Fault diagnosis method and health evaluation method for mud pump water containment sensing system

Abstract

The application provides a mud pump sealing water sensing system fault diagnosis method and health evaluation method, by inputting the current sensing parameter value of the mud pump sealing water sensing system into the virtual sensor model based on PINN, the first sealing water pressure prediction value output by the virtual sensor model is obtained, the virtual sensor model is obtained by training the sealing water pressure dynamic output equation obtained by establishing a physical information model as a physical constraint; according to the actual power, pump efficiency and flow of the sealing water pump, the sealing water pressure dynamic output equation is used to calculate the second sealing water pressure prediction value; the mud pump sealing water pressure sensing value of the mud pump sealing water sensing system is compared with the first sealing water pressure prediction value and the second sealing water pressure prediction value, whether the mud pump sealing water sensing system is faulty is determined. The application makes the fault diagnosis understandable and interpretable by physical principles, and the diagnosis logic is more intuitive and credible.

Description

Technical Field

[0001] This invention relates to the field of dredging and mud pump technology, and in particular to a fault diagnosis method and health assessment method for a mud pump sealing water sensing system. Background Technology

[0002] The main function of the mud pump sealing water system is to provide appropriate sealing water to prevent mud particles from flowing back into the gap between the impeller and the liner, the pump shaft seal, and the suction end seal, thereby reducing wear on the mud pump liner and impeller. However, the system operates in harsh environments year-round, leading to unstable data transmission from the sensors in its sensing system, making it difficult to control the real-time operating status of the mud pump. On the other hand, although artificial intelligence models introduced in recent years have made some progress in predictive performance, the degradation and decreased reliability of the mud pump sealing water sensing system originate internally but manifest externally. Therefore, it generally suffers from severe "black box" problems, uninterpretable model outputs, and unverifiable results.

[0003] Currently, fault diagnosis and health assessment of sensors in large rotating machinery often employ data-driven methods such as machine learning and deep learning. In sensor fault diagnosis, the data is first categorized into fault labels. Then, the self-learning capability of neural networks is utilized to optimize the structure and parameters of the neural network model, increase the number of iterations, and converge the loss curve, thereby completing the identification and classification of fault data and ultimately improving the accuracy of prediction and classification. In sensing system health assessment, the focus is typically on modeling the normal operating state and quantifying degradation trends. Neural networks are often used to learn from multi-sensor time-series signals in a large amount of fault-free historical operating data. By setting statistical thresholds or learning to predict residual distributions, the model can calculate the deviation between the real-time sensor data stream and the learned "health baseline," quantifying this deviation as a continuous health index or remaining service life estimate. While these methods offer high accuracy in calculation results, their "lack of interpretability" severely impacts user trust and acceptance of the diagnostic results.

[0004] Numerous studies have been conducted on fault diagnosis and health assessment of sensing systems. Among these, mainstream methods for fault diagnosis rely on deep neural networks, particularly Long Short-Term Memory (LSTM) networks or Gated Recurrent Units (GRUs) for complex temporal signals, and Convolutional Neural Networks (CNNs) adept at capturing spatial features. The core process involves first labeling historical sensor data with faults based on expert experience or specific rules, and then training the network using the labeled dataset. During training, hyperparameters such as the number of network layers, nodes, and optimizers are adjusted to increase the number of training iterations, while monitoring the convergence of the loss function. The goal is for the network to automatically learn discriminative features from complex multi-sensor, multi-channel input data to differentiate between different fault types, ultimately outputting a high-confidence classification prediction of the fault category of the current data. This method often achieves high classification accuracy and recognition precision / recall when there is sufficient labeled data and complete fault mode coverage.

[0005] Health status assessment is widely used to evaluate overall system degradation or predict remaining lifespan (RUL). Unsupervised or self-supervised neural network models, such as autoencoders and their variants (e.g., denoising autoencoders), are widely applied. The core idea is to train the autoencoder using a large amount of sensor data from the system's normal operating state (health state), enabling it to learn the ability to reconstruct the original input and capture the potential distribution and core features of the "healthy" data. During deployment, real-time sensor data is input, and the model calculates its reconstruction error or prediction error (for time-series prediction models). When the reconstruction error significantly deviates from the statistical range of the training set (health baseline) (e.g., exceeding a set threshold), it is considered that the system's health state is abnormal or degraded. By continuously monitoring the magnitude or trend of the error, continuous quantitative assessment of the system's health level (e.g., the Health Index) or early warning of anomalies can be achieved.

[0006] The existing problems and shortcomings include:

[0007] (1) Extreme dependence on data quantity, quality, and annotation:

[0008] High accuracy heavily relies on large-scale, high-quality labeled data: Fault diagnosis models require a large amount of accurately labeled data covering all possible fault modes, operating condition changes, and degradation stages for training. However, obtaining sufficient, balanced, and accurate fault label data is extremely difficult and costly in reality (especially for novel or rare faults). For health assessment, although fault labels are not required, the requirements for the amount of baseline data representing "health" and the coverage of the data (covering all normal operating conditions) are extremely high.

