Industrial pump interpretable intelligent fault detection method based on anti-fact learning

Abstract

The invention discloses an industrial pump interpretable intelligent fault detection method based on anti-fact learning, and relates to the technical field of industrial pump fault diagnosis, and the method comprises the steps: constructing a fault detection model; constructing an anti-fact search space for limiting a physical feasible range for generating an anti-fact sample; in the anti-fact search space, an anti-fact sample is generated through an optimization process, the optimization process minimizes the disturbance amplitude between the original working condition and the anti-fact sample, and meanwhile, the output judgment result of the fault detection model on the anti-fact sample is forced to be opposite to the output judgment result of the original sample; and outputting a structured interpretable result, the structured interpretable result including physical variable difference analysis of the anti-fact sample and the original sample, sensitivity ranking of the key physical variables and abnormal source inference, to provide engineering executable maintenance suggestions. According to the method, high-reliability and engineering executable pump fault diagnosis and explanation capability is realized.

Description

Technical Field

[0001] This invention relates to the field of industrial pump fault diagnosis technology, and in particular to an interpretable intelligent fault detection method for industrial pumps based on counterfactual learning. Background Technology

[0002] Industrial pumps are core power equipment in industrial process systems, and their operating status directly affects the stability, safety, and energy efficiency of the production process. A malfunction can lead to production interruptions, equipment damage, safety accidents, and significant economic losses.

[0003] With the development of sensor technology and data acquisition systems, modern industrial pumps are generally equipped with various types of sensors such as temperature, vibration, pressure, flow, and speed, which can capture operating data in real time and accurately, realize all-round dynamic monitoring, and provide support for the discovery of potential faults.

[0004] However, existing industrial pump fault diagnosis technologies have significant shortcomings. Traditional diagnostic methods, such as manual experience, threshold rules, or traditional statistical methods, rely on fixed rules and preset thresholds to determine faults. Industrial production environments are complex and variable, with frequent nonlinear characteristics in operating conditions. Fixed rules struggle to adapt to such changes, leading to false alarms and missed alarms, thus affecting production stability. Furthermore, while pump fault prediction models based on machine learning or deep learning can predict faults and output classification results (e.g., "fault status"), they are black-box models, unable to explain the predicted causes or provide root causes of faults, nor can they generate actionable engineering recommendations to avoid faults (e.g., "which physical quantities should be adjusted to avoid faults"), limiting their practical application. Engineers need to understand the key influencing factors of faults and feasible adjustment schemes, such as identifying the most critical risk sources among variables like vibration, pressure, temperature, or flow rate. However, some existing data-based attribution methods (e.g., LIME, SHAP) can only rank the importance of variables, failing to generate actionable recommendations that satisfy the physical laws of pump operation. They may also produce interpretations that violate equipment characteristic curves or do not meet safety limits, making them unsuitable for engineering maintenance decisions.

[0005] Therefore, there is an urgent need in the field of industrial pump fault diagnosis technology for a new technology that can simultaneously meet the requirements of accurate prediction, physical constraint-based approach, and provide executable adjustment solutions, in order to provide a scientific basis for operation and maintenance decisions and improve operational reliability and maintenance efficiency. Summary of the Invention

[0006] In view of this, the present invention proposes an interpretable intelligent fault detection method for industrial pumps based on counterfactual learning, in order to solve the problem that existing industrial pump fault diagnosis methods are difficult to provide actionable and accurate operation and maintenance suggestions.

[0007] The specific technical solution of this invention is as follows: An interpretable intelligent fault detection method for industrial pumps based on counterfactual learning includes: A fault detection model is constructed. The fault detection model takes the multi-source physical characteristics of the industrial pump during operation as input and outputs a binary classification judgment result on whether the industrial pump needs maintenance. A counterfactual search space is constructed, which is based on the mechanical characteristics, fluid dynamics laws and equipment operation constraints of industrial pumps, and is used to limit the physical feasible range for generating counterfactual samples. Within the counterfactual search space, counterfactual samples are generated through an optimization process. The optimization process minimizes the perturbation amplitude between the original operating conditions and the counterfactual samples, while forcing the fault detection model to make the opposite judgment on the counterfactual samples to the original samples. It also outputs structured and interpretable results, including physical variable difference analysis between counterfactual samples and original samples, sensitivity ranking of key physical variables, and inference of anomaly sources, to provide engineering-executable maintenance recommendations.

