Emission monitoring method and system for automobile internal combustion engine

Abstract

The invention discloses an emission monitoring method and system for an automobile internal combustion engine, and the method comprises the steps: collecting a multi-dimensional fusion sensing data set according to the operation state of the automobile internal combustion engine; for the multi-dimensional fusion sensing data set, generating a dynamic emission state feature vector set representing the current combustion efficiency, the wear state of the key part and the emission generation path; according to the dynamic emission state feature vector set, an emission health state map reflecting the current comprehensive emission health state of the engine is constructed and updated in real time; according to the emission health state map, generating an emission abnormity diagnosis report containing a specific fault mode; and according to the emission abnormality diagnosis report and the historical emission health state map sequence, predicting the deterioration trend of a specific emission path of the engine, and generating a targeted preventive engine maintenance strategy. By utilizing the embodiment of the invention, the real-time accurate diagnosis and trend prediction of the engine emission health condition can be realized, and the initiative and maintenance efficiency of emission exception handling are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of vehicles, and in particular to an emission monitoring method and system for an automobile internal combustion engine. Background Art

[0002] Traditional automotive internal combustion engine emissions monitoring methods typically rely on a single sensor (such as an oxygen sensor or exhaust gas analyzer) to determine thresholds, making it difficult to fully reflect the complex engine emissions profile. Existing technologies lack the ability to collaboratively analyze combustion efficiency, component wear, and emission generation pathways, leading to delayed fault diagnosis or misdiagnosis. Furthermore, discrete monitoring data cannot dynamically correlate with engine health status, nor can it predict emission deterioration trends, making preventive maintenance strategies less targeted. Summary of the Invention

[0003] The purpose of the present invention is to provide an emission monitoring method and system for an automobile internal combustion engine to address the deficiencies in the prior art, enable real-time and accurate diagnosis and trend prediction of the engine emission health status, and improve the initiative and maintenance efficiency of emission anomaly handling.

[0004] One embodiment of the present application provides a method for monitoring emissions of an automobile internal combustion engine, the method comprising: Based on the operating status of the vehicle's internal combustion engine, a multi-dimensional fusion sensor data set including exhaust gas composition spectral data, combustion chamber pressure fluctuation waveform, in-cylinder temperature gradient, crankcase blowby flow rate, and exhaust back pressure is collected in real time and synchronously; Performing spatiotemporal alignment and feature extraction on the multi-dimensional fusion sensor data set to generate a dynamic emission state feature vector set representing current combustion efficiency, wear status of key components, and emission generation path; Based on the dynamic emission status feature vector set, a pre-built emission health map model is used to map the feature vectors to map nodes, and an emission health state map reflecting the current comprehensive emission health status of the engine is constructed and updated in real time; Based on the emission health state map, perform map node state deviation analysis and path anomaly detection processing to identify the root cause components or processes that lead to potential or actual emission violations, and generate an emission anomaly diagnosis report containing specific failure modes; According to the emission abnormality diagnosis report and the historical emission health status map sequence, a prediction algorithm based on map evolution is applied to predict the deterioration trend of the engine's specific emission path and generate a targeted preventive engine maintenance strategy.

[0005] Optionally, the real-time synchronous acquisition of a multi-dimensional fusion sensing data set including exhaust gas component spectral data, combustion chamber pressure fluctuation waveform, in-cylinder temperature gradient, crankcase blowby flow rate, and exhaust back pressure based on the operating state of the automobile internal combustion engine includes: Based on the phase-synchronized pulse signal output by the crankshaft position sensor, the sampling clocks of the cylinder pressure sensor, exhaust gas spectrometer, temperature array, blowby gas flowmeter, and back pressure sensor are synchronized to obtain the original signal stream with time base alignment. Based on the original signal stream aligned with the time base, the combustion cycle is segmented, and the compression stroke and power stroke data slices are divided according to the crankshaft angle to obtain the combustion stage grouped signal set; According to the grouped signal sets of the combustion stage, the engine geometric space mapping is performed. Based on the piston kinematics, the cylinder wall vibration signal is mapped to the position of the piston ring-cylinder liner friction pair to obtain the spatial registration vibration feature matrix. According to the spatial registration vibration feature matrix, multi-source noise cancellation processing is performed, exhaust spectrum, temperature and flow data are fused, and ignition electromagnetic interference is suppressed to obtain a multi-dimensional fusion sensing data set synchronized in time and space.

