expert experience and considers the reliabilities of the plurality of risk dimensions. Even if the final priorities are the same, the ranking can be further refined according to the reliabilities of the risk dimensions and the scores of the corresponding dimensions, the failure modes can be finely ranked, and the product development, management, test, etc. have good guiding significance.

[0083] In combination with Figure 2 , the steps are described in detail as follows.

[0084] S1: Obtain a plurality of failure modes of a product, initial scores of each failure mode in a plurality of risk dimensions, and text data of the product.

[0085] Specifically, there are four kinds of risk dimensions, which are occurrence probability (O), occurrence severity (S), detectability (D), and management cost (M). Specifically, O represents the frequency of occurrence of the problem related to the failure mode; S represents the degree of influence that the failure mode may bring once it occurs; D represents the possibility of the failure mode being observed and detected in actual management; and M represents the cost and feasibility of managing the failure mode once it occurs.

[0086] S2: Split the text data into sentences according to the punctuation marks, classify the plurality of sentences according to the content of each sentence and the interpretation of each failure mode in the plurality of risk dimensions, and obtain the text information corresponding to each failure mode in the plurality of risk dimensions.

[0087] Specifically, first, all collected failure mode text information is split into logically complete sentences according to punctuation marks. Second, the generative language model is pre-trained, and in the input of the generative language model, several examples are first given as the execution reference of the generative language model. Specifically, we need to input 3 types of information to the generative language model: A. The task goal of the execution text information (i.e. the text information classification of the failure mode according to the four categories of O, S, D and M, while the specific connotation of O, S, D and M is given). B. All logically complete sentences. C. Given text information classification examples. Finally, the generative language model will output n text information of failure modes in different risk dimensions according to the task setting. The generative language model is trained using the small sample learning method, which helps to provide more high-quality classification results and has strong reasoning ability, which is more objective and accurate than manual classification operation. Finally, the generative language model is used to correspond the logically complete sentences to the different risk dimensions (O, S, D, M) of the failure mode, and the risk information classification of the failure mode is performed. Using the generative language model to split and classify text data can avoid the dependence on expert identification experience and classification experience, making the identification and classification more objective and improving the classification efficiency.

[0088] S3: Utilize the generative language model to vectorize the text information corresponding to each risk dimension to obtain the vector of each failure mode in multiple risk dimensions.

[0089] Specifically, the generative language model is used to vectorize the text information, and the textual description of the occurrence probability (O), occurrence severity (S), detectability (D) and management cost (M) of n failure modes FM is vectorized. Considering the operability of data processing, all word vectors in a sentence or text segment are averaged to obtain a vector representing the entire text. Finally, the vectorization results of n failure modes FM in four risk dimensions are output.

[0090] S4: According to the vector of each failure mode in multiple risk dimensions, analyze the similarity between multiple failure modes. Specifically, it includes:

[0091] S410: Optionally, two failure modes, according to the vector of one failure mode in one risk dimension and the vector of another failure mode in the same risk dimension, use the cosine similarity algorithm to obtain the similarity of the two failure modes in the same risk dimension.

[0092] S420: Loop to analyze all failure modes to obtain the similarity of any two failure modes in multiple risk dimensions.

[0093] For two failure modes on the same risk dimension, the vector is (x 11 , x 12 , x 13 ,..., x 1k ) and (x 21 , x 22 , x 23 ,..., x 2k ), the similarity calculation formula is as follows:

[0094]

[0095] Wherein, x 1p represents the pth element of the vector of one failure mode on one risk dimension, x 2p represents the pth element of the vector of another failure mode on the same risk dimension, and k represents the total number of elements of the vector corresponding to the risk dimension.

[0096] For example, the occurrence severity (S) of failure mode FM1 needs to be predicted. First, the occurrence severity (S) of failure mode FM1 and other n-1 failure modes is vectorized, and then the similarity between the occurrence severity (S) vector of failure mode FM1 and the occurrence severity (S) vector of each of the other n-1 failure modes is calculated.

