A high-pressure fuel distribution pump head state monitoring and fault prediction method and system based on deep learning

Abstract

The application discloses a high-pressure fuel distribution pump head state monitoring and fault prediction method and system based on deep learning, and comprises the following steps: acquiring multi-source sensing data of the pump head through an innovative sensor layout scheme corresponding to subsequent graph neural network (GNN) model nodes; inputting the data into a physical information neural network (PINN) model constructed based on an enhanced physical model to obtain physical residuals of the quantitative data and the deviation degree of the physical law, wherein the enhanced physical model considers the nonlinear characteristics of the fuel, such as pressure-compressibility, temperature-viscosity and the like; abstracting the physical structure of the pump head into a GNN model containing multiple component nodes; updating the physical residuals to the features of the corresponding component nodes in the GNN model based on a mapping rule; and performing reasoning through the GNN model to output fault diagnosis information. The application solves the problems of poor interpretability, dependence on massive fault data and weak generalization ability of traditional deep learning methods, and realizes the diagnosis and prediction of early weak faults of the pump head.

Description

Technical Field

[0001] This invention relates to the field of mechanical equipment condition monitoring and predictive maintenance technology based on artificial intelligence, and more specifically, to a method and system for high-pressure fuel distribution pump head condition monitoring and fault prediction based on deep learning. Background Technology

[0002] The high-pressure fuel distribution pump head, as the core power source of the high-pressure fuel injection system, plays a crucial role in the internal combustion engine system. Its main function is to draw fuel from the low-pressure fuel line, precisely pressurize it to hundreds or even thousands of atmospheres through the reciprocating motion of its internal precision mechanical structure, and then, according to the engine's real-time operating requirements, strictly deliver the high-pressure fuel to the injectors of each cylinder according to a set time, sequence, and quantity. This component operates continuously under extreme conditions of high temperature, high pressure, high-frequency alternating loads, and fuel scouring. Its internal precision components, such as the plunger and plunger sleeve, the delivery valve and valve seat, as well as key transmission and control parts such as the camshaft, tappets, and springs, are subjected to enormous mechanical and thermal stresses, making it one of the weakest links in the entire engine system with a high failure rate.

[0003] A malfunction in the high-pressure fuel distribution pump head—such as internal leakage due to wear, cracks due to fatigue, jamming due to fuel contamination, or control failure due to spring failure—will trigger a series of chain reactions. Minor malfunctions may lead to decreased fuel atomization quality and incomplete combustion, resulting in insufficient engine output, abnormally high fuel consumption, and excessive emissions. Severe, sudden malfunctions can directly cause engine stalling or even irreversible serious consequences due to mechanical damage, posing a significant threat to equipment safety, transportation efficiency, and even personal safety. Therefore, this field requires a technology capable of real-time and accurate monitoring of the health status of the high-pressure fuel distribution pump head and early prediction of potential malfunctions. This technology has irreplaceable engineering value and economic significance for ensuring the safe, reliable, and economical operation of modern internal combustion engine drive systems.

[0004] Currently, the condition monitoring and fault diagnosis technologies for high-pressure fuel distribution pump heads can be broadly categorized based on their technological evolution. The first category is traditional diagnostic techniques based on signal processing. This approach typically involves installing vibration acceleration sensors or acoustic sensors externally to the pump body to collect vibration or noise signals generated during operation. Technicians then analyze the signal's characteristic changes in the time, frequency, or time-frequency domains using signal processing algorithms such as Fourier Transform (FFT), Wavelet Transform (WT), or Empirical Mode Decomposition (EMD). For example, by observing amplitude changes at specific frequencies (such as the pump oil frequency and its harmonics) or by calculating statistical indicators such as kurtosis and margin of the signal and comparing them with preset health status thresholds, an anomaly can be determined. The advantages of this type of method are its intuitive principle and low computational load. However, it largely relies on the prior knowledge and experience of diagnostic experts, is insensitive to early, subtle fault characteristics with low signal-to-noise ratios, and struggles to cope with complex and variable actual operating conditions, resulting in limited automation and intelligence in the diagnosis.

[0005] The second category is diagnostic techniques based on traditional machine learning. To overcome the over-reliance on expert experience in traditional signal processing techniques, this field has begun to introduce machine learning algorithms. The core of this approach is feature engineering + classifier. First, a series of statistical or physical features deemed effective in distinguishing different health states are manually extracted from the raw signal, forming a high-dimensional feature vector. Then, machine learning models such as Support Vector Machines (SVM), Artificial Neural Networks (ANN), decision trees, or Bayesian classifiers are used to train and learn these feature vectors, establishing a mapping relationship from features to fault types. This method improves the automation and accuracy of diagnosis to some extent, but its performance is firmly capped by the quality of feature engineering. The feature selection process still requires a great deal of expertise and repeated experimentation, and a carefully designed feature set is often only effective for specific equipment and operating conditions, lacking good transferability and generalization ability.

[0006] The third category represents a current research hotspot: diagnostic techniques based on standard deep learning. Deep learning models, represented by Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and their variant Long Short-Term Memory Networks (LSTMs), have brought new breakthroughs to the field of fault diagnosis due to their powerful end-to-end feature self-learning capabilities. These methods typically convert the original one-dimensional time-series signal into a two-dimensional time-spectrum graph using techniques such as Short-Time Fourier Transform (STFT), then treat it as an "image" and input it into a CNN for classification; or directly input the time series into an LSTM, utilizing its ability to capture temporal dependencies for diagnosis. This approach eliminates the tedious and subjective manual feature extraction process, and under conditions of sufficient data, it often achieves higher diagnostic accuracy than traditional machine learning.

[0007] However, despite the significant achievements of standard deep learning methods, their application in real-world scenarios, such as critical industrial equipment like high-pressure fuel distribution pump heads, still reveals some fundamental and insurmountable shortcomings. First, there is the "black box" nature of the model and its lack of interpretability. Deep neural networks make decisions through complex nonlinear combinations of millions of parameters, and their internal reasoning process is almost entirely opaque to humans. When the model outputs a diagnostic conclusion, it cannot provide a clear, physically logical explanation that engineers can understand and trust. For example, it can identify a fault, but it cannot answer "which specific part, such as plunger number 2 or the delivery valve, and what physical reason, such as wear-induced leakage or jamming, caused the current abnormal signal?" This vague diagnostic result significantly diminishes its value in guiding subsequent precise maintenance, spare parts management, and operational decisions.