[0006] Optionally, performing spatiotemporal alignment and feature extraction on the multi-dimensional fusion sensor data set to generate a dynamic emission state feature vector set characterizing current combustion efficiency, wear state of key components, and emission generation path includes: Based on the multi-dimensional fusion sensor data set, the cylinder pressure waveform variational mode decomposition is performed to extract the pressure rise rate inflection point and the indicated mean effective pressure energy distribution characteristics, and the combustion stability feature vector is obtained; Based on the exhaust spectral data, the chemical bond absorption peaks are correlated, and the emission concentration ratio and transformation path are analyzed to obtain the pollutant generation path characteristic vector; Based on the spatial registration vibration characteristic matrix, the harmonic wavelet packet entropy algorithm is used to quantify the energy attenuation gradient in the 2-5kHz frequency band to obtain the piston ring wear state characteristic vector; The combustion stability feature vector, pollutant generation path feature vector and piston ring wear state feature vector are tensor-fused to construct a three-dimensional dynamic emission state tensor. According to the three-dimensional dynamic emission state tensor, non-negative constrained tensor decomposition is performed to extract the low-rank feature subspace that characterizes emission degradation and obtain the dynamic emission state feature vector set.

[0007] Optionally, the method includes mapping the feature vectors to map nodes based on the dynamic emission state feature vector set using a pre-built emission health map model, and constructing and updating an emission health state map reflecting the current comprehensive emission health status of the engine in real time, including: Based on the dynamic emission state feature vector set, graph attention network matching is performed to map the nearest nodes in the pre-built emission health knowledge graph to obtain the initial node mapping relationship; According to the initial node mapping relationship, the node state diffusion algorithm is used to calculate the semantic similarity weight between the dynamic emission state feature vector and the node, and the weighted edge connection relationship is obtained; Based on the weighted edge connection relationship, a temporal graph convolutional network is used to fuse the historical health status sequence and update the key node scores to obtain a real-time health score matrix; Based on the real-time health score matrix and the abnormal path backtracking mechanism, the causal chain of combustion domain, component domain and emission domain is reconstructed to obtain an emission health status map that reflects the current comprehensive emission health status of the engine.

[0008] Optionally, the emission health state map is used to perform map node state deviation analysis and path anomaly detection processing to identify the root cause components or processes that lead to potential or actual emission violations, and generate an emission anomaly diagnosis report containing specific failure modes, including: According to the emission health status map, the node deviation is calculated to quantify the KL divergence between the key node status and the health benchmark, and the node deviation vector is obtained; Based on the node deviation vector, the minimum causal path search algorithm is used to locate the core fault propagation chain that causes excessive emissions. Based on the core fault propagation chain, thermodynamic constraint verification is performed to screen out false paths that violate the temperature-pressure-flow coupling equation and obtain the verified fault chain; According to the verified fault chain, the fault mode is decoupled to obtain a fault mode list; Based on the failure mode list, component failure probabilities are fused and combined with real-time data to generate an emission anomaly diagnosis report with confidence intervals.

[0009] Optionally, the application of a prediction algorithm based on the evolution of the map based on the emission abnormality diagnosis report and the historical emission health status map sequence to predict the deterioration trend of a specific emission path of the engine and generate a targeted preventive engine maintenance strategy includes: Based on the emission anomaly diagnosis report and the historical emission health status graph sequence, the graph evolution spatiotemporal tensor is constructed, and the node state time slices are stored to obtain the graph evolution spatiotemporal tensor; According to the spatiotemporal tensor of the graph evolution, the spatiotemporal graph neural network is used to predict the state trajectory of key nodes in the future preset period to obtain the predicted state trajectory; Based on the predicted state trajectory, a three-objective Pareto optimization process is performed to find the optimal balance point among emission risk, maintenance cost, and downtime, and to obtain the optimal intervention plan. According to the optimal intervention plan, the vehicle operation plan is integrated and combined with the database to output a preventive maintenance instruction set.

[0010] Another embodiment of the present application provides an emission monitoring system for an automobile internal combustion engine, the system comprising: The acquisition module is used to synchronously collect multi-dimensional fusion sensor data sets including exhaust gas component spectral data, combustion chamber pressure fluctuation waveform, cylinder temperature gradient, crankcase blowby flow rate and exhaust back pressure in real time according to the operating status of the automobile internal combustion engine; an extraction module for performing spatiotemporal alignment and feature extraction on the multi-dimensional fusion sensor data set to generate a dynamic emission state feature vector set representing current combustion efficiency, wear status of key components, and emission generation path; A construction module is used to map the feature vectors to map nodes based on the dynamic emission state feature vector set using a pre-built emission health map model, and to construct and update the emission health state map reflecting the current comprehensive emission health status of the engine in real time; a detection module configured to perform, based on the emission health state map, state deviation analysis of map nodes and path anomaly detection processing, identify the root cause components or processes that lead to potential or actual emission violations, and generate an emission anomaly diagnosis report containing specific failure modes; The prediction module is used to apply a prediction algorithm based on the evolution of the map based on the emission abnormality diagnosis report and the historical emission health status map sequence to predict the deterioration trend of the specific emission path of the engine and generate a targeted preventive engine maintenance strategy.