[0008] Specifically, the fault detection model is implemented using a multi-layer feedforward neural network. The multi-layer feedforward neural network performs normalization preprocessing on the multi-source physical features of the input and trains fixed parameters through supervised learning to output the probability judgment result of whether the industrial pump needs maintenance or not.

[0009] Specifically, the multi-source physical features include temperature, vibration amplitude, spatial pressure, flow rate, and rotational speed. These features are collected in real time from industrial pump sensors and used as vector inputs to the fault detection model.

[0010] Specifically, the counterfactual search space includes physical feasible boundary constraints, which limit the range of variation of each physical variable within the safe operating range of the equipment. These constraints include the material's allowable temperature range for temperature, the equipment's standard vibration limit for vibration amplitude, the system design pressure range for pressure, the pump's allowable flow range for flow rate, and the upper and lower limits of the drive equipment's rotational speed.

[0011] Specifically, the counterfactual search space includes data-driven relevant constraints. These constraints are based on statistical analysis of industrial pump operating data, defining the coupling relationships between physical variables, including the correlation between flow rate and pressure pump characteristic curves, the mutual influence between speed and flow rate, pressure and vibration, and the statistical correlation characteristics between temperature and vibration and pressure. The joint variation magnitude of counterfactual samples is limited by a coupling offset threshold.

[0012] Specifically, the counterfactual search space includes a minimum perturbation constraint, which requires that the difference between the generated counterfactual sample and the original operating condition be minimized. The perturbation magnitude is measured by norm distance to ensure that the counterfactual sample is as close as possible to the original state while satisfying the physical feasible boundary and relevant constraints.

[0013] Specifically, the optimization process includes a perturbation minimization term, which is achieved by minimizing the norm distance between the original operating conditions and the counterfactual samples, in order to quantify and reduce the overall magnitude of change in physical variables.

[0014] Specifically, the optimization process includes predicting the flip term, which forces the fault detection model to output the opposite result to the original judgment for counterfactual samples by constructing a flip loss function, and ensures that the output probability exceeds the confidence threshold to achieve reliable flipping.

[0015] Specifically, the optimization process includes a physical constraint preservation term, which introduces violations of physical feasible boundary constraints and data-driven related constraints through a penalty function to ensure that the counterfactual samples conform to the mechanical characteristics and fluid dynamics of industrial pumps.

[0016] Specifically, the structured interpretable results include: calculating the magnitude of disturbances in each physical quantity based on physical variable difference analysis and ranking them by sensitivity to identify key influencing variables; generating actionable maintenance recommendations based on anomaly sources, which describe how to adjust physical variables to eliminate failure risks, including recommendations for reducing vibration amplitude, increasing pressure, or controlling temperature.

[0017] The beneficial effects of this invention are as follows: (1) By constructing a fault detection model based on a feedforward neural network, the collaborative perception of multiple physical characteristics such as temperature, vibration, pressure, flow rate and speed is realized, which significantly improves the accuracy of industrial pump fault state judgment and establishes a highly reliable decision basis for subsequent interpretable analysis.

[0018] (2) Based on the mechanical characteristics and fluid dynamics of industrial pumps, a physical constraint space is constructed. Combined with the statistical correlation characteristics of data and the principle of minimum disturbance, the generated counterfactual samples strictly conform to the safety operation boundary of the equipment. This fundamentally avoids the misleading conclusions caused by the violation of physical laws in traditional interpretation methods, and makes the interpretation results feasible on site.

[0019] (3) By integrating the three-dimensional optimization objectives of disturbance minimization, predictive flipping forced and physical constraint maintenance, the minimum necessary physical variable changes that cause the fault criterion to reverse are accurately located, the core sensitive parameters affecting the equipment status are directly revealed, and the efficiency of fault root cause location is greatly improved.