[0008] Secondly, there is a strong dependence on massive, balanced, and labeled fault samples. The superior performance of deep learning models is predicated on supervised learning on large-scale datasets. However, in real industrial production environments, high-value equipment operates in good health most of the time, and faults, especially early faults for which data can be collected, are scarce events. Collecting a dataset that covers all potential fault modes, with tens of thousands of labeled samples for each mode, is extremely costly, even unacceptable, in terms of time, economic cost, and even safety risks.

[0009] Finally, there is the problem of insufficient generalization ability and robustness of the model. Purely data-driven models learn the statistical regularities of data under a specific distribution. Once the operating conditions of the equipment (such as engine speed, load, fuel grade, and ambient temperature) change, the data distribution of the sensor signals will also change accordingly. For the model, it is difficult to distinguish between signal fluctuations caused by changes in normal operating conditions and those caused by real faults, which easily leads to a large number of false alarms and false negatives, causing its performance to deteriorate sharply in the dynamic environment of real-world applications.

[0010] In summary, existing technologies, when applied to condition monitoring and fault prediction of complex and critical industrial equipment such as high-pressure fuel distribution pump heads, still face a series of pressing technical challenges, including poor model interpretability, reliance on massive amounts of fault data, and insufficient robustness to varying operating conditions. Therefore, there is an urgent need in this field for a novel technical solution that can overcome the limitations of purely data-driven approaches and deeply couple the nonlinear fitting capabilities of deep learning with the inherent physical operating mechanisms of the equipment. Summary of the Invention

[0011] The purpose of this invention is to provide a method and system for monitoring the condition and predicting the faults of high-pressure fuel distribution pump heads based on deep learning, so as to solve the technical pain points of existing deep learning models in industrial applications, such as the black box problem, strong dependence on massive fault data, and insufficient generalization ability under varying operating conditions.

[0012] To achieve the above objectives, this invention provides a deep learning-based method for monitoring the condition and predicting the faults of a high-pressure fuel distribution pump head. First, the method's initial step involves acquiring multi-source sensor data from the high-pressure fuel distribution pump head through an innovative sensor layout scheme tightly coupled with the subsequent inference model structure. This allows for comprehensive and multi-dimensional capture of the equipment's operational status information, rather than relying solely on a single physical quantity. Next, in the second step, the acquired multi-source sensor data is input into a pre-defined Physics-Informed Neural Network (PINN) model. The core task of this step is to use the PINN model to calculate and output an index that quantitatively characterizes the degree of deviation between the sensor data and the equipment's inherent physical operating laws—the physical residual. It is noteworthy that the Physics-Informed Neural Network model used in this invention is not based on idealized standard physical equations, but rather on an enhanced physical model that incorporates the nonlinear physical characteristics of fuel under extreme operating conditions. This makes the model closer to the real physical process, thereby improving the sensitivity and accuracy of detection.

[0013] After identifying system-level operational anomalies using the PINN model, the third and fourth steps of the method trace and locate the anomalies. Specifically, the third step mathematically abstracts the physical structure of the high-pressure fuel distribution pump head into a graph neural network (GNN) model containing multiple component nodes. Each node in the graph corresponds to a specific physical component inside the pump head, such as a plunger or delivery valve, and the connections between nodes represent the physical interactions between components. In the fourth step, based on a pre-defined mapping rule, the method updates the node features of the corresponding component nodes in the graph neural network model with the physical residuals output from the PINN model in the previous step. This step is a crucial bridge connecting system-level anomaly detection and component-level fault location. Finally, in the fifth step, by running the graph neural network model for information propagation and inference, a detailed fault diagnosis report of the high-pressure fuel distribution pump head, containing fault location information, is output, clearly identifying the most likely root cause component of the failure.

[0014] Building upon the aforementioned method, another aspect of the present invention provides a deep learning-based high-pressure fuel distribution pump head condition monitoring and fault prediction system for executing the method. This system includes a data acquisition module, a physical analysis module, and a fault reasoning module. First, the data acquisition module performs the first step of the aforementioned method, implementing the innovative sensor layout and acquiring multi-source sensor data from the pump head. Second, the physical analysis module internally deploys and runs the aforementioned physical information neural network model. It receives data from the data acquisition module and executes the second step of the method, outputting key physical residuals. Finally, the fault reasoning module acts as the system's fault locator. Internally, it constructs and maintains a graph neural network model isomorphic to the pump head's physical structure and is configured to execute the third, fourth, and fifth steps of the method: receiving physical residuals, updating the GNN model through mapping rules, and finally outputting a diagnostic result containing fault location information through reasoning.

[0015] This invention, through the aforementioned technical solution, constructs an intelligent diagnostic scheme that deeply integrates data-driven approaches with physical mechanisms. Its core technical concept lies in utilizing a Physical Information Neural Network (PINN) as a highly sensitive deviation detector. This detector uses the inherent, corrected physical operating laws of the equipment as a benchmark to evaluate whether real-time operating data meets requirements. Any minute, persistent deviation from physical laws caused by early faults will be significantly amplified in the form of physical residuals. Subsequently, a Graph Neural Network (GNN) is used as a powerful logical reasoning engine to perform structured, causal attribution analysis on this abstract, system-level non-compliance information through a graph model isomorphic to the equipment's physical topology, ultimately focusing the fault risk onto specific physical components.

[0016] Compared to existing technologies, the method and system proposed in this invention bring significant technological advancements and beneficial effects. First, by introducing physical laws as strong constraints for the neural network, this invention improves the accuracy of fault diagnosis and its sensitivity to early, subtle faults, because its judgment is no longer based solely on the statistical correlation of data, but rather on the consistency between the data and physical laws. Second, through the attribution and localization of faults using GNNs, this invention fundamentally solves the "black box" problem of traditional deep learning models, making the diagnostic results clearly interpretable and providing a direct and reliable basis for subsequent maintenance decisions. Third, since the physical model itself contains a large amount of prior knowledge, the method of this invention no longer relies on massive amounts of labeled fault data, and can even perform effective training and anomaly detection with only healthy data, greatly lowering the barrier to its application in industry. Finally, because the model is built based on physical equations that are universally applicable to operating conditions, this invention is more robust to changes in tooling such as speed and load, and possesses better generalization capabilities. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present 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 only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the overall process of a method according to an embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram of the overall system architecture according to an embodiment of the present invention.

[0020] Figure 3 This is a schematic diagram of a sensor optimization layout scheme according to an embodiment of the present invention.

