Abrasion online identification and compensation control system and method for high-pressure distribution plunger and barrel assembly

Abstract

The invention discloses a wear online identification and compensation control system and method for a high-pressure distribution plunger and barrel assembly, and the system comprises a data collection module which is used for collecting working condition data and state data; a health state dynamic baseline model obtained through offline training is arranged in the data processing module, and the wear level of the plunger and barrel assembly is recognized by calculating the residual error between actual state data and health state data predicted by the model under the current working condition; and the compensation control module is used for generating a self-adaptive compensation control instruction according to the identified wear grade and the current working condition and sending the self-adaptive compensation control instruction to an execution mechanism. According to the method, the dynamic base line synchronous with the real-time working condition is constructed, early weak wear characteristics can be separated from strong working condition interference, the recognition sensitivity and accuracy are improved, the influence of wear on the system performance is actively counteracted through closed-loop adaptive compensation control, the service life of equipment is expected to be prolonged, and the system performance is improved. And technical support is provided for predictive wear monitoring management.

Description

Technical Field

[0001] This application relates to the fields of high-end industrial equipment monitoring and control, artificial intelligence and online compensation technology, specifically to an online wear identification and compensation control system and method for high-pressure distribution plunger components. Background Technology

[0002] High-pressure distribution plunger assemblies are key power and control components in high-end industrial equipment such as high-pressure common rail fuel systems and precision hydraulic servo systems. The fit clearance and surface morphology between the plunger and sleeve have a decisive impact on the stability, response speed, and control accuracy of the entire system's output pressure and flow rate. During long-term service, the plunger assemblies undergo progressive surface wear under the combined effects of high-frequency, high-speed reciprocating motion and high-pressure fluid erosion. This wear leads to an increase in the fit clearance of the plunger assemblies, disrupting the original hydrodynamic sealing characteristics, and consequently causing a series of problems such as increased system leakage, decreased effective output flow rate, increased pressure fluctuations, and increased energy consumption. In severe cases, it may even lead to system failure. Figure 9 The diagram illustrates that the plunger assembly has a micron-level precision fit clearance in a healthy state (left image), while in a worn state (right image), the fit clearance increases significantly, the surface becomes rough, and thus serious internal leakage occurs.

[0003] Therefore, real-time online monitoring of the wear condition of high-pressure distribution plunger assemblies is crucial for ensuring the reliable operation of equipment. However, in practical industrial applications, the operating conditions of equipment, such as engine load and speed, or target pressure and load torque of the hydraulic system, are dynamically changing. Fluctuations in these operating parameters directly cause corresponding changes in system output characteristics (e.g., outlet pressure, flow rate). The normal system response signal caused by changes in operating conditions is highly similar to and coupled with the state characteristic signal generated by early minor wear of the plunger assemblies. The normal system response signal generated by fluctuations in operating conditions creates strong background noise, which masks the weak state characteristic signal generated by early minor wear.

[0004] Existing online monitoring methods, such as those based on signal processing (spectral analysis), wavelet transform, or simple threshold alarms, struggle to effectively distinguish between two signals from different sources but exhibiting similar characteristics. These methods typically achieve some monitoring effectiveness only in the mid-to-late stages of wear when the signal-to-noise ratio is high, but their sensitivity and accuracy in identifying early wear are insufficient, easily leading to missed and false alarms. This is because existing technologies have failed to establish a dynamic benchmark model that accurately reflects the system's health status under different operating conditions, thus failing to effectively separate normal system responses caused by changes in operating conditions from the monitoring signals. Therefore, accurately separating and identifying the state characteristic signals caused by early, weak wear from complex and changing operating conditions under strong background interference remains a significant technical challenge in the field of online monitoring of high-pressure distribution plunger assemblies. Summary of the Invention

[0005] To address the technical problems existing in the background art, the present invention provides an online wear identification and compensation control system and method for high-pressure distribution plunger components. This scheme can separate wear characteristic signals from strong operating condition interference background by constructing a dynamic baseline of system health status synchronized with real-time operating conditions, thereby achieving highly sensitive online identification of early and weak wear of plunger components, and performing closed-loop compensation control based on the identification results to maintain the stability of system performance.

[0006] One aspect of the present invention provides an online wear identification and compensation control system for a high-pressure distribution plunger assembly. The system includes a data acquisition module, a data processing module, and a compensation control module. The data acquisition module, deployed on the monitored equipment system, is used to collect in real-time operating condition data reflecting the operating conditions of the plunger assembly and status data reflecting its working state. The data processing module, electrically connected to the data acquisition module, integrates a dynamic baseline model of health status and a wear status identification model. It is used to receive and process the data, identify the current wear level of the plunger assembly, and predict its development trend. The compensation control module, electrically connected to the data processing module, is used to generate adaptive compensation control commands based on the identified wear level and current operating conditions, and send them to the actuators of the equipment system.