[0009] Environmental interference can easily lead to model failure: The harsh environments in which sensors operate, such as high wear, vibration, and corrosion, result in raw data containing a large amount of noise, drift, transient anomalies, and intermittent failures. Purely data-driven models cannot distinguish whether data anomalies are caused by system failures, sensor problems, or environmental interference, making them highly susceptible to "dirty data." This leads to the model learning noisy patterns, poor robustness, false alarms / false negatives, and a significant performance degradation when data quality fluctuates.

[0010] Feature learning is highly black-boxed: the features learned by the model to "distinguish faults" or "health patterns" are often abstract, high-dimensional representations resulting from complex transformations at deep network layers. These features lack clear physical meaning or engineering interpretation, making it difficult for engineers to understand why the model classifies certain data patterns as specific faults or health degradation. When the model malfunctions, it is difficult to trace the root cause of the error and make targeted improvements.

[0011] (2) Weak generalization ability and difficulty in transfer:

[0012] The model is "helpless" when faced with unseen operating conditions or new faults: It is highly dependent on the distribution of the training data (the operating conditions and fault types it covers). When the system operates under new operating conditions not covered by the training data (such as extreme loads, new materials or components) or encounters entirely new fault modes not included in the training (such as multiple faults occurring concurrently or complex degradation paths), the model's performance usually drops precipitously, and may even give completely unreliable diagnostic results or health assessments.

[0013] Ignoring potential failure mechanisms: Because the method completely disregards the inherent physical degradation mechanisms of the sensor system, it only learns statistical correlations between input and output labels or health indicators, rather than physical causal relationships. This prevents the model from fundamentally understanding or predicting the root causes of sensor failure, weakening its ability to specifically predict and prevent potential failure modes. Summary of the Invention

[0014] This invention provides a fault diagnosis method and a health assessment method for a mud pump sealing water sensing system, which addresses the shortcomings of existing research on fault diagnosis and health assessment of mud pump sealing water sensing systems, such as poor interpretability, low generalization, and severe black box nature. It realizes an interpretable fault diagnosis and health assessment method for mud pump sealing water sensing systems, thereby improving the reliability, interpretability, and versatility of mud pump sealing water sensing systems.

[0015] This invention provides a fault diagnosis method for a mud pump sealing water sensing system, comprising:

[0016] The current sensing parameter values ​​of the mud pump sealing water sensing system are input into the PINN-based virtual sensor model to obtain the first sealing water pressure prediction value output by the virtual sensor model. The virtual sensor model is trained by using the sealing water pressure dynamic output equation obtained from the physical information model as a physical constraint.

[0017] Based on the actual power, pump efficiency, and flow rate of the sealing pump, the predicted value of the second sealing pressure is calculated using the dynamic output equation of the sealing pressure.

[0018] The mud pump sealing pressure sensing value of the mud pump sealing pressure sensing system is compared with the first sealing pressure prediction value and the second sealing pressure prediction value to determine whether the mud pump sealing pressure sensing system is faulty.

[0019] According to the fault diagnosis method of a mud pump sealing water sensing system provided by the present invention, the current sensing parameter values ​​of the mud pump sealing water sensing system are input into a PINN-based virtual sensor model to obtain a first predicted value of sealing water pressure output by the virtual sensor model, including:

[0020] The historical sensing parameter values ​​of the mud pump sealing pressure sensing system were analyzed using XGBoost and Random Forest algorithms, and a preset number of sensing parameters most relevant to the mud pump sealing pressure were selected.

[0021] The MCNN-Transformer model is used to predict the current sensing parameter values ​​of the selected sensing parameters, and the first predicted value of the sealing pressure output by the MCNN-Transformer model is obtained.

[0022] According to the fault diagnosis method of a mud pump sealing water sensing system provided by the present invention, before inputting the current sensing parameter values ​​of the mud pump sealing water sensing system into a PINN-based virtual sensor model to obtain the first sealing water pressure prediction value output by the virtual sensor model, the method further includes:

[0023] Based on the difference between the first sealing pressure predicted by the virtual sensor model for the sample data and the corresponding actual measured sealing pressure, the data loss term of the virtual sensor model is determined.

[0024] Based on the difference between the predicted second sealing pressure value of the sample data and the corresponding actual measured sealing pressure value according to the sealing pressure dynamic output equation, the physical loss term of the virtual sensor model is determined.

[0025] Based on the physical loss term and data loss term of the virtual sensor model, the total loss function of the virtual sensor model is determined, and the virtual sensor model is trained based on the total loss function.