[0021] Figure 4 This is a schematic diagram of the PINN and GNN fusion mechanism according to an embodiment of the present invention.

[0022] Figure 5 This is a schematic diagram of a GNN fault reasoning process according to an embodiment of the present invention.

[0023] Figure 6 This is a performance analysis chart comparing the application effect of an embodiment of the present invention with that of the prior art. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0025] Example 1

[0026] This embodiment details the specific execution steps of the deep learning-based high-pressure fuel distribution pump head condition monitoring and fault prediction method proposed in this invention. (Refer to...) Figure 1 This method constitutes a closed-loop intelligent processing flow from data to decision-making.

[0027] The first step of this method, S101, is to acquire multi-source sensor data from the high-pressure fuel distribution pump head. In modern industrial systems, sensor data from a single type or single measuring point often only reflects one aspect of the equipment's status and is easily affected by local interference or fluctuations in operating conditions, leading to information loss or misjudgment. To construct a comprehensive and three-dimensional profile of the equipment's status, this embodiment employs a multi-source heterogeneous data acquisition strategy. Furthermore, this invention proposes an innovative sensor layout scheme. The core idea is to directly map the physical layout of the sensors to the logical structure of the subsequent graph neural network model, thereby providing high-quality data input with clear physical directionality for accurate fault location.

[0028] Reference Figure 3 This sensor layout scheme is designed around the component nodes of the graph neural network (GNN) model constructed in subsequent step S103. Specifically, the core of this scheme lies in localized monitoring. Unlike traditional methods that randomly or empirically select a measuring point on the pump body to represent the overall vibration, the scheme of this invention requires pre-setting physical locations on the pump head associated with key component nodes in the GNN model, and strategically placing sensors there. A preferred embodiment is:

[0029] For the "plunger" component node in the GNN model, a vibration sensor is installed on the outer wall (V1) of its physical corresponding pump body plunger sleeve. The sensor at this location can most directly and sensitively capture dynamic events such as friction, impact, and radial knocking that may occur due to increased wear clearance during the reciprocating motion of this specific plunger pair. The data collected provides directly relevant evidence for the health status assessment of the "plunger" node.

[0030] For the "outlet valve" component node in the GNN model, a vibration sensor is installed near the valve seat (V2) of its physical counterpart on the pump body. The opening and closing of the outlet valve are two instantaneous high-frequency impact events, and the sensor at this location can clearly pick up the signals of these two events. Once the outlet valve becomes stuck, experiences delayed closure, or fails to seal properly, it will cause changes in the waveform, amplitude, or subsequent vibration attenuation mode of the impact signal, thus providing crucial information for fault diagnosis of the "outlet valve" node.

[0031] For the "camshaft" component node in the GNN model, a vibration sensor is installed near the pump body cam bearing housing (V3) where it physically corresponds. The vibration signal at this location mainly reflects the smoothness of camshaft rotation, the operating condition of the bearing, and the health of the cam profile.

[0032] By implementing this one-to-one layout strategy corresponding to GNN nodes, the system no longer collects fuzzy, globally mixed vibration signals, but rather multi-channel, decoupled signals that directly reflect the local dynamic behavior of specific components. This is a key difference between this invention and existing technologies. Traditional methods typically select one or a few locations on the pump body to collect mixed vibration signals representing the overall state of the equipment. These signals are the result of the propagation and superposition of multiple vibration sources, and the independent state information of each component within them is severely coupled and submerged. When an early fault occurs, the weak signal characteristics it generates are easily masked by signals from other strong vibration sources, resulting in an extremely low signal-to-noise ratio and making detection difficult. In contrast, the layout scheme of this invention, by placing dedicated sensors near each key component, naturally provides data with a clear physical spatial orientation. For example, the data from channel V1 mainly reflects the state of the plunger assembly, while the data from channel V2 mainly reflects the state of the oil valve. This localized and decoupled data can provide higher signal-to-noise ratio and more directional inputs for the initial feature vectors of each component node in the subsequent GNN model, reducing the ambiguity and uncertainty of GNN in fault tracing and inference. It is a necessary prerequisite and physical basis for achieving high-precision fault location.

[0033] In addition to the aforementioned local vibration monitoring, this scheme also includes monitoring of global performance and boundary conditions. For example, a pressure sensor is installed at the main high-pressure oil outlet (P1) to monitor the final output performance of the entire pump head system, reflecting its overall hydrodynamic state; a temperature sensor is installed on the pump casing (T1) to obtain the overall operating temperature of the pump. These data provide global constraints and necessary parameter inputs for the subsequent physical model.

[0034] During the data acquisition and preprocessing stage, a high-speed data acquisition card supporting multi-channel synchronous acquisition is required to ensure that all sensor data are strictly aligned in the time dimension. The sampling frequency setting must follow the Nyquist sampling theorem, typically taking a value greater than 2.56 times the upper limit of the sensor's effective frequency response range. The acquired raw digital signal sequence also needs to undergo a series of preprocessing operations to provide clean and standardized input for subsequent deep learning models. These operations include, but are not limited to, wavelet packet thresholding denoising, min-max normalization, and sliding window segmentation with overlap. A typical preprocessing workflow includes: First, using methods such as wavelet packet thresholding denoising to filter out Gaussian white noise and power frequency interference in the signal, while retaining as much as possible the non-stationary impact components caused by faults. Second, for data acquired from different sensors with different physical dimensions and numerical ranges, min-max normalization or Z-score standardization is performed to uniformly map them to the interval [-1, 1] to eliminate the adverse effects of dimensional differences on the model training convergence speed and stability. Finally, the continuous time series data is segmented into overlapping data segments using a fixed-length sliding window (e.g., a window length of 2048 data points with an overlap rate of 50%). Each data segment is treated as an independent sample instance and input into the subsequent model for analysis.

[0035] The second step of this method, namely step S102, involves inputting the multi-source sensor data into a preset physical information neural network model (PINN) to obtain a physical residual that characterizes the degree of deviation between the multi-source sensor data and the preset physical model. The physical residual refers to the deviation between the output of the preset enhanced physical model and the theoretical value (which should be 0) after substituting the multi-source sensor data into it. This step is the core of this invention for achieving high-sensitivity anomaly detection.

[0036] Specifically, this step first requires constructing an enhanced physical model that can accurately describe the operating mechanism of the high-pressure fuel distribution pump head under healthy conditions. This embodiment emphasizes constructing a nonlinear model that more closely resembles real-world operating conditions. This model mainly consists of control equations describing the dynamic changes in fuel pressure within the plunger chamber.