[0097] S5: According to the similarity and the initial score of multiple failure modes on multiple risk dimensions, the predicted score of multiple failure modes on multiple risk dimensions is obtained. Specifically, it includes:

[0098] S510: Select a failure mode, compare the similarity corresponding to the selected failure mode on the current risk dimension, and determine the maximum similarity corresponding to the selected failure mode on the current risk dimension, wherein the similarity corresponding to the selected failure mode is the similarity between the selected failure mode and other failure modes, and the other failure modes are each of the other failure modes except the selected failure mode among all failure modes, for example, the similarities between FM1 and other failure modes (FM2, FM3 and FM4) are 0.6, 0.7 and 0.4 respectively.

[0099] S520: According to the similarity corresponding to the selected failure mode on the current risk dimension and the maximum similarity corresponding to the selected failure mode on the current risk dimension, a plurality of similarity absolute differences corresponding to the selected failure mode on the current risk dimension is obtained, wherein the similarity absolute differences corresponding to all failure modes on the current risk dimension constitute a similarity absolute difference set.

[0100] Suppose there are four failure modes FM1, FM2, FM3 and FM4, and the occurrence severity (S) is selected as the current analysis risk dimension. First, FM1 is selected, and suppose the similarities between FM1 and the other three failure modes are 0.6, 0.7 and 0.4 in turn. The maximum similarity corresponding to FM1 is 0.7, and the absolute difference in similarity corresponding to FM1 is obtained by subtracting each similarity from the maximum similarity, which is 0.1, 0 and 0.3 in turn. Then, FM2 is selected, and suppose the similarities between FM2 and the other three failure modes are 0.6, 0.5 and 0.8 in turn. It can be seen that the maximum similarity corresponding to FM2 is 0.8, and the absolute difference in similarity corresponding to FM2 is 0.2, 0.3 and 0 in turn. Similarly, suppose the similarities between FM3 and the other three failure modes are 0.7, 0.5 and 0.9 in turn. It can be seen that the maximum similarity corresponding to FM3 is 0.9, and the absolute difference in similarity corresponding to FM3 is 0.2, 0.4 and 0 in turn. Suppose the similarities between FM4 and the other three failure modes are 0.4, 0.8 and 0.9 in turn. It can be seen that the maximum similarity corresponding to FM4 is 0.9, and the absolute difference in similarity corresponding to FM4 is 0.5, 0.1 and 0 in turn. Thus, 12 absolute difference values are obtained, and repeated absolute difference values can be omitted in subsequent calculation process. All the obtained absolute difference values form an absolute difference set of similarity, for example, (0.1, 0, 0.3, 0.2, 0.4, 0.5), and it can be seen from the above that the maximum similarity corresponding to each failure mode can be different.

[0101] S530: Each absolute difference in the absolute difference set of similarity corresponding to the current risk dimension is taken as a candidate threshold, and a plurality of screening thresholds are determined according to the plurality of candidate thresholds and the maximum similarity corresponding to the failure mode to be predicted.

[0102] Specifically, the maximum similarity corresponding to the failure mode to be predicted is subtracted from the candidate threshold to obtain a screening threshold, and each candidate threshold corresponds to a screening threshold. For example, the occurrence severity S of the predicted failure mode FM1, the failure mode to be predicted is FM1, the candidate threshold is 0.2, and the maximum similarity corresponding to FM1 is 0.7. Thus, the screening threshold is 0.5. Similarly, each candidate threshold can be calculated to obtain a plurality of screening thresholds.

[0103] S540: According to the similarity between the failure mode to be predicted and other failure modes in the current risk dimension, each screening threshold of the current risk dimension and the initial score of the plurality of failure modes in the current risk dimension, the predicted value of the failure mode to be predicted corresponding to each candidate threshold is obtained.

[0104] Specifically, similar failure modes to the failure mode to be predicted can be screened out by each screening threshold, the average of the initial scores of all similar failure modes is calculated to obtain the prediction value of the failure mode to be predicted, the prediction values of other failure modes are similarly obtained, and finally the prediction value of each failure mode is subtracted from the initial score (true value) to obtain the error sum of all failure modes, and the prediction value corresponding to the minimum error sum is taken as the final prediction score.

[0105] In another embodiment, S540 specifically comprises:

[0106] S541: According to the similarity between the failure mode to be predicted and other failure modes in the current risk dimension and each screening threshold of the current risk dimension, determine the similar failure modes corresponding to the failure mode to be predicted under each screening threshold, wherein the other failure modes are the other failure modes in the plurality of failure modes except the current failure mode.