[0026] According to the fault diagnosis method of the mud pump sealing water sensing system provided by the present invention, the physical loss term of the virtual sensor model is determined by the difference between the predicted value of the second sealing water pressure of the sample data and the corresponding actual measured value of the sealing water pressure based on the dynamic output equation of the sealing water pressure using the following formula. :

[0027]

[0028] in, The number of data entries in the sample data. For the first The actual power of the sealing pump in the sample data. For the first The pump efficiency of the sealing pump in the sample data. For the first The flow rate of the sealing pump in the sample data. For the first The actual measured value of the sealing pressure corresponding to each sample data point.

[0029] According to the fault diagnosis method of the mud pump sealing water sensing system provided by the present invention, the total loss function of the virtual sensor model is determined by the following formula based on the physical loss term and data loss term of the virtual sensor model. :

[0030]

[0031] in, This is the physical loss term of the virtual sensor model. This is the data loss term of the virtual sensor model. It is an adjustment parameter within the range of [0,1].

[0032] According to the present invention, a fault diagnosis method for a mud pump sealing water sensing system compares the mud pump sealing water pressure sensing value of the mud pump sealing water sensing system with a first sealing water pressure prediction value and a second sealing water pressure prediction value to determine whether the mud pump sealing water sensing system is faulty, including:

[0033] The first and second predicted sealing pressure values ​​are combined to obtain the combined sealing pressure value.

[0034] Residual threshold analysis is performed on the mud pump sealing pressure sensing value and the fused sealing pressure value to determine whether the mud pump sealing pressure sensing system is faulty.

[0035] According to a fault diagnosis method for a mud pump sealing water sensing system provided by the present invention, the mud pump sealing water pressure sensing value of the mud pump sealing water sensing system is compared with the first sealing water pressure prediction value and the second sealing water pressure prediction value using the following formula to determine whether the mud pump sealing water sensing system is faulty:

[0036]

[0037]

[0038] in, It is the combined value of sealing water pressure. This is the predicted value of the first sealing pressure. This is the predicted value of the second sealing pressure. and These are the weighting coefficients. It is the pressure sensing value of the mud pump seal water. It is the residual between the mud pump sealing water pressure sensing value and the sealing water pressure fusion value;

[0039] If the residual is greater than a first preset threshold, the mud pump sealing water sensing system is determined to be faulty.

[0040] According to the fault diagnosis method of the mud pump sealing water sensing system provided by the present invention, after comparing the mud pump sealing water pressure sensing value of the mud pump sealing water sensing system with the first sealing water pressure prediction value and the second sealing water pressure prediction value to determine whether the mud pump sealing water sensing system is faulty, the method further includes:

[0041] If the duration of the failure of the mud pump sealing water sensing system reaches a second preset threshold, the mud pump sealing water sensing system is determined to be a transient failure.

[0042] If the duration of the failure of the mud pump sealing water sensing system is greater than the second preset threshold and less than or equal to the third preset threshold, the mud pump sealing water sensing system is determined to be an intermittent failure, and the third preset threshold is greater than the second preset threshold.

[0043] If the duration of the fault in the mud pump sealing water sensing system exceeds the third preset threshold, the mud pump sealing water sensing system is determined to be a permanent fault.

[0044] This invention also provides a health assessment method for a mud pump sealing water sensing system based on the above-mentioned fault diagnosis method for the mud pump sealing water sensing system, comprising:

[0045] Determine the total number of failures and the duration of each failure in the mud pump sealing water sensing system;

[0046] The health degradation index of the mud pump sealing water sensing system is determined based on the total number of failures and the duration of each failure.

[0047] The health status of the mud pump sealing water sensing system is determined based on the aforementioned health degradation indicators.

[0048] According to the health assessment method of a mud pump sealing water sensing system provided by the present invention, the health degradation index of the mud pump sealing water sensing system is determined by the following formula based on the total number of failures and the duration of each failure. :

[0049]

[0050]

[0051]

[0052] in, The total number of failures, This represents the sum of the fault durations corresponding to the total number of faults. For the first Duration of the second fault. The fault intensity coefficient is determined based on the mean time between failures of the mud pump sealing water sensing system under rated operating conditions.

[0053] The fault diagnosis method and health assessment method for the mud pump sealing water sensing system provided by this invention deeply integrates prior physical knowledge with data-driven approaches. By constructing physical constraints and embedding them into model training, the model is guided to conform to physical mechanisms. Physical laws are explicitly used as constraints, making the prediction and diagnosis results physically consistent. The results are easier to understand and explain using known physical principles. The theoretical prediction values ​​output by the physical model, the prediction values ​​of virtual sensors, and the actual sensing data are compared for diagnosis, making the basis for fault diagnosis clearer and the diagnostic logic more intuitive and reliable. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0055] Figure 1 This is a flowchart illustrating the fault diagnosis method for the mud pump sealing water sensing system provided by the present invention.