[0037] Assuming the fuel pressure in the plunger chamber is P(t), the volume is V(t), the inflow rate is Qin(t), the outflow rate is Qout(t), and the bulk modulus of the fuel is K, the basic pressure-flow relationship can be described by the following differential equation:

[0038] dP(t) / dt = (K / V(t)) * (Qin(t) - Qout(t) - dV(t) / dt)

[0039] Traditional models typically treat the bulk modulus K and fuel viscosity μ as constants. However, in high-pressure fuel systems, these constants are highly subjective. This embodiment corrects this:

[0040] A pressure-compressibility nonlinear model is introduced: the compressibility of fuel oil is not constant. Under high pressure, its bulk modulus changes significantly. This invention introduces a bulk modulus function K(P) that varies with pressure P to replace the constant K. This function can be obtained by fitting empirical formulas or experimental data; for example, a commonly used form is:

[0041] K(P) = K0 + aP + bP 2

[0042] Where K0 is the bulk modulus at standard atmospheric pressure, and a and b are coefficients calibrated through high-pressure fluid experiments. This correction allows the model to more accurately describe the pressure build-up process during the high-pressure phase.

[0043] A temperature-viscosity nonlinear model is introduced: fuel viscosity μ is strongly dependent on temperature T, and viscosity directly affects leakage (leakage is an important characteristic of early failures), thus affecting the calculation of Qout(t). This invention utilizes the collected pump body temperature T to dynamically calculate the fuel viscosity μ(T) using a nonlinear temperature-viscosity model (such as the Vogel model):

[0044] μ(T) = A * exp(B / (T - C))

[0045] Where A, B, and C are the physical property constants of a specific fuel. This closely links the leakage model to the actual operating temperature.

[0046] Integrating the above nonlinear models, we obtain a more accurate set of nonlinear ordinary differential equations f concerning pressure P(t), coupled with the effects of temperature and pressure, in the form:

[0047] f(t, P(t), dP(t) / dt, ω, T) = 0

[0048] Where ω is the engine speed (affecting Qin and V(t)), and T is the pump body temperature. This system of equations f represents the physical constraints of PINN.

[0049] After setting up this enhanced physical model, the PINN model can be constructed. PINN is essentially a deep neural network, specifically implemented as a fully connected network (MLP). In a preferred embodiment, the MLP can be designed to contain 5 to 10 hidden layers, each containing 64 to 256 neurons. To ensure the network can approximate smooth physical function solutions well and facilitate the calculation of their derivatives, the activation functions of the hidden layers are preferably hyperbolic tangent (tanh) or Swish functions. This choice of network depth and width reflects a balance between ensuring the model has sufficient nonlinear expressive power to capture complex physical processes and avoiding overfitting due to too many parameters while also reducing training costs. The network's input layer receives time point t and the corresponding operating parameters (speed ω, temperature T), while the output layer outputs the predicted system state at that time, i.e., the high-pressure oil rail pressure P̂(t).

[0050] Network input: time point t, and the corresponding operating parameters (speed ω, temperature T).

[0051] Network output: The predicted system state at this point in time, mainly the high-pressure oil rail pressure P̂(t).

[0052] Network structure: A fully connected network (MLP) with multiple hidden layers is used. For example, it can be a 5-layer network with 128 neurons in each layer. The hyperbolic tangent function (tanh) is used as the activation function to ensure the smoothness of the output and its derivative.

[0053] The training process of PINN is optimized using a composite loss function L_total:

[0054] L_total = λ_d * L_data + λ_p * L_physics

[0055] The data loss (L_data) is used to ensure that the network's predicted value P̂(t) is as close as possible to the actual pressure value P_m(t) measured by the sensor. It is calculated using the mean squared error (MSE): L_data = (1 / N) *Σ[ (P̂(t_i) - P_m(t_i)) 2 ], where N is the number of data points in the training samples.

[0056] Physics Loss (L_physics): This part is the core of PINN, used to penalize violations of our constructed augmented physics model f by the network output. The network output P̂(t) and its derivative dP̂(t) / dt, calculated using automatic differentiation, are substituted into the augmented physics equations f. Ideally, the equations should hold, i.e., f = 0. The physics loss is the mean square value of the residuals from these equations.

[0057] L_physics = (1 / M) * Σ[ f(t_j, P̂(t_j), dP̂(t_j) / dt, ω_j, T_j) 2 ]

[0058] Where M is the number of placement points selected across the entire solution domain to enforce physical constraints. These points can even be taken in locations without sensor data, which is one of the reasons why PINN can perform few-shot learning.

[0059] λ_d and λ_p are two hyperparameters used to balance the weights of data-driven and physics-driven approaches during model training.

[0060] By using the backpropagation algorithm and minimizing L_total, PINN can learn a function approximator that can both fit the measurement data and strictly adhere to the underlying physical laws.

[0061] After training, for any new set of input data, not only can the predicted pressure P̂(t) be obtained, but more importantly, the physical residual r(t) = f(t, P_m(t), ...) can be calculated. Under healthy conditions, the actual data P_m(t) should perfectly satisfy the physical equations, therefore the value of r(t) will be very small, close to zero. When an early fault occurs in the equipment (such as a minor leak causing an abnormal pressure build-up process), even if the change in P_m(t) is very weak and insufficient to be detected by traditional thresholding methods, its deviation from physical laws will cause r(t) to exhibit a significant and persistent non-zero value. This physical residual r(t) is the core input sent to the next stage of the GNN module; it is a highly sensitive indicator of faults with underlying physical meaning.

[0062] The third step of this method, namely step S103, is to abstract the physical structure of the high-pressure fuel distribution pump head into a graph neural network model containing multiple component nodes. Preferably, this invention deconstructs the pump head into a directed attribute graph G = (V, E, X, A) to achieve a high degree of isomorphism between the mathematical model and the physical entity. In this definition:

[0063] Here, V represents the set of nodes in the diagram, where each node v_i ∈ V corresponds to a core physical component inside the pump head. For example, the node set of a six-cylinder pump can be defined as V = {v_cam, v_tappet_1..6,v_plunger_1..6, v_sleeve_1..6, v_valve_1..6, v_spring_1..6}, representing the camshaft, 6 tappets, 6 plungers, 6 plunger sleeves, 6 delivery valves, and 6 springs, respectively.

