CN-121997137-A - Industrial pump interpretable intelligent fault detection method based on inverse fact learning

CN121997137ACN 121997137 ACN121997137 ACN 121997137ACN-121997137-A

Abstract

The invention discloses an industrial pump interpretable intelligent fault detection method based on inverse fact learning, which relates to the technical field of industrial pump fault diagnosis and comprises the following steps of constructing a fault detection model; the method comprises the steps of generating a ground truth sample in a ground truth search space, constructing a ground truth search space for limiting a physical feasible range for generating the ground truth sample, generating the ground truth sample through an optimization process in the ground truth search space, wherein the optimization process minimizes disturbance amplitude between an original working condition and the ground truth sample, meanwhile forces an output judging result of a fault detection model on the ground truth sample to be opposite to an output judging result of the original sample, and outputting a structured interpretable result, wherein the structured interpretable result comprises physical variable difference analysis of the ground truth sample and the original sample, sensitivity ordering of key physical variables and abnormal source inference so as to provide engineering executable maintenance suggestions. The invention realizes the high reliability and the engineering executable pump fault diagnosis and interpretation capability.

Inventors

CHEN CONG
CHEN XIANBING
YUAN WEIWEI
GU YUXUAN
ZHAO JIE

Assignees

中国船舶集团有限公司第七一九研究所
南京航空航天大学

Dates

Publication Date: 20260508
Application Date: 20260126

Claims (10)

1. An industrial pump interpretable intelligent fault detection method based on counterfactual learning, comprising: constructing a fault detection model, wherein the fault detection model takes multi-source physical characteristics of an industrial pump during operation as input and outputs a two-classification judging result of whether the industrial pump needs maintenance or not; constructing a counterfactual search space defined based on mechanical properties, fluid dynamics rules and equipment operation constraints of the industrial pump for defining a physically viable range for generating counterfactual samples; Generating a counterfactual sample in the counterfactual search space through an optimization process, wherein the optimization process minimizes the disturbance amplitude between the original working condition and the counterfactual sample, and meanwhile forces the output judging result of the fault detection model on the counterfactual sample to be opposite to the output judging result of the original sample; And outputting a structured interpretable result including a physical variable variance analysis of the counterfactual sample and the original sample, a sensitivity ordering of key physical variables, and anomaly source inference to provide engineering executable maintenance recommendations.
2. The method for intelligent fault detection interpretable by an industrial pump based on inverse fact learning according to claim 1, wherein the fault detection model is implemented by a multi-layer feedforward neural network, the multi-layer feedforward neural network performs normalization preprocessing on input multi-source physical characteristics, and trains fixed parameters through supervised learning so as to output a probability determination result that the industrial pump needs maintenance or does not need maintenance.
3. The method for intelligent fault detection interpretable by an industrial pump based on inverse facts learning of claim 1, wherein the multisource physical characteristics include temperature, vibration amplitude, spatial pressure, flow and rotational speed, which are collected from industrial pump sensors in real time and input as vectors into the fault detection model.
4. The method of claim 1, wherein the counterfactual search space includes a physically feasible boundary constraint defining a variation interval for each physical variable within a safe operating range of the device, including a material allowable temperature interval for temperature, a device standard vibration limit for vibration amplitude, a system design pressure range for pressure, a pump allowable flow interval for flow, and a driving device rotational speed upper and lower limits for rotational speed.
5. The method of claim 1, wherein the counterfactual search space includes data-driven correlation constraints based on statistical analysis of industrial pump operational data, defining coupling relationships between physical variables including pump characteristic curve correlation of flow and pressure, rotational speed and flow and pressure and vibration interactions, statistical correlation characteristics of temperature and vibration and pressure, and limiting joint variation amplitude of counterfactual samples by coupling offset thresholds.
6. The industrial pump interpretable intelligent fault detection method based on counterfactual learning of claim 1, wherein the counterfactual search space includes a minimum disturbance constraint that requires that the magnitude of the difference between the generated counterfactual sample and the original operating conditions be minimized, and that the counterfactual sample be closest to the original condition while meeting the physical viable boundaries and related constraints is ensured by measuring the magnitude of the disturbance by a norm distance.
7. The method for intelligent fault detection interpretable by an industrial pump based on counterfactual learning of claim 1, wherein the optimization process includes a perturbation min term implemented by minimizing a norm distance between the original operating conditions and the counterfactual samples to quantify and reduce the overall magnitude of the change in the physical variables.
8. The method of claim 1, wherein the optimization process includes a predictive rollover term that forces the failure detection model to output a result to the counterfactual sample that is opposite to the original decision by constructing a rollover loss function and ensures that the output probability exceeds a confidence threshold to achieve reliable rollover.
9. The method for intelligent fault detection interpretable by an industrial pump based on counterfactual learning of claim 1, wherein the optimization process includes a physical constraint maintenance term that introduces violations of physically feasible boundary constraints and data driven related constraints through a penalty function, ensuring that counterfactual samples conform to mechanical properties and fluid dynamics rules of the industrial pump.
10. The method of claim 1, wherein the structured interpretable result comprises calculating a disturbance magnitude for each physical quantity based on the physical variable differential analysis and sensitivity ordering to identify key influencing variables, and generating executable maintenance recommendations based on the anomaly source inferences, the maintenance recommendations describing how to adjust physical variables to eliminate risk of failure, including vibration amplitude reduction ratios, pressure increase values, or temperature control recommendations.