[0011] Yet another embodiment of the present application provides a storage medium, wherein the storage medium stores a computer program, wherein the computer program is configured to execute any of the above methods when run.

[0012] Yet another embodiment of the present application provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute any of the above methods.

[0013] Compared with the prior art, the present invention provides an emission monitoring method for an automobile internal combustion engine. According to the operating status of the automobile internal combustion engine, a multi-dimensional fusion sensor data set is synchronously collected in real time; for the multi-dimensional fusion sensor data set, a dynamic emission state feature vector set is generated to characterize the current combustion efficiency, wear status of key components and emission generation path; according to the dynamic emission state feature vector set, an emission health state map reflecting the current comprehensive emission health status of the engine is constructed and updated in real time; according to the emission health state map, an emission abnormality diagnosis report containing specific fault modes is generated; according to the emission abnormality diagnosis report and the historical emission health state map sequence, the deterioration trend of the specific emission path of the engine is predicted, and a targeted preventive engine maintenance strategy is generated, thereby realizing real-time and accurate diagnosis and trend prediction of the engine emission health status, and improving the initiative of emission abnormality handling and maintenance efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 A hardware structure block diagram of a computer terminal for an automobile internal combustion engine emission monitoring method provided by an embodiment of the present invention; Figure 2 A schematic flow chart of an emission monitoring method for an automobile internal combustion engine provided by an embodiment of the present invention; Figure 3 A schematic structural diagram of an emission monitoring system for an automobile internal combustion engine provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0015] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limiting the present invention.

[0016] The embodiment of the present invention first provides a method for monitoring emissions of an automobile internal combustion engine. The method can be applied to electronic devices such as computer terminals, specifically ordinary computers.

[0017] The following describes it in detail by taking running on a computer terminal as an example. Figure 1 The hardware structure block diagram of a computer terminal for an automobile internal combustion engine emission monitoring method provided by an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, a memory, and a network interface connected via a system bus, wherein the memory may include a non-volatile storage medium and an internal memory.

[0018] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions, which, when executed, can cause a processor to execute any one of the emission monitoring methods for an automobile internal combustion engine.

[0019] The processor is used to provide computing and control capabilities and support the operation of the entire computer equipment.

[0020] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor, the processor can execute any emission monitoring method for an automobile internal combustion engine.

[0021] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art will understand that Figure 1 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0022] It should be understood that the processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0023] See also Figure 2 An embodiment of the present invention provides an emission monitoring method for an automobile internal combustion engine, which may include the following steps: S201, based on the operating state of the automobile internal combustion engine, real-time synchronous acquisition of a multi-dimensional fusion sensor data set including exhaust gas component spectral data, combustion chamber pressure fluctuation waveform, cylinder temperature gradient, crankcase blowby flow rate, and exhaust back pressure; Specifically, the sampling clocks of the cylinder pressure sensor, exhaust gas spectrometer, temperature array, blowby gas flowmeter, and back pressure sensor can be synchronized based on the phase-synchronized pulse signal output by the crankshaft position sensor to obtain the original signal stream with time base alignment. The core synchronization mechanism of the Engine Control Unit (ECU) relies on the crankshaft position sensor (CPS). This sensor, mounted on the end of the crankshaft or near the flywheel, uses magnetoresistive or Hall effect principles to generate a phase-synchronized pulse signal (PSPS) every time the crankshaft rotates by a specified angle (e.g., 6 degrees or 1 degree). The rising or falling edge of this pulse signal corresponds to the precise physical position of the crankshaft angle (e.g., 30 degrees before top dead center). The ECU is equipped with a dedicated clock synchronization module (CSM). This module uses the PSPS as a global time reference (GTR) and sends a unified sampling trigger command (STC) to all sensors. For example, the cylinder pressure sensor (CyPS), mounted in the combustion chamber spark plug hole or a dedicated pressure tap, originally runs independently at a 1 megahertz (MHz) sampling clock (SCLK). When the CSM detects a rising edge in the PSPS, it immediately sends an STC signal to the CyPS, forcing its next sampling point to align with the crankshaft angle. The sampling period of an exhaust gas spectrometer (EGS) is typically in the millisecond range. The CSM calculates the EGS's equivalent sampling time window (STW) based on the PSPS interval (e.g., one pulse every 6 degrees), ensuring that the start of each spectral scan is strictly synchronized with the crankshaft angle.