[0064] E represents the set of directed edges in the graph, where each edge e_ij ∈ E represents a defined physical interaction relationship from node v_i to node v_j. For example, (v_cam, v_tappet_i) represents the driving relationship between the cam and the i-th tappet, and (v_plunger_i, v_valve_i) represents the effect of the high-pressure oil pumped by the i-th plunger on the outlet valve.

[0065] X represents the set of node features, where the feature vector x_i of each node is a digital description of the state of the component. This vector consists of two parts: one part is the static or quasi-static properties of the component, such as design parameters, materials, etc.; the other part is the dynamic observation value strongly correlated with the node, specifically, the statistical features (such as root mean square, kurtosis, frequency band energy, etc.) obtained after preprocessing the local sensing signals collected by the innovative sensor layout scheme in step S101.

[0066] A is the adjacency matrix of the graph, which is derived from the set of edges E. It is a mathematical representation of the graph topology and is used to define the neighborhood relationships between nodes in the computation of GNN.

[0067] Based on the above definitions, this embodiment constructs a mathematical model that is highly isomorphic to the physical entity of the device in terms of structure, relationships, and states.

[0068] The fourth step of this method, namely step S104, involves updating the physical residual to the node features of the corresponding component nodes in the graph neural network model based on a preset mapping rule. This is the bridge connecting PINN and GNN, and is also key to achieving interpretable diagnostics. To this end, this invention designs a feature update mechanism based on physical prior knowledge, which includes two core steps:

[0069] The first step is to decompose the total physical residual r(t) output by PINN. Because the enhanced physical model f constructed in this invention is itself composed of multiple sub-equations with clear physical meaning (e.g., the continuity equation describing flow conservation, the Newton's second law equation describing plunger motion, etc.), the total physical residual r(t) can therefore be decomposed into residual components corresponding to different physical sub-equations, such as the flow residual r_flow(t) and the dynamic residual r_dynamics(t).

[0070] In a preferred embodiment, the total physical loss L_physics is defined at design time as a weighted sum of multiple physical loss components:

[0071] L_physics = w_flow * L_flow + w_dynamics * L_dynamics + ...

[0072] Where L_flow is the loss that is only related to the sub-equation f_flow = 0 which describes the conservation of fluid mass (i.e., flow continuity); L_dynamics is the loss that is only related to the sub-equation f_dynamics = 0 which describes the motion of the plunger and transmission components (i.e., Newton's second law); w_flow and w_dynamics are the corresponding weighting coefficients.

[0073] Therefore, the physical residual decomposition step is not a fuzzy split of a scalar total residual, but rather, during online monitoring, the sensor measurement data is substituted into these predefined, independent physical sub-equations to obtain a set of physical residual component vectors r(t) = [r_flow(t), r_dynamics(t), ...]. Here, r_flow(t) = f_flow(t, P_m(t), ...), and r_dynamics(t) = f_dynamics(t, P_m(t), ...). Each residual component is an independent scalar time series, and its amplitude change directly and uniquely reflects the degree of deviation of the real system from the theoretical model at the specific physical stage described by that physical sub-equation. This decomposition method is deterministic and programmable, providing a foundation for subsequent precise mapping.

[0074] The second step involves injecting these residual components, based on their physical meaning, into the feature vectors of the nodes in the GNN graph most likely to cause such bias. A preferred embodiment is as follows: if the flow residual r_flow(t) remains consistently high, this strongly suggests an abnormal leak or poor flow in the system. Therefore, the mapping rule adds the magnitude of r_flow(t) as a new, dynamic feature to the end of the feature vector x_i of all nodes directly related to sealing and flow control (i.e., all v_plunger, v_sleeve, and v_valve nodes). Similarly, if the dynamic residual r_dynamics(t) remains consistently high, indicating potential abnormal wear or impact in the force transmission chain, this residual value is updated in the feature vectors of nodes such as v_cam, v_tappet, and v_spring in the force transmission chain.

[0075] The mapping rule here abandons vague qualitative descriptions and instead adopts a structured association mechanism, for example, it can be implemented as a 'residual-component association matrix' or a lookup table. In a specific embodiment, the mapping rule can be defined as follows:

[0076] Rule 1: Traffic Residual Mapping

[0077] Triggering condition: The moving average of r_flow(t) exceeds the preset dynamic baseline threshold.

[0078] Physical logic: A persistently high flow residual strongly suggests, physically, an abnormal leak (such as due to wear) or poor flow (such as due to jamming) in the system. These phenomena must occur in the core components that constitute the fluid seal and control.

[0079] Operation: The current value of r_flow(t) is added as a new feature to the end of the feature vector x_i of all nodes in the GNN graph that are directly related to sealing and flow control. These nodes specifically include all plunger nodes (v_plunger_1..6), plunger sleeve nodes (v_sleeve_1..6), and outlet valve nodes (v_valve_1..6).

[0080] Rule 2: Dynamic Residual Mapping

[0081] Triggering condition: The kurtosis index or energy value of r_dynamics(t) exceeds the preset dynamic baseline threshold.

[0082] Physical logic: The persistently high dynamic residuals indicate that there may be abnormal impacts (such as excessive gaps) or abnormal forces (such as wear and fatigue) in the mechanical force transmission chain.

[0083] Operation: The current value of r_dynamics(t) is added as a new feature to the end of the feature vector x_i of all nodes that constitute the mechanical force transmission chain in the GNN graph. These nodes specifically include the camshaft node (v_cam), all tappet nodes (v_tappet_1..6), and all spring nodes (v_spring_1..6).

[0084] By implementing the aforementioned explicit and quantifiable mapping rules, this invention transforms system-level, abstract physical biases into component-level, concrete feature anomalies in a traceable and programmable manner. When the GNN model receives these anomalous features with precisely labeled physical origins, its subsequent inference process has a solid, causal-logical physical foundation. Through this mechanism, system-level, abstract physical biases are transformed into component-level, concrete feature anomalies, providing precise targets for the GNN's causal inference.

[0085] Reference Figure 4 This diagram illustrates the fusion mechanism of PINN and GNN. The left side of the diagram shows the output of the PINN physical analysis module, where the total physical residual r(t) is decomposed into multiple physically meaningful components, such as flow residual and kinetic residual. These residual components, like 'probes' carrying information, are directionally injected into the GNN graphical model on the right side of the diagram according to predefined mapping rules. For example, the update arrow for the flow residual (dashed line in the diagram) points to the 'plunger' and 'outlet valve' nodes related to flow and sealing, while the update arrow for the kinetic residual (dotted line in the diagram) points to the 'cam' node related to force transmission. The GNN graphical model itself is an abstraction of the pump head's physical structure, and the connections between nodes represent the actual drive and pumping relationships. This fusion process essentially transforms the unstructured system-level anomalies discovered by PINN into structured feature anomalies of specific nodes in the GNN model, thus completing a crucial step from fault phenomena to attribution hypotheses, providing high-quality, interpretable input for subsequent GNN inference.

