An engine state monitoring method, device and storage medium

Abstract

The application provides an engine state monitoring method and device and a storage medium, in the method, edges are defined through a physical action path, node association without physical constraints is avoided, and physical interpretability of a graph structure is ensured; through dynamic edge weight driven by a working condition, connection strength of the edges changes in real time with a combustion phase, a rotating speed and the like, a dynamic process of physical action strength between measuring points is truly reflected, and a limitation that a traditional static graph structure cannot adapt to transient working conditions is broken. A dynamic graph sequence constructed in this way can accurately depict an evolution law of physical coupling relations inside an engine, causes the graph network model to learn a real physical law instead of data noise in training and reasoning, and significantly improves recognition ability and diagnosis precision of the model on early weak faults and gradual degradation of the engine.

Description

Technical Field

[0001] This invention relates to the field of engine technology, and specifically to an engine condition monitoring method, device, and storage medium. Background Technology

[0002] Currently, engine condition monitoring mainly encompasses physical model-based methods and data-driven methods. Physical model-based methods rely on high-fidelity simulations (such as multibody dynamics or fluid simulations) to construct accurate engine system models, and engine condition monitoring is achieved by comparing the residuals between the model output and the actual monitoring data. Data-driven methods, on the other hand, utilize machine learning algorithms to directly extract fault features from historical monitoring data. In particular, in recent years, graph neural networks (GNNs) have been introduced into this field, constructing a physical information graph by treating sensors as nodes and data correlations as edges, and combining the network structure to extract temporal features to achieve condition monitoring.

[0003] However, model-based technologies heavily rely on accurate system models, which are often costly or difficult to obtain in engineering practice. Physical information maps (PIMs) constructed using data-driven methods are typically static, meaning their structure is fixed and they cannot intuitively reflect how changes in the system's internal state affect the correlations between sensor signals during fault evolution—in other words, they lack the ability to interpret the "process." Furthermore, they lack adaptability to early and progressive faults, resulting in low early fault identification rates. Summary of the Invention

[0004] In view of this, the present invention aims to provide an engine condition monitoring method, device and storage medium to solve the problems existing in traditional related technologies when performing engine condition monitoring based on physical models or data-driven methods.

[0005] A first aspect of the present invention provides an engine condition monitoring method, comprising: Acquire data collected by multiple source sensors and define graph nodes based on the data; Directed edges are defined by direct physical interaction paths between nodes, and these direct physical interaction paths are determined based on the engine's fundamental physical model. Using the engine's basic physical model as a constraint, edge weights are constructed based on a dynamic weight function driven by transient operating conditions; A dynamic graph structure is constructed based on the graph nodes, the directed edges, and the edge weights. A graph network model is then used to process the dynamic graph sequence formed by the dynamic graph structure to monitor the engine status.

[0006] In one possible implementation of the present invention, defining graph nodes based on the data includes: Based on the engine's fundamental physical model, time-varying features with physical significance are extracted from the data; Construct a physical field correlation matrix based on the correlation strength of the physical fields to which different sensors belong; Based on the physical field correlation matrix, the time-varying features are weighted and fused between fields to obtain the collaborative features corresponding to each sensor node. Graph nodes are defined based on the sensor's physical type, location coordinates, and the aforementioned collaborative features.

[0007] In one possible implementation of the present invention, the directed edges defined by the direct physical interaction paths between nodes, wherein the direct physical interaction paths are determined based on the engine's fundamental physical model, include: Based on the engine's basic physical model, determine whether there are direct physical interaction paths between nodes. These direct physical interaction paths include force flow, vibration waves, and heat conduction. When there is a direct physical interaction path between any two nodes, a directed edge is formed between the two nodes.

[0008] In one possible implementation of the present invention, the step of constructing edge weights based on a dynamic weight function driven by transient operating conditions, using the engine's fundamental physical model as a constraint, includes: Using the fundamental physical model of the engine as a constraint, the first relationship between amplitude gain and speed is determined; A second relationship between phase delay and rotational speed is determined based on node spacing and structural wave velocity; By combining the Gaussian window function, a third relationship between the angle modulation function and the crankshaft angle is determined based on the combustion period of the cylinder in the engine. The static coupling coefficient is determined based on the geometric distance between nodes and the structural connectivity. A dynamic weight function is constructed by combining the first, second, and third relationships and the static coupling coefficient, and the dynamic weight function is quantized according to the rotational speed and crankshaft angle in the data to construct the edge weights.

[0009] In one possible implementation of the present invention, the multi-source sensor includes a cylinder pressure sensor, a cylinder head vibration sensor, a main bearing housing vibration sensor, and a body surface vibration sensor; the dynamic graph structure includes a multi-layer dynamic graph structure, which includes a combustion excitation layer, a crankshaft force transmission layer, and a structure propagation layer. The combustion excitation layer includes cylinder pressure sensor nodes and cylinder head vibration sensor nodes, the crankshaft force transmission layer includes cylinder pressure sensor nodes and main bearing housing vibration sensor nodes, and the structure propagation layer includes cylinder head vibration sensor nodes, main bearing housing vibration sensor nodes, and body surface vibration sensor nodes. Corresponding sensor nodes between layers are interconnected.

[0010] In one possible implementation of the present invention, the fundamental physical model of the engine is constructed in the following manner: The kinematic equations of the engine are constructed based on its basic geometric dimensions. Based on the kinematic equations, the force analysis of the engine is carried out, and the dynamic equations of the engine are constructed. Based on the kinematic and dynamic equations, and combined with the parameter values ​​determined by the engine model, a basic physical model of the engine is constructed.

[0011] In one possible implementation of the present invention, the graph network model includes multiple spatial-temporal convolutional layers, attention pooling layers, and classification output layers; The process of using a graph network model to process the dynamic graph sequence composed of the dynamic graph structure for engine status monitoring includes: Multi-layer spatial-temporal convolutional layers are used to convolve the dynamic graph sequence composed of the dynamic graph structure to obtain the spatiotemporal fusion features of all nodes at each time step. Attention pooling layers are used to aggregate spatiotemporal fusion features to obtain global graph features; A classification output layer is used to classify the global graph features to obtain the fault probabilities of different fault categories; Based on the edge weights of directed edges and the attention weights of different nodes determined in the attention pooling layer, combined with the fault probability of the fault category, fault node localization and fault path identification are performed.

[0012] In one possible implementation of the present invention, the method further includes: A graph network model is used to process the dynamic graph sequence during normal engine operation to obtain normal global graph features; Determine the engine's health baseline based on normal global graph features; The engine is given a graded early warning system based on the global graph features of the engine at the current moment and the deviation from the health benchmark.

[0013] A second aspect of the present invention also provides an engine condition monitoring device, the device comprising: The node definition module is used to acquire data collected by multiple source sensors and define graph nodes based on the data; The edge definition module is used to define directed edges based on the direct physical interaction path between nodes, wherein the direct physical interaction path is determined according to the engine's basic physical model. The edge weight construction module is used to construct edge weights based on a dynamic weight function driven by transient operating conditions, with the engine's basic physical model as a constraint. The graph construction and status monitoring module is used to construct a dynamic graph structure based on the graph nodes, the directed edges, and the edge weights, and to process the dynamic graph sequence formed by the dynamic graph structure using a graph network model to monitor the engine status.

[0014] A third aspect of the present invention provides an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to cause the at least one processor to perform an engine condition monitoring method as described in the first aspect and any possible implementation thereof.

[0015] A fourth aspect of the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement an engine condition monitoring method as described in the first aspect and any possible implementation thereof.

[0016] A fifth aspect of the present invention provides a computer program product comprising: a computer program that, when executed by a processor, implements an engine condition monitoring method as described in the first aspect and any possible implementation thereof.

[0017] This invention provides an engine condition monitoring method, device, and storage medium. The method defines edges through physical action paths, avoiding node associations without physical constraints and ensuring the physical interpretability of the graph structure. By using dynamic edge weights driven by operating conditions, the connection strength of edges changes in real time with combustion phase, engine speed, and other operating conditions, realistically reflecting the dynamic process of physical interaction strength between measuring points. This overcomes the limitation of traditional static graph structures in adapting to transient operating conditions. The resulting dynamic graph sequence can accurately depict the evolution of internal physical coupling relationships within the engine, allowing the graph network model to learn real physical laws rather than data noise during training and inference. This significantly improves the model's ability to identify and diagnose early, minor engine faults and progressive degradation. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the specific embodiments or related technologies of the present invention, the drawings used in the description of the specific embodiments or related technologies 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.

[0019] Figure 1 Flowchart of the engine condition monitoring method provided in the embodiments of the present invention Figure 1 .

[0020] Figure 2 This is a schematic diagram illustrating the connection relationship between nodes and edges based on physical effects, provided for an embodiment of the present invention.

[0021] Figure 3 Flowchart of the engine condition monitoring method provided in the embodiments of the present invention Figure 2 .

[0022] Figure 4 This is a schematic diagram of the engine condition monitoring device provided in an embodiment of the present invention.

[0023] Figure 5 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present invention. 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] When using a data-driven approach for engine condition monitoring, the physical information maps (PIMs) employed are all static. However, condition monitoring based on static PIMs suffers from the following problems: Static PIM structures, built upon fixed physical priors, cannot accurately reflect the dynamic changes in the propagation paths and coupling relationships of fault vibration signals under transient operating conditions such as different engine speeds and loads, leading to decreased diagnostic accuracy under complex real-world conditions. Furthermore, constructing an effective PIM relies heavily on accurate and complete modeling of internal physical interactions within the system (such as interference between sensors), which is often costly or difficult to obtain in engineering practice. Moreover, while PIMs incorporate physical knowledge, their fixed structure fails to intuitively reflect how changes in the system's internal state affect the correlations between sensor signals during fault evolution, lacking the ability to interpret the "process." Additionally, the fixed PIM structure struggles to model the characteristics of engine performance degradation over its lifespan, exhibiting weak adaptability to condition degradation and thus limiting the identification of early and progressive faults.