[0020] (4) By presenting the sensitivity ranking of key physical quantities and the inference of abnormal sources in a structured manner, maintenance personnel can intuitively understand the fault formation mechanism, guide the formulation of maintenance plans for specific physical quantity abnormalities, effectively reduce the frequency of unnecessary equipment disassembly and inspection, and reduce maintenance costs. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating the interpretable intelligent fault detection method for industrial pumps based on counterfactual learning, as described in this invention. Detailed Implementation

[0023] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0024] This invention proposes an interpretable intelligent fault detection method for industrial pumps based on counterfactual learning, aiming to solve the challenges of fault prediction and root cause explanation in industrial pump operation status monitoring. This method combines an intelligent fault detection model with a counterfactual learning framework to not only determine whether the pump requires maintenance but also generate an executable explanation that can reverse the prediction result with minimal changes to physical variables, assisting maintenance personnel in locating the root cause of the fault. The specific implementation is as follows: Step 1: Construct an industrial pump fault detection model. This model uses physical characteristics such as temperature, vibration amplitude, spatial pressure, flow rate, and rotational speed as input vectors to determine whether the pump needs maintenance. The input feature vector is denoted as... ; in, For temperature, The amplitude of the vibration. For space pressure, For traffic, The rotational speed is used as the reference value. These features are preprocessed with normalization to ensure that different physical quantities are input on a uniform scale, thereby improving the comparability and interpretability of subsequent counterfactual optimization. The fault detection model uses a multilayer feedforward neural network (MLP) as the discriminant model, denoted as . It can be represented as: ; ; … ; ; in, It is a non-linear activation function (i.e., ReLU). For all trainable parameters of an MLP, This represents the probability of whether maintenance is needed in the current state. With threshold (Usually set to 0.5) The comparison will yield the final binary classification result: ; in, This indicates that "maintenance is required". This indicates "no maintenance required." The model training phase uses cross-entropy loss and supervised learning on a large number of labeled samples to obtain fixed parameters. During the online operation phase, the fault detection module maintains its parameters. Unchanged, for each input condition The output determines whether or not to maintain the product, providing a basis for subsequent counterfactual interpretations.

[0025] Step 2: After establishing the fault detection model, a counterfactual search space is constructed to ensure that the generated counterfactual samples meet engineering feasibility. This invention constructs a counterfactual search space for pump equipment based on the mechanical structural characteristics, fluid dynamics, and other physical laws of industrial pumps, as well as equipment operating constraints. This constraint is used to limit the scope of the counterfactual perturbation, ensuring that all generated candidate samples meet the physical feasibility requirements. The search space is primarily defined by three types of constraints, as follows: First, set physical feasibility boundary constraints, limiting each physical quantity to only vary within the safe range allowed by the equipment. For example, temperature. The material and bearing must meet the allowable temperature range; vibration amplitude The vibration must not exceed the maximum vibration limit specified in the equipment standard; pressure The flow rate must be greater than zero and within the system's designed operating pressure range; The flow rate must be between the pump's minimum and maximum allowable flow rates; the rotational speed... The upper and lower speed limits of the drive motor or frequency converter should be met. Formally, for any generated counterfactual feature... ,have in, and The lower and upper bounds of the physical parameters given in the equipment technical manual or operating specifications.

[0026] Industrial pumps have five categories of physical quantities with clearly defined physical upper limits and coupling relationships, such as the operating range of flow rate and pressure, the correlation between vibration and rotational speed, and the requirement that temperature must not exceed the equipment's allowable upper limit. This invention incorporates the upper and lower bounds of temperature, vibration, pressure, flow rate, and rotational speed, the correlation of pump characteristics, and the statistical coupling matrix into counterfactual optimization, ensuring that the generated counterfactual samples always conform to the pump's operating mechanism. Therefore, the output interpretation not only has mathematical significance but also possesses industrial feasibility and safety.