[0024] The temperature array (TA) consists of multiple miniature thermocouples (TCs) or resistance temperature detectors (RTDs) distributed at key locations on the cylinder head, cylinder wall, or exhaust manifold. Each probe has an independent sampling clock, and the CSM achieves synchronization through the Distributed Timestamp Protocol (DTP). When a PSPS pulse arrives, the CSM broadcasts a synchronization frame (SF) with a timestamp (TS) to all TA nodes. Each node dynamically adjusts the sampling instant (SI) based on the deviation (Clock Skew, CSK) between its local clock and the TS. Blow-by flow meters (BFMs) typically use hot-film or ultrasonic principles, and their flow sampling period is related to the crankshaft speed. The CSM dynamically calculates the BFM sampling interval (SIV) based on the PSPS frequency (e.g., 33.3 microseconds at 3000 rpm), ensuring that each flow data point corresponds to a fixed crank angle interval (e.g., every 10 degrees). The back-pressure sensor (BPS) is installed at the end of the exhaust manifold, and its signal is susceptible to interference from exhaust pulsation. The CSM utilizes the PSPS to trigger the BPS's sample-and-hold circuit (SHC), collecting valid pressure values ​​only during the stable phase of exhaust valve closing (e.g., 240-480 degrees crank angle), thus minimizing pulsation noise.

[0025] Ultimately, all sensor data is transmitted to the ECU via the Controller Area Network (CAN) bus or a dedicated high-speed data link (such as Ethernet). The ECU's Signal Alignment Buffer (SAB) allocates a separate buffer for each sensor, sorting data by PSPS timestamp index (TIDX). For example, after the Nth PSPS pulse is triggered, the CyPS uploads the cylinder pressure value (in bar) at that moment within 0.1 milliseconds, the EGS uploads the spectral data (absorbance matrix) for that window within 5 milliseconds, the TA uploads the temperature at each point (in degrees Celsius) within 2 milliseconds, the BFM uploads the blowby flow rate (in liters / minute), and the BPS uploads the backpressure value (in kilopascals). The SAB packages all data from the same pulse trigger into a time-base aligned raw signal stream (TBARSS) using the TIDX as the index. Each packet in this data stream contains a global time tag (GTT) accurate to the microsecond level, ensuring comparability of all physical quantities in subsequent analysis.

[0026] Based on the original signal stream aligned with the time base, the combustion cycle is segmented, and the compression stroke and power stroke data slices are divided according to the crankshaft angle to obtain the combustion stage grouped signal set; The ECU's Cycle Segmentation Engine (CSE) parses the TBARSS data stream, first identifying engine cycle (EC) boundaries. Each complete cycle corresponds to 720 degrees of crankshaft rotation (for a four-stroke engine) and is marked by the Top Dead Center Signal (TDCS). The TDCS is signaled by specific pulses in the PSPS sequence (e.g., the 0-degree crank angle pulse). Based on the TDCS, the CSE segments the TBARSS into consecutive, independent Cycle Data Blocks (CDBs). Each CDB contains all sensor data packets within that cycle, sorted by TIDX. The CSE then further subdivides each CDB into four strokes using the Crankshaft Angle Resolver (CAR): intake stroke (0-180 degrees), compression stroke (180-360 degrees), power stroke (360-540 degrees), and exhaust stroke (540-720 degrees). Among them, the compression and power strokes are the core of combustion analysis.

[0027] For the compression stroke (CS), CSE extracts a data subset from 180-360 degrees of crankshaft angle. During this range, cylinder pressure data (CyPS) exhibit a monotonically increasing trend, and the temperature array (TA) indicates an increasing in-cylinder gas temperature gradient. EGS typically does not collect data during this period (due to exhaust valve closure). CSE resamples (RS) the CS interval data at a preset angular resolution (e.g., every 1 degree) to generate evenly spaced compression stroke data slices (CSDS). These slices contain key parameters such as the cylinder pressure curve, cylinder wall temperature distribution, and crankcase blowby flow (BFM). Data processing for the power stroke (PS) is more complex: CSE extracts data from 360-540 degrees and identifies combustion feature points (CFPs), such as the location of the cylinder pressure peak (typically 10-20 degrees after top dead center). Based on CFP, CSE subdivides PS into the pre-combustion period (360 degrees to the moment of spark ignition), the main combustion period (ignition to the cylinder pressure peak), and the after-combustion period (peak to 540 degrees), generating a high-resolution power stroke data slice (PSDS), which includes cylinder pressure fluctuation details, EGS transient emission spectrum, and in-cylinder temperature field evolution.

[0028] Finally, CSE packages the CSDS and PSDS of each engine cycle into a Combustion Phase Grouped Signal Set (CPGSS). This set is a structured data container, for example: Compression stroke group: contains the cylinder pressure array (CyPS_Array), cylinder head temperature distribution map (TA_Map), and blowby flow sequence (BFM_Seq) for every degree from 180 to 360 degrees; Power stroke group: includes the cylinder pressure micro-fluctuation waveform (CyPS_Wave) every 0.5 degrees within 360-540 degrees, the nitrogen oxide (NOx) spectrum absorption peak of EGS (EGS_NOxPeak), and the coordinates of the high-temperature hotspot in the cylinder (TA_HotSpot).