[0107] For example, in the similarity between the failure mode FM1 and other failure modes, the failure modes corresponding to the similarity greater than or equal to the screening threshold are screened, which are the similar failure modes of the failure mode FM1. According to the above example, when the candidate threshold is 0.2 and the screening threshold is 0.5, FM2 and FM3 are screened as the similar failure modes of FM1.

[0108] S542: According to the similarity of the similar failure modes corresponding to each screening threshold, determine the similarity weight corresponding to each similar failure mode. The similarity weight of a similar failure mode is the similarity of the similar failure mode divided by the sum of the similarities of all failure modes.

[0109] S543: According to the similarity weight and the initial score of each similar failure mode in the current risk dimension, obtain the prediction value of the failure mode to be predicted corresponding to each screening threshold. Since the screening threshold and the candidate threshold correspond to each other, the prediction value of the failure mode to be predicted corresponding to each candidate threshold is obtained.

[0110] The occurrence severity (S) of the failure mode FM1 is the weighted sum of the initial scores of the occurrence severity (S) of the similar failure modes of FM1, wherein the similarity weight is proportional to their similarity score. Assuming that the initial scores of the occurrence severity corresponding to FM1, FM2, FM3 and FM4 are 4, 3, 6 and 8, then when the candidate threshold is 0.2, the prediction value of FM1 in the occurrence severity dimension is 0.6x3 / (0.6+0.7)+0.7x6 / (0.6+0.7)=4.62.

[0111] S550: Loop analysis to determine the predicted value of all failure modes corresponding to each candidate threshold on the current risk dimension. Similarly, similar failure modes of the remaining failure modes can be screened, and the initial scores of the similar failure modes are weighted and summed to obtain the predicted value of each failure mode on each risk dimension. This method can weight and sum the initial scores of expert evaluation, not directly use the initial scores, reduce the error caused by the subjectivity of expert scoring, and avoid misleading the final ranking due to personal subjectivity.

[0112] S560: According to the predicted value of all failure modes corresponding to each candidate threshold on the current risk dimension and the initial score of all failure modes on the current risk dimension, the mean absolute error of all failure modes corresponding to each candidate threshold is obtained.

[0113] For example, when the candidate threshold is 0.2, the absolute error of FM1 on the severity dimension is |4.62-4|=0.62. Similarly, when the candidate threshold is 0.2, the absolute errors of FM2, FM3 and FM4 are calculated, which are assumed to be 0.2, 0.1 and 0.5. Then the mean absolute error of all failure modes corresponding to the candidate threshold 0.2 is (0.62+0.2+0.1+0.5) / 4=0.355.

[0114] S570: Compare the mean absolute errors of all failure modes corresponding to all candidate thresholds, and take the predicted value corresponding to the minimum mean absolute error as the predicted score of multiple failure modes on the current risk dimension.

[0115] S580: Loop analysis to determine the predicted score of multiple failure modes on each risk dimension, wherein the risk dimensions include occurrence probability, occurrence severity, detectability and management cost.

[0116] Specifically, expert scoring has a certain reference, but individual expert scoring also introduces too much personal subjectivity, so expert scoring is not directly used for analysis and ranking, but weighted sum is used to reduce the influence of personal subjectivity of expert scoring, and management cost is added to the evaluation system when considering risk dimensions, which can provide more operable guidance for risk management activities in complex product development process, and is conducive to the comprehensive management of risk activities and cost allocation in product development process.

[0117] S6: According to the predicted score of each failure mode on multiple risk dimensions and the correction weight of multiple risk dimensions, the correction score of multiple failure modes on multiple risk dimensions is obtained. Specifically, the following steps are included:

[0118] S610: Normalize the prediction scores of the plurality of failure modes on the plurality of risk dimensions to obtain standardized scores, and form a standardized score matrix H as follows. The standardized score of a failure mode on a risk dimension is obtained by dividing the modified score of the failure mode on the risk dimension by the square root of the sum of squares of the modified scores of all failure modes on the same risk dimension.

