CN-121980941-A - FMEA risk level prediction system based on machine learning

Abstract

The invention discloses a machine learning-based FMEA risk level prediction system which comprises the steps of collecting FMEA historical data and running data, completing preprocessing to form a standardized basic data set, constructing an FMEA semantic relation structure, calculating path potential energy values to generate structural risk features, generating a structural channel and a distribution channel, respectively outputting structural and distribution risk sequence values, constructing an improved TabDDPM model, supplementing a risk sample, correcting the distribution risk sequence values, fusing the two types of risk sequence values, calculating fusion weights, outputting candidate risk level results, constructing a risk feasible domain, and outputting final FMEA risk level prediction results by applying a multi-layer constraint mechanism. According to the invention, by constructing the FMEA semantic relation structure and fusing the causal path analysis and the distribution modeling mechanism, the stable and credible prediction of the failure mode risk level under the complex working condition is realized.

Inventors

• YANG RAN

Assignees

• 辽宁中恩科技有限责任公司

Dates

Publication Date

20260505

Application Date

20260123

Claims (8)

1. 1. An FMEA risk level prediction system based on machine learning, comprising: the data preprocessing module is used for acquiring FMEA historical data and operation data corresponding to the object to be evaluated, and preprocessing the FMEA historical data and the operation data to form a basic data set; the semantic modeling module is used for constructing an FMEA semantic relation structure based on the basic data set, calculating path potential energy values by combining FMEA historical data and generating structural risk characteristics; The sequence generation module is used for constructing a structural channel and a distribution channel based on the structural risk characteristics, the structural channel is used for carrying out risk intensity sequencing on the failure modes according to the path potential energy values corresponding to the failure modes and generating structural risk sequence values, and the distribution channel is used for mapping the risk characteristics of the current failure modes to the corresponding historical distribution intervals and generating distribution risk sequence values based on the statistical distribution of the historical risk characteristic samples in the basic data set; the distribution modeling module is used for carrying out distribution modeling on the risk characteristic samples in the basic dataset by using the improved TabDDPM model, generating a supplementary risk sample and carrying out distribution consistency correction on the distribution risk sequence values; The fusion prediction module is used for calculating fusion weights based on the structural risk sequence values and the distribution risk sequence values corrected by the distribution consistency, and fusing the two types of risk sequence values to generate candidate risk level prediction results; And the feasible domain correction module is used for constructing an FMEA risk feasible domain based on the FMEA historical data, the path potential energy value and the candidate risk level prediction result, carrying out consistency correction on the candidate risk level prediction result and outputting a final FMEA risk level prediction result.
2. 2. The machine learning based FMEA risk level prediction system of claim 1 wherein the FMEA history data includes failure modes, failure causes, failure outcomes, control measures, detection modes, operating conditions corresponding to the analyzed object, and severity scores, occurrence scores, detectivity scores, historical risk level results, and production operation records and quality detection records corresponding to the failure modes formed based on historical analysis, the operation data including equipment operation parameter data, process parameter data, environmental state parameter data, quality detection result data, fault and maintenance record data corresponding to the FMEA analyzed object.
3. 3. The machine learning based FMEA risk level prediction system of claim 1 wherein the preprocessing of FMEA history data and operational data includes performing missing value completion processing, outlier identification and culling processing, data format unification processing, data time alignment processing, data normalization processing, and associative mapping of data corresponding to failure modes, failure causes, failure outcomes, control measures, detection modes, and operating conditions to form a structured dataset.
4. 4. The machine learning based FMEA risk level prediction system of claim 1 wherein said semantic modeling module comprises: Based on a basic data set, carrying out structural analysis on failure modes, failure reasons, failure results, control measures, detection modes and working conditions in FMEA historical data, extracting each type of information into independent semantic entities respectively, and distributing unique identifiers for each semantic entity to form a multi-type semantic entity set; Based on the corresponding relation recorded in the FMEA historical data and the operation data associated with the FMEA historical data, respectively constructing a reason association relation, a result association relation, a control constraint relation, a detection association relation and a working condition influence relation between semantic entities, and representing the different types of relations in a directed association form to form an FMEA semantic relation structure containing multiple relation types; Identifying an failure cause entity, an failure result entity, a control measure entity and a working condition entity which are directly or indirectly related to the failure mode entity aiming at each failure mode entity in the FMEA semantic relation structure, constructing a causal path set according to the sequence that the failure cause points to the failure mode and the failure mode points to the failure result, and identifying each causal path as an independent risk conduction unit; Based on a causal path set, severity score, occurrence score and detection score in FMEA historical data corresponding to each causal path are respectively obtained, and the risk conduction intensity of the causal path is quantitatively represented by combining a control measure entity and a working condition entity associated with the path to obtain path potential energy values reflecting the risk conduction differences of different causal paths; And summarizing path potential energy values aiming at a plurality of causal paths related to the same failure mode entity to form a structured risk feature for describing the overall risk conduction characteristic of the failure mode.
5. 5. The machine learning based FMEA risk level prediction system of claim 1 wherein said order generation module comprises: Based on the structural risk characteristics, summarizing a corresponding causal path set, control measure information, detection mode information and working condition information aiming at each failure mode, and extracting historical risk characteristic samples which are the same as the failure modes from a basic data set to form input data for parallel processing; Constructing a structural channel, carrying out deterministic sorting on a causal path set of each failure mode according to risk conduction intensity, sequentially comparing the risk intensity of a single path, the number of uncontrolled paths, the number of paths which are difficult to cover, the result severity category of path coverage and the path length by adopting a hierarchical comparison rule, automatically transferring into the next comparison item when the last comparison item is parallel until a structured sorting result without parallel items is obtained, and converting the sorting result into a structural risk sequence value; establishing a distribution channel, carrying out layered division on historical risk characteristic samples in a basic data set according to working conditions and detection modes, establishing reference distribution based on frequency and time windows in each layer, mapping risk characteristics of a current failure mode to a corresponding interval according to the accumulated position of the reference distribution, generating a distribution risk sequence value, marking samples falling outside the reference distribution as samples exceeding a supporting range, and marking rare sample identifiers for samples in a low-occurrence frequency interval; And carrying out consistency constraint processing on the outputs of the structural channels and the distribution channels, generating channel reliability indication information according to causal path coverage integrity, historical sample matching degree, working condition layering matching degree and detection mode matching degree, marking the sequence bit output of the corresponding channel as low-reliability sequence bit when the reliability indication information of any channel is lower than a threshold value, and reserving an original sequence bit value.