[0029] All data are labeled with a crank angle label (CAL) to facilitate subsequent correlation analysis. CPGSS is stored by cycle number, forming a two-dimensional time-angle indexed database.

[0030] According to the grouped signal sets of the combustion stage, the engine geometric space mapping is performed. Based on the piston kinematics, the cylinder wall vibration signal is mapped to the position of the piston ring-cylinder liner friction pair to obtain the spatial registration vibration feature matrix. The cylinder wall vibration signal (CWVS) is collected by a vibration accelerometer (VA) mounted on the outside of the cylinder block. Its raw data only reflects the overall vibration energy. To localize wear, Engine Geometric Space Mapping (EGSM) is required. The EGSM module first loads the engine's 3D Geometric Model (GM), which includes parameters such as the cylinder bore diameter (CBD), piston pin height (PPH), and connecting rod length (CRL). Based on the piston kinematics equation (PKE), the instantaneous piston position (PIP) is calculated as a function of crankshaft angle (CAL). For example, when CAL = 360 degrees (top dead center), the PIP is at its highest point, while when CAL = 540 degrees (bottom dead center), the PIP is at its lowest point.

[0031] The EGSM module parses CWVS data from CPGSS (typically at a sampling rate of 50 kHz) and extracts vibration waveform segments (VWS) by CAL slice. Based on the PIP calculation results, the piston ring-liner contact zone (PRLCZ) corresponding to each CAL is determined. For example, when the piston is at top dead center, the first piston ring corresponds to the cylinder liner top at coordinate Z = Max; when the piston descends to 400 degrees of crankshaft rotation, the ring's position is Z = Max - 30 mm (the specific value is calculated by PKE). EGSM then models the spatial relationship between the physical installation position of the VA (e.g., at cylinder height Z = 80 mm) and the PRLCZ as a transfer function matrix (TFM). The TFM is solved using the inverse convolution algorithm (ICA), which back-projects (BP) the energy components in the original vibration signal to the actual location of the PRLCZ.

[0032] The spatially registered vibration feature matrix (SRVFM) generated after projection is a two-dimensional structure: Row dimension: Z-Axis Coordinate (ZAC) of the cylinder liner, with a resolution of 1 mm and a range covering the piston stroke (e.g., 0-100 mm); Column dimension: Vibration Feature Parameters (VFP), including time domain indicators (RMS, crest factor) and frequency domain indicators (2-5 kHz band energy, harmonic distortion rate THD).

[0033] For example, an abnormally high RMS value of SRVFM at Z = 45 mm indicates piston ring scraping wear at this height, while a 5 kHz energy decay at Z = 10 mm suggests microcracks on the cylinder liner top. This matrix transforms the vibration signal from "global noise" into a "spatially localized wear fingerprint."

[0034] According to the spatial registration vibration feature matrix, multi-source noise cancellation processing is performed, exhaust spectrum, temperature and flow data are fused, and ignition electromagnetic interference is suppressed to obtain a multi-dimensional fusion sensing data set synchronized in time and space.

[0035] The core of Multi-source Noise Cancellation (MNC) is to separate the effective wear component from the interference noise in the cylinder wall vibration signal. The main noise sources include: Ignition Electromagnetic Interference (IEMI): High-frequency radiation (frequency band > 100 MHz) generated by spark plug discharge is coupled to the vibration sensor circuit; Valve Impact Noise (VIN): Mechanical impact when the intake and exhaust valves seat (frequency range 1-3 kHz); Crankshaft Bearing Vibration (CBV): Low-frequency broadband vibration (frequency band <800 Hz).

[0036] The MNC module adopts the Reference Sensor Cooperative Filtering (RSCF) strategy: an additional ignition current probe (ICP) is deployed to collect IEMI waveforms, an acoustic emission sensor (AES) is installed on the cylinder head to capture VIN characteristics, and a low-frequency accelerometer (LFA) is installed on the crankcase to monitor the CBV.