[0119]

[0120] wherein h nj represents the standardized score of the nth failure mode on the jth risk dimension, and q represents the total number of failure modes. In this embodiment, the matrix H has four dimensions, which are occurrence probability (O), severity of occurrence (S), detectability (D), and management cost (M).

[0121] S620: Determine the variability of each risk dimension according to the standardized score on each risk dimension.

[0122] The variability is:

[0123]

[0124] wherein G j represents the variability of the jth risk dimension, and a smaller variability means that the jth risk dimension can provide more certain and reliable information, so the jth risk dimension should be given a larger weight in weight distribution; h nj represents the standardized score of the nth failure mode on the jth risk dimension, and q represents the total number of failure modes.

[0125] S630: Determine the modified weight of each risk dimension according to the variability of the plurality of risk dimensions.

[0126] The modified weight is:

[0127]

[0128] wherein t j represents the modified weight of the jth risk dimension, and G j represents the variability of the jth risk dimension.

[0129] The correction weight is inversely proportional to the variability, that is, the greater the variability, the smaller the correction weight; on the contrary, the greater the correction weight. The more reliable the information of the risk dimension is, the greater the correction weight is. Because the information of some risk dimensions may not be accurate, if the scores made according to accurate information and inaccurate information are regarded as equivalent, the score made according to inaccurate information will mislead the result, in order to reduce the interference degree of uncertain information, the score of this part of risk dimension can be given a smaller weight in correction, and the proportion in the final calculation of priority is reduced. Thus, the accuracy of the final ranking is ensured, and the value of the final ranking is improved.

[0130] S640: According to the correction weight and the standardized score, the correction score of the plurality of failure modes in the plurality of risk dimensions is obtained.

[0131] The correction score is:

[0132] U nj =t j h nj ,

[0133] Wherein, u nj represents the correction score of the nth failure mode in the jth risk dimension, h nj represents the standardized score of the nth failure mode in the jth risk dimension.

[0134] S7: According to the correction score of the plurality of failure modes in the plurality of risk dimensions, the priority of each failure mode is obtained. Specifically, it includes:

[0135] According to the correction score of the plurality of failure modes in the plurality of risk dimensions, the maximum correction score and the minimum correction score in each risk dimension are determined.

[0136] According to the maximum correction score and the minimum correction score in each risk dimension and the correction score of each failure mode in the plurality of risk dimensions, the priority of each failure mode is determined.

[0137] The priority is:

[0138]

[0139] Wherein, l n represents the priority of the nth failure mode, a represents the maximum or minimum of the correction score in the risk dimension, represents the maximum correction score in the jth risk dimension, represents the minimum correction score in the jth risk dimension, and U nj represents the correction score of the nth failure mode in the jth risk dimension.

[0140] S8: The plurality of failure modes are ranked according to the priority to guide the management decision of the product.

[0141] Specifically, the priority is ranked in descending order, if the priority of two failure modes is equal, the predicted score corresponding to the risk dimension with the largest correction weight is compared, and the one with the larger predicted score is ranked in front; if they are still equal, the predicted scores on the risk dimensions corresponding to the correction weights are compared in descending order of the correction weights, until the two have a size relationship.

[0142] In summary, compared with the prior art, the following beneficial effects are possessed:

[0143] 1、In this application, the text data of the product is first split and classified using a language model, and in the failure mode identification stage, it does not depend on the experience of experts, reduces the introduction of individual experience of experts, and improves the classification accuracy, objectivity and classification efficiency. Then analyze the similarity of the text information of each failure mode in each risk dimension, then according to the similarity, screen similar failure modes, determine the optimal predicted value as the predicted score according to the size of the average absolute error between the predicted value and the true value, do not directly use the initial score evaluated by experts, but based on the similarity of the text vector, combined with the initial score of multiple similar failure modes to get the predicted score, avoid the adverse effects caused by the misjudgment of experts or the lack of experience of experts. Then calculate the variability of each failure mode in each risk dimension, so as to obtain the correction weight, give different correction weights to different risk dimensions, and can see the primary and secondary relationship of multiple risk dimensions, improve the accuracy and reference of the final priority, and also help to refine the final ranking. The method of the present application can be applied to data analysis in the product development stage, and can also be applied to data analysis in the product testing, product improvement, product production and qualified inspection stages. This method is not only suitable for analyzing the data of simple products, but also can be applied to the data analysis of complex products, which can better reflect the advantages of the method, and can analyze the priority of a large number of failure modes of complex products, which is beneficial to refining the development focus of complex products and has great guiding and promoting significance for the development process.