[0027] Secondly, considering the coupling relationship between physical quantities in actual operation, this invention introduces data-driven relevant constraints to avoid unreasonable counterfactual combinations. For example, flow rate. With pressure The speed follows the general rules of the pump characteristic curve; Changes will simultaneously affect traffic. With pressure and the vibration level Effects; temperature Features often exhibit a certain correlation with changes in vibration and pressure. Therefore, a correlation matrix between features can be obtained through statistical analysis of all samples. And thereby constrain counterfactual disturbances. The combined magnitude of change: ; in This is the original working condition vector. The counterfactual working condition vector, The allowed coupling offset threshold is used to ensure that the combined changes of the physical quantities still conform to the statistical correlation characteristics of the pump in actual operation.

[0028] Finally, to reflect the interpretability principle that minimal changes can lead to prediction reversal, this invention introduces a minimum perturbation constraint under the aforementioned physical boundary and coupling constraints, requiring the counterfactual sample to be as close as possible to the original operating condition. Specifically, this can be achieved by minimizing the difference between the original and counterfactual operating conditions. Norm implementation: ; in, Either 1 or 2 can be chosen. Under the premise of satisfying the physical boundary and coupling constraints, by solving the above minimization problem, the minimum necessary perturbation that flips the model's prediction can be obtained. In summary, this invention defines the counterfactual search space as: ; The counterfactual samples generated in this space not only conform to the physical feasibility of temperature, vibration, pressure, flow rate and speed, but also ensure that the disturbance amplitude is as small as possible, thus providing a credible and engineering-guided counterfactual explanation for industrial pump failure prediction.

[0029] Step 3, Construct an interpretable counterfactual learning model for fault detection: After completing the counterfactual search space After the initial setup is complete, the counterfactual generation module searches within the constraint space for the smallest perturbation sample that can reverse the model's prediction. Based on the previously constructed and trained MLP fault detection model, the counterfactual generation module aims to construct a new operating condition vector. To reverse the prediction results of MLP with minimal changes to the original physical quantities, i.e. ; Therefore, this invention models counterfactual generation as a constrained optimization problem, given the current sample and trained parameters In this case, numerical optimization algorithms are used to achieve the following: Search for the optimal counterfactual sample The overall optimization objective consists of three parts: a perturbation minimization term, a prediction reversal term, and a physical constraint preservation term. The perturbation minimization term ensures that the counterfactual samples are as close as possible to the original state; this invention uses the L2 norm to characterize the magnitude of the perturbation. ; The predicted flip term forces the MLP to make a judgment that the counterfactual sample is opposite to the original sample. This invention is based on the output probability. Constructing the flip loss: in, The desired confidence threshold is used to ensure that the prediction result is truly reversed rather than near the decision boundary. The physical constraint preservation term ensures that the counterfactual operating condition still satisfies the physical boundaries and coupling relationships of temperature, vibration, pressure, flow rate, and rotational speed. This invention introduces a physical constraint loss in the form of a penalty function: ; in, Physical quantities used to punish those that exceed their limits. The correlation matrix is ​​obtained based on physical feature estimation. Here are the weighting coefficients. Combining the above three parts, the counterfactual generation optimization problem of this invention can be expressed as: ; in, These are hyperparameters used to adjust the balance between the magnitude of the perturbation, the predicted flipping effect, and physical feasibility. Through the above optimization, the optimal counterfactual condition that satisfies the constraints can be obtained. Then, the magnitude of the perturbation is calculated for each physical quantity: ; This is used to characterize how much each physical variable needs to be changed to alter the maintenance decision of the MLP. Variables with larger absolute values ​​of disturbances have less impact on the model's decision-making, while variables that can cause a prediction reversal with a small disturbance are considered key sensitive variables, used for subsequent interpretability demonstrations and root cause analysis of failures.