[0026] Based on this, in order to address the problems of static graph structures being unable to adapt to the dynamic changes in the propagation path and coupling relationship of fault vibration signals under transient operating conditions such as engine speed and load, leading to a decrease in diagnostic accuracy under complex operating conditions; the strong dependence on high-fidelity physical models or complete prior knowledge when constructing "physical information graphs," which are often costly or difficult to obtain in engineering practice; the inability of fixed graph structures to intuitively reflect the impact of changes in the internal state of the system on the correlation of sensor signals during fault evolution, lacking the ability to interpret the "process"; and the difficulty of fixed graph structures in modeling the characteristics of engine performance degradation over lifespan, and the inability to modulate graph connection strength through continuous variables (such as health status), thus limiting the ability to identify early and progressive faults, this invention proposes the following inventive concept: First, data collected from multiple sensors is acquired, and graph nodes are defined based on this data. Then, directed edges are defined using the direct physical interaction paths between nodes, determined according to the engine's fundamental physical model. Using the engine's fundamental physical model as a constraint, edge weights are constructed based on a dynamic weight function driven by transient operating conditions. A dynamic graph structure is constructed based on the graph nodes, directed edges, and dynamic edge weights. A graph network model is then used to process the dynamic graph sequence to monitor engine status. Thus, by constructing a dynamic graph, the problems existing in related technologies using fixed graph structures are solved.

[0027] Exemplary methods Figure 1 This is a flowchart illustrating the engine condition monitoring method provided in an embodiment of the present invention. The executing entity in this embodiment can be a monitoring system. For example... Figure 1 As shown, the method includes: S101: Acquire data collected by multiple source sensors and define graph nodes based on the data.

[0028] Specifically, the multi-source sensors in this embodiment include a crankshaft position sensor for providing crankshaft angle (°CA) and rotational speed, a cylinder pressure sensor for acquiring cylinder pressure, a vibration sensor for acquiring cylinder head vibration, main bearing housing vibration and engine block surface vibration, a camshaft position sensor for acquiring camshaft angle phase signal, a common rail pressure sensor for acquiring common rail pressure, and an exhaust temperature sensor for acquiring exhaust manifold temperature.

[0029] The presence or absence of a cylinder pressure sensor in the engine can be determined based on the specific circumstances. If no cylinder pressure sensor is installed, the cylinder pressure can be inverted using the common rail pressure and injection signal. Additionally, the camshaft position sensor's camshaft angle phase signal can be used for phase verification, such as providing an intake / exhaust valve timing reference, and determining phase deviation by combining it with the set phase. Since constant rail pressure is a prerequisite for the operation of an electronic fuel injection system, the pressure signal collected by the common rail pressure sensor can serve as the core closed-loop signal of the fuel system.

[0030] Of the aforementioned multi-source sensors, the crankshaft position sensor, camshaft position sensor, common rail pressure sensor, and exhaust temperature sensor are all built-in sensors inherent to the engine's electronic control system and require no additional modification. The cylinder head vibration sensor, engine block surface vibration sensor, and main bearing housing vibration sensor need to be installed according to the engine's structural layout, such as using "non-intrusive" sensors for the cylinder head and engine block surfaces. Furthermore, the cylinder pressure sensor, cylinder head vibration sensor, and main bearing housing vibration sensor need to be configured according to the number of cylinders and bearings. For example, a certain engine model with 6 cylinders and 7 bearings requires 6 cylinder pressure sensors, 6 cylinder head vibration sensors, and 7 main bearing housing vibration sensors. In this embodiment, the distribution of the multi-source sensors is shown in Table 1 below: Table 1 Distribution of Multi-Source Sensors By deploying the aforementioned multi-source sensors within the engine, the extracted state data can cover the entire chain from "combustion excitation to mechanical transmission to structural response to thermal state," providing a raw data foundation for physical path modeling and solving the problems of "incomplete signal dimensions and lack of physical meaning" in traditional methods. Simultaneously, during engine operation, the engine speed changes under different operating conditions, and the rate of change of the crankshaft angle changes accordingly. The multi-source sensors synchronously collect data corresponding to the operating conditions, following the crankshaft angle sequence. For example, the crankshaft angle can be used as a unified sampling benchmark, and state data can be collected in segments at fixed crankshaft angle intervals (e.g., 10°CA).

[0031] Specifically, in this embodiment, when constructing the graph structure, the graph nodes are defined based on the data collected from multiple source sensors. For example, according to Table 1, 24 sensors are set at 24 locations on the engine, so the graph structure includes 24 graph nodes. Then, the node attributes are determined based on the data collected by each sensor to realize the node definition.

[0032] S102: Define directed edges by direct physical interaction paths between nodes, wherein the direct physical interaction paths are determined based on the engine's basic physical model.

[0033] After defining each node in the graph structure, instead of connecting all the nodes to form edges in the graph structure, we determine whether there is a physical interaction path between any two nodes. If there is a physical interaction path, we connect the two nodes to form an edge between the nodes.

[0034] In this embodiment, the existence of physical interaction paths between nodes is determined based on the engine's fundamental physical model. This fundamental physical model is constructed through basic kinematic and dynamic analysis of the engine, while multi-source sensors are distributed at key measurement points of the engine to collect corresponding operational status data. Based on the fundamental physical model, the transmission links of energy, vibration, and force within the engine can be derived, thereby determining whether connection edges are configured between sensor nodes corresponding to different measurement points. Since the physical excitation and energy transfer within the engine have unidirectional conduction characteristics, the connections constructed between each node are all directed edges.

[0035] S103: Using the engine's basic physical model as a constraint, construct edge weights based on a dynamic weight function driven by transient operating conditions.

[0036] Specifically, during the actual operation of the engine, the data collected by various sensors continuously changes with the crankshaft angle position, real-time speed and load conditions. After the weight quantization of each directed edge is completed by the dynamic weight function driven by transient operating conditions, the edge weight value will be dynamically adjusted in sync with the operating conditions and crankshaft phase.

[0037] Meanwhile, in the entire process of constructing edge weights using a dynamic weight function driven by transient operating conditions, the fundamental physical model of the engine, based on kinematics and dynamics, is used as the benchmark. The construction of the dynamic weight function follows the objective mechanical laws inherent in the model, such as mechanical transmission, energy transfer, and vibration attenuation, ensuring that the trend of weight changes does not deviate from the physical mechanism. For example, when the cylinder is in the combustion power range, the combustion excitation energy increases significantly, and the weight of the corresponding combustion excitation edge increases sharply. When the engine is in a high-speed operating condition, due to the influence of oil film damping and structural vibration loss, the loss of vibration transmitted over long distances along the engine body increases, and the weight of the corresponding structural attenuation-type connection edge decreases accordingly, making the weight change law conform to the actual physical operating characteristics of the engine.

[0038] S104: Construct a dynamic graph structure based on the graph nodes, the directed edges, and the edge weights, and use a graph network model to process the dynamic graph sequence formed by the dynamic graph structure to perform engine status monitoring.

[0039] In this embodiment, the dynamic graph structure is composed of nodes and directed edges with attributes (edge ​​weights). The edge weights change in real time, which drives the correlation strength of the overall graph topology to change synchronously. Unlike traditional static graphs with fixed weights, this ultimately forms a dynamic graph sequence that can adapt to the engine's operating status in real time.

[0040] Specifically, after constructing and summarizing the time-series dynamic graphs into a sequence, the sequence is fed into a pre-built graph network model. The model processes each frame of the dynamic graph corresponding to each crankshaft rotation angle moment in the sequence step-by-step, exploring the evolution of the state between preceding and subsequent moments, and outputting the engine state characterization results time-by-time. Compared to the static graph modeling approach, the dynamic graph structure fully preserves the real-time changes in component coupling relationships under different operating conditions and combustion phases, allowing the network to fully capture subtle data fluctuations during equipment operation. This significantly improves the model's ability to identify early latent faults and progressive degradation faults, and allows for the early detection of potential slow degradation of components.

[0041] In one embodiment of the present invention, the basic physical model of the engine is constructed using the following steps: S1001: Construct the engine's kinematic equations based on the engine's basic geometric dimensions.

[0042] In this embodiment, the kinematic equations of the engine are constructed based on the classic crank-slider structure model. The required basic geometric dimensions of the engine include the crank radius R and the connecting rod length. l The ratio λ of crank radius to connecting rod length λ = R / l The cylinder bore (B) and piston stroke (S=2R) are given. Based on these basic geometric dimensions, kinematic equations including piston displacement, piston velocity, piston acceleration, and connecting rod angle can be constructed (all equations are expressed in terms of crankshaft rotation angle). (as independent variable) The piston displacement is expressed as: Piston speed is expressed as: Piston acceleration is expressed as: The linkage angle is expressed as: S1002: Based on the kinematic equations, the force analysis of the engine is carried out to construct the dynamic equations of the engine.

[0043] Specifically, when constructing the dynamic equations, the engine is subjected to force analysis based on kinematics to obtain the engine's dynamic equations. These dynamic equations specifically include gas force, reciprocating inertial force, piston lateral force, connecting rod force, tangential force, single-bar output torque, total output torque, and total overturning moment.

[0044] Wherein, gas force is expressed as: In the formula, A curve showing the change in cylinder gas pressure with crankshaft angle (Pa) (to be obtained from experiments or simulations); This represents the piston top area. ( ).

[0045] The reciprocating inertial force is expressed as: In the formula, It indicates the mass of the reciprocating motion (kg) (including the mass of the piston assembly and the small end of the connecting rod).

[0046] The piston lateral force (connecting rod thrust) is expressed as: The link force is expressed as: Tangential force (driving torque) is expressed as: The output torque of a single cylinder is expressed as: In an engine, the number of cylinders, firing order, and phase difference determine how the forces and torques generated by each cylinder are superimposed over time (crankshaft angle), directly affecting the engine's smoothness (vibration) and the uniformity of torque output.