[0144] 2、Using the generative language model to classify and vectorize the text data can more accurately understand the text content and semantics, which helps to improve the use effect of objective data and provides better classification and vectorization results. At the same time, it can quickly set instructions and tasks, and can avoid the shortcomings of deep learning algorithms that rely too much on a large amount of training data and complex model operations.

[0145] 3) By increasing the measurement and evaluation of failure mode management cost, more operable guidance can be provided for risk management activities in complex product development processes, which is conducive to the integrated management of risk activities and cost allocation in product development processes, and also fully utilizes the advantages of new generation information technology.

[0146] As shown in Figure 3 The embodiment of the present application provides a failure mode data analysis system based on a generative language model, which comprises:

[0147] A data acquisition module 100 is configured to acquire a plurality of failure modes of a product, initial scores of each failure mode in a plurality of risk dimensions, and text data of the product.

[0148] An information classification module 200 is configured to split the text data into sentences according to punctuation marks, classify a plurality of sentences according to the content of each sentence and the interpretation of each failure mode in the plurality of risk dimensions, and obtain text information corresponding to each failure mode in the plurality of risk dimensions.

[0149] An information vectorization module 300 is configured to vectorize the text information corresponding to each risk dimension by using a generative language model, and obtain vectors of each failure mode in the plurality of risk dimensions.

[0150] A similarity analysis module 400 is configured to analyze the similarity between the plurality of failure modes according to the vectors of each failure mode in the plurality of risk dimensions.

[0151] A score prediction module 500 is configured to obtain predicted scores of the plurality of failure modes in the plurality of risk dimensions according to the similarity and the initial scores of the plurality of failure modes in the plurality of risk dimensions.

[0152] A score correction module 600 is configured to obtain corrected scores of the plurality of failure modes in the plurality of risk dimensions according to the predicted scores of each failure mode in the plurality of risk dimensions and correction weights of the plurality of risk dimensions.

[0153] A priority analysis module 700 is configured to obtain the priority of each failure mode according to the corrected scores of the plurality of failure modes in the plurality of risk dimensions.

[0154] A failure mode sorting module 800 is configured to sort the plurality of failure modes according to the priority, so as to guide the management decision of the product.

[0155] In the embodiment, the failure mode data analysis system based on the generative language model has similar beneficial effects to the failure mode data analysis method based on the generative language model, which will not be repeated here.

[0156] An electronic device provided by an embodiment of the present application includes a memory and a processor; the memory is configured to store a computer program; the processor is configured to, when executing the computer program, implement the failure mode data analysis method based on the generative language model as described above.

[0157] A computer readable storage medium provided by an embodiment of the present application has a computer program stored thereon, and when the computer program is executed by a processor, the failure mode data analysis method based on the generative language model as described above is implemented.

[0158] In the present embodiment, the electronic device and the computer readable storage medium have similar advantages to the failure mode data analysis method based on the generative language model as described above, and will not be described here again.

[0159] Now, an electronic device that can be a server or a client of the present application will be described, which is an example of a hardware device that can be applied to various aspects of the present application. The electronic device is intended to represent a variety of forms of digital electronic computing devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections, and relationships, and their functions, are meant only as examples, and are not meant to limit implementations of the present application described and / or claimed herein.

[0160] The electronic device includes a computing unit that can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) or a computer program loaded from a storage unit into a random access memory (RAM). In the RAM, various programs and data required for device operation can also be stored. The computing unit, the ROM, and the RAM are connected to each other through a bus. An input / output (I / O) interface is also connected to the bus.

[0161] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The program can be stored in a computer readable storage medium. When the program is executed, the processes of the above-mentioned embodiment methods can be included. The storage medium can be a magnetic disc, an optical disc, a read-only memory (ROM) or a random access memory (RAM), etc. In the present application, the units described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., they can be located in one place, or they can be distributed on multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiments of the present application. In addition, the functional units in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0162] It should be noted that, in the present document, relational terms such as first and second and the like can be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by "comprises a... " does not, without more constraints, exclude the existence of additional identical elements in the process, method, article, or apparatus that comprises the element.