[0086] The final step of this method, step S105, involves reasoning through the graph neural network model to output fault diagnosis information for the high-pressure fuel distribution pump head. (Refer to...) Figure 5Upon receiving a graph updated with abnormal features, the GNN model (preferably a Graph Attention Network, GAT) begins iterative information propagation and aggregation. After several layers of propagation, the node-level classifier at the top layer of the GNN calculates the probability that each component node is the root cause of the fault. The node with the highest probability value is determined as the fault location result. Further, to determine the specific type of fault, this invention includes a subsequent step. First, a graph pooling layer aggregates the feature vectors of all nodes in the last layer of the GNN into a unique, fixed-dimensional graph feature vector h_G, which characterizes the overall health status of the entire pump head system. The graph feature vector, a fixed-dimensional vector formed by aggregating the feature vectors of all nodes in the last layer of the graph neural network through a graph pooling layer (such as global average pooling), is used to represent the global state of the entire graph. Then, this graph feature vector is input into a pre-trained lightweight classifier (such as an MLP with two hidden layers). The classifier's output layer uses the Softmax function, outputting a probability distribution vector. Each dimension of the vector corresponds to the probability of a predefined fault type (such as "normal," "plunger wear," "oil valve jamming," "spring fatigue," etc.). Finally, the system integrates the localization results from the GNN and the qualitative results from the classifier to output a structured and comprehensive fault diagnosis report. This report preferably includes, but is not limited to, the following fields: equipment ID, diagnosis timestamp, overall equipment status (e.g., normal, warning, fault), fault root cause location (e.g., oil valve #3), fault type judgment (e.g., early-stage minor jamming), and confidence scores associated with these judgments (e.g., localization confidence 97%, classification confidence 94%). This detailed report provides equipment maintenance personnel with clear and actionable decision support.

[0087] Example 2

[0088] This embodiment details the specific structure and implementation of the deep learning-based high-pressure fuel distribution pump head condition monitoring and fault prediction system that implements the above method. (Refer to...) Figure 2 As a complete technical entity, the system can be logically divided into four closely cooperating functional modules: data acquisition module 10, physical analysis module 20, fault reasoning module 30, and diagnosis and output module 40.

[0089] First, the data acquisition module 10 is responsible for interacting with the physical world and serves as the data source for all subsequent analyses. Its function is to comprehensively and accurately capture multi-dimensional physical signals from the high-pressure fuel distribution pump head during operation. In its implementation, the data acquisition module 10 consists of two parts: hardware and embedded software. The hardware part mainly includes three types of devices: First, a sensor array. This array is deployed strictly according to the innovative sensor layout scheme associated with the GNN model node machine described in Embodiment 1. This array preferably consists of multiple piezoelectric accelerometers installed at specific locations on the pump body, a dynamic pressure sensor installed in the high-pressure oil circuit, and at least one temperature sensor for measuring the pump body temperature, forming a three-dimensional sensing network capable of covering key physical quantities such as vibration, pressure, and temperature. Second, a signal conditioning circuit. Since the raw electrical signals output from the sensors are often very weak and may contain noise, they need to be amplified, filtered, and anti-aliasing processed by the signal conditioning circuit to improve signal quality and signal-to-noise ratio. Third, a high-speed data acquisition card (DAQ). This is a high-performance analog-to-digital converter (ADC) responsible for digitizing conditioned analog signals at a set high sampling rate (e.g., 51.2kHz) and high precision (e.g., 24-bit), converting them into a digital signal stream that a computer can process. The module's embedded software controls the DAQ card's startup and shutdown, channel configuration, and sampling parameter settings. It also packages the acquired data stream, adds timestamps, and transmits it in real-time to subsequent processing modules via bus interfaces such as USB, Ethernet, or PCIe.

[0090] Secondly, the physical analysis module is responsible for performing the first level of intelligent analysis based on physical laws. Its core function is to execute step S102 in Embodiment 1, that is, to detect anomalies in the system's operating state with high sensitivity. This module is typically deployed on a hardware platform with strong computing power, such as an edge computing gateway or a backend server. Its core is a statically trained Physical Information Neural Network (PINN) model instance. The physical analysis module 20 can be further subdivided into three sub-units. The first is the data receiving and preprocessing unit, which is responsible for receiving the real-time data stream from the data acquisition module 10 and performing preprocessing operations such as noise reduction, normalization, and sliding window segmentation as detailed in Embodiment 1, converting the raw data into a standard input format acceptable to the PINN model. The second is the PINN forward inference unit, which is the core computing unit of this module. It inputs the preprocessed data samples into the PINN model for a forward computation. This computation process utilizes pre-trained network weights to predict the system behavior in the current state. The third is the physical residual calculation unit, which receives the prediction results from PINN. More importantly, it substitutes the actual measurement data into the enhanced physical model associated with PINN to calculate the physical residual r(t). This calculation unit is crucial for achieving the high-sensitivity detection of this invention. Finally, the physical analysis module 20 transmits the calculated, time-aligned physical residual sequence as its final output to the fault reasoning module 30.

[0091] Furthermore, the fault reasoning module 30 is responsible for performing in-depth, structured causal reasoning on the anomalies discovered by the physical analysis module 20. Its core function is to execute steps S103, S104, and S105 in Embodiment 1 to achieve accurate fault tracing and location. Similar to the physical analysis module 20, this module is also deployed on a computing server, and is particularly suitable for running in environments equipped with graphics processing units (GPUs) because graph neural network computation has high parallelism. The fault reasoning module 30 mainly includes the following key parts in its implementation. First, it is a predefined device graph model library. This library stores pre-built graph neural network (GNN) structure definition files for different models of high-pressure fuel distribution pump heads. This file describes in detail all the elements of the directed attribute graph G = (V, E, X, A) defined in Embodiment 1. Second, it is a feature update engine. This engine is responsible for executing the PINN and GNN fusion mechanism described in step S104 of Embodiment 1. It receives the physical residual stream from the physical analysis module 20 and, according to a preset, physics-based mapping rule, updates the residual values ​​in real time and directionally to the feature vectors of the corresponding nodes in the GNN graph. Third, it is a GNN inference engine. This engine loads pre-trained Graph Attention Network (GAT) model weights. Upon receiving a feature update instruction, it triggers a complete GNN forward inference process, including multi-layer information propagation, attention calculation, and node state updates. The final result of the inference is to output the confidence score of each component node as the root cause of the fault, thereby completing fault localization. Optionally, the engine also generates a graph feature vector of the global state of the features through graph pooling operations.