[0007] In another aspect, the present invention provides a method for online wear identification and compensation control of a high-pressure distribution plunger assembly, the method comprising the following steps: Step 1: Multi-source heterogeneous time-series data acquisition and synchronous preprocessing. The purpose of this step is to comprehensively and in real-time acquire various information reflecting the operating conditions and working status of the plunger assembly. Specifically, two types of sensor arrays are deployed on the monitored equipment system. The first type is operating condition sensors, used to collect operating condition data characterizing the system's operating conditions. Preferably, operating condition sensors include speed sensors for measuring engine or drive motor speed, pressure or torque sensors for measuring system load, temperature sensors for measuring hydraulic oil or fuel temperature, and control signal acquisition units for acquiring controller output commands. The second type is status sensors, used to collect physical signals characterizing the working status of the plunger assembly as status data. Preferably, status sensors include high-frequency dynamic pressure sensors installed at the outlet of the system's high-pressure pipeline, and broadband vibration acceleration sensors or acoustic emission sensors installed at key locations on the plunger assembly housing. All acquired sensor signals are time-series data. A high-precision data acquisition system ensures that data from all channels are acquired at a uniform sampling frequency, and timestamp alignment technology is applied to achieve strict time synchronization of the multi-source heterogeneous data. The synchronized data is preprocessed, including using digital filters to remove high-frequency noise and DC bias from the signal, and using the max-min normalization method or Z-score normalization method to scale the data of different physical dimensions to a uniform numerical range, so as to facilitate subsequent input into the deep learning model for processing.

[0008] Step 2: Dynamic baseline modeling of health status based on a conditional deep generative model. The purpose of this step is to construct a deep learning model capable of accurately predicting the normal operating state signal that a healthy plunger assembly should exhibit under any given operating condition. This model is called the dynamic baseline model of health status. This invention achieves the technical effect of separating early weak fault features from strong operating condition disturbances by constructing a dynamic baseline model of health status based on a conditional variational autoencoder and using it to generate residual signals. Preferably, the dynamic baseline model of health status adopts a conditional variational autoencoder (CVAE) architecture, which consists of an encoder, a decoder, and a latent space. First, a comprehensive operating condition traversal experiment is conducted on a brand-new, unworn reference plunger assembly, collecting operating condition data and corresponding state data covering all normal operating ranges of the equipment, forming a pure health dataset. Subsequently, the CVAE model is trained using this health dataset. During training, the operating condition data sequence collected and preprocessed in Step 1 is used as the conditional vector C, and the corresponding state data sequence is used as the target input X of the model. Guided by a conditional vector C, the encoder compresses the input state data sequence X into a latent variable z that follows a specific prior distribution (e.g., a Gaussian distribution). The decoder, also guided by the same conditional vector C, attempts to reconstruct the original state data sequence X from the latent variable z. The model's optimization objective is to minimize the reconstruction error between the original state data sequence X and the decoder-reconstructed state data sequence X_hat, while simultaneously constraining the posterior distribution of the latent variable z using the Kullback-Leibler divergence loss function to approximate the preset prior distribution. After sufficient training, this CVAE model learns the high-dimensional nonlinear mapping relationship of the healthy plunger assembly under different operating conditions. Furthermore, to enhance the model's generative capability, the decoder can employ a recurrent neural network structure based on Long Short-Term Memory (LSTM) or Gated Recurrent Unit (GRU) to better capture the temporal dependencies of the state signals.

[0009] This invention also achieves online calibration of the dynamic baseline model for health status by introducing a model adaptive update and transfer learning mechanism. To address the non-wearing, slow state drift of the equipment system caused by factors such as changes in hydraulic oil viscosity or sensor zero-point drift, an online model calibration and update module is added to the system. This module continuously monitors the statistical characteristics of the residual sequences generated in subsequent steps. When the system operates continuously under multiple operating conditions, if the energy or amplitude of the residual sequence remains below a health confidence threshold determined based on the residual statistical characteristics of the health dataset for an extended period, the system determines that the current plunger assembly is still in a healthy state. At this time, the system can automatically collect operating condition-state data pairs for this stage as new health samples. After accumulating a predetermined number of new samples, transfer learning technology is used to fine-tune the parameters of the pre-trained CVAE model to adapt it to the latest health benchmark state of the current system. This online adaptive model update mechanism ensures that the dynamic baseline accurately reflects the current health benchmark of the equipment, avoiding the risk of misjudgment caused by slow system drift.

[0010] Step 3: Online residual signal generation and feature extraction based on model prediction. The health status dynamic baseline model trained and calibrated in Step 2 is deployed to the online monitoring system. During actual equipment operation, the system collects the current operating condition data sequence C(t) and the status data sequence X(t) in real time. The real-time operating condition data sequence C(t) is input into the condition terminal of the trained CVAE model, and the model immediately generates a predicted health status data sequence X_hat(t), which is the dynamic baseline. Subsequently, the residual sequence R(t) between the actually measured status data sequence X(t) and the health baseline data sequence X_hat(t) generated by the model is calculated using the formula R(t) = X(t) - X_hat(t). Since the health baseline data sequence X_hat(t) has reproduced all the effects of the current operating condition C(t) on the system state, the normal system fluctuations caused by changes in operating conditions are canceled out from the original signal through the subtraction operation. If the plunger assembly is in a healthy state, the residual sequence R(t) should be close to zero-mean white noise; if the plunger assembly is worn, the abnormal disturbance information introduced by this change in physical structure will be completely preserved in the residual sequence R(t). To quantify the wear information in the residual sequence, a sliding time window is used to segment the residual sequence R(t), and a set of statistical features is extracted from the residual data of each window. Preferably, this set of statistical features includes the root mean square value (RMS), kurtosis, margin factor, waveform factor, etc., thus forming a residual feature vector that varies with time. (Refer to...) Figure 5The above figure visually demonstrates the technical effects of this invention. In the upper subfigure, when the plunger assembly is in a healthy state, although the actual state signal X(t) fluctuates drastically due to changes in operating conditions, the healthy baseline signal X_hat(t) generated by the model accurately follows its changes, making the residual signal R(t) obtained by subtracting the two close to zero-mean white noise. In the lower subfigure, when the plunger assembly is in an early wear state, the actual state signal X(t) is superimposed with weak wear pulses, while the healthy baseline model X_hat(t) still only reproduces the influence of operating conditions. Therefore, the residual signal R(t) clearly highlights the periodic pulses related to wear, and the signal-to-noise ratio is greatly improved, proving that the method of this invention can effectively separate early weak wear characteristics from strong operating condition interference.