[0056] Figure 2 This is a complete flowchart of the fault diagnosis and health assessment method for the mud pump sealing water sensing system provided by the present invention;

[0057] Figure 3 This is a schematic diagram of the overall structure of the virtual sensor model based on PINN in the fault diagnosis method of the mud pump sealing water sensing system provided by the present invention.

[0058] Figure 4 This is a schematic diagram of the data-based sealing pressure prediction framework in the mud pump sealing water sensing system fault diagnosis method provided by the present invention;

[0059] Figure 5 This is a schematic diagram of the MCNN-Transformer neural network model structure in the mud pump sealing water sensing system fault diagnosis method provided by the present invention;

[0060] Figure 6 This is a flowchart illustrating the health assessment method for the mud pump sealing water sensing system provided by the present invention.

[0061] Figure 7 This is a schematic diagram of the structure of the health assessment model of the mud pump sealing water sensing system in the health assessment method of the mud pump sealing water sensing system provided by the present invention. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0063] The following is combined Figure 1 A fault diagnosis method for a mud pump sealing water sensing system according to the present invention includes:

[0064] Step 101: Input the current sensing parameter values ​​of the mud pump sealing water sensing system into the virtual sensor model based on PINN (Physics-Informed Neural Networks) to obtain the first sealing water pressure prediction value output by the virtual sensor model. The virtual sensor model is obtained by training the sealing water pressure dynamic output equation obtained by establishing the physical information model as a physical constraint.

[0065] Step 102: Calculate the predicted value of the second sealing pressure using the sealing pressure dynamic output equation based on the actual power, pump efficiency, and flow rate of the sealing pump.

[0066] Step 103: Compare the mud pump sealing pressure sensing value of the mud pump sealing water sensing system with the first sealing pressure prediction value and the second sealing pressure prediction value to determine whether the mud pump sealing water sensing system is faulty.

[0067] Interpretable fault diagnosis methods for mud pump sealing water sensing systems include prior physical information models, PINN-based virtual sensor models, and fault diagnosis models based on physical-virtual data superposition residuals. A complete flowchart is shown below. Figure 2 As shown.

[0068] The prior physical information model is used to construct a physical model related to the operating mechanism of the sensing parameters of the mud pump sealing system. Its core lies in extracting prior physical knowledge (such as fluid dynamics equations, thermodynamic equations, and kinetic equations) from the system based on expert knowledge, and coupling out interpretable physical constraints for subsequent model training and constraints. This aims to ensure that the results of fault diagnosis and assessment have physical consistency and interpretability, thereby effectively alleviating the "black box" problem caused by data-driven models.

[0069] The core of the PINN-based virtual sensor model includes: an innovative MCNN-Transformer neural network prediction model, incorporating a loss constraint function based on the system's physical laws, and a feedback mechanism to enhance adaptability, such as... Figure 3 As shown, this model fully utilizes the neural network structure to capture pattern features in sensor data; simultaneously, by explicitly embedding constraint terms constructed from physical laws into the loss function, it guides the model to learn mapping relationships that conform to physical mechanisms. The data fitting error term and the physical constraint error term in the model are jointly trained and optimized, thereby synergistically driving the model to converge quickly and significantly improving its generalization ability (robustness) and the interpretability of the decision-making process in fault diagnosis tasks.

[0070] The core function of the fault diagnosis model based on physical-virtual data superposition residuals is to accurately identify transient, intermittent, and permanent sensor faults by calculating and analyzing the residuals between actual sensed data and fused predicted values ​​in real time.

[0071] In this implementation case, the pressure sensor in the dredging vessel mud pump sealing water sensing system can be selected as the typical sensor experimental object:

[0072] First, a priori physical information model is constructed, and then the dynamic output equation of the sealing pressure is obtained. Based on historical empirical formulas and the physical mechanism of centrifugal pumps, the outlet pressure of the sealing pump is the sealing pressure. As a type of centrifugal pump, the outlet pressure of the sealing pump is related to the actual power, pump efficiency, and flow rate, satisfying… ,in For sealing water pressure, This is the actual power. For pump efficiency, For traffic, This is a dynamic disturbance term.

[0073] By leveraging prior physical knowledge to empower neural network models and embedding physical constraints into these models, we can achieve collaborative modeling of mechanisms and data, thereby improving the accuracy, generalization, stability, and interpretability of fault diagnosis models. This approach has significant engineering implications and academic value.

[0074] This embodiment deeply integrates prior physical knowledge with data-driven approaches. By constructing physical constraints and embedding them into model training, it guides the model to conform to physical mechanisms, explicitly using physical laws as constraints. This ensures that the prediction and diagnosis results have physical consistency, making the results easier to understand and explain using known physical principles. By comparing the theoretical predictions output by the physical model, the predictions from virtual sensors, and the actual sensing data, the basis for fault diagnosis becomes clearer, and the diagnostic logic becomes more intuitive and reliable.