Description

Industrial pump interpretable intelligent fault detection method based on inverse fact learning Technical Field The invention relates to the technical field of industrial pump fault diagnosis, in particular to an industrial pump interpretable intelligent fault detection method based on inverse fact learning. Background Industrial pumps are the core power plant of industrial process systems, the operating state of which directly affects the stability, safety and energy efficiency level of the production process. Once a failure occurs, it may lead to production interruption, equipment damage, safety accidents, and huge economic losses. With the development of sensor technology and data acquisition systems, modern industrial pumps are commonly provided with multiple types of sensors such as temperature, vibration, pressure, flow and rotating speed, and can accurately capture operation data in real time, realize omnibearing dynamic monitoring and provide support for finding fault hidden danger. However, existing industrial pump fault diagnosis techniques have significant drawbacks. Traditional diagnostic methods such as manual experience, threshold rules or traditional statistical methods rely on fixed rules and preset thresholds to determine faults. The industrial production environment is complex and changeable, the nonlinear characteristic of the working condition is frequent, the fixed rule is difficult to adapt to the change, the problems of false alarm and missing report are easy to occur, and the production stability is affected. In addition, although the pump failure prediction model based on machine learning or deep learning can predict a failure and output a classification result (such as "whether the failure is a failure") belongs to a black box model, the prediction cause cannot be explained or the failure source cannot be provided, and engineering executable suggestions for avoiding the failure (such as "which physical quantities should be adjusted if the failure is to be avoided") cannot be generated, so that the practical application is limited. Engineering personnel need to know key influencing factors of faults and executable adjustment schemes, such as determining the most key risk sources in variables such as vibration, pressure, temperature or flow, but some existing attribution type methods (such as LIME, SHAP and the like) based on data can only sort the importance of the variables, can not generate movable suggestions meeting the physical rule of pump operation, can also generate interpretation results violating the characteristic curve of equipment or not meeting the safety range, and are difficult to use for engineering maintenance decision. Therefore, a new technology which can simultaneously meet the requirements of accurate prediction, physical constraint-based and executable adjustment schemes is urgently needed in the technical field of industrial pump fault diagnosis, so that scientific operation and maintenance decision basis is provided, and the operation reliability and maintenance efficiency are improved. Disclosure of Invention In view of this, the invention proposes an industrial pump interpretable intelligent fault detection method based on counterfactual learning to solve the problem that the industrial pump fault diagnosis method in the prior art is difficult to give executable accurate operation and maintenance suggestions. The specific technical scheme of the invention is as follows: Industrial pump interpretable intelligent fault detection method based on counterfactual learning, comprising: Constructing a fault detection model, wherein the fault detection model takes multi-source physical characteristics of the industrial pump during operation as input, and outputs a two-classification judging result of whether the industrial pump needs maintenance or not; Constructing a counterfactual search space defined based on mechanical characteristics, fluid dynamics rules and equipment operation constraints of the industrial pump for defining a physical feasible range for generating counterfactual samples; Generating a counterfactual sample in the counterfactual search space through an optimization process, wherein the optimization process minimizes the disturbance amplitude between the original working condition and the counterfactual sample, and meanwhile forces the output judging result of the failure detection model on the counterfactual sample to be opposite to the output judging result of the original sample; And outputting a structured interpretable result including physical variable difference analysis of the counterfactual sample and the original sample, sensitivity ordering of key physical variables, and anomaly source inference to provide engineering executable maintenance suggestions. Specifically, the fault detection model is realized by adopting a multi-layer feedforward neural network, the multi-layer feedforward neural network carries out normalization pretreat