6. 6. The machine learning based FMEA risk level prediction system of claim 1 wherein said distribution modeling module comprises: Extracting a risk feature sample based on a basic data set, wherein the historical risk feature sample comprises structural risk features corresponding to failure modes, working conditions, detection modes and hierarchical identification information of a distribution channel, so as to form a training sample set for distribution modeling; generating a diffusion stage sample sequence for the training sample set, and inputting working conditions, detection modes and distribution channel layering identification information as conditions to be aligned with each diffusion stage sample sequence; An improvement TabDDPM model is built, the improvement TabDDPM model comprising a conditional alignment layer, a structural interaction layer, and a range agreement layer, wherein: The condition alignment layer encodes the working condition, the detection mode and the distribution channel layering identification information into a condition vector and aligns the condition vector with the sample sequence of each diffusion stage; the structural interaction layer carries out field-level interaction combination on the condition vector and the structural risk feature to output the denoised risk feature; the range consistency layer performs field legal range check and type consistency check on the denoised risk characteristics, and performs projection correction on the fields which do not meet the check constraint to obtain reverse denoised input of the next diffusion stage; Training the improved TabDDPM model based on a diffusion stage sample sequence and condition input, wherein in the training process, the diffusion stage sample sequence is used as input, the noiseless target characteristics of the corresponding stage are used as supervision targets, and the iterative updating parameters enable the improved TabDDPM model to learn the joint distribution of the historical risk characteristic samples under the constraint of working conditions and detection modes; And generating a supplementary risk feature sample in a layering corresponding to each working condition and the detection mode by utilizing the improved TabDDPM model after training, combining the supplementary risk feature sample and the historical risk feature sample to form an extended sample set, reconstructing the reference distribution of the distribution channel based on the extended sample set to execute distribution consistency correction on the distribution risk sequence value, and obtaining a corrected distribution risk sequence value.
7. 7. The machine learning based FMEA risk level prediction system of claim 1 wherein said fusion prediction module comprises: aiming at each failure mode to be evaluated, acquiring a structural risk sequence value and a distribution risk sequence value corrected by distribution consistency, and simultaneously retrieving channel credibility indication information; based on the structural risk sequence value of the structural channel, calculating the ordering stability information of the failure modes in the structural channel according to the relative positions of adjacent sequence intervals, the local sequence stability condition and the similar failure modes; Calculating distribution matching degree information of the failure mode in the distribution channel according to historical sample coverage conditions, whether the marks exceeding the supporting range exist and whether rare sample marks exist or not based on layered reference distribution of the distribution channel and distribution risk sequence values of the failure mode; Comparing the structural risk sequence value with the distribution risk sequence value, and judging the structural risk sequence value to be consistent or conflict according to a preset sequence difference range: Under the consistent condition, respectively distributing fusion weights for the structural channels and the distribution channels according to the sorting stability information and the distribution matching degree information, so that the sum of the weights of the two channels is defined as one and is respectively defined between zero and one; under the conflict situation, determining a main channel according to the causal path coverage integrity, the number of uncontrolled high-risk paths, rare sample identification and channel reliability indication information, and adopting a single-channel fusion mode to only reserve the weight corresponding to the main channel, wherein the weight of the other channel is reset to zero, so that the addition of the weights of the two channels is equal to one; and carrying out weighted fusion on the structural risk rank value and the distributed risk rank value according to the determined fusion weight, and generating a candidate risk level prediction result.
8. 8. The machine learning based FMEA risk level prediction system of claim 1 wherein said feasible region correction module comprises: Based on severity score, occurrence score and detection score in FMEA historical data, constructing a basic risk constraint layer, and limiting the risk level corresponding to each failure mode within a constraint range which satisfies that the risk level is not reduced when the severity is higher, the risk level is not reduced when the occurrence is higher, and the risk level is not reduced when the detection difficulty is higher, so that a first-layer risk feasible constraint is formed; Constructing a path risk constraint layer based on a causal path set and path potential energy value corresponding to the failure mode, setting corresponding risk level lower bound constraint for failure modes of which the causal path number exceeds a preset threshold and which are uncontrolled or which cover a plurality of failure outcome categories, and forming a second layer of risk feasible constraint; Based on a risk distribution result obtained by modeling an improved TabDDPM model in a distribution channel and a main channel judgment result obtained in a channel fusion stage, a distribution consistency constraint layer is constructed, corresponding risk level lower bound constraint is set for failure modes of which risk characteristics fall into a preset risk distribution interval, are marked as rare samples or are marked as exceeding a historical support range, and a third layer of risk feasible constraint is formed; Combining a first-layer risk feasible constraint, a second-layer risk feasible constraint and a third-layer risk feasible constraint, determining a risk feasible domain lower bound corresponding to each failure mode, comparing a candidate risk level prediction result with the risk feasible domain lower bound, adjusting the risk level to the risk feasible domain lower bound when the candidate risk level prediction result is lower than the risk feasible domain lower bound, and outputting a final FMEA risk level prediction result, otherwise, directly outputting the candidate risk level prediction result.