[0037] The MNC performs three steps to process the vibration data in SRVFM: IEMI suppression: Cross-correlation analysis (CCA) is performed between the interference waveform (Ignition Noise Template, INT) collected by the ICP and the vibration signal, and adaptive subtraction (AS) is performed after calculating the delay compensation from the original signal. VIN / CBV separation: Using the Blind Source Separation (BSS) algorithm, with the AES and LFA signals as reference inputs, the vibration signal is subjected to Independent Component Analysis (ICA) to extract the pure piston ring friction component. Frequency band enhancement: Wavelet Threshold Denoising (WTD) is used to enhance the 2-5 kHz frequency band (piston ring wear characteristic frequency band) and suppress the energy of other frequency bands. The purified vibration characteristics are spatially and temporally fused with other data in CPGSS: Exhaust Gas Spectrometry (EGS): Aligns the NOx absorption peak area (unit: millivolt-second mV·s) with the cylinder pressure curve during the power stroke to establish a correlation between combustion temperature and emission generation Temperature Array Data (TA): Extracts the axial temperature gradient of the cylinder liner (unit: °C / mm) and superimposes it with the RMS vibration value of the SRVFM at the same position to identify wear hot spots caused by overheating Blowby Flow (BFM): Correlates the peak vibration energy corresponding to the sudden increase in flow (in liters per minute) and piston ring blowby during the compression stroke. Ultimately, a multi-dimensional fused sensor dataset synchronized in time and space is generated. This dataset has a five-dimensional structure: time dimension (crankshaft angle), space dimension (cylinder liner height), sensor dimension (pressure / vibration / temperature / spectrum / flow), cycle dimension (engine cycle number), and feature dimension (physical quantities + statistical indicators), providing comprehensive input for subsequent emissions health analysis.

[0038] This step uses multiple sensors to collaboratively collect key physical and chemical parameters during engine operation. These include exhaust gas composition characteristics captured by a spectrometer, combustion dynamics recorded by a high-frequency cylinder pressure sensor, in-cylinder thermal field distribution measured by a distributed temperature probe, blowby gas volume monitored by a flow meter, and exhaust resistance characteristics captured by a backpressure sensor. All data is synchronized at the microsecond level using the crankshaft phase signal to ensure temporal and spatial consistency. This overcomes the limitations of traditional single-point monitoring and builds a multi-dimensional data base reflecting the engine's true operating state. This provides comprehensive, synchronized raw data support for subsequent precise diagnosis, effectively addressing analytical errors caused by asynchronous sampling.

[0039] S202, performing spatiotemporal alignment and feature extraction on the multi-dimensional fusion sensor data set to generate a dynamic emission state feature vector set representing current combustion efficiency, wear status of key components, and emission generation path; Specifically, the cylinder pressure waveform variational mode decomposition can be performed based on the multi-dimensional fusion sensor data set to extract the pressure rise rate inflection point and the indicated mean effective pressure energy distribution characteristics to obtain the combustion stability feature vector; The system receives a Multidimensional Fusion Sensor Dataset (MFSD) from the pre-processing stage. This dataset has been spatially and temporally aligned to ensure synchronization of signals such as the combustion chamber pressure waveform (referred to as the cylinder pressure waveform) and the exhaust gas spectrum. The cylinder pressure waveform is the core input, acquired during each combustion cycle by a high-frequency cylinder pressure sensor (typically sampling at a rate of 100 kHz, or 100,000 samples per second). It is represented by a pressure curve that varies with crankshaft angle. To analyze combustion stability, the raw cylinder pressure waveform is first processed using the Variational Mode Decomposition (VMD) algorithm. VMD uses the principle of adaptive frequency segmentation to decompose the non-stationary cylinder pressure signal into several Intrinsic Mode Functions (IMFs). During specific implementation, the modal number K is set to 5 (empirical value) and the penalty factor α is set to 2000 (control bandwidth). The cylinder pressure signal of a single cycle is iteratively optimized and solved, and sub-signal components representing different physical processes are finally separated: for example, IMF1 corresponds to high-frequency combustion oscillations, IMF2 reflects the main combustion pressure wave, and IMF3 contains information on the pressure rise rate.

[0040] Among the decomposed modes, IMF2 and IMF3 are analyzed in particular. The Pressure Rise Rate Inflection Point (PRR_IP) is extracted as follows: The first-order derivative of the IMF3 component is calculated to obtain the pressure rise rate curve, and the sliding window standard deviation algorithm is used to detect the location of the sudden change. An inflection point is identified when the slope change of five consecutive sampling points (corresponding to 0.1 degrees of crankshaft angle) exceeds the threshold Δslope = 0.5 bar / degree (Bar / Degree). This inflection point marks the transition from premixed combustion to diffusion combustion. A deviation of more than ±2 degrees from the standard value (e.g., 8 degrees after top dead center) indicates combustion instability. The Indicated Mean Effective Pressure Energy Distribution Feature (IMEP_EDF) is extracted from the IMF2 component. The work ratio (WR) of the integrated area of ​​the IMF2 during the compression stroke (crankshaft angle -180° to 0°) and the power stroke (0° to 180°) is calculated. The energy ratio (ER) of the IMF2 signal in the 3-10 kHz frequency band is then calculated. WR reflects the effective work efficiency, while ER represents the intensity of pressure oscillations. Together, these two constitute the core dimensions of IMEP_EDF.