[0092] Preferably, the system also includes a diagnostic and output module 40, used to transform complex analysis results into clear and actionable diagnostic information. This module integrates the outputs of preceding modules and presents them in a user-friendly manner. In implementation, the diagnostic and output module 40 can be a software application running on a server or user terminal. Internally, it first includes an optional fault classification unit. This unit receives graph feature vectors from the fault reasoning module 30 and inputs them into a lightweight multilayer perceptron (MLP) classifier to determine the specific type of fault (e.g., wear, jamming). Then, a report generation unit summarizes all the analysis results, including fault location information (node ​​confidence from the GNN), fault type information (from the MLP classifier), and physical residual curves and raw signal segments used as diagnostic evidence. Finally, through a human-machine interface (HMI), this information is clearly displayed to equipment maintenance personnel or operation engineers in visual formats such as text, charts, and dashboards. In addition, the module can be configured with an alarm interface. When the diagnostic results meet the preset severity level, the alarm information and diagnostic report can be automatically pushed to the enterprise's Manufacturing Execution System (MES), Equipment Health Management System (PHM), or cloud platform via Industrial Ethernet, OPC UA protocol, or MQTT protocol to trigger automated maintenance work orders or decision-making processes.

[0093] Example 3

[0094] This embodiment will demonstrate, in conjunction with a specific industrial application scenario, how the method and system described in this invention can play a role in the fault prediction task of a high-pressure fuel distribution pump head (model: inline six-cylinder pump) on a test bench.

[0095] The goal of this application scenario is to predict, as early and accurately as possible, the early failures of the pump head's internal core components (especially the plunger assembly and the delivery valve assembly), such as internal leakage caused by minor wear or slight jamming caused by oil contamination, before the pump head experiences a significant performance degradation or complete failure.

[0096] To implement this invention, the test bench was configured as follows: The computing platform consisted of an edge computing server equipped with an NVIDIA Tesla T4 GPU. The data acquisition system used the NI CompactRIO platform, equipped with an NI 9234 four-channel dynamic signal acquisition module, with a global synchronous sampling frequency set to 51.2 kHz. Regarding sensors, following the sensor layout scheme proposed in Embodiment 1 of this invention, a total of 12 Kistler 8702B500 wideband accelerometers were installed on the outer walls of the six plunger sleeves and at the corresponding six outlet valve seats; a WIKA A-10 high-frequency dynamic pressure sensor was installed on the main high-pressure outlet oil pipe of the pump head; and a K-type armored thermocouple was attached to the pump body casing.

[0097] The fault prediction process includes:

[0098] 1. Health Model Training Phase: During the initial installation of the pump head, the system collected health status data for 100 hours of continuous operation under different speeds (from 800 rpm to 2000 rpm) and loads (from 25% to 100%). Using this pure health data, the system trained the PINN physical analysis module offline. Due to the strong constraints of the augmented physical model, PINN was able to learn the precise physical behavior patterns of the pump head under various normal operating conditions. Simultaneously, using this health data and the corresponding equipment graph model, the GNN fault reasoning module was also baseline-trained, enabling it to understand the normal distribution of health status graph information.

[0099] 2. Online Monitoring and Early Warning: After the pump head has been in long-term operation, the system enters online monitoring mode. During a routine check after approximately 1500 hours of operation, on-site maintenance personnel did not find any abnormalities. However, the system's backend issued an early warning signal. The physical analysis module 20 detected that in the output of the physical residual r(t), the residual component r_flow(t) related to the fluid continuity equation exhibited a small, stable, and persistent positive spike during the depressurization phase in each cycle corresponding to the pump oil in cylinder 3. This phenomenon was judged by the PINN model as a high-confidence deviation from the healthy physical model, physically indicating an abnormal, delayed fluid shutdown event.

[0100] 3. Fault Origin and Diagnosis: Upon receiving the abnormal physical residual, the fault reasoning module 30 immediately starts. According to the preset mapping rules, the outlier value of r_flow(t) is directionally updated to all nodes in the GNN graph directly related to fuel flow and sealing, namely all six "plunger" nodes and six "outlet valve" nodes. Simultaneously, the local vibration data provided by the data acquisition module 10 also shows that the kurtosis index of the sensor signal installed on the valve seat of outlet valve 3 (V2_3) shows a slight increase of 0.8%. This information is used as input and propagated in the GNN inference engine. Because the "outlet valve 3" node (v_valve_3) simultaneously receives strongly correlated physical residual evidence from PINN and direct dynamic behavior anomaly evidence from local sensors, its anomaly weight is significantly amplified under the attention mechanism of GAT. After two layers of information propagation, the final output of GNN is: the fault root cause confidence of the v_valve_3 node is as high as 0.97, while the confidence of all other nodes is below 0.01. The diagnosis and output module 40 further inputs the image feature vector at this time into the classifier, determines the fault type as "stuck", and finally generates a diagnosis report: "Warning: There are early slight signs of sticking in oil outlet valve No. 3. It is recommended to perform an endoscopic examination at the next shutdown."

[0101] 4. Verification and Effect Evaluation: Based on the diagnostic report, the maintenance team disassembled and inspected the No. 3 delivery valve of the pump head during the next planned maintenance. They discovered scratches on the valve seat caused by tiny oil sludge particles, almost invisible to the naked eye, confirming a slight closure issue. Without the system's early warning, this fault could very likely have developed into severe jamming or sealing problems within the next few hundred hours, leading to abnormal fuel injection in that cylinder and causing a series of problems such as engine vibration and power loss.