[0163] The above embodiments are only used to illustrate the technical solutions of the present application, but not limit the technical solutions of the present application; although the technical solutions of the present application are described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for failure mode data analysis based on a generative language model, characterized by, The method comprises the following steps: obtaining a plurality of failure modes of a product, initial scores of each failure mode on a plurality of risk dimensions, and text data of the product; splitting the text data into sentences according to punctuation marks, classifying the plurality of sentences according to the content of each sentence and the interpretation of each failure mode on the plurality of risk dimensions, and obtaining text information corresponding to each failure mode on the plurality of risk dimensions; vectorizing the text information corresponding to each risk dimension by using a generative language model to obtain vectors of each failure mode on the plurality of risk dimensions; analyzing the similarity between the plurality of failure modes according to the vectors of each failure mode on the plurality of risk dimensions; obtaining predicted scores of the plurality of failure modes on the plurality of risk dimensions according to the similarity and the initial scores of the plurality of failure modes on the plurality of risk dimensions; obtaining corrected scores of the plurality of failure modes on the plurality of risk dimensions according to the predicted scores of each failure mode on the plurality of risk dimensions and correction weights of the plurality of risk dimensions; obtaining the priority of each failure mode according to the corrected scores of the plurality of failure modes on the plurality of risk dimensions; ranking the plurality of failure modes according to the priority to guide the management decision of the product; The method for obtaining the predicted scores of the plurality of failure modes on the plurality of risk dimensions according to the similarity and the initial scores of the plurality of failure modes on the plurality of risk dimensions comprises the following steps: selecting a failure mode, comparing the similarity corresponding to the selected failure mode on the current risk dimension, and determining the maximum similarity corresponding to the selected failure mode on the current risk dimension; obtaining a plurality of absolute differences of the similarity corresponding to the selected failure mode on the current risk dimension according to the similarity corresponding to the selected failure mode on the current risk dimension and the maximum similarity corresponding to the selected failure mode on the current risk dimension, wherein the absolute differences of the similarity corresponding to all failure modes on the current risk dimension constitute an absolute difference set of the similarity; taking each absolute difference in the absolute difference set of the similarity corresponding to the current risk dimension as a candidate threshold, and determining a plurality of screening thresholds according to the plurality of candidate thresholds and the maximum similarity corresponding to the failure mode to be predicted; obtaining the predicted value of the failure mode to be predicted corresponding to each screening threshold according to the similarity between the failure mode to be predicted and other failure modes on the current risk dimension, each screening threshold of the current risk dimension, and the initial scores of the plurality of failure modes on the current risk dimension; recursively analyzing and determining the predicted value of all failure modes corresponding to each candidate threshold on the current risk dimension; obtaining the mean absolute error corresponding to each candidate threshold according to the predicted value of all failure modes corresponding to each candidate threshold on the current risk dimension and the initial scores of all failure modes on the current risk dimension; comparing the mean absolute errors corresponding to all candidate thresholds, and taking the predicted value corresponding to the minimum mean absolute error as the predicted score of the plurality of failure modes on the current risk dimension; recursively analyzing and determining the predicted score of the plurality of failure modes on each risk dimension, wherein the risk dimensions include occurrence probability, occurrence severity, detectability, and management cost.

2. The generative language model-based failure mode data analysis method of claim 1, wherein, Classifying the plurality of sentences by using a generative language model.

3. The generative language model-based failure mode data analysis method of claim 1, wherein, The analyzing the similarity between the plurality of failure modes according to the vector of each failure mode on the plurality of risk dimensions comprises: Optionally, two failure modes, according to the vector of one failure mode on one risk dimension and the vector of another failure mode on the same risk dimension, use a cosine similarity algorithm to obtain the similarity of the two failure modes on the same risk dimension; Cyclically analyzing all failure modes to obtain the similarity of any two failure modes on the plurality of risk dimensions.