[0041] The final output, the Combustion Stability Feature Vector (CSFV), is a five-dimensional array: [PRR_IP angle deviation value, WR value, ER value, maximum pressure rise rate, pressure cyclic variation coefficient]. The pressure cyclic variation coefficient is calculated (in percentage) by taking the standard deviation of the cylinder pressure peak values ​​over 50 consecutive cycles. For example, under certain operating conditions, the vector might be: [-1.5, 0.82, 0.15, 4.3, 3.8%], indicating an inflection point delay of 1.5 degrees, power efficiency of 82%, a high-frequency oscillation energy share of 15%, a maximum pressure rise rate of 4.3 bar / degree, and a cyclic variation coefficient of 3.8%. This vector is transmitted in real time to the subsequent tensor fusion module, and any numerical anomalies can be directly linked to combustion issues such as ignition delay and air-fuel ratio imbalance.

[0042] Based on the exhaust spectral data, the chemical bond absorption peaks are correlated, and the emission concentration ratio and transformation path are analyzed to obtain the pollutant generation path characteristic vector; Exhaust Gas Spectrum Data (EGSD) is acquired from a wide-band UV-IR spectrometer (covering wavelengths from 200 nm to 5000 nm) at a rate of 100 frames per second. The raw spectra are first baseline corrected using Adaptive Iteratively Reweighted Penalized Least Squares (AIRPLS) to eliminate scattering noise, and then normalized for pathlength effects using Standard Normal Variate (SNV). The preprocessed spectra are then input into a chemical bond absorption peak correlation module, which has a built-in database of pollutant characteristic peaks. For example, nitric oxide (NO) has a strong absorption band at 5.3 μm (mid-infrared), the C-H bond stretching vibration peak of hydrocarbons (HC) is located at 3.4 μm, and nitrogen oxides (NO2) have a characteristic absorption band at 400 nm (UV).

[0043] Emission concentration analysis is achieved using a partial least squares regression (PLSR) model. A mapping relationship between the spectral absorbance matrix and the concentrations measured by the gas chromatograph (GC) is established through bench testing and calibration. During real-time analysis, the absorbance sequence of the measured spectrum in the characteristic wavelength band (e.g., NO: 5.2-5.4 μm) is input into the PLSR model, which outputs concentration values ​​(parts per million) for NO, HC, CO, CO₂, and NO₂. Conversion path analysis relies on reaction kinetic feature extraction: calculation of the NOx Formation Index (NFI) = (NO concentration + NO2 concentration) / (in-cylinder peak temperature × oxygen concentration). An index exceeding the threshold of 0.25 indicates that thermal NOx is dominant. Simultaneously, the peak asymmetry (Asymmetry Index, AI) of the HC spectrum at 2900-3000 wavenumber cm⁻¹ is analyzed. If AI>1.2, it is determined that unburned fuel droplets are present, indicating HC emissions due to the wetting wall effect.

[0044] The generated pollutant generation path feature vector (PGPFV) consists of a six-tuple: [NO concentration, HC concentration, NOx generation index, HC asymmetry, soot transmittance, ammonia-to-nitrogen ratio]. The soot transmittance is calculated as the transmittance at a wavelength of 660 nm (unit: percentage), reflecting the level of particulate matter formation. The ammonia-to-NOx ratio (ANR) = (theoretical ammonia production corresponding to urea injection amount) / (NO concentration × exhaust flow rate) is used to monitor the efficiency of the selective catalytic reduction system. For example, the vector [120, 35, 0.31, 1.5, 88%, 0.9] indicates: NO 120 ppm, HC 35 ppm, predominantly thermal NOx, strong wall wetting, moderate soot concentration, and insufficient urea injection. This vector provides a chain of evidence for subsequent diagnosis, reflecting the chemical reaction dimension.

[0045] Based on the spatial registration vibration characteristic matrix, the harmonic wavelet packet entropy algorithm is used to quantify the energy attenuation gradient in the 2-5kHz frequency band to obtain the piston ring wear state characteristic vector; The Spatially Registered Vibration Feature Matrix (SRVFM) is derived from pre-processing. Its rows correspond to the engine cylinder numbers (e.g., a four-cylinder engine has four rows), its columns represent the crankshaft angle position (with 0.5-degree resolution), and its elements represent the vibration acceleration values ​​(unit: g) at specific cylinder-angle points. This matrix focuses on the piston ring-liner friction pair (PRLFP) region, mapping the cylinder wall vibration sensor signals to the piston ring top dead center (TDC) and bottom dead center (BDC) positions using the piston kinematic model. For example, the vibration acceleration at TDC (0° crankshaft angle) of cylinder 2 is extracted as a key observation point.