[0011] Step 4: Wear State Grading and Trend Prediction Based on Temporal Deep Learning. This step aims to achieve accurate grading of wear states through analysis of the residual feature vector sequence. This invention achieves accurate grading of wear states by utilizing a temporal convolutional network to perform deep analysis of the temporal evolution patterns of residual features. Preferably, a temporal convolutional network (TCN) is used as the wear state recognition model. The continuously generated residual feature vector sequence in Step 3 is used as the input to the TCN model. This model learns the deep patterns of residual feature evolution over time through multiple layers of causal convolution, dilated convolution, and residual connections. The output layer of the model is a classifier, such as a fully connected layer followed by a softmax activation function, used to map the input feature sequence to predefined wear levels. The wear levels include at least: healthy state, early wear, intermediate wear, and severe wear. Optionally, the wear level classification is based on physical indicators. As a non-limiting example, early wear corresponds to an increase of 5-10 micrometers in the clearance of the plunger assembly, intermediate wear corresponds to an increase of 10-20 micrometers, and severe wear corresponds to an increase of more than 20 micrometers or a drop in system pressure exceeding a predetermined threshold. The training data for this TCN classification model can be obtained through a full lifecycle wear experiment on the plunger assembly, recording data from new to failure, and accurately labeling different stages through offline analysis (such as periodic disassembly and measurement). Furthermore, based on identifying the current wear level, the future development trend of wear can be predicted. By establishing an Autoregressive Integrated Moving Average (ARIMA) model or an independent recurrent neural network, the historical wear level sequence or the time series of key residual features is modeled to predict the evolution trajectory of the wear state over a future period. (Refer to...) Figure 7This demonstrates the effectiveness of the residual features extracted using the method of this invention. In the two-dimensional feature space composed of the root mean square (RMS) value and kurtosis, sample points from different wear levels (healthy condition, early wear, intermediate wear, severe wear) form well-defined and mutually separated clusters. This good separability proves that the extracted features can effectively characterize different wear states, laying the foundation for subsequent high-precision classification. (Refer to...) Figure 6 The figure compares the identification accuracy of the method of this invention with that of the traditional threshold method under different wear levels. As can be seen from the figure, the method of this invention maintains an extremely high identification accuracy of over 95% in all wear stages, including early wear. In contrast, the traditional threshold method has an identification accuracy of less than 20% in the early wear stage, making it difficult to provide effective early warning. This comparison highlights the significant technical advantages of this invention in early, subtle fault identification.

[0012] Step 5: Adaptive Compensation Closed-Loop Control Based on Fuzzy Inference. When the identification result in Step 4 is an unhealthy state (i.e., early wear, intermediate wear, or severe wear), the system activates the compensation control module. This invention achieves proactive compensation for performance degradation caused by wear by constructing a dual-input adaptive compensation control system based on fuzzy inference—operating condition and wear level. Preferably, this compensation control module is implemented based on a fuzzy inference system (FIS). This fuzzy inference system uses the current operating condition (e.g., speed, load) and the wear level identified in Step 4 as two input variables. The system internally presets a fuzzy rule base constructed from domain expert knowledge. A typical fuzzy rule can be expressed as: if (wear level is "early wear") and (current load is "high"), then (compensation control quantity is "positive fine-tuning"). The system fuzzifies the precise input values ​​(such as engine speed and quantified wear level) into linguistic variables (such as "high", "medium", and "low") using membership functions. Then, it matches rules from the rule base using a fuzzy inference engine and performs fuzzy calculations. Finally, it calculates a precise compensation control quantity using a defuzzification method (such as the center-of-gravity method). This compensation control quantity is sent to the actuators of the equipment system, for example, to fine-tune the duty cycle of the solenoid valve control signal of the plunger drive mechanism, or to change the power output of the drive motor, thereby compensating for internal leakage caused by wear to a certain extent and restoring the system's macroscopic output performance to the predetermined target range. For example, in a high-pressure common rail fuel system, this compensation control quantity could be a duty cycle increment used to fine-tune the control signal of the solenoid valve driving the plunger pump. By increasing the duty cycle, the effective pumping stroke of fuel can be increased, thereby compensating for internal leakage caused by wear to a certain extent, allowing the system's macroscopic output performance (such as common rail pressure) to remain stable near the predetermined target value even under wear conditions, achieving closed-loop compensation control of system performance. (Refer to...) Figure 8 This demonstrates the practical effect of the adaptive compensation control of the present invention. When the plunger assembly has undergone mid-term wear, without compensation (pressure curve before compensation in the figure), the system outlet pressure will continuously decrease and fluctuate more significantly over time, deviating severely from the target pressure. When the compensation control of the present invention is activated (pressure curve after compensation in the figure), the system can actively adjust the control quantity, effectively offsetting the performance degradation caused by wear, and keeping the outlet pressure stable near the target pressure value for a long period, thereby ensuring the stable and reliable operation of the equipment.