[0047] For an inline multi-cylinder engine, the crankshaft pins of each cylinder are on the same crankshaft, but their angular positions are different. For an n-cylinder four-stroke engine, the ignition interval angle is... 。 This means that when cylinder 1 is at a certain angle... At that time, the effective crankshaft rotation angle of the kth cylinder is Add a fixed phase offset .

[0048] The firing order determines which cylinder fires at which phase, and must be matched with the physical layout of the crankshaft (crankshaft pin angle) to ensure uniform firing intervals and engine balance.

[0049] The total output torque is the algebraic sum of the driving torques generated by each cylinder in the direction of crankshaft rotation.

[0050] Total output torque is expressed as follows: The more cylinders n, the better. The more times the torque is superimposed, the smaller the amplitude of torque fluctuation (pulsation) is usually, and the smoother the output is.

[0051] A uniform firing interval (ensuring proper cylinder count and firing order) is essential for... Minimal fluctuations are key. Inconsistencies can lead to severe vibrations and unstable idling.

[0052] Overturning moment is the lateral force of each cylinder. The torque that attempts to "roll" the engine crankshaft axis. The lever arm, i.e., the distance from the cylinder centerline to the engine reference fulcrum, needs to be considered in the calculation.

[0053] Let the horizontal distance from the centerline of cylinder k to the overall engine center or a reference axis (such as the vertical projection of the crankshaft centerline) be . In an inline engine, the cylinders are typically arranged in a straight line. It can be simplified to the cumulative distance starting from the first cylinder.

[0054] The total overturning moment is expressed as: in For each cylinder at the current turning angle Lateral force. The direction will vary with the crankshaft angle. Positive and negative changes, therefore It is also a fluctuating quantity.

[0055] S1003: Based on kinematic and dynamic equations, and combined with parameter values ​​determined by the engine model, a basic physical model of the engine is constructed.

[0056] Specifically, the aforementioned kinematic and dynamic equations are constructed based on the analysis of a general crank-slider mechanism model and can serve as universal equations. Therefore, for different engine models, their specific parameters can be substituted into the above equations to obtain the corresponding engine's basic physical model.

[0057] As a specific implementation method, a parameter configuration file (such as JSON or YAML) can be created for each engine model, depending on actual needs, such as Isuzu 6HK1XDHAG-01-C3, Scania DC16 076A, Cummins QSL9 (Stage V), etc. This data originates from engine technical manuals, CAD drawings, 3D model measurements, or supplier information. The process of automatic equation generation and output is automated through a model compilation script (such as Python / MATLAB). The specific application process is as follows: 1. Select template and model: The user enters or selects the target engine model (e.g., Scania_DC16_076A).

[0058] 2. Load parameters: The script automatically reads the JSON parameter file corresponding to the model.

[0059] 3. Symbol substitution: The script substitutes specific parameter values ​​( Substitute (etc.) into the above general equation.

[0060] 4. Generate specific equations: that is, generate all kinematic and dynamic equations to calculate the theoretical vibration and pressure characteristics for different engine models.

[0061] By establishing the aforementioned parameter database, high-precision basic physical models can be generated quickly and consistently for any engine model, greatly improving modeling efficiency and ensuring model traceability.

[0062] As an embodiment of the present invention, step 101 above, which defines graph nodes based on the data, may specifically include the following steps: S1011: Extract physically meaningful time-varying features from the data based on the engine's basic physical model.

[0063] Specifically, the data acquired in this embodiment comes from various multi-source sensors arranged in the engine. Each acquired signal changes periodically with the crankshaft angle. When constructing the graph structure, the original acquired signals of each sensor can be directly set as the basic attributes of the corresponding graph nodes. However, the original sensor signals are mixed with irrelevant components such as operating noise and environmental interference. Relying solely on the original values ​​cannot correspond to the actual mechanics and combustion laws of engine components. Simply using the original signals as node attributes makes it difficult to determine the underlying physical characteristics of the equipment, which is not conducive to subsequent feature analysis and fault tracing based on fault mechanisms.

[0064] Therefore, physical feature extraction based on first principles is performed on the state signals collected by various sensors. The original signals are deconstructed based on the fundamental physical laws of engine kinematics and dynamics, and irrelevant interference components are removed. Effective features are selected from physical dimensions such as vibration, pressure, heat transfer, and mechanical transmission, ensuring that the extracted nodal physical features correspond to objective physical changes such as changes in component stress and combustion state. In other words, the features extracted from the data are physically meaningful, rather than meaningless numerical values ​​obtained through statistical or mathematical calculations. Specifically, the physically meaningful features extracted in this embodiment directly correspond to the actual mechanical operating laws of the engine. Each feature matches objective physical processes such as in-cylinder gas force, piston inertial force, and structural vibration transmission, and their numerical changes correspond one-to-one with the actual mechanical processes occurring during engine operation. This achieves a direct binding between the extracted features and the actual mechanical operating mechanism of the engine. For example, the pressure rise rate feature extracted from the cylinder pressure sensor data physically represents the combustion rate; a higher pressure rise rate indicates a faster combustion rate.

[0065] Specifically, for cylinder pressure sensors, cylinder head vibration sensors, main bearing housing vibration sensors, engine block surface vibration sensors, crankshaft position sensors, camshaft position sensors, common rail pressure sensors, and exhaust temperature sensors, time-varying features with clear physical meanings are extracted based on the physical laws in the engine's basic physical model, thereby achieving a direct correlation with the fault mechanism.

[0066] (1) Extract the cylinder pressure sensor ( )feature.

[0067] Cylinder pressure signal It directly reflects the combustion process (pressure rise rate, combustion phase, heat release rate) and is the core basis for diagnosing fuel injection, combustion, and valve malfunctions.

[0068] 1.1, Extracting the pressure rise rate (reflecting the combustion rate) as the first characteristic of the cylinder pressure sensor. .

[0069] Extraction method: Cylinder pressure signal To calculate the derivative of the crankshaft rotation angle, the five-point central difference method is used to improve accuracy. in, Sampling interval for crankshaft angle (e.g.) Take the combustion period (after top dead center). The maximum slope is used as a feature. .

[0070] Physical meaning: The rate of pressure rise is positively correlated with the combustion rate. If the rate is too fast, it is easy to cause knocking, while if it is too slow, afterburning will increase.

[0071] Corresponding faults: Increased when fuel injection is advanced, decreased when combustion atomization is poor, decreased when valve leakage.

[0072] 1.2, Extracting the center of gravity of combustion ( (Reflecting the combustion phase) is the second characteristic of the cylinder pressure sensor. .

[0073] Extraction method: Calculate the crankshaft angle when the accumulated heat release reaches 50%. The specific extraction formula is as follows: in, This is the ignition start point (the point where the pressure rise rate suddenly increases). The exhaust point is at top dead center, and the solution is obtained through numerical integration. As a feature .

[0074] Physical meaning: The normal range is after the top dead center. A backward movement indicates delayed combustion, while an forward movement indicates advanced combustion.

[0075] Corresponding faults: incorrect fuel injection timing (advanced / delayed), abnormal fuel cetane number (delayed).

[0076] 1.3 Extracting pressure fluctuation entropy (reflecting combustion stability) as the third feature of the cylinder pressure sensor. .

[0077] Extraction method: for combustion period ( )of Perform wavelet packet decomposition to extract the energy proportion of the 3-5kHz frequency band. Calculate the entropy value: in, The energy of the kth sub-band is given. The larger the entropy value, the more disordered the pressure fluctuation (and the less stable the combustion).

[0078] Physical significance: Pressure fluctuation entropy increases significantly during detonation and fire.

[0079] Corresponding faults: Increased when knocking occurs, increased when misfires in a single cylinder occurs, increased when fuel quality is poor.

[0080] (2) Extract the cylinder head vibration sensor ( )feature.

[0081] Cylinder head vibration signals include combustion impact (high frequency), inertial force (medium frequency), and structural resonance (low frequency), and features need to be extracted by separating them according to the excitation source frequency band.

[0082] 2.1 Extracting combustion impact frequency band energy ( (500~2000Hz) is the first characteristic of the cylinder head vibration sensor.

[0083] Extraction method: The vibration signal was decomposed into k modal components using Variational Mode Decomposition (VMD), and modes with center frequencies between 500 and 2000 Hz were selected. Energy percentage was calculated. in, For VMD modal components, For target frequency band mode index.

[0084] Physical meaning: The energy of combustion impact (such as detonation, pre-ignition) is concentrated in this frequency band.

[0085] Corresponding faults: Increased when knocking occurs, increased when fuel injection is advanced, increased when piston rings leak air.

[0086] 2.2 Extracting inertial force frequency band energy ( (100~500Hz) is the second characteristic of the cylinder head vibration sensor.

[0087] Extraction method: Similarly, VMD modes with center frequencies of 100~500Hz are selected, and their energy percentages are calculated. .

[0088] Physical meaning: It reflects the transmission of piston-connecting rod inertial force (first / second order), and is related to component wear (increased clearance).

[0089] Corresponding faults: Increased when piston pin is loose, increased when connecting rod bearing is worn, increased when crankshaft is unbalanced.

[0090] 2.3 Extract the vibration signal kurtosis (reflecting transient impact) as the third feature of the cylinder head vibration sensor.

[0091] Extraction method: Calculate vibration signal kurtosis value: Physical meaning: Kurtosis is sensitive to transient shocks (such as cylinder scoring, valve seat failure), <3 under normal operating conditions, and >10 during shock.

[0092] Corresponding faults: Increased during the initial stage of cylinder scoring, increased when valve springs break, and increased when foreign objects enter the combustion chamber.

[0093] (3) Extract the vibration sensor from the main bearing housing ( (Measurement of crankshaft-bearing force transmission characteristics).

[0094] The vibration of the main bearing reflects the crankshaft torsional / bending vibration and bearing wear (changes in oil film clearance). The focus is on extracting frequency harmonics and high-frequency noise.