[0046] The Harmonic Wavelet Packet Entropy (HWPE) algorithm is performed in three steps: Step 1 - Harmonic Wavelet Packet Decomposition: For each critical position (e.g., top dead center), the vibration signal is decomposed using a seven-layer wavelet packet decomposition using a harmonic wavelet basis function (with adjustable center frequency). The 2-5 kHz frequency band (corresponding to the 4th and 5th sub-bands) is selected, as this frequency range is sensitive to piston ring wear.

[0047] Step 2 - Energy Attenuation Gradient Calculation: Calculate the energy value (EV) for each crankshaft angle interval within the target frequency band. For example, divide the compression stroke (-30° to 0°) into 10 equal segments, and calculate the EV for each segment by the sum of the squares of the wavelet packet coefficients. Energy Attenuation Gradient (EAG) = (EV of the first segment - EV of the tenth segment) / (angular span). As wear increases, lubrication deteriorates, causing high-frequency vibration energy to decay more slowly in the later stages of the compression stroke, resulting in a decrease in the absolute value of EAG.

[0048] Step 3 - Entropy Analysis: Calculate the permutation entropy (PE) of the wavelet packet coefficients in the target frequency band to measure signal complexity. As piston ring wear increases, vibration randomness increases, and the PE value rises.

[0049] The output Piston Ring Wear State Feature Vector (PRWSFV) contains four parameters: [energy decay gradient in the 2-5kHz frequency band at top dead center, permutation entropy in the same frequency band at bottom dead center, total energy value across all frequency bands, and wear consistency index]. The Wear Consistency Index (WCI) is calculated as (standard deviation of EAG across cylinders) / (mean) and is used to detect eccentric wear. For example, the vector [-0.25, 0.78, 15.3, 0.18] indicates a gentle energy decay at top dead center (gradient -0.25 g² / degree), high vibration complexity at bottom dead center (PE = 0.78), a total high-frequency energy value of 15.3 g², and an 18% wear variation across cylinders. This vector directly quantifies the degree of piston ring sealing degradation.

[0050] The combustion stability feature vector, pollutant generation path feature vector and piston ring wear state feature vector are tensor-fused to construct a three-dimensional dynamic emission state tensor. The three input feature vectors are strictly aligned in time and space: they are all generated based on the same combustion cycle. The Combustion Stability Feature Vector (CSFV) has a dimension of 5 (pressure-related parameters), the Pollutant Generation Path Feature Vector (PGPFV) has a dimension of 6 (emission chemistry parameters), and the Piston Ring Wear State Feature Vector (PRWSFV) has a dimension of 4 (mechanical wear parameters). For each cycle, the system constructs a Feature Matrix (FM): rows correspond to feature categories (3 rows total), and columns expand to specific parameters. For example, the first row contains the five parameters of the CSFV, the second row contains the six parameters of the PGPFV (zero-padded if missing), and the third row contains the four parameters of the PRWSFV.

[0051] Tensor Fusion is achieved through high-order tensor expansion. The three dimensions of the 3D Dynamic Emission State Tensor (3D-DEST) are defined as follows: Dimension 1 (combustion dimension): contains all five elements of CSFV and represents the state of the thermodynamic process; Dimension 2 (emission dimension): contains all 6 elements of PGPFV, characterizing the chemical reaction state; Dimension 3 (component dimension): contains all four elements of PRWSFV and characterizes the mechanical wear state.

[0052] Tensor element values ​​undergo feature normalization: For each feature parameter, the mean and standard deviation over the last 1000 cycles are independently calculated, and then normalized using the Z-score (current value - mean) / standard deviation. For example, if the original value of the cylinder pressure cyclic variation coefficient is 3.8%, and the historical mean is 4.0% and the standard deviation is 0.5%, the normalized value is (3.8 - 4.0) / 0.5 = -0.4.

[0053] The resulting three-dimensional dynamic emissions state tensor (3D-DEST) measures 5×6×4 (120 elements total). Its physical significance lies in establishing cross-domain correlations: for example, the tensor element (3,2,1) represents the coupling between the maximum pressure rise rate (the third parameter in the combustion dimension) and the NO concentration (the second parameter in the emissions dimension) under the conditions of the top dead center energy decay gradient (the first parameter in the component dimension). This tensor is updated in real time (once per cycle) and fed into the subsequent eigendecomposition module, providing a structured data foundation for global emissions health analysis.

[0054] According to the three-dimensional dynamic emission state tensor, non-negative constrained tensor decomposition is performed to extract the low-rank feature subspace that characterizes emission degradation and obtain the dynamic emission state feature vector set.

[0055] Non-negative constrained tensor factorization (NTF) employs the PARAFAC model (Parallel Factor Analysis