[0030] Step 4: After completing fault detection and counterfactual condition generation, this invention further outputs structured and interpretable results to maintenance personnel. First, based on the judgment of the feedforward neural network model, the system provides a final prediction result of "the pump needs maintenance" or "the pump does not require maintenance." Then, the counterfactual interpretation module, based on the optimal counterfactual condition... Compared with the original working conditions The differences provide maintenance personnel with intuitive explanatory information, such as: "If the vibration amplitude is reduced by 12%, the model prediction will change from 'maintenance required' to 'maintenance not required'." Simultaneously, the magnitude of the system's disturbance to physical quantities (temperature, vibration, pressure, flow rate, rotational speed) is also considered. The system sorts these variables to identify the key variables most sensitive to the model's decisions. Based on the direction and magnitude of these variable changes, the system can further infer potential sources of anomalies, such as indicating that "the main influencing factor may be excessive pressure fluctuations inducing increased vibration," or "abnormal rise in pump body temperature combined with low flow rate leads to an increased risk of failure." Through this explanatory information, maintenance personnel can understand the physical causes determined by the model and implement targeted maintenance strategies accordingly.

[0031] The beneficial effects of this invention are as follows: (1) In step 1, by normalizing the key physical quantities of industrial pumps such as temperature, vibration, pressure, flow rate and speed, different physical signals are input into the model under a unified scale, thereby ensuring that the counterfactual disturbances are comparable and interpretable in the optimization process, and providing a stable input basis for subsequent reasoning; by constructing a fault detection model, it is possible to accurately determine whether the pump is in a maintenance-required state. (2) In step 2, a counterfactual learning framework is introduced to combine the multivariate monitoring data of the pump with physical feasible constraints. By constructing a counterfactual search space, minimum disturbance optimization, and executable operating condition adjustment suggestions, interpretable, operable, and implementable pump fault diagnosis is realized. The clear modeling of physical boundaries and coupling relationships ensures that the interpretation instructions will not exceed the actual operating conditions that the pump equipment can achieve. By generating counterfactual operating conditions that satisfy physical boundary conditions and variable coupling relationships, maintenance personnel can not only know "whether there is a fault", but also clarify "why there is a fault" and "how to eliminate the fault risk". (3) In step 3, optimize the target design so that the difference in output is This results in quantifiable and actionable recommendations for correcting operating conditions. For example, if a reduction in vibration of only 0.5 mm / s² can overturn a prediction, it directly indicates that "vibration" is the most sensitive physical quantity for the pump's current state. This is achieved by minimizing disturbances in counterfactual optimization. At the same time, it forces the prediction to change from "fault" to "normal", making the counterfactual sample It can intuitively provide the minimum physical adjustments required for the pump to recover to health under current operating conditions; minimum perturbation counterfactual optimization makes the interpretation "operable and critically sensitive". Minimum perturbation can not only find the minimum executable change that causes the prediction to reverse, but also indicate the sensitive variables near the discrimination boundary, so that maintenance personnel can understand the true risk level of the pump. For example, when the temperature of a pump is only slightly perturbed from the abnormal range, while the vibration needs to change significantly to change the model judgment, minimum perturbation can point out that "the temperature is close to the critical risk point", thereby strengthening the early warning capability of anomalies. This effect comes entirely from the mathematical properties of counterfactual optimization, rather than contribution methods such as SHAP, because traditional importance methods cannot measure "how much change will lead to a change in the prediction". (4) In step 4, the counterfactual disturbance is converted into an operating condition adjustment suggestion that is easy for maintenance personnel to understand through the interpretable output module, realizing a closed-loop output from "model judgment" to "engineering instructions". For example, output "If the vibration is reduced by about 8%, the pump can be restored to normal from the fault judgment", or "The pressure needs to be increased by at least 12 kPa to avoid potential risks". These explanations can be directly used for industrial maintenance decisions, which greatly improves maintenance efficiency and positioning accuracy.

[0032] In summary, this invention achieves "explainable, executable, physically consistent, and directly guides maintenance" diagnostic capabilities in the field of industrial pump fault diagnosis, which is significantly superior to traditional attribution-based interpretable methods and black-box models.