[0075] Based on the above embodiments, this embodiment inputs the current sensing parameter values ​​of the mud pump sealing water sensing system into a PINN-based virtual sensor model to obtain the first sealing water pressure prediction value output by the virtual sensor model, including:

[0076] The historical sensing parameter values ​​of the mud pump sealing pressure sensing system were analyzed using XGBoost and Random Forest algorithms, and a preset number of sensing parameters most relevant to the mud pump sealing pressure were selected.

[0077] The MCNN-Transformer model is used to predict the current sensing parameter values ​​of the selected sensing parameters, and the first predicted value of the sealing pressure output by the MCNN-Transformer model is obtained.

[0078] The construction of a PINN-based virtual sensor model is based on the following assumptions:

[0079] Assumption 1: The system contains a sensor array. Each sensor The fault state is determined by a binary variable. ( (This represents a fault). In fault diagnosis, it is assumed that the accuracy of the virtual sensor (based on feature correlation and regression models) depends on the absence of faults in other associated sensors, as described by the following mathematical description:

[0080]

[0081]

[0082] In the formula, It is a sensor The associated sensor set, It is a regression prediction function based on feature correlation and regression model. It is a predicted value. It is a sensor The perceived value, if Then it is necessary to ensure .

[0083] Then, a virtual sensor model based on PINN is constructed. Data is collected using a multi-source sensor sensing system, and feature analysis is performed by superimposing XGBoost and Random Forest algorithms. Ten sensing parameters strongly correlated with the sealing pressure are selected. The MCNN-Transformer model is then used to predict the parameters with high feature correlation. That is, the values ​​of the selected sensing parameters are sequentially input into the MCNN and Transformer models to obtain the predicted values ​​of the sealing pressure. The specific prediction process is as follows: Figure 4 and Figure 5 As shown.

[0084] Based on the above embodiments, this embodiment further includes the following step before inputting the current sensing parameter values ​​of the mud pump sealing water sensing system into the PINN-based virtual sensor model to obtain the first sealing water pressure prediction value output by the virtual sensor model:

[0085] Based on the difference between the first sealing pressure predicted by the virtual sensor model for the sample data and the corresponding actual measured sealing pressure, the data loss term of the virtual sensor model is determined.

[0086] Based on the difference between the predicted second sealing pressure value of the sample data and the corresponding actual measured sealing pressure value according to the sealing pressure dynamic output equation, the physical loss term of the virtual sensor model is determined.

[0087] Based on the physical loss term and data loss term of the virtual sensor model, the total loss function of the virtual sensor model is determined, and the virtual sensor model is trained based on the total loss function.

[0088] Based on the above embodiments, this embodiment uses the following formula to determine the physical loss term of the virtual sensor model based on the difference between the predicted second sealing pressure value and the corresponding actual measured sealing pressure value of the sample data according to the sealing pressure dynamic output equation. :

[0089]

[0090] in, The number of data entries in the sample data. For the first The actual power of the sealing pump in the sample data. For the first The pump efficiency of the sealing pump in the sample data. For the first The flow rate of the sealing pump in the sample data. For the first The actual measured value of the sealing pressure corresponding to each sample data point.

[0091] The dynamic output equation of sealing pressure is embedded into the Transformer model as a physical constraint, and the physical constraint loss term is constructed as shown in formula (3).

[0092] Based on the above embodiments, this embodiment determines the total loss function of the virtual sensor model according to the physical loss term and data loss term of the virtual sensor model using the following formula. :

[0093]

[0094] in, This is the physical loss term of the virtual sensor model. This is the data loss term of the virtual sensor model. It is an adjustment parameter within the range of [0,1].

[0095] It varies with the number of iterations, especially in the early stages of iterative training. Larger values ​​make the model focus on recognizing prior physical knowledge. In the later stages of training, to improve iteration speed... By gradually decreasing the value, the model focuses on understanding the training results of the neural network, thereby improving the overall prediction accuracy, convergence speed, and interpretability.

[0096] Subsequently, by introducing a physical constraint mechanism to optimize the virtual sensor model of the sealing pressure, a prediction value can be obtained that balances the interpretability of prior physical knowledge with the accuracy of neural network algorithms under big data.

[0097] Based on the above embodiments, this embodiment compares the mud pump sealing pressure sensing value of the mud pump sealing pressure sensing system with the first sealing pressure prediction value and the second sealing pressure prediction value to determine whether the mud pump sealing pressure sensing system is faulty, including:

[0098] The first and second predicted sealing pressure values ​​are combined to obtain the combined sealing pressure value.

[0099] Residual threshold analysis is performed on the mud pump sealing pressure sensing value and the fused sealing pressure value to determine whether the mud pump sealing pressure sensing system is faulty.