[0095] 3.1, Extract 1 / 2 times the frequency conversion amplitude ( The first characteristic of the main bearing housing vibration sensor is that it reflects the second-order inertial force.

[0096] Extraction method: Perform FFT on the vibration signal to extract the rotational speed frequency. (n is the rotational speed in rpm) 1 / 2 harmonic amplitude .

[0097] Physical meaning: The second-order inertial force theory of the inline six-cylinder engine is in equilibrium. An increase in the amplitude indicates that the balance weight has failed or the connecting rod has deformed.

[0098] Corresponding faults: Increased when the connecting rod is bent, increased when the balance shaft is faulty.

[0099] 3.2 Extracting high-frequency noise energy ( The second characteristic of the main bearing housing vibration sensor is that the Hz frequency (>2000Hz, reflecting bearing wear) is 2000Hz.

[0100] Extraction method: Wavelet thresholding is used to separate high-frequency components >2000Hz, and the energy percentage is calculated. .

[0101] Physical meaning: After the bearing wears out, the collision of balls / raceways generates broadband noise, and the energy is positively correlated with the degree of wear.

[0102] Corresponding faults: Increased when the main bearing is burned out, increased when the bearing clearance is too large.

[0103] (4) Extract the vibration sensor from the surface of the machine body ( (Measurement of structural coupled vibration characteristics).

[0104] Vibration signals on the machine body surface reflect the resonance of the entire machine structure (such as cylinder bending mode), and the energy shift of the resonance frequency band is extracted. (Reflecting changes in structural stiffness) is a characteristic of vibration sensors on the body surface.

[0105] Extraction method: Under healthy conditions, calibrate the body's resonant frequency band (e.g., 800~1200Hz) and calculate the energy of the current signal in that frequency band. Compared with the reference energy The deviation (i.e., the average energy within this frequency band): Physical significance: Cracks in the body and loose bolts will reduce the structural stiffness, leading to a shift in the resonant frequency band energy.

[0106] Corresponding faults: Increased when there are rigid body cracks, increased when anchor bolts are loose.

[0107] (5) Extract the crankshaft position sensor (CKP, providing...) and )feature.

[0108] Crankshaft angle It is the "time reference" for the engine cycle, the engine speed. These are the core parameters of the operating conditions, extracting instantaneous fluctuations and torsional harmonics.

[0109] 5.1 Extracting instantaneous speed fluctuations ( The first characteristic of the crankshaft position sensor is that it reflects the imbalance in the work done by each cylinder.

[0110] Extraction method: for speed signal Perform a low-pass filter (cutoff frequency 10Hz) and calculate the average rotational speed. Deviation: Physical meaning: Misfire in a single cylinder or uneven fuel injection can lead to increased speed fluctuations.

[0111] Corresponding faults: Increased when a single-cylinder injector is clogged, increased when valve timing is incorrect.

[0112] 5.2 Extracting the amplitude of torsional harmonics ( The second characteristic of the crankshaft position sensor is that it reflects the stiffness of the crankshaft system.

[0113] Extraction method: Perform FFT on the speed fluctuation signal to extract 2 to 6 times the rotational frequency. The harmonic amplitude.

[0114] Physical meaning: When crankshaft torsional vibration modes (such as the second and fourth order) are excited, the amplitude increases sharply, which is related to the flywheel inertia and coupling stiffness.

[0115] Corresponding faults: loose flywheel bolts (increased by level 2), crankshaft cracks (increased by level 4).

[0116] (6) Extract the features of the camshaft position sensor (CMP, verification phase).

[0117] Provides intake / exhaust valve timing references and phase deviations. Specifically, it extracts valve timing deviations (…). The first characteristic of the camshaft position sensor is that it reflects a fault in the valve train mechanism.

[0118] Extraction method: Compare the phase difference between the CMP signal and the CKP signal with the calibration value (e.g., intake valve advance). )Compare: Physical meaning: Timing belt / chain skipping teeth and cam wear can cause phase deviation.

[0119] Corresponding faults: Increased when timing is misaligned, increased when cam assembly is worn.

[0120] (7) Extract the features of the common rail pressure sensor (CRP, on the common rail).

[0121] 7.1 Extracting rail pressure control residuals ( The first characteristic of a common rail pressure sensor is that it reflects the quality of ECU closed-loop control.

[0122] Extraction method: Through signal analysis, the dynamic error sequence between the target rail pressure set by the ECU and the actual rail pressure acquired by CRP is extracted. Physical significance: Constant rail pressure is a prerequisite for the operation of electronic fuel injection systems, reflecting the dynamic balance between fuel supply and consumption.

[0123] Corresponding faults: Decrease when the low-pressure oil pump supply is insufficient, decrease when the pressure relief valve is stuck and normally open, decrease when the injector is worn and leaking.

[0124] 7.2 Extracting the rail pressure pulsation amplitude ( The second characteristic of a common rail pressure sensor is that it reflects the resonance state of the liquid column in the common rail tube.

[0125] Extraction method: Bandpass filtering is applied to the original high-frequency CRP signal to remove the low-frequency trend term, and the peak-to-peak value of the high-frequency pressure is extracted. Physical significance: It reflects the resonance intensity of the common rail fluid column under the action of injection pressure wave, and is related to the pipeline volume, fuel compressibility, and damping element effectiveness.

[0126] Corresponding faults: common rail carbon buildup / blockage, high-pressure fuel line joint loose / damaged, excessive air mixed in with fuel.

[0127] 7.3 Extracting the overshoot rate of the step response ( The third characteristic of a common rail pressure sensor is that it reflects the hysteresis of the system's thermodynamic and mechanical response.

[0128] Extraction method: Monitor the rail pressure under rapid acceleration or heavy load conditions, from steady state... The target value set by the instruction The transient pressure surge process occurs during which the actual rail pressure reaches its highest peak due to the dynamic hysteresis of the system. ,based on and Calculate the overshoot rate of the step response, that is, extract the percentage of overshoot and rise time of the step response: Physical significance: It reflects the dynamic response characteristics of the high-pressure oil pump regulating valve and the viscosity-temperature characteristics of the fuel itself, and embodies the matching degree between the control algorithm and the actuator.

[0129] Corresponding faults: fuel metering unit stuck / slow response, abnormal fuel viscosity (such as poor quality fuel), excessively high fuel temperature causing a decrease in density.

[0130] 7.4, Extracting sensor signal drift rate / noise entropy ( The fourth characteristic of a common rail pressure sensor is that it reflects the aging and electrical health of the sensor itself.

[0131] Extraction method: During the pressure holding phase after the engine is shut down and the high-pressure pump stops working, extract the rate of change (drift rate) of the CRP output DC component over time. The normal drift rate is <0.5 bar / min; or, under steady-state conditions, calculate the approximate entropy / sample entropy of the signal (noise complexity), and the healthy-state noise entropy. 0.8bit.

[0132] Physical significance: Characterizes the aging of the CRP sensor itself (bridge creep) and the health of the electrical system (wiring harness interference).

[0133] Corresponding faults: aging of the internal bridge of the sensor / diaphragm fatigue, water ingress into the connector / damage to the wire harness shielding layer, and loose grounding of the sensor.

[0134] (8) Extract the exhaust temperature sensor ( (Features on the exhaust manifold).

[0135] Exhaust temperature signal Located upstream of the turbocharger, it directly reflects the residual heat energy and air-fuel ratio after combustion in the cylinder, and is a key thermodynamic parameter for diagnosing combustion efficiency, aftertreatment system and turbine operating condition.

[0136] 8.1, Extract the average exhaust temperature (reflecting the overall heat load and air-fuel ratio) as the first feature of the exhaust temperature sensor. .

[0137] Extraction method: within a complete engine duty cycle or a specified time window Within this process, an arithmetic average is performed on the exhaust temperature signal. Considering the influence of exhaust pulses, a moving average filter is used to eliminate high-frequency noise. in This represents the number of sampling points within the window. The mean value under steady-state conditions is taken as... .

[0138] Physical meaning: Average exhaust temperature is a direct reflection of the air-fuel ratio. An excessively lean mixture... This leads to an increase in combustion temperature, which in turn causes a rise in exhaust temperature; the air-fuel mixture is too rich. Incomplete combustion leads to a decrease in exhaust temperature. This value also directly reflects the turbine's thermal load.

[0139] Corresponding faults: Decrease when incomplete combustion is caused by poor fuel injector physicochemical properties, or decrease when abnormal exhaust temperature is caused by excessively high EGR (exhaust gas recirculation) rate.

[0140] 8.2 Extracting the exhaust temperature gradient (reflecting combustion stability and catalytic converter ignition) as the second feature of the exhaust temperature sensor. .

[0141] Extraction method: The rate of change of exhaust temperature over time was calculated, and the transient changes were captured using the first-order backward difference method. Take the maximum rate of change during the cold start phase or the rapid acceleration phase as the characteristic. Alternatively, take the standard deviation of the fluctuation under steady state. .

[0142] Physical significance: An excessively large temperature gradient can lead to thermal fatigue in the exhaust manifold and turbocharger; a gradient that is too small or negative may indicate combustion interruption (misfire) or the catalytic converter not reaching its ignition temperature. This feature can effectively capture the transient stability of combustion.

[0143] Corresponding faults: Decrease when spark plug gap is too large / erosion causes insufficient ignition energy; decrease when high-voltage coil failure causes intermittent misfire; increase or fluctuate drastically when catalyst carrier rupture causes abnormal heat transfer.

[0144] In summary, the physically meaningful time-varying features extracted from different sensors provide high-value input for subsequent physical path map construction and dynamic weight calculation, which is the foundation for achieving interpretable monitoring.

[0145] S1012: Construct a physical field correlation matrix based on the correlation strength of the physical fields to which different sensors belong.

[0146] Specifically, step a1 above extracts one or more features from different sensors. In this embodiment, a physical field correlation matrix is ​​used to achieve a collaborative representation of the features from multiple sensors, thereby quantifying the coupling relationship between different physical fields such as pressure, vibration, or rotational speed. The physical field correlation matrix represents the correlation strength between different physical fields.