[0013] This invention constructs a dynamic baseline model of health status that changes in real time with operating conditions. This effectively separates early, weak wear signals from strong background interference, improving the sensitivity and accuracy of early wear identification and avoiding false alarms and missed detections common in traditional methods. The deep learning-based dynamic model can adapt to complex, nonlinear, and time-varying industrial operating conditions, giving this method good robustness and reliability in various application scenarios. This invention not only identifies wear but also actively offsets some of the negative impacts of wear on system performance through closed-loop adaptive compensation control, extending the service life of core components and the overall lifespan of equipment. Simultaneously, it integrates state classification identification and trend prediction functions, providing technical support for predictive and health management of equipment and enhancing its intelligence level. Attached Figure Description

[0014] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0015] Figure 1 This is a schematic diagram of the structure of an online wear identification and compensation control system for a high-pressure distribution plunger assembly provided in an embodiment of the present invention.

[0016] Figure 2 This is a flowchart of an online wear identification and compensation control method for a high-pressure distribution plunger assembly provided in an embodiment of the present invention.

[0017] Figure 3 This is a schematic diagram illustrating the principle of dynamic baseline generation of health status and extraction of residual signals based on conditional variational autoencoder in an embodiment of the present invention.

[0018] Figure 4 This is a block diagram of the adaptive compensation control module based on fuzzy inference in an embodiment of the present invention.

[0019] Figure 5 This is a comparison diagram of the residual signal generation effect in the healthy state and the early wear state in an embodiment of the present invention.

[0020] Figure 6 This is a comparison chart of the recognition accuracy of the method of this invention and the traditional threshold method under different degrees of wear.

[0021] Figure 7 This is a schematic diagram showing the distribution of different wear levels in the residual feature space in an embodiment of the present invention.

[0022] Figure 8 This is a verification diagram of the effect of adaptive compensation control under wear conditions in an embodiment of the present invention.

[0023] Figure 9 This is a schematic diagram comparing the structure of the high-pressure distribution plunger assembly under healthy and worn conditions. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0025] Example 1 This embodiment provides a method for online wear identification and compensation control of high-pressure distribution plunger assemblies, referring to... Figure 2 This method is mainly applied to Figure 1 The detailed steps in the system shown are as follows: Step S100: Acquisition and synchronous preprocessing of multi-source heterogeneous time-series data.

[0026] In this step, to comprehensively obtain information reflecting the operating conditions and working status of the plunger assembly, multiple types of sensors are deployed on the monitored equipment system. Specifically, refer to... Figure 1 The operating condition data is acquired from operating condition sensors, which may include a speed sensor 101 for measuring the drive motor speed, a pressure sensor 102 for measuring the system load, and a temperature sensor 103 for measuring the hydraulic oil temperature. Simultaneously, current control commands, such as target pressure or flow commands, are directly obtained from the equipment system's controller via a control signal acquisition unit 104 and are also included as part of the operating condition data. The status data is acquired from status sensors, which may include a high-frequency dynamic pressure sensor 105 installed at the system's high-pressure pipeline outlet to capture pressure pulsation details; and a broadband vibration acceleration sensor 106 installed at a key location on the plunger assembly housing to monitor structural vibrations caused by internal impacts and fluid excitation.

[0027] All signals acquired by the sensors are time-series data. A high-precision data acquisition system ensures that data from all channels are acquired at a uniform and sufficiently high sampling frequency. By applying timestamp alignment techniques, such as Network Time Protocol (NTP) or Precise Time Protocol (PTP), strict time synchronization of data from different sources is achieved, forming a unified time reference. Subsequently, the synchronized raw data undergoes preprocessing. Specifically, a Butterworth digital low-pass filter is used to remove high-frequency noise introduced by electromagnetic interference, and a moving average method is used to filter out the DC bias of the signal. To eliminate the influence of different physical dimensions on subsequent model training, a Z-score normalization method is further employed to scale the data from different sources (e.g., rotational speed, pressure, vibration acceleration) to a uniform numerical range with a mean of 0 and a standard deviation of 1, facilitating effective processing by the deep learning model.

[0028] Step S200: Dynamic baseline modeling of health status based on conditional deep generative model.

[0029] The core objective of this step is to construct a dynamic baseline model of the health state that can accurately predict the normal operating state signal that a healthy plunger assembly should have under any given operating condition. This model preferably employs a Conditional Variational Auto-Encoder (CVAE) architecture.