4. The generative language model-based failure mode data analysis method of claim 1, wherein, The obtaining the predicted value of the to-be-predicted failure mode corresponding to each screening threshold according to the similarity between the to-be-predicted failure mode and other failure modes on the current risk dimension, each screening threshold of the current risk dimension, and the initial score of the plurality of failure modes on the current risk dimension comprises: According to the similarity between the to-be-predicted failure mode and other failure modes on the current risk dimension and each screening threshold of the current risk dimension, determining the similar failure mode corresponding to the to-be-predicted failure mode under each screening threshold, wherein the other failure modes are other failure modes in the plurality of failure modes except the current failure mode; According to the similarity of the similar failure mode corresponding to each screening threshold, determining the similarity weight corresponding to each similar failure mode; According to the similarity weight and the initial score of each similar failure mode on the current risk dimension, obtaining the predicted value of the to-be-predicted failure mode corresponding to each screening threshold.

5. The generative language model-based failure mode data analysis method of claim 1, wherein, The obtaining the corrected score of the plurality of failure modes on the plurality of risk dimensions according to the predicted score of each failure mode on the plurality of risk dimensions and the correction weight of the plurality of risk dimensions comprises: Normalizing the predicted score of the plurality of failure modes on the plurality of risk dimensions to obtain a standardized score; According to the standardized score on each risk dimension, determining the variability of each risk dimension; According to the variability of the plurality of risk dimensions, determining the correction weight of each risk dimension; According to the correction weight and the standardized score, obtaining the corrected score of the plurality of failure modes on the plurality of risk dimensions.

6. The generative language model-based failure mode data analysis method of claim 5, wherein, The variability is: ， wherein, represents the variability of the jth risk dimension, represents the normalized score of the nth failure mode on the jth risk dimension, q represents the total number of failure modes; The correction weight is: ， wherein, represents the modified weight of the jth risk dimension, represents the variability of the jth risk dimension; The corrected score is: ， wherein, denotes the modified score of the nth failure mode on the jth risk dimension, denotes the normalized score of the nth failure mode on the jth risk dimension.

7. The generative language model-based failure mode data analysis method of claim 1, wherein, The obtaining the priority of each failure mode according to the corrected score of the plurality of failure modes on the plurality of risk dimensions comprises: According to the corrected score of the plurality of failure modes on the plurality of risk dimensions, determining the maximum corrected score and the minimum corrected score on each risk dimension; According to the maximum corrected score and the minimum corrected score on each risk dimension and the corrected score of each failure mode on the plurality of risk dimensions, determining the priority of each failure mode.

8. The generative language model-based failure mode data analysis method of claim 7, wherein, The priority is: ， wherein, represents the priority of the nth failure mode, a represents the maximum class of the modification score in the risk dimension, represents the modification score of the nth failure mode in the jth risk dimension, represents the maximum modification score in the jth risk dimension, represents the minimum modification score in the jth risk dimension.

9. A system for failure mode data analysis based on generative language models, the system comprising: The system for implementing the failure mode data analysis method based on the generative language model according to any one of claims 1-8 comprises: A data acquisition module configured to acquire a plurality of failure modes of a product, an initial score of each failure mode on a plurality of risk dimensions, and text data of the product; An information classification module configured to split the text data into sentences according to punctuation, classify the plurality of sentences according to the content of each sentence and the interpretation of each failure mode on the plurality of risk dimensions, and obtain text information corresponding to each failure mode on the plurality of risk dimensions; An information vectorization module is configured to vectorize the text information corresponding to each risk dimension by using the generative language model to obtain vectors of each failure mode in multiple risk dimensions; A similarity analysis module is configured to analyze the similarity between multiple failure modes according to the vectors of each failure mode in multiple risk dimensions; A score prediction module is configured to obtain predicted scores of multiple failure modes in multiple risk dimensions according to the similarity and initial scores of multiple failure modes in multiple risk dimensions; A score correction module is configured to obtain corrected scores of multiple failure modes in multiple risk dimensions according to the predicted scores of each failure mode in multiple risk dimensions and correction weights of multiple risk dimensions; A priority analysis module is configured to obtain the priority of each failure mode according to the corrected scores of multiple failure modes in multiple risk dimensions; A failure mode ranking module is configured to rank multiple failure modes according to the priority to guide the management decision of the product.

Images