[0147] For example, build 3D physical field correlation matrix C , of which elements This represents the correlation strength, or functional coupling strength, between the physical fields to which sensors i and j belong, and is a value in... The scalar values ​​between them. This correlation strength can be quantified and determined based on historical data of a healthy engine under various steady-state operating conditions through signal processing and statistical analysis, to determine the dynamic dependence between sensor signals (e.g., the correlation strength of the "pressure-vibration" field is set to 0.8, and the correlation strength of the "temperature-speed" field is set to 0.2).

[0148] S1013: Based on the physical field correlation matrix, time-varying features are weighted and fused between fields to obtain the collaborative features corresponding to each sensor node. Specifically, the extracted feature vectors... The inter-field weighted fusion process is specifically represented by the following formula: in, For i, it is a set of sensors that share the same physical field or a strongly correlated field.

[0149] Specifically, a homogeneous physical field represents physical quantities that are exactly the same or highly homogeneous as those measured by the sensors. Variations in this field are primarily driven by differences in the behavior of the same physical process at different spatial locations. A strongly correlated field, on the other hand, represents physical quantities that are measured by the sensors but interact through inherent, strong physical coupling mechanisms within the engine (such as force-thermal-fluid coupling), resulting in their dynamic signals exhibiting a high degree of statistical dependence.

[0150] Through the weighted fusion process of this feature, the physical meaning of combustion-mechanical-thermal multi-physics field features is fused, rather than data dimension stacking, providing collaborative feature input for graph node definition.

[0151] S1014: Graph nodes are defined based on the sensor's physical type, location coordinates, and cooperative features. Specifically, in addition to the cooperative features obtained through weighted fusion, the sensor's physical type and location are also determined when defining graph nodes. That is, physical type, location, and cooperative features constitute node attributes, thereby achieving a one-to-one correspondence between nodes and engine physical entities. Subsequent edge definitions and weight calculations can be directly traced back to specific components, solving the black-box problem of traditional graph neural network nodes having no physical meaning. This lays the foundation for the interpretability of the graph structure.

[0152] In this embodiment, the node attributes of each graph node include at least the following: Node ID: (like This indicates the pressure sensor node for cylinder 1. (Indicates the vibration node of cylinder head 3). Physical types: Combustion (P), Vibration (V), Rotation speed (R), Temperature (T); Location coordinates: (Based on the installation location of the engine CAD model, used to calculate structural distances); Feature vector: (Collaborative features after data fusion).

[0153] In one embodiment of this application, in step S102 above, directed edges are defined based on direct physical interaction paths between nodes. These direct physical interaction paths are determined according to the engine's fundamental physical model and may specifically include: S1021: Determine whether there are direct physical interaction paths between nodes based on the engine's basic physical model. Direct physical interaction paths include force flow, vibration waves, and heat conduction.

[0154] In this embodiment, the existence of an edge connection between any two nodes is determined by the existence of a direct physical interaction path between them. Based on the physical laws contained in the engine's fundamental physical model, the existence of a direct physical interaction path between any two nodes can be determined, such as force flow, vibration waves, and heat conduction. When there is no physical interaction path between two nodes, there is no edge connection between them, thus avoiding redundant connections without physical meaning and achieving isomorphism between the graph structure and the physical system. For example, there is no edge connection between the exhaust temperature sensor and the cylinder 1 vibration sensor.

[0155] It should be noted that the direct physical action path in this embodiment can be understood as a physical transmission link formed by the unidirectional conduction of energy and load between the corresponding installation points of two sensors during engine operation without the need for other measuring points. Specifically, it can be represented by force flow, vibration waves, and heat conduction. Among them, the force flow path represents the direct transmission of mechanical loads such as combustion pressure and inertial force from one end of a component to another; the vibration wave path represents the path of mechanical vibration directly transmitted from one point (sensor installation point) to another along solid structures such as the engine body and bearings; the heat conduction path represents the direct transfer of heat from high-temperature media and heat-generating components to adjacent points. Only when there is the above-mentioned objective transmission relationship between two points without the need for a third measuring point can the two have a direct physical action path, which serves as the basis for determining the construction of directed edges between nodes.

[0156] S1022: When there is a physical interaction path between any two nodes, a directed edge is formed between the two nodes.

[0157] Specifically, the physical action paths of the multi-source sensor provided according to the present invention include the six types listed in Table 2 below: Table 2 Definition of Directed Edges The directed edges formed in this embodiment follow the engine's physical topology, ensuring that the information transmission path in the graph neural network is consistent with the actual force / vibration propagation path, thus achieving "model reasoning = physical path reproduction". The connection relationship between nodes and edges based on the physical action definition is as follows: Figure 2 As shown. Thin solid lines represent the transmission of physical forces and the linkage of mechanical structures (such as the piston driving the connecting rod, and vibration being transmitted from the cylinder head to the main bearing); dashed lines represent electronic synchronization signals and control relationships between sensors (such as crankshaft position synchronizing cylinder pressure signals); thick solid lines represent strongly coupled physical impacts and resonances (such as combustion pressure directly impacting the cylinder head and causing vibration, and the core main bearing causing resonance on the engine block surface).

[0158] In one embodiment of the present invention, step S103 above, using the engine's basic physical model as a constraint, constructs edge weights based on a dynamic weight function driven by transient operating conditions, which may specifically include the following steps: S1031: Determine the first relationship between amplitude gain and speed using the engine's basic physical model as a constraint.

[0159] Specifically, after forming directed edges between nodes, edge weights are calculated as edge attributes to form a graph structure. To overcome the limitations of static edge weights, this embodiment uses physical equation parameterization to dynamically adjust the weight parameters during calculation. That is, under the constraints of the engine's basic physical model, the weights change with the engine's real-time operating conditions (such as speed, crankshaft angle, and load), accurately reflecting the dynamic characteristics of the intensity of physical forces (such as force transmission efficiency and vibration attenuation).

[0160] Among them, the magnitude gain used in calculating the weights It is a key parameter characterizing the variation of energy transmission amplitude of vibration and force along the physical transmission path of the engine with rotational speed. Under the constraints of the engine's basic physical model, it can determine the fundamental laws of the reciprocating inertial force of the engine piston, the in-cylinder combustion gas force, and the torsional dynamics of the crankshaft system. It can be determined through finite element analysis or experimental modal analysis. .For example, The gain varies with The changes reflect variations in oil film stiffness and structural damping. In this embodiment, the relationship between amplitude gain and rotational speed is obtained through experimental fitting and expressed by the following formula (i.e., the first relationship): S1032: Determine the second relationship between phase delay and rotational speed based on node spacing and structural wave velocity.

[0161] Specifically, phase delay Used to characterize the phase lag of pressure waves and vibration waves along the mechanical transmission path from node i to node j, expressed as signal propagation time. ( The node spacing, The structural wave velocity and system phase lag are characterized. In this embodiment, the phase delay can be expressed by the following formula (i.e., the second relationship): S1033: Combining the Gaussian window function, a third relationship between the angle modulation function and the crankshaft angle is determined based on the combustion period of the cylinder in the engine.

[0162] Specifically, the angle modulation function This is a key parameter for achieving adaptive opening and closing of dynamic edge weights based on the cylinder combustion phase. It is determined by the range of top dead center angles of each cylinder and the crankshaft angle during actual combustion. In other words, this angle modulation function is used to quantify the weights of combustion-related edges. Furthermore, according to its definition, this angle modulation function is activated during the combustion period, or in other words, it is non-zero during the combustion period of the source cylinder. Based on this, this embodiment uses a Gaussian window function to determine the angle modulation function, which is expressed by the following formula (i.e., the third relation): in, For the current corner Top dead center of combustion The difference, that is .

[0163] S1034: Determine the static coupling coefficient based on the geometric distance between nodes and structural connectivity; specifically, this static coupling coefficient... Represents the structural connection strength between each pair of sensors, as a unit in... The scalar value between them. In practical applications, the static coupling coefficient can be determined by quantifying the inherent ability of sensor positions to transmit information through mechanical structures based on prior knowledge of geometric distance and structural connectivity.

[0164] Geometric distance represents the physical distance between two sensor measuring points in the engine's physical space. It is a spatial quantitative indicator characterizing the degree of attenuation of vibration, heat, and mechanical force as they are transmitted along the structure. The greater the geometric distance between the measuring points, the greater the loss during physical signal transmission, and the weaker the physical coupling between the two points. Structural connectivity characterizes whether there is a continuous, uninterrupted solid mechanical conduction path between the two sensor mounting points. It reflects whether vibration, load, and heat energy can be directly transmitted through the engine block, bearings, pipelines, and other metal structures. If the two points are separated by gaskets, partitions, etc., the structural connectivity is poor, the transmission of physical effects is hindered, and the degree of coupling is reduced.

[0165] S1035: Construct a dynamic weight function by combining the first relationship, the second relationship, the third relationship and the static coupling coefficient, and quantize the dynamic weight function according to the rotational speed and crankshaft angle in the data to construct the edge weights.

[0166] Specifically, based on the amplitude gain, phase delay, and angle modulation function determined by the first, second, and third relationships mentioned above, and combined with the static coupling coefficient, the dynamic weighting function can be expressed by the following formula: In a graph structure, edges are treated as directed edges. During actual computation, their scalar values ​​can be calculated using a dynamic weight function, and then the edge's direction can be added to obtain its properties. This dynamic weight function can be further expressed by the following formula: The scalar calculated using the dynamic weighting function is used to quantify the instantaneous effect strength of the edge in real time, while the vector with added direction can comprehensively describe the inherent properties and dynamic changes of the edge.

[0167] Below is 6HK1 ( arrive The specific calculation of the edge weight is explained below: The physical path of this directed edge includes: Cylinder 1 combustion pressure → Piston → Connecting rod → Crank pin → First & Second crank arms → Second main bearing → Sensor .