[0102] To quantify the advantages of this invention, a set of comparative experiments were conducted. Fault data from the same batch were input into the system of this invention and a diagnostic system based on the prior art, which uses CNN to classify the time-spectrum graphs of vibration signals. Figure 6 The performance comparison chart shown indicates that the experimental results are as follows:

[0103] Regarding fault detection accuracy, both methods achieve over 95% accuracy for obvious faults that have developed to a certain extent. However, in the crucial indicator of early fault detection rate (defined as the probability of successful detection when the fault feature signal-to-noise ratio is below 3dB), the system of this invention achieves 92%, while the CNN method, unable to effectively distinguish subtle texture changes from a noisy background, has a detection rate of only 48%. In terms of positioning accuracy, the CNN method, due to its "black box" nature, cannot output component-level positioning information, hence this indicator is 0%; while the present invention, benefiting from GNN and a matching sensor layout, achieves a positioning accuracy of 97%. In the variable operating condition false alarm rate test, when the engine speed suddenly changes, the false alarm rate of the CNN method jumps to 25%, while the present invention, due to the inherent description of the operating condition variables by a physical model, maintains a false alarm rate below 2%.

[0104] In summary, this embodiment fully demonstrates the feasibility and effectiveness of the present invention in practical industrial applications. The present invention not only achieves earlier and more accurate fault prediction than existing technologies, but more importantly, it transforms the powerful capabilities of deep learning into a truly usable and traceable decision-making basis for engineers through an interpretable and traceable approach.

[0105] It should be noted that in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. The above embodiments are merely illustrative of the technical solutions of the present invention and not intended to limit it; the present invention has been described in detail only with reference to preferred embodiments. Those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications and substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for monitoring the condition and predicting the fault of a high-pressure fuel distribution pump head based on deep learning, characterized in that, Includes the following steps: Multi-source sensing data of the high-pressure fuel distribution pump head is acquired, wherein sensors are arranged at preset physical locations of the high-pressure fuel distribution pump head associated with component nodes in the graph neural network model to acquire local dynamic sensing data directly related to the state of the component nodes. The preset physical locations include the outer wall of the pump body plunger sleeve corresponding to the plunger component node, the vicinity of the pump body outlet valve seat corresponding to the outlet valve component node, and the vicinity of the pump body cam bearing seat corresponding to the camshaft component node. The predicted state of the high-pressure fuel distribution pump head is obtained based on a physical information neural network model trained with health condition data. Then, based on the predicted state, the multi-source sensor data, and the operating parameters, the physical residuals characterizing the deviation between the multi-source sensor data and the enhanced physical model are calculated by substituting these into the corresponding physical sub-equations of the enhanced physical model. The physical information neural network model is constructed based on the enhanced physical model, which incorporates nonlinear physical characteristics of fuel. The enhanced physical model includes a pressure-compressibility model describing the nonlinear relationship between fuel pressure and compressibility, and a temperature-viscosity model describing the nonlinear relationship between fuel temperature and viscosity. The physical structure of the high-pressure fuel distribution pump head is abstracted as a graph neural network model containing multiple component nodes, including plunger component nodes, plunger sleeve component nodes, delivery valve component nodes, camshaft component nodes, tappet component nodes, and spring component nodes. The physical residual is decomposed into residual components corresponding to different physical sub-equations in the enhanced physical model, and the residual components are updated to the node features of the corresponding component nodes in the graph neural network model based on a preset residual-component association matrix or lookup table; wherein, the flow residual component is updated to at least one of the plunger component node, plunger sleeve component node, and oil outlet valve component node, and the dynamic residual component is updated to at least one of the camshaft component node, tappet component node, and spring component node; The fault diagnosis information of the high-pressure fuel distribution pump head is output through reasoning by the graph neural network model. The fault diagnosis information includes fault location information, which is used to characterize the component node that is the root cause of the fault among the multiple component nodes.

2. The method according to claim 1, characterized in that, After inferring the fault diagnosis information of the high-pressure fuel distribution pump head through the graph neural network model, the method further includes: Graph feature vectors representing the entire graph state are extracted using graph pooling layers; The graph feature vector is input into a preset classifier to obtain the fault type information of the high-pressure fuel distribution pump head; The fault type information includes at least one of plunger wear, delivery valve sticking, and spring fatigue.

3. A high-pressure fuel distribution pump head condition monitoring and fault prediction system based on deep learning, characterized in that, include: A data acquisition module is used to acquire multi-source sensor data of the high-pressure fuel distribution pump head; wherein, the data acquisition module includes multiple sensors respectively arranged at preset physical positions of the high-pressure fuel distribution pump head associated with component nodes in the graph neural network model, the multiple sensors are used to acquire local dynamic sensor data directly related to the state of the component nodes, the preset physical positions include the outer wall of the pump body plunger sleeve corresponding to the plunger component node, the vicinity of the pump body delivery valve seat corresponding to the delivery valve component node, and the vicinity of the pump body cam bearing seat corresponding to the camshaft component node; The physical analysis module is used to obtain the predicted state of the high-pressure fuel distribution pump head based on a physical information neural network model trained with health condition data, and to calculate the physical residual, which characterizes the degree of deviation between the multi-source sensor data and the enhanced physical model, by substituting the predicted state, the multi-source sensor data, and the operating parameters into the physical sub-equations corresponding to the enhanced physical model. The physical information neural network model is constructed based on the enhanced physical model, which includes nonlinear physical characteristics of fuel. The enhanced physical model includes a pressure-compressibility model describing the nonlinear relationship between fuel pressure and compressibility, and a temperature-viscosity model describing the nonlinear relationship between fuel temperature and viscosity. A fault reasoning module is used to run a graph neural network model. This model is abstracted from the physical structure of the high-pressure fuel distribution pump head and includes multiple component nodes, such as plunger assembly nodes, plunger sleeve assembly nodes, delivery valve assembly nodes, camshaft assembly nodes, tappet assembly nodes, and spring assembly nodes. The fault reasoning module is configured to decompose the physical residuals into residual components corresponding to different physical sub-equations in the enhanced physical model, and update the residual components to the node features of the corresponding component nodes in the graph neural network model based on a preset residual-component association matrix or lookup table. Specifically, the flow residual component is updated to at least one of the plunger assembly nodes, plunger sleeve assembly nodes, and delivery valve assembly nodes, and the dynamic residual component is updated to at least one of the camshaft assembly nodes, tappet assembly nodes, and spring assembly nodes. The module performs reasoning through the graph neural network model and outputs fault diagnosis information for the high-pressure fuel distribution pump head. The fault location information is used to characterize the component node among the multiple component nodes that serves as the root cause of the fault.

4. The system according to claim 3, characterized in that, It also includes a diagnostic and output module, which is used to transform complex analysis results into clear and actionable diagnostic information. This is a software application that runs on a server or user terminal. Internally, it first contains a fault classification unit, which receives graph feature vectors from the fault reasoning module and inputs them into a lightweight multilayer perceptron classifier to determine the specific type of fault.

Images