[0033] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An interpretable intelligent fault detection method for industrial pumps based on counterfactual learning, characterized in that, include: A fault detection model is constructed, which takes the multi-source physical characteristics of the industrial pump during operation as input and outputs a binary classification judgment result on whether the industrial pump needs maintenance. A counterfactual search space is constructed, which is based on the mechanical characteristics, fluid dynamics laws and equipment operation constraints of industrial pumps, and is used to limit the physical feasible range for generating counterfactual samples; Within the counterfactual search space, counterfactual samples are generated through an optimization process. This optimization process minimizes the perturbation amplitude between the original operating conditions and the counterfactual samples, while simultaneously forcing the fault detection model to output the opposite judgment result of the counterfactual samples to the original samples. It also outputs structured and interpretable results, including physical variable difference analysis between the counterfactual sample and the original sample, sensitivity ranking of key physical variables, and inference of anomaly sources, to provide engineering-executable maintenance recommendations.

2. The industrial pump interpretable intelligent fault detection method based on counterfactual learning as described in claim 1, characterized in that, The fault detection model is implemented using a multi-layer feedforward neural network. The multi-layer feedforward neural network performs normalization preprocessing on the input multi-source physical features and trains fixed parameters through supervised learning to output the probability judgment result of whether the industrial pump needs maintenance or not.

3. The industrial pump interpretable intelligent fault detection method based on counterfactual learning as described in claim 1, characterized in that, The multi-source physical features include temperature, vibration amplitude, spatial pressure, flow rate, and rotational speed. These features are collected in real time from industrial pump sensors and used as vector inputs to the fault detection model.

4. The industrial pump interpretable intelligent fault detection method based on counterfactual learning as described in claim 1, characterized in that, The counterfactual search space includes physical feasible boundary constraints, which limit the range of variation of each physical variable within the safe operating range of the equipment. These constraints include the material's allowable temperature range for temperature, the equipment's standard vibration limit for vibration amplitude, the system design pressure range for pressure, the pump's allowable flow range for flow rate, and the upper and lower limits of the drive equipment's rotational speed.

5. The industrial pump interpretable intelligent fault detection method based on counterfactual learning as described in claim 1, characterized in that, The counterfactual search space includes data-driven relevant constraints, which are based on statistical analysis of industrial pump operating data. These constraints define the coupling relationships between physical variables, including the correlation between pump characteristic curves of flow rate and pressure, the mutual influence between speed and flow rate, pressure and vibration, and the statistical correlation characteristics between temperature and vibration and pressure. The combined variation of counterfactual samples is limited by a coupling offset threshold.

6. The industrial pump interpretable intelligent fault detection method based on counterfactual learning as described in claim 1, characterized in that, The counterfactual search space includes a minimum perturbation constraint, which requires that the difference between the generated counterfactual sample and the original operating condition be minimized. The perturbation magnitude is measured by norm distance to ensure that the counterfactual sample is as close as possible to the original state while satisfying the physical feasible boundary and relevant constraints.

7. The industrial pump interpretable intelligent fault detection method based on counterfactual learning as described in claim 1, characterized in that, The optimization process includes a perturbation minimization term, which is achieved by minimizing the norm distance between the original operating conditions and the counterfactual samples, in order to quantify and reduce the overall variation of physical variables.

8. The industrial pump interpretable intelligent fault detection method based on counterfactual learning as described in claim 1, characterized in that, The optimization process includes predicting a flip term, which forces the fault detection model to output results opposite to the original judgment for counterfactual samples by constructing a flip loss function, and ensures that the output probability exceeds the confidence threshold to achieve reliable flipping.

9. The industrial pump interpretable intelligent fault detection method based on counterfactual learning as described in claim 1, characterized in that, The optimization process includes a physical constraint preservation term, which introduces violations of physical feasible boundary constraints and data-driven related constraints through a penalty function to ensure that counterfactual samples conform to the mechanical characteristics and fluid dynamics of industrial pumps.

10. The industrial pump interpretable intelligent fault detection method based on counterfactual learning as described in claim 1, characterized in that, The structured interpretable results include: calculating the magnitude of disturbances in each physical quantity based on the difference analysis of the physical variables, and ranking them by sensitivity to identify key influencing variables; inferring executable maintenance recommendations based on the sources of anomalies, the maintenance recommendations describing how to adjust physical variables to eliminate the risk of failure, including recommendations for reducing vibration amplitude, increasing pressure, or controlling temperature.

Images