[0100] The fused prediction value is calculated from the following two parts using adaptive weighting: 1) a predictive value with a certain degree of interpretability output from a virtual sensor model constructed based on a Physical Information Neural Network (PINN); 2) a theoretical prediction value directly calculated based on the system's inherent physical conservation equations. The diagnostic method, by setting interpretability residual threshold boundaries and combining this with quantitative analysis of the duration of residual signals exceeding the threshold, classifies and diagnoses the specific type (transient, intermittent, or permanent) of sensor faults according to preset temporal characteristic judgment rules.

[0101] The prerequisite for constructing a fault diagnosis model based on physical-virtual data superposition residuals is that the following assumptions must be met:

[0102] Assumption 2: Use a rectangular function to represent the fault activation state at a certain time period:

[0103]

[0104] set up For the first The moment the fault began, it continued Seconds. Furthermore, the three fault characteristics of the rectangular sensor conform to the following mathematical description:

[0105]

[0106] in, For the first The abnormal amplitude of the secondary fault, when the pulse width If it lasts for 3 seconds, it indicates a transient fault; when the pulse width... If the fault lasts for more than 3 seconds and is not permanent, it indicates an intermittent fault; when the pulse width... If the duration is infinitely long, it will manifest as a permanent failure.

[0107] Based on the above embodiments, this embodiment compares the mud pump sealing pressure sensing value of the mud pump sealing pressure sensing system with the first sealing pressure prediction value and the second sealing pressure prediction value using the following formula to determine whether the mud pump sealing pressure sensing system is faulty:

[0108]

[0109]

[0110] in, It is the combined value of sealing water pressure. This is the predicted value of the first sealing pressure. This is the predicted value of the second sealing pressure. and These are the weighting coefficients. It is the pressure sensing value of the mud pump seal water. It is the residual between the mud pump sealing water pressure sensing value and the sealing water pressure fusion value;

[0111] If the residual is greater than a first preset threshold, the mud pump sealing water sensing system is determined to be faulty.

[0112] The data of the sensing system is collected in real time by the data acquisition system, including key parameters such as sealing pressure, actual power, pump efficiency and flow rate. The actual sealing pressure sensor data is compared and analyzed with the predicted value of physical-virtual data superposition, and residual threshold is set for fault diagnosis, as shown in formula (7) and formula (8).

[0113] in Through The physical prediction value calculated by this formula The predicted value is obtained through a virtual sensor model based on PINN. We can assign a weight of 0.5 to each of the two values ​​to obtain... Later compared with the actual perceived sealing pressure Perform residual threshold analysis, if If the value exceeds the first preset threshold, such as 0.25, the current sensing data is determined to be abnormal data. This abnormality is recorded and transmitted to the mud pump sealing water sensing system health assessment model for assessing the health status of the mud pump sealing water sensing system.

[0114] Based on the above embodiments, this embodiment, after comparing the mud pump sealing pressure sensing value of the mud pump sealing pressure sensing system with the first sealing pressure prediction value and the second sealing pressure prediction value to determine whether the mud pump sealing pressure sensing system is faulty, further includes:

[0115] If the duration of the failure of the mud pump sealing water sensing system reaches a second preset threshold, the mud pump sealing water sensing system is determined to be a transient failure.

[0116] If the duration of the failure of the mud pump sealing water sensing system is greater than the second preset threshold and less than or equal to the third preset threshold, the mud pump sealing water sensing system is determined to be an intermittent failure, and the third preset threshold is greater than the second preset threshold.

[0117] If the duration of the fault in the mud pump sealing water sensing system exceeds the third preset threshold, the mud pump sealing water sensing system is determined to be a permanent fault.

[0118] like Figure 6 As shown, this embodiment also provides a health assessment method for a mud pump sealing water sensing system based on the mud pump sealing water sensing system fault diagnosis method in any of the above embodiments, including:

[0119] Step 601: Determine the total number of failures and the duration of each failure in the mud pump sealing water sensing system;

[0120] Step 602: Determine the health degradation index of the mud pump sealing water sensing system based on the total number of failures and the duration of each failure;

[0121] Step 603: Determine the health status of the mud pump sealing water sensing system based on the health degradation indicators.

[0122] like Figure 7 As shown, the health assessment model of the mud pump sealing water sensing system mainly includes: a fault frequency assessment module, a fault cumulative duration assessment module, and a system reliability calculation module. A fault diagnosis model based on physical-virtual data fusion residuals, deployed on a real-time processing platform, generates time-series fault diagnosis results. Based on these diagnostic results, the health assessment model performs fault event statistics (including total frequency and cumulative duration). Based on this statistical data, the system reliability calculation module constructs a dynamic reliability assessment model, calculates the state curve characterizing the performance degradation of the sensing system, and determines the severity level of the current fault accordingly.