[0168] In its angle modulation function, The burning period is approximately ,therefore The value is relatively large within this range. Simultaneously, harmonic response analysis can be performed using the finite element model of the engine to calculate the frequency response function of the second main bearing housing under combustion force excitation, thus obtaining... and Additionally, since the firing order is 1-5-3-6-2-4, when cylinder 5 fires (cylinder 1 cranks at 240°), it will also excite the second main bearing through the crankshaft, but the path is different, and the weight should be calculated separately.

[0169] After determining the graph nodes, node attributes, directed edges between nodes, and edge attributes in the graph structure, a complete graph structure is formed. For more refined modeling, this graph structure can be further divided into a bipartite graph or a multi-layer graph structure. This embodiment uses a multi-layer graph structure, specifically including a combustion excitation layer, a crankshaft force transmission layer, and a structural propagation layer.

[0170] The combustion excitation layer includes cylinder pressure sensor nodes and cylinder head vibration sensor nodes, i.e., the nodes are... and The edge represents the direct impact of combustion in the cylinder onto the cylinder head. The crankshaft force transmission layer includes cylinder pressure sensor nodes and main bearing housing vibration sensor nodes, i.e., it contains... and Nodes, where edges represent force transmission through the crankshaft system, have highly dynamic weights that change with the load. and Changes. The structural propagation layer includes cylinder head vibration sensor nodes, main bearing housing vibration sensor nodes, and engine block surface vibration sensor nodes; that is, the nodes are all vibration sensors. , , The edges represent vibrations propagating through the body structure, with relatively stable weights, but affected by structural modes (and...). (Related) impact. Corresponding nodes in different layers (e.g. Layers 1 and 3 are interconnected, representing the exchange of energy between subsystems.

[0171] As a specific implementation method, the dynamic graph sequence is constructed using the following process: S1, original signal segmentation and feature extraction.

[0172] Acquiring multi-sensor synchronization timing signals ,in This represents the raw signal (such as vibration acceleration, cylinder pressure) of the i-th sensor (node) at time t. Then, it is calculated according to the crankshaft rotation angle. (or time window) The signal is segmented, with each segment containing N sampling points (corresponding to one work cycle or a fixed duration). Physically meaningful features are extracted from each segment (avoiding black-box features), such as: kurtosis and RMS features in the time domain for vibration signals; 1x / 2x RPM amplitude features in the frequency domain; and VMD modal energy features in the time-frequency domain. Pressure rise rate is extracted from cylinder pressure signals. Maximum burst pressure and the duration of continuous combustion Features such as these are used to extract instantaneous speed fluctuations from the speed signal. Features such as... and record the feature vector of the i-th node at time t. (d is the feature dimension).

[0173] S2, Dynamic Graph Structure Definition.

[0174] For each signal segment, calculate the node features and node set using the method described above. V Directed edge set and dynamic neighboring order matrix .

[0175] Among them, node set ， N represents the total number of sensors (e.g., for a 6-cylinder engine, this includes 6 cylinder pressure sensors + 6 cylinder head vibration sensors + 7 main bearing vibration sensors + 1 engine block surface vibration sensor + 1 crankshaft position sensor + 1 camshaft position sensor + 1 common rail pressure sensor + 1 exhaust temperature sensor = 24 nodes). This indicates the cylinder pressure sensor for cylinder 6. The corresponding node This indicates the cylinder head vibration sensor for cylinder 3. (Corresponding nodes). Edge set Represents the physical interaction path between nodes i and j (e.g.) (For the same-cylinder combustion excitation path), the edge attributes are calculated by the dynamic weight function to obtain the dynamic calculation result of the edge weight (scalar), which is used to quantify the instantaneous effect strength of the edge in real time; while the edge weight can be understood as the complete representation of the edge attributes (vector), which is used to comprehensively describe the inherent attributes and dynamic changes of the edge.

[0176] S3, Dynamic Graph Sequence Generation.

[0177] Arrange the graph sequence in chronological order to generate a dynamic graph sequence. Each diagram Contains a set of nodes V Directed edge set Dynamic neighbor matrix This enables a unified representation of three-dimensional information encompassing space, time, and physics. Consequently, this dynamic graph sequence fully records the real-time changes in the physical action path under varying operating conditions, providing data structure support for capturing the weak and time-varying characteristics of early faults and solving the problem that traditional static graphs cannot adapt to changing operating conditions.

[0178] As a specific embodiment of the present invention, the graph network model includes multiple spatial-temporal convolutional layers, attention pooling layers, and classification output layers; in step S104 above, the graph network model is used to process the dynamic graph sequence composed of the dynamic graph structure to perform engine status monitoring, which may specifically include the following steps: S1041: Multi-layer spatial-temporal convolutional layers are used to convolve the dynamic graph sequence composed of the dynamic graph structure to obtain the spatiotemporal fusion features of all nodes at each time step.

[0179] Specifically, to simultaneously capture spatial correlation (sensor network topology) and temporal evolution (the transmission of fault characteristics with operating conditions), a spatiotemporal graph convolutional network architecture is adopted in the graph network model. This architecture consists of a series of spatial graph convolutions and temporal convolutions, and incorporates the physical characteristics of the signal into the convolution kernel of the temporal convolution. That is, the convolution kernel of the temporal convolution is not fixed, but dynamically changes with the engine operating conditions.

[0180] In spatial convolutional layers, for graph sequences ( Each time step t within a time window is processed independently to achieve physical path aggregation. Specifically, the hidden state of node i at time t is considered. Updated via weighted neighborhood aggregation: In the formula, The set of neighbors of node i (defined by the edge set E, such as...) Neighbors , ). This represents the dynamic edge weights (see the preceding definition, which integrates amplitude, phase, and angle modulation). , Indicates the first l The layer can learn the weight matrix and bias; Indicates the characteristics of the splicing node itself. With aggregated neighbor information; This represents the activation function (such as ReLU).

[0181] It should be noted that the hidden state ,in, The hidden layer dimension, i.e., the length of the node features, determines the feature extraction capacity of the network. The hidden layer refers to the intermediate computational layer in a spatial-temporal convolutional network that performs node state updates layer by layer along the constructed physical topology graph. The output of each layer is called the hidden state. According to the hidden state update formula, the hidden state represents the weight of node i via its incoming edge at time t. The internal representation vector obtained after aggregating the information of adjacent nodes.

[0182] Among them, the aggregation path of nodes follows the physical action chain (e.g. For the direct path of combustion impact, weight (maximum during combustion period); in terms of weight middle, Reflects the change in structural stiffness / damping with rotational speed. Reflects signal propagation delay, Overall, it reflects the "efficiency of force / vibration transmission along the physical path".

[0183] The process of calculating the hidden state of a single node layer by layer, as described above, can also be represented in the following matrix form: in: (Add a self-loop); for The degree matrix; These are learnable weights, but the initial weights are set according to the sensor type (e.g., the initial weight of a vibration sensor is higher than that of a temperature sensor, reflecting the difference in their sensitivity). The feature set of each node i after aggregation in the spatial domain can be represented as: In the temporal convolutional layer, for each node i, its sequence on the temporal axis of the spatial convolution output is... Temporal convolution is performed. Simultaneously, physical properties are incorporated into the convolution kernel of this temporal convolution, thereby achieving one-dimensional convolution with physical constraints. Specifically, the convolution kernel in this temporal convolution... (K is the kernel size) is determined by the following physical variables: In the formula, This represents the average rotational speed (rpm) within the current time window. Indicates the starting crankshaft angle (°CA) of the current time window; An embedding vector representing the sensor type (vibration, pressure, rotational speed, etc.); The physical modulation function, which dynamically adjusts its weights based on the engine's real-time operating conditions, is expressed by the following formula: In the formula, Indicates the engine's rated speed; This indicates the reference crankshaft angle. It is usually set to 0 degrees (top dead center of combustion) or the injection start angle. This represents the time constant, used to control the width of the Gaussian window. In this formula, the first term (Gaussian window, i.e.) Emphasis on reference angle (e.g., top dead center of combustion) The moments with strong correlation are because combustion events occur at specific turning angles. The second term (speed scaling, i.e.) This indicates that the time span of the convolution kernel should vary with the rotation speed. At higher rotation speeds, the duration of physical events is shorter, and the convolution kernel should be more "compact." The third item (sensor type embedding) The expression indicates that vibration signals focus on high-frequency transients, and the convolution kernel emphasizes differential characteristics; pressure signals focus on low-frequency cycles, and the convolution kernel emphasizes integral characteristics. This is a basic weight matrix independent of real-time operating conditions.

[0184] Based on the convolution kernel determined by the above formula, the output of node i at time t in the temporal convolutional layer can be expressed as: in, The feature of node i is fused with spatiotemporal information. The calculation of this formula is equivalent to an adaptive time-varying digital filter, whose filtering characteristics (cutoff frequency, phase delay) are adjusted with the real-time engine speed and crankshaft angle to match the actual physical changes of the signal under different operating conditions.

[0185] Stack L spatial-temporal convolutional layers, with the output of each layer serving as the input to the next layer, and add residual connections to prevent gradient vanishing: in, Includes spatial convolution (generating) ) and temporal convolution (generating This involves two steps. Therefore, the spatial convolution result of each layer is the input to the temporal convolution of the next layer, and the final spatiotemporal fusion feature is the accumulation of multiple layers of spatial-temporal convolution.

[0186] Final output: The spatiotemporal fusion features of all nodes at time t are obtained: S1042: Attention pooling layers are used to aggregate spatiotemporal fusion features to obtain global graph features.

[0187] Specifically, an attention pooling mechanism is used to aggregate the spatiotemporal fusion features of all nodes at time t, and the features of N nodes are combined. Aggregation into global graph features The system automatically learns the importance weights of each node to the system state through the Softmax function. : in, As a learnable attention vector, the contribution weights of each node to the system state are automatically learned (e.g., the vibration node weight of the faulty cylinder). higher).