[0030] First, a pure health dataset needs to be constructed. Specifically, a comprehensive bench test is conducted on a brand-new, confirmed, wear-free reference plunger assembly. During the test, the equipment system is programmed to traverse all its designed normal operating ranges, such as rotating at a constant speed from minimum to maximum, gradually loading the load from no load to full load, and covering different temperature ranges. Simultaneously, complete operating condition data and corresponding status data are collected according to the method in step S100 to form the health dataset.

[0031] Subsequently, the CVAE model was trained offline using this health dataset. (See reference...) Figure 3 The CVAE model consists of an encoder 301, a decoder 303, and a latent space 302. During training, the working condition data sequence collected and preprocessed in step S100 is used as the condition vector C, and the corresponding state data sequence is used as the target input X of the model. Guided by the condition vector C, the encoder 301 compresses the input state data sequence X into a latent variable z that follows a specific prior distribution (e.g., a standard Gaussian distribution). Guided by the same condition vector C, the decoder 303 attempts to reconstruct the original state data sequence from the latent variable z, denoted as the reconstructed sequence X_hat. The optimization objective of the model is to minimize the reconstruction error between the original state data sequence X and the reconstructed sequence X_hat, which can be measured using the mean squared error (MSE). Simultaneously, the posterior distribution of the latent variable z is constrained by the Kolb-Leibler divergence loss function to make it as close as possible to the preset prior distribution, thus ensuring the model's generative capability. After sufficient training, the CVAE model essentially learns a high-dimensional nonlinear mapping relationship from the operating condition space to the health state signal space. Furthermore, to enhance the model's ability to model time-series data, the decoder 303 can employ a Long Short-Term Memory (LSTM) network or a gated recurrent unit (GRU) internal structure, as these recurrent neural network structures can effectively capture the time dependence in state signals (such as pressure waveforms), thereby generating a more accurate and realistic dynamic baseline.

[0032] Furthermore, to address the non-wear-related slow state drift that may occur in equipment systems during long-term operation, such as viscosity changes due to hydraulic oil aging or sensor zero-point drift, this method also includes a model adaptive update and transfer learning mechanism. The system continuously monitors the statistical characteristics of the residual sequence generated in subsequent step S300. When the system operates continuously under multiple typical operating conditions, if the energy of the residual sequence (e.g., the root mean square value) remains below a health confidence threshold for a relatively long period (e.g., 100 consecutive working hours), this threshold is pre-set based on the energy statistical distribution of the residual signals generated by the health dataset under various operating conditions (e.g., the 99% confidence upper limit of the root mean square value). The system then determines that the current plunger assembly is still in a healthy state and automatically collects and marks the operating condition-state data pairs of this stage as new healthy samples. After accumulating a predetermined number of new samples, transfer learning technology is used to fine-tune the parameters of the pre-trained CVAE model, enabling it to quickly adapt to the latest health baseline state of the current system, thereby ensuring the long-term accuracy of the dynamic baseline.

[0033] Step S300: Online residual signal generation and feature extraction based on model prediction.

[0034] The health status dynamic baseline model trained and calibrated in step S200 is deployed to the online monitoring system. During the actual operation of the equipment, the system collects the current operating condition data sequence C(t) and the status data sequence X(t) in real time. (Refer to...) Figure 3 In the online application phase, the real-time operating condition data sequence C(t) is input into the condition end of the trained CVAE model, and the model immediately generates a predicted health status data sequence X_hat(t), which is the dynamic baseline synchronized with the current operating condition.

[0035] Subsequently, using subtractor 304, the residual sequence R(t) between the actual measured state data sequence X(t) and the model-generated healthy baseline data sequence X_hat(t) is calculated, with the formula R(t) = X(t) - X_hat(t). Since the healthy baseline data sequence X_hat(t) accurately reproduces the entire impact of the current operating condition C(t) on the system state, the normal system response fluctuations caused by the change in operating condition are effectively canceled out from the original signal through the subtraction operation. If the plunger assembly is in a healthy state, the amplitude of the residual sequence R(t) will be close to zero, and its statistical characteristics will be close to zero-mean white noise. Conversely, if the plunger assembly is worn, this abnormal disturbance introduced by changes in physical structure, such as internal leakage pulses caused by increased clearance or abnormal vibrations caused by wear particles, cannot be reproduced by the healthy model and will therefore be completely preserved in the residual sequence R(t), greatly improving the signal-to-noise ratio. (Refer to...) Figure 5The above figure visually demonstrates the technical effects of this invention. In the upper subfigure, when the plunger assembly is in a healthy state, although the actual state signal X(t) fluctuates drastically due to changes in operating conditions, the healthy baseline signal X_hat(t) generated by the model accurately follows its changes, making the residual signal R(t) obtained by subtracting the two close to zero-mean white noise. In the lower subfigure, when the plunger assembly is in an early wear state, the actual state signal X(t) is superimposed with weak wear pulses, while the healthy baseline model X_hat(t) still only reproduces the influence of operating conditions. Therefore, the residual signal R(t) clearly highlights the periodic pulses related to wear, and the signal-to-noise ratio is greatly improved, proving that the method of this invention can effectively separate early weak wear characteristics from strong operating condition interference.