Description

FMEA risk level prediction system based on machine learning Technical Field The invention relates to the technical field of machine learning, in particular to an FMEA risk level prediction system based on machine learning. Background The failure mode and influence analysis is a risk assessment method widely applied to product design, manufacturing process and system operation stages, and the risk is assessed in a grading manner by identifying potential failure modes, analyzing failure reasons and possible influences caused by the potential failure modes and combining indexes such as severity, occurrence degree and detection degree, so that basis is provided for risk control and improvement measures. The conventional FMEA method is generally based on expert experience and rules, adopts a manual or semi-automatic mode to evaluate risk levels, has the characteristics of clear flow and strong engineering interpretability, but in complex systems and multi-working condition scenes, the evaluation result is highly dependent on subjective judgment of the expert, and the real evolution characteristics of the risk under the multi-factor coupling condition are difficult to fully reflect. With the development of data acquisition technology and computing power, some prior art starts to try to introduce a machine learning method into an FMEA risk assessment process, and by modeling historical failure data and operation data, automatic prediction of risk level is achieved. The method generally utilizes a statistical learning or supervised learning model to fit the relationship between the failure mode and the risk level, so that the artificial participation degree is reduced to a certain extent, and the evaluation efficiency is improved. However, most of the existing FMEA methods based on machine learning consider failure modes as independent samples, focus on the mapping relationship between numerical characteristics and risk results, lack systematic modeling on causal relationships among failure reasons, failure results and control measures, and are difficult to characterize the conduction process of risks in complex causal links. In the prior art, when expert rules and data driving results are fused, a simple weighting or threshold correction mode is generally adopted, a dynamic consistent fusion mechanism is difficult to establish between historical data distribution and engineering logic constraint, and the situation that a predicted result is inconsistent with FMEA basic logic easily occurs, for example, a low-risk structure is predicted to be high risk, or a critical failure path is underestimated. Under the scene that high risk samples are scarce or the working condition changes frequently, the adaptability of the existing method to abnormal distribution and few sample conditions is insufficient, so that the stability and the reliability of a risk level prediction result are low, and the actual requirement of a complex engineering system on refined risk assessment is difficult to meet. Therefore, how to provide a FMEA risk level prediction system based on machine learning is a problem that needs to be solved by those skilled in the art. Disclosure of Invention The invention aims to provide a FMEA risk level prediction system based on machine learning, which is used for intelligently improving a traditional FMEA risk assessment process by introducing technologies such as semantic modeling, causal path analysis, generation type distribution modeling and the like. According to the invention, the semantic relation structure between failure modes is constructed by comprehensively utilizing FMEA historical data and operation data, risk conduction paths among failure reasons, failure results and control measures are described to form structural risk characteristics, dynamic fusion of expert engineering logic and historical data distribution is realized through dual-channel risk sequence processing and distribution consistency correction, and consistency correction is further carried out on a prediction result by combining risk feasible domain constraint, so that reliable prediction of FMEA risk level is realized. The invention can stably output the risk grade result conforming to the FMEA basic logic under the complex causal relationship and multiple working conditions, and has the advantages of high reliability, good stability, strong adaptability to high-risk scarce scenes and the like of the prediction result. According to an embodiment of the invention, an FMEA risk level prediction system based on machine learning comprises: the data preprocessing module is used for acquiring FMEA historical data and operation data corresponding to the object to be evaluated, and preprocessing the FMEA historical data and the operation data to form a basic data set; the semantic modeling module is used for constructing an FMEA semantic relation structure based on the basic data set, calculating path potential energy values by combin