[0188] S1043: A classification output layer is used to classify the global graph features to obtain the fault probability of different fault categories.

[0189] Specifically, the input to the Softmax layer in the classification output layer is the feature vector representing the current engine state, output from the previous network. Its internal weight parameters calculate the "matching degree" between this feature vector and each fault category, generating k raw scores. A higher score indicates a greater similarity between the current feature and the pattern of that fault category. This process can be represented by the following formula: In the formula, The input feature vector of the Softmax layer at time t (representing the engine state) is used. This is the weight vector corresponding to the kth fault category; This is the bias term for the k-th fault category; Total number of fault categories; Let t be the probability that the fault belongs to the k-th type at time t.

[0190] S1044: Based on the edge weights of the directed edges and the attention weights of different nodes determined in the attention pooling layer, combined with the fault probabilities of the fault categories, fault node localization and fault path identification are performed.

[0191] Specifically, after calculating the failure probabilities of different failure categories through the above steps, the graph neural network model... The fault category with the highest output probability is found using the following formula: in, For the index of predicted fault categories (such as in the formula "") "Highest point indicates a 3-cylinder fault", that is ); This means finding, for the independent variable k, that... The largest k value.

[0192] Based on fault category probabilities, combined with attention weights and edge weights, key node localization and associated path identification can be performed. Specifically, by analyzing attention weights... This can identify the sensor features that contribute most to the model's current state judgment. For example, if The highest reading typically indicates that the cylinder pressure sensor for cylinder 3 provides crucial evidence in the current fault diagnosis, but this needs to be considered in conjunction with the fault category. A comprehensive assessment is needed to determine whether it is the root cause of the fault.

[0193] By analyzing the edge weights in graph convolution This can identify the strongest dependencies between features. For example, consider the edge weights from "cylinder 3 compression" to "cylinder 3 head vibration". ( → A significant increase indicates a strong correlation between "abnormal combustion pressure in cylinder 3" and "abnormal cylinder head vibration in cylinder 3," providing a key clue to the fault propagation path.

[0194] It should be noted that for the graph neural network model in this embodiment, which consists of multiple spatial-temporal convolutional layers, attention pooling layers, and classification output layers, a preset total loss function can be used for pre-training to jointly optimize monitoring accuracy and physical consistency, ensuring that the model learns physical laws rather than data noise. The total loss function is expressed by the following formula: In the formula, , as well as These represent the losses of the main task. Physical constraint loss The weighting coefficients or hyperparameters of the active regularization term. The main task loss can be set to either classification or regression depending on the application scenario of the model. When corresponding to engine fault category identification in the above embodiment, classification loss is used; when corresponding to engine health status numerical prediction in other embodiments, regression loss can be used. Specifically, the classification loss can use the cross-entropy function, and the regression loss is calculated using the mean squared error (MSE). Thus, the classification loss and regression loss are expressed as follows: In the formula, t represents the time sequence sampling time, and K represents the total number of fault categories. This represents the true label of the k-th type of fault at time t. This represents the output of the model's classification output layer, i.e., the probability that the model predicts the fault type k at time t. T represents the total time-series sample length. This represents the model's predicted output value. It represents the actual measured value in physical measurements.

[0195] The settings ensure that the model output conforms to engine dynamics, such as the consistency between vibration signals and inertial forces. This physical constraint loss can be expressed by the following formula: in From physical equations (such as The theoretical characteristics of node i are calculated.

[0196] The setting can prevent overfitting, and this regularization term can be L2 regularization, etc.

[0197] As an embodiment of the present invention, the engine condition monitoring method further includes the following steps: S201: A graph network model is used to process the dynamic graph of the engine during normal operation to obtain normal global graph features.

[0198] S202: Determine the engine health baseline based on normal global graph features.

[0199] It should be noted that the engine health benchmark established in this embodiment is based on the principle of combining physical topology priors with data-driven verification, constructing a dynamic monitoring boundary that is linked to operating parameters in real time. This process does not rely on any high-fidelity numerical model, but completes the calibration solely through physical structure definition and short-term experimental data.

[0200] As a specific implementation method, the following process is adopted when establishing the engine health benchmark in this embodiment: Step b1, Physical topology skeleton preset (i.e., parameterless initialization). Specifically, in the initialization phase, no specific physical state of any node (such as damping, oil film thickness) is preset; only the graph structure of the system is defined, i.e., the definitions of nodes, directed edges, and dynamic weights mentioned above, without including specific mechanical numerical solutions.

[0201] Step b2, short-term bench data acquisition and feature extraction. During the engine's factory break-in or bench acceptance phase, only short-term (e.g., 30 minutes to one hour) health operation data needs to be collected, covering typical operating conditions such as idling, medium load, and full load. First, the raw signal is processed to obtain an initial feature vector. Then, operating condition alignment is performed, and the corresponding operating condition parameters are recorded synchronously. This is used for subsequent calculation of dynamic weights.

[0202] Step b3, Physically Guided Feature Mapping. This step inputs the processed data stream into the graph neural network model. The graph neural network model automatically learns the dynamic coupling relationships between the engine components through stacked spatiotemporal convolutional layers and outputs the hidden feature representations of each sensor node at the current moment, denoted as the node feature set. .

[0203] Attention pooling is used to aggregate node features, combining the features of N nodes. Aggregation into global graph features The Softmax function automatically learns the importance weights of each node to the system state. : During the offline phase, all collected normal operation data samples are traversed to calculate the corresponding global graph feature vector sequence. Based on this sequence, the mean of its statistical distribution can be directly calculated. Covariance Matrix : The above obtained and They jointly defined the multidimensional feature space envelope of the engine in a healthy state, namely the healthy operating space or the health benchmark.

[0204] S203: Provide graded early warnings for the engine based on the deviation of the engine's current global graph features and health baseline.

[0205] Specifically, in this embodiment, Mahalanobis distance is used to calculate the current global features. Deviation from health benchmarks: in, This represents the inverse indicator of the Comprehensive Health Index (HI) (the higher the value, the worse the health). The health index is defined as: 1 represents perfect health.

[0206] After determining the health index based on the deviation, the engine is given a graded warning based on the relationship between the health index and the graded warning threshold. This graded warning threshold can be set based on historical fault data or reliability targets. In this embodiment, the thresholds in Table 3 below are specifically used for graded warnings: Table 3. Tiered Early Warning Figure 3 This is a flowchart illustrating the engine condition monitoring method provided in an embodiment of the present invention. Figure 3 As shown, the engine condition monitoring method is implemented using the following process: S301: Data acquisition from multi-source sensors and physical equations.

[0207] Data collected using a crankshaft position sensor ), vibration sensor ( ), cylinder pressure sensor ( ), and other sensors such as temperature and pressure collect measured state data; at the same time, theoretical physical data are obtained by calculating kinematic equations, dynamic equations and thermodynamic equations.

[0208] S302: Definition and operation of graph elements.

[0209] Feature extraction and weighted fusion are performed on the measured state data to obtain collaborative features. These features are then combined with physical type and location to obtain graph node attributes, thus defining the nodes. Directed edges between nodes are constructed based on the existence of a physical interaction path between any two nodes. Subsequently, amplitude gain is used... Phase delay and angle modulation function Perform dynamic edge weight calculation. Construct a dynamic graph structure based on the node, directed edges, and dynamic edge weights.

[0210] S303: Graph Neural Network Model Processing.

[0211] The dynamic sequence of graphs showing changes with crankshaft angle is input into a graph neural network model for information aggregation to obtain global features and failure probabilities. Simultaneously, initial data consisting of short-term operational data from the factory is also input into the pre-trained graph neural network model to obtain a health baseline.

[0212] S304: Online assessment and diagnosis.

[0213] Based on the health benchmarks established during the manufacturing phase, and bidirectionally compared with the real-time global features extracted during the online operation phase, the system ultimately completes health status assessment, fault warning and diagnosis, feeds back the diagnostic results, and performs state modulation.

[0214] This invention employs a first-principles physical information graph neural network, embedding engine dynamics equations (such as the kinematic / dynamic equations of the crankshaft-connecting rod-piston system and the in-cylinder combustion thermodynamic equations) as hard constraints or soft regularization terms into the training process of the graph neural network model. Simultaneously, in the constructed dynamic graph, the connection weights are no longer simple distance functions or generated by reinforcement learning, but rather dynamic functions synchronized in real-time with the crankshaft angle (phase). Furthermore, this invention combines causal reasoning with graph neural networks, achieving root cause analysis and prognostic prediction of faults through causal discovery and reasoning of early, minor faults. Thus, this invention constructs a complete intelligent operation and maintenance system from data acquisition to decision feedback.

[0215] Exemplary device Figure 4 This is a schematic diagram of the engine condition monitoring device provided in an embodiment of the present invention. Figure 4 As shown, the engine condition monitoring device is applied to a monitoring system and includes: The node definition module 401 is used to acquire data collected by multiple source sensors and define graph nodes based on the data; The edge definition module 402 is used to define directed edges based on direct physical interaction paths between nodes, wherein the direct physical interaction paths are determined according to the engine's basic physical model. The edge weight construction module 403 is used to construct edge weights based on a dynamic weight function driven by transient operating conditions, with the engine's basic physical model as a constraint. The graph construction and status monitoring module 404 is used to construct a dynamic graph structure based on the graph nodes, the directed edges, and the edge weights, and to process the dynamic graph sequence formed by the dynamic graph structure using a graph network model to monitor the engine status.

[0216] In one or more embodiments of the present invention, the node definition module is specifically used for: extracting physically meaningful time-varying features from the data based on the engine's basic physical model; constructing a physical field correlation matrix based on the correlation strength of the physical fields to which different sensors belong; performing inter-field weighted fusion on the time-varying features based on the physical field correlation matrix to obtain the collaborative features corresponding to each sensor node; and defining graph nodes based on the physical type, location coordinates, and collaborative features of the sensors.