[0036] To quantify wear information in the residual sequence, a sliding time window is used to segment the residual sequence R(t). For example, the window length is set to 1 second, and the sliding step size is 0.5 seconds. A set of statistical features characterizing the signal properties is extracted from the residual data of each window. These statistical features preferably include: root mean square (RMS) value, reflecting the signal energy; kurtosis, reflecting the signal's impulsiveness; kurtosis, reflecting the sharpness of the signal distribution; and dimensionless indices such as margin factor and waveform factor. These features together constitute a time-varying residual feature vector for subsequent wear state identification.

[0037] Step S400: Wear status classification and trend prediction based on time-series deep learning.

[0038] This step aims to achieve accurate classification of the wear state of the plunger assembly by performing in-depth analysis on the residual feature vector sequence generated in step S300. Preferably, a Temporal Convolutional Network (TCN) is used as the wear state identification model.

[0039] The continuously generated residual feature vector sequence in step S300 is used as the input to the TCN model. Through its unique multi-layer causal convolution, dilated convolution, and residual connection structure, the TCN model can efficiently learn the evolution patterns and long-term dependencies of residual features over time, which is particularly effective for identifying progressive wear processes that range from weak early stages to significant later stages. The model's output layer is a fully connected layer followed by a Softmax activation function, forming a classifier that maps the input feature sequence to predefined wear levels. These wear levels include at least four categories: healthy state, early wear, intermediate wear, and severe wear. These levels are calibrated based on explicit physical indicators. As a non-limiting example, through offline experimental calibration, an increase of 5-10 micrometers in the clearance between the plunger and sleeve can be defined as early wear, an increase of 10-20 micrometers as intermediate wear, and an increase of more than 20 micrometers or causing a drop in outlet pressure of more than 10% under rated operating conditions can be defined as severe wear. The training data for this TCN classification model was obtained through accelerated wear experiments on multiple plunger components throughout their entire lifecycle. Data was recorded synchronously during the experiments, and the plunger components were periodically disassembled and measured to accurately label the data at different wear stages. (Refer to...) Figure 7 This demonstrates the effectiveness of the residual features extracted using the method of this invention. In the two-dimensional feature space composed of the root mean square (RMS) value and kurtosis, sample points from different wear levels (healthy condition, early wear, intermediate wear, severe wear) form well-defined and mutually separated clusters. This good separability proves that the extracted features can effectively characterize different wear states, laying the foundation for subsequent high-precision classification. (Refer to...) Figure 6 The figure compares the identification accuracy of the method of this invention with that of the traditional threshold method under different wear levels. As can be seen from the figure, the method of this invention maintains an extremely high identification accuracy of over 95% in all wear stages, including early wear. In contrast, the traditional threshold method has an identification accuracy of less than 20% in the early wear stage, making it difficult to provide effective early warning. This comparison highlights the significant technical advantages of this invention in early, subtle fault identification.

[0040] Furthermore, based on identifying the current wear level, future wear trends can be predicted. One possible approach is to establish an Autoregressive Integrated Moving Average (ARIMA) model to model the time series of historical wear level sequences (e.g., quantifying the four levels as 0, 1, 2, 3) or a key residual feature (such as the root mean square value). By analyzing historical data, the ARIMA model can predict the wear level that may be reached in the future, such as the next 50 working hours, thus providing decision support for predictive maintenance.

[0041] Step S500: Adaptive compensation closed-loop control based on fuzzy inference.

[0042] When the identification result of step S400 is an unhealthy state, the system activates the compensation control module to actively compensate for the system performance degradation caused by wear. Preferably, this compensation control module is implemented based on a fuzzy inference system (FIS). (Refer to...) Figure 4 The fuzzy inference system uses the current operating conditions (e.g., speed, load) and the wear level identified in step S400 as two input variables.

[0043] Specifically, the system internally pre-defines a fuzzy rule base 403 constructed from domain expert knowledge. A typical fuzzy rule can be expressed as: "If (wear level is 'intermediate wear') and (current load is 'high'), then (compensation control quantity is 'positive medium adjustment')". First, the precise input values ​​are fuzzified using a membership function 401. For example, a speed value of 1500 rpm and a wear level quantized as "2" are converted into linguistic variables such as "medium speed" and "intermediate wear" and their membership degrees. Then, the fuzzy inference engine 402 performs inference calculations based on these fuzzy inputs and the rule base. Finally, a precise compensation control quantity is calculated using a defuzzification module 404 (e.g., using the center of gravity method).

[0044] This compensation control quantity is ultimately sent to the actuators of the equipment system. For example, in a high-pressure common rail fuel system, this compensation control quantity could be a duty cycle increment used to fine-tune the control signal of the solenoid valve driving the plunger pump. By increasing the duty cycle, the effective pumping stroke of the fuel can be increased, thereby compensating to some extent for internal leakage caused by wear. This allows the system's macroscopic output performance (such as common rail pressure) to remain stable near the predetermined target value even under wear conditions, achieving closed-loop compensation control of system performance. (Refer to...) Figure 8 This demonstrates the practical effect of the adaptive compensation control of the present invention. When the plunger assembly has undergone mid-term wear, without compensation (pressure curve before compensation in the figure), the system outlet pressure will continuously decrease and fluctuate more significantly over time, deviating severely from the target pressure. When the compensation control of the present invention is activated (pressure curve after compensation in the figure), the system can actively adjust the control quantity, effectively offsetting the performance degradation caused by wear, and keeping the outlet pressure stable near the target pressure value for a long period, thereby ensuring the stable and reliable operation of the equipment.