[0123] This embodiment assesses health status based on the frequency and duration of faults, ensuring consistency, coherence, and dynamism between diagnostic results and health status assessments, thus providing more accurate and comprehensive decision support for predictive maintenance.

[0124] Based on the above embodiments, this embodiment determines the health degradation index of the mud pump sealing water sensing system using the following formula, based on the total number of failures and the duration of each failure. :

[0125]

[0126]

[0127]

[0128] in, The total number of failures, This represents the sum of the fault durations corresponding to the total number of faults. For the first Duration of the second fault. The fault intensity coefficient is determined based on the mean time between failures of the mud pump sealing water sensing system under rated operating conditions.

[0129] After completing the sensor fault diagnosis, the final step is to assess the health status of the mud pump's sealing water sensing system in real time. Each fault occurrence is transmitted to the fault frequency assessment module and recorded. The duration of each fault occurrence is defined as the fault duration. Input the frequency and duration of fault occurrences into the fault cumulative duration evaluation module, and output the sum of the products of the total number of faults and the duration of each fault. As shown in formula (10).

[0130] Then and As input data, it is transmitted to the system reliability calculation module, and its health degradation index (HDI) is defined as shown in formula (9). The calculation is shown in formula (11).

[0131] A performance degradation statistical model is constructed based on historical data throughout the sensor's lifecycle. The degradation process is analyzed as a state evolution comprising four significant stages: In the initial healthy state, the sensor failure rate approaches zero; if occasional minor anomalies occur, the system can usually recover to normal output through self-correction mechanisms. This is followed by the incipient fault state, characterized by extremely low-frequency, short-duration transient abnormal signal output. Subsequently, it transitions to the functional degradation state, where the failure rate increases significantly, manifesting as intermittent abnormal output with prolonged duration and intermittent recovery characteristics. Finally, it reaches the failure state, where the sensor experiences persistent failure and outputs persistent abnormal data, indicating equipment failure and the need for replacement. Based on this, the frequency of fault occurrence and cumulative anomaly duration monitored in real-time by the sensor are used as covariate inputs and coupled with a fitted performance degradation probability curve for analysis, thereby quantitatively assessing its current health status and reliability level.

[0132] Compared to the fault diagnosis and health assessment of sensors in traditional mud pump sealing water sensing systems, the main advantages and technical effects of this invention are as follows:

[0133] 1. Significantly improves explainability and alleviates the black box problem:

[0134] This solution deeply integrates prior physical knowledge with data-driven methods. By constructing physical constraints and embedding them into model training (such as the PINN loss function), it guides the model to conform to physical mechanisms. Explicitly using physical laws as constraints ensures physical consistency between prediction and diagnostic results, making the results easier to understand and interpret using known physical principles. The residuals are generated by superimposing theoretical predictions from the physical model with virtual sensor predictions, clarifying the basis for fault diagnosis (residual magnitude and duration) and making the diagnostic logic more intuitive and reliable.

[0135] 2. Effectively enhances generalization ability and robustness:

[0136] The virtual sensor model built on PINN forces the model to learn a mapping relationship that conforms to physical laws by jointly optimizing data fitting error and physical constraint error in the loss function, rather than simply fitting the training data. Furthermore, the innovative MCNN-Transformer structure helps capture complex spatiotemporal patterns. This physically-guided training method significantly improves the model's generalization ability and robustness in scenarios such as changes in data distribution, partial sensor failure, or the emergence of new fault modes. This makes the diagnostic and evaluation results more stable and reliable, adaptable to a wider range of real-world conditions, and solves the problem of low generalization in traditional methods.

[0137] 3. A closed-loop evaluation system for deep fusion of multi-source information has been implemented:

[0138] This solution constructs a complete, information-integrated closed-loop technology chain: from prior physical knowledge modeling → virtual sensor prediction combining physics and data-driven approaches → accurate fault diagnosis (instantaneous, intermittent, and permanent) based on physical-virtual data superposition residuals → dynamic reliability calculation and health status assessment based on quantitative fault statistics (frequency and duration). This ensures the consistency, coherence, and dynamism of the diagnostic results and health status assessment, providing more accurate and comprehensive decision support for predictive maintenance.

[0139] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A fault diagnosis method for a mud pump sealing water sensing system, characterized in that, The method comprises the following steps: inputting the current sensing parameter value of the mud pump sealing water sensing system into the PINN-based virtual sensor model to obtain a first sealing water pressure prediction value output by the virtual sensor model, wherein the virtual sensor model is trained by taking a sealing water pressure dynamic output equation obtained by establishing a physical information model as a physical constraint; calculating a second sealing water pressure prediction value by using the sealing water pressure dynamic output equation according to the actual power, pump efficiency and flow of the sealing water pump; comparing the mud pump sealing water pressure sensing value of the mud pump sealing water sensing system with the first sealing water pressure prediction value and the second sealing water pressure prediction value to determine whether the mud pump sealing water sensing system is faulty; inputting the current sensing parameter value of the mud pump sealing water sensing system into the PINN-based virtual sensor model to obtain a first sealing water pressure prediction value output by the virtual sensor model, comprising: performing feature analysis on historical sensing parameter values of the mud pump sealing water sensing system by using XGBoost and Random Forest algorithms, and selecting a preset number of sensing parameters most related to the mud pump sealing water pressure; predicting the current sensing parameter value of the selected sensing parameters by using an MCNN-Transformer model to obtain a first sealing water pressure prediction value output by the MCNN-Transformer model; before inputting the current sensing parameter value of the mud pump sealing water sensing system into the PINN-based virtual sensor model to obtain a first sealing water pressure prediction value output by the virtual sensor model, further comprising: determining a data loss term of the virtual sensor model according to the difference between the first sealing water pressure prediction value of the virtual sensor model for sample data and the corresponding actual sealing water pressure measurement value; determining a physical loss term of the virtual sensor model according to the difference between the second sealing water pressure prediction value of the sealing water pressure dynamic output equation for the sample data and the corresponding actual sealing water pressure measurement value; determining a total loss function of the virtual sensor model according to the physical loss term and the data loss term of the virtual sensor model, and training the virtual sensor model according to the total loss function; a difference between a second sealing water pressure predicted value of the sample data and a corresponding sealing water pressure actual measurement value is determined according to the sealing water pressure dynamic output equation by the following formula : ; in, The number of data entries in the sample data. For the first The actual power of the sealing pump in the sample data. For the first The pump efficiency of the sealing pump in the sample data. For the first The flow rate of the sealing pump in the sample data. For the first The actual measured value of the sealing pressure corresponding to each sample data point.

2. The method of claim 1, wherein, A total loss function of the virtual sensor model is determined according to the physical loss term and the data loss term of the virtual sensor model by the following equation : ; wherein, is a physical loss term of the virtual sensor model, is a data loss term of the virtual sensor model, is a tuning parameter in the range [0, 1].

3. The mud pump water containment system fault diagnostic method of claim 1 or 2, wherein, comparing the mud pump sealing water pressure sensing value of the mud pump sealing water sensing system with the first sealing water pressure prediction value and the second sealing water pressure prediction value to determine whether the mud pump sealing water sensing system is faulty, comprising: fusing the first sealing water pressure prediction value and the second sealing water pressure prediction value to obtain a sealing water pressure fusion value; performing residual threshold analysis on the mud pump sealing water pressure sensing value and the sealing water pressure fusion value to determine whether the mud pump sealing water sensing system is faulty.

4. The method of claim 3, wherein, The mud pump sealing water pressure sensing value of the mud pump sealing water sensing system is compared with the first sealing water pressure prediction value and the second sealing water pressure prediction value by the following formula to determine whether the mud pump sealing water sensing system is faulty: ; ; wherein, is a sealing water pressure fusion value, is a first sealing water pressure prediction value, is a second sealing water pressure prediction value, and is a weight coefficient, is a mud pump sealing water pressure perception value, is a residual error of the mud pump sealing water pressure perception value and the sealing water pressure fusion value; if the residual is greater than a first preset threshold, it is determined that the mud pump sealing water sensing system is faulty.

5. The mud pump water containment system fault diagnostic method of claims 1 or 2, wherein, In the comparison of the mud pump sealing water pressure sensing value of the mud pump sealing water sensing system with the first sealing water pressure prediction value and the second sealing water pressure prediction value to determine whether the mud pump sealing water sensing system is malfunctioning, the method further comprises: In the case where the failure duration of the mud pump sealing water sensing system reaches a second preset threshold, it is determined that the mud pump sealing water sensing system is a transient failure; In the case where the failure duration of the mud pump sealing water sensing system is greater than the second preset threshold and less than or equal to a third preset threshold, it is determined that the mud pump sealing water sensing system is an intermittent failure, and the third preset threshold is greater than the second preset threshold; In the case where the failure duration of the mud pump sealing water sensing system is greater than the third preset threshold, it is determined that the mud pump sealing water sensing system is a permanent failure.

6. A method for health assessment of a mud pump seal water detection system according to any one of claims 1-5, characterized in that, It comprises: Determine the total number of failures and the duration of each failure of the mud pump sealing water sensing system; According to the total number of failures and the duration of each failure, determine the health degradation index of the mud pump sealing water sensing system; According to the health degradation index, determine the health status of the mud pump sealing water sensing system.

7. The method of claim 6, wherein, A health degradation indicator of the mud pump seal water detection system is determined from the total number of failures and the duration of each failure according to the following equation : ; ; ; wherein, is the total number of failures, is the total duration of failures corresponding to the total number of failures, is the duration of the th failure, is the failure intensity coefficient determined according to the average failure-free time of the mud pump water sealing sensing system under the rated working condition.

Images