[0217] In one or more embodiments of the present invention, the edge definition module is specifically used to: determine whether there is a direct physical interaction path between nodes based on the engine's basic physical model, wherein the direct physical interaction path includes force flow, vibration wave and heat conduction; when there is a direct physical interaction path between any two nodes, a directed edge is formed between the two nodes.

[0218] In one or more embodiments of the present invention, the edge weight construction module is specifically used for: determining a first relationship between amplitude gain and rotational speed using finite element analysis or modal analysis, constrained by the engine's basic physical model; determining a second relationship between phase delay and rotational speed based on node spacing and structural wave velocity; determining a third relationship between angle modulation function and crankshaft angle based on the cylinder combustion period in the engine, combined with a Gaussian window function; determining a static coupling coefficient based on the geometric distance between nodes and structural connectivity; constructing a dynamic weight function by combining the first, second, and third relationships and the static coupling coefficient, and quantizing the dynamic weight function according to the rotational speed and crankshaft angle in the data, so as to construct edge weights.

[0219] In one or more embodiments of the present invention, the multi-source sensor includes a cylinder pressure sensor, a cylinder head vibration sensor, a main bearing housing vibration sensor, and a body surface vibration sensor; the dynamic graph structure includes a multi-layer dynamic graph structure, which includes a combustion excitation layer, a crankshaft force transmission layer, and a structure propagation layer. The combustion excitation layer includes cylinder pressure sensor nodes and cylinder head vibration sensor nodes, the crankshaft force transmission layer includes cylinder pressure sensor nodes and main bearing housing vibration sensor nodes, and the structure propagation layer includes cylinder head vibration sensor nodes, main bearing housing vibration sensor nodes, and body surface vibration sensor nodes. Corresponding sensor nodes between layers are interconnected.

[0220] In one or more embodiments of the present invention, the basic physical model of the engine is constructed in the following manner: the kinematic equations of the engine are constructed based on the basic geometric dimensions of the engine; the dynamic equations of the engine are constructed based on the force analysis of the engine based on the kinematic equations; and the basic physical model of the engine is constructed based on the kinematic equations and the dynamic equations, combined with the parameter values ​​determined by the engine model.

[0221] In one or more embodiments of the present invention, the graph network model includes multiple spatial-temporal convolutional layers, attention pooling layers, and a classification output layer; the graph construction and state monitoring module is specifically used for: performing convolution processing on the dynamic graph sequence composed of the dynamic graph structure using multiple spatial-temporal convolutional layers to obtain the spatiotemporal fusion features of all nodes at each time step; performing aggregation processing on the spatiotemporal fusion features using attention pooling layers to obtain global graph features; performing classification processing on the global graph features using a classification output layer to obtain the fault probabilities of different fault categories; and performing fault node localization and fault path identification based on the edge weights of directed edges and the attention weights of different nodes determined in the attention pooling layer, combined with the fault probabilities of the fault categories.

[0222] In one or more embodiments of the present invention, the device further includes an early warning module, specifically used for: processing the dynamic graph of the engine during normal operation using a graph network model to obtain normal global graph features; determining the engine's health baseline based on the normal global graph features; and providing graded early warnings for the engine based on the deviation between the engine's current global graph features and the health baseline.

[0223] Exemplary device Figure 5 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present invention. For example... Figure 5 As shown, the electronic device of this embodiment includes a processor 501 and a memory 502.

[0224] The memory 502 stores computer-executed instructions; the processor 501 executes the computer-executed instructions stored in the memory to implement the various steps performed by the electronic device in the above embodiments. For details, please refer to the relevant descriptions in the foregoing method embodiments.

[0225] Alternatively, the memory 502 can be either standalone or integrated with the processor 501.

[0226] When the memory 502 is set up independently, the electronic device also includes a bus 503 for connecting the memory 502 and the processor 501.

[0227] Exemplary media and products This invention also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the aforementioned engine status monitoring method.

[0228] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described engine condition monitoring method.

[0229] In the several embodiments provided by this invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or modules, and may be electrical, mechanical, or other forms.

[0230] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to implement the solution of this embodiment according to actual needs.

[0231] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The unit composed of the above modules can be implemented in hardware or in the form of hardware plus software functional units.

[0232] The integrated modules described above, implemented as software functional modules, can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute certain steps of the methods of the various embodiments of the present invention.

[0233] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor.

[0234] The memory may include high-speed RAM, and may also include non-volatile storage (NVM), such as at least one disk storage device, and may also be a USB flash drive, external hard drive, read-only memory, disk or optical disc, etc.

[0235] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0236] The aforementioned storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage medium can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0237] An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of the processor. Both the processor and the storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and storage medium can exist as discrete components in an electronic device or host device.

[0238] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0239] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An engine condition monitoring method, characterized in that, The method includes: Acquire data collected by multiple source sensors and define graph nodes based on the data; Directed edges are defined by direct physical interaction paths between nodes, and these direct physical interaction paths are determined based on the engine's fundamental physical model. Using the engine's basic physical model as a constraint, edge weights are constructed based on a dynamic weight function driven by transient operating conditions; A dynamic graph structure is constructed based on the graph nodes, the directed edges, and the edge weights. A graph network model is then used to process the dynamic graph sequence formed by the dynamic graph structure to monitor the engine status.

2. The engine condition monitoring method according to claim 1, characterized in that, The graph nodes defined based on the data include: Based on the engine's fundamental physical model, time-varying features with physical significance are extracted from the data; Construct a physical field correlation matrix based on the correlation strength of the physical fields to which different sensors belong; Based on the physical field correlation matrix, the time-varying features are weighted and fused between fields to obtain the collaborative features corresponding to each sensor node; Graph nodes are defined based on the sensor's physical type, location coordinates, and the cooperative features.

3. The engine condition monitoring method according to claim 1, characterized in that, The directed edges are defined by direct physical interaction paths between nodes, and these direct physical interaction paths are determined based on the engine's fundamental physical model, including: Based on the engine's basic physical model, it is determined whether there are direct physical interaction paths between nodes. These direct physical interaction paths include force flow, vibration waves, and heat conduction. When there is a direct physical interaction path between any two nodes, a directed edge is formed between the two nodes.

4. The engine condition monitoring method according to claim 1, characterized in that, The construction of edge weights based on a dynamic weight function driven by transient operating conditions, constrained by the engine's fundamental physical model, includes: Using the aforementioned basic physical model of the engine as a constraint, the first relationship between amplitude gain and rotational speed is determined; A second relationship between phase delay and rotational speed is determined based on node spacing and structural wave velocity; By combining the Gaussian window function, a third relationship between the angle modulation function and the crankshaft angle is determined based on the combustion period of the cylinder in the engine. The static coupling coefficient is determined based on the geometric distance between nodes and the structural connectivity. A dynamic weight function is constructed by combining the first relationship, the second relationship, the third relationship, and the static coupling coefficient. The dynamic weight function is then quantized based on the rotational speed and crankshaft angle in the data to construct edge weights.

5. The engine condition monitoring method according to claim 1, characterized in that: The multi-source sensors include a cylinder pressure sensor, a cylinder head vibration sensor, a main bearing housing vibration sensor, and a machine body surface vibration sensor. The dynamic graph structure includes a multi-layer dynamic graph structure, which includes a combustion excitation layer, a crankshaft force transmission layer, and a structure propagation layer. The combustion excitation layer includes cylinder pressure sensor nodes and cylinder head vibration sensor nodes. The crankshaft force transmission layer includes cylinder pressure sensor nodes and main bearing housing vibration sensor nodes. The structure propagation layer includes cylinder head vibration sensor nodes, main bearing housing vibration sensor nodes, and engine block surface vibration sensor nodes. Corresponding sensor nodes between layers are interconnected.

6. The engine condition monitoring method according to claim 1, characterized in that, The fundamental physical model of the engine is constructed in the following manner: The kinematic equations of the engine are constructed based on its basic geometric dimensions. Based on the kinematic equations, the force analysis of the engine is performed, and the dynamic equations of the engine are constructed. Based on the aforementioned kinematic and dynamic equations, and combined with the parameter values ​​determined by the engine model, a basic physical model of the engine is constructed.

7. The engine condition monitoring method according to claim 1, characterized in that, The graph network model includes multiple spatial-temporal convolutional layers, attention pooling layers, and a classification output layer. The process of using a graph network model to process the dynamic graph sequence composed of the dynamic graph structure for engine status monitoring includes: The dynamic graph sequence composed of the dynamic graph structure is processed by the multi-layer spatial-temporal convolutional layer to obtain the spatiotemporal fusion features of all nodes at each time step. The attention pooling layer is used to aggregate the spatiotemporal fusion features to obtain global graph features; The global graph features are classified using the classification output layer to obtain the fault probabilities of different fault categories; Based on the edge weights of the directed edges and the attention weights of different nodes determined in the attention pooling layer, combined with the fault probabilities of the fault categories, fault node localization and fault path identification are performed.

8. The engine condition monitoring method according to claim 7, characterized in that, The method further includes: A graph network model is used to process the dynamic graph sequence during normal engine operation to obtain normal global graph features; The engine's health baseline is determined based on the aforementioned normal global graph features; The engine is given a graded early warning based on the global graph features of the engine at the current moment and the deviation of the health benchmark.

9. An engine condition monitoring device, characterized in that, The device includes: The node definition module is used to acquire data collected by multiple source sensors and define graph nodes based on the data; The edge definition module is used to define directed edges based on the direct physical interaction path between nodes, wherein the direct physical interaction path is determined according to the engine's basic physical model. The edge weight construction module is used to construct edge weights based on a dynamic weight function driven by transient operating conditions, with the engine's basic physical model as a constraint. The graph construction and status monitoring module is used to construct a dynamic graph structure based on the graph nodes, the directed edges, and the edge weights, and to process the dynamic graph sequence formed by the dynamic graph structure using a graph network model to monitor the engine status.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the engine condition monitoring method as described in any one of claims 1 to 8.

Images