[0045] Example 2 This embodiment provides an online wear identification and compensation control system for high-pressure distribution plunger assemblies, referring to... Figure 1 The system includes: a data acquisition module, a data processing module, and a compensation control module.

[0046] The data acquisition module 110 is deployed on the monitored equipment system and is used to execute step S100 in Embodiment 1. This module specifically includes an array of operating condition sensors (such as a speed sensor 101, a pressure sensor 102, and a temperature sensor 103), an array of status sensors (such as a high-frequency dynamic pressure sensor 105 and a vibration acceleration sensor 106), a control signal acquisition unit 104, and a high-precision data acquisition card. This module is responsible for real-time acquisition of operating condition data reflecting the operating conditions of the plunger assembly and status data reflecting its working state, and for time synchronization and preprocessing.

[0047] The data processing module 120, electrically connected to the data acquisition module 110, is typically an industrial computer or embedded controller with sufficient data processing capabilities. The module's internal memory stores the health status dynamic baseline model (CVAE model) and wear condition identification model (TCN model) described in Embodiment 1. The data processing module 120 receives preprocessed data from the data acquisition module 110 and performs the calculation tasks in steps S200 to S400 of Embodiment 1, namely, generating residual signals, extracting features, identifying the current wear level of the plunger assembly, and predicting its development trend.

[0048] The compensation control module 130 is electrically connected to the data processing module 120. This module receives the wear level identified by the data processing module 120 and the current operating condition data provided by the data acquisition module 110. Internally, this module implements the fuzzy inference system described in step S500 of Embodiment 1. Based on the input wear level and current operating condition, the module generates an adaptive compensation control command and sends this command to the actuator 140 of the equipment system, such as a solenoid valve controller or a motor driver, thereby achieving closed-loop compensation for system performance.

[0049] This invention, through the aforementioned system and method, constructs a dynamic baseline model of health status that changes in real time with operating conditions. This model effectively separates early, weak wear signals from strong background interference, significantly improving the sensitivity and accuracy of early wear identification. Simultaneously, through closed-loop adaptive compensation control, it proactively offsets some of the negative impacts of wear on system performance, extending the service life of core components and the overall lifespan of the equipment.

[0050] Example 3 This embodiment uses a high-pressure common rail fuel injection system applied to large construction machinery as an example to further illustrate the specific application process of the present invention. The operating characteristics of such systems are that their load and speed undergo frequent and drastic changes during operation. For example, when digging or lifting heavy objects, the engine load increases instantaneously. The normal fluctuations in state signals such as fuel common rail pressure and pump vibration caused by these drastic changes in operating conditions are highly coupled in amplitude and frequency characteristics with the abnormal signal characteristics generated by the early, slight wear of the plunger assembly, and may even far exceed the latter in signal strength. Therefore, traditional signal processing or fixed threshold alarm methods are insufficient to effectively distinguish between the two, easily misjudging normal operating condition responses as wear faults, or completely ignoring early wear signals drowned out by strong background noise, leading to monitoring failure. The method proposed in this invention can effectively address this technical challenge.

[0051] In this application case, firstly, following step one, sensors are deployed on the fuel system. Operating condition data includes engine speed, load (obtained via injection pulse width or torque sensors), and fuel temperature; status data includes high-frequency signals from the common rail pressure sensor and vibration acceleration signals mounted on the high-pressure fuel pump housing. Subsequently, following step two, a comprehensive operating condition traversal test is conducted on a test bench using a brand-new high-pressure fuel pump, collecting its health status data under different speed and load combinations. This data is then used to train a conditional variational autoencoder (CVAE) as a dynamic baseline model for the health status. This model learns the nonlinear mapping relationship from the "speed-load" operating condition to the "common rail pressure waveform-pump vibration signal".

[0052] In the actual online monitoring phase, the system was deployed on engineering machinery. The system collects the engine's current speed and load C(t) in real time and inputs it into a pre-trained CVAE model. The model instantly generates the theoretical state signal X_hat(t) that a healthy oil pump should exhibit under this operating condition. Simultaneously, the system collects the actual common rail pressure and vibration signal X(t). Following step three, the residual signal R(t) is obtained by calculating the difference between X(t) and X_hat(t). This residual signal effectively isolates normal pressure and vibration fluctuations caused by drastic changes in engine load, thus highlighting subtle internal leakage or abnormal impact signals caused by wear of the plunger assembly.

[0053] Next, following step four, the system extracts features from the continuous residual signals and inputs the feature sequence into a pre-trained Temporal Convolutional Network (TCN) model. Assuming that after approximately 2000 hours of device operation, the TCN model identifies a pattern in the residual feature sequence that matches the "early wear" level calibrated in the database, the system then classifies the current wear state as early wear and activates compensation control.

[0054] Finally, following step five, the compensation control module is activated. This module takes the current operating conditions (e.g., "high load," "medium speed") and the identified "early wear" level as input. Based on its internal fuzzy rule base, such as "if the wear level is early and the load is high, increase the duty cycle of the high-pressure fuel pump drive solenoid valve by 1.2%), the system calculates a precise compensation command. This command is sent to the engine control unit, which fine-tunes the fuel pump control signal. This increases the effective fuel delivery, thereby compensating for internal leakage caused by early wear, ensuring the common rail pressure remains stable at the target value, and guaranteeing the engine's power output performance.

[0055] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for online wear identification and compensation control of a high-pressure distribution plunger assembly, characterized in that, Includes the following steps: Step 1: Collect the operating condition data and status data of the monitored equipment system. The operating condition data is used to characterize the system operating conditions, and the status data is used to characterize the working status of the plunger assembly. The operating condition data and the status data are then preprocessed for time synchronization and normalization. Step 2: Based on the health dataset, train an offline dynamic baseline model of health status using a conditional variational autoencoder architecture; the model learns a high-dimensional mapping relationship between the working condition data and the health status data using the working condition data as a condition. Step 3: During the online operation of the equipment, the working condition data sequence collected and preprocessed in real time in Step 1 is input into the dynamic baseline model of health status trained in Step 2 to generate a predicted health status data sequence synchronized with the current working condition; the difference between the real-time collected status data sequence and the predicted health status data sequence is calculated to obtain the residual sequence; and wear features are extracted based on the residual sequence. Step four: Input the wear features continuously extracted in step three into the pre-trained wear state recognition model, and the wear state recognition model outputs the wear level that characterizes the current wear degree of the plunger assembly; Step 5: Determine whether the wear level is in an unhealthy state. If so, activate the compensation control module. The compensation control module generates an adaptive compensation control quantity based on the current operating data and the wear level, and sends it to the actuator of the equipment system to compensate for the performance degradation caused by wear.

2. The method according to claim 1, characterized in that, In step one, the operating condition data is acquired from at least one of a speed sensor, a pressure or torque sensor, a temperature sensor, and a control signal acquisition unit; the status data is acquired from at least one of a high-frequency dynamic pressure sensor, a broadband vibration acceleration sensor, or an acoustic emission sensor.

3. The method according to claim 1, characterized in that, The method for extracting the wear features in step three is as follows: the residual sequence is segmented using a sliding time window, and a set of statistical features is extracted from the residual data of each window to form a residual feature vector; the wear state recognition model in step four is a temporal convolutional network model, which learns the evolution pattern of the residual feature vector over time through its internal multi-layer causal convolution, dilated convolution and residual connection structure, and maps the input residual feature vector to a predefined wear level; The predefined wear levels include: healthy condition, early wear, intermediate wear, and severe wear.

4. The method according to claim 3, characterized in that, The wear levels are classified based on physical indicators. Early wear corresponds to an increase of 5-10 micrometers in the clearance between the plunger components, medium wear corresponds to an increase of 10-20 micrometers in the clearance, and severe wear corresponds to an increase of more than 20 micrometers in the clearance or a drop in system pressure exceeding a predetermined threshold.

5. The method according to claim 1, characterized in that, The compensation control module in step five is implemented based on a fuzzy inference system. The fuzzy inference system uses the current operating condition data and the wear level as two input variables. Through its internal preset fuzzy rule base, it performs fuzzification, fuzzy inference, and defuzzification operations to calculate the accurate adaptive compensation control quantity.

6. The method according to claim 5, characterized in that, The adaptive compensation control quantity is sent to the actuator of the equipment system to fine-tune the duty cycle of the solenoid valve control signal of the piston drive mechanism, or to change the power output of the drive motor.

7. The method according to claim 1, characterized in that, Following step three, a model adaptive update step is also included: continuously monitoring the statistical characteristics of the residual sequence; when the energy or amplitude of the residual sequence runs continuously for more than a preset duration under multiple operating conditions and remains below the health confidence threshold determined based on the residual statistical characteristics of the health dataset for a long period of time, it is determined that the current plunger assembly is still in a healthy state, and the operating condition-state data pairs of this stage are automatically collected as new health samples; after accumulating a predetermined number of new samples, the parameters of the pre-trained dynamic baseline model of the health state are fine-tuned using transfer learning technology.

8. The method according to claim 1, characterized in that, The decoder of the conditional variational autoencoder in step two adopts a recurrent neural network structure based on a long short-term memory network or a gated recurrent unit.

9. The method according to claim 3, characterized in that, Following step four, the method further includes: establishing an autoregressive integral moving average model or an independent recurrent neural network to model the time series of the historical wear level sequence or the preset statistical features in the residual feature vector, thereby predicting the evolution trajectory of the wear state in the future.

10. A wear online identification and compensation control system for a high-pressure distribution plunger assembly, characterized in that, include: The data acquisition module is used to collect working condition data reflecting the operating conditions of the plunger assembly and status data reflecting its working status in real time, and to perform time synchronization and preprocessing. The data processing module is electrically connected to the data acquisition module. It integrates a health status dynamic baseline model and a wear status identification model trained according to the method of claim 1. It is used to receive the working condition data and the status data to generate a residual signal for characterizing the wear information of the plunger assembly, and to identify the current wear level of the plunger assembly based on the residual signal. The compensation control module, electrically connected to the data processing module, is used to generate an adaptive compensation control command based on the wear level identified by the data processing module and the current working condition data provided by the data acquisition module, and send the command to the actuator of the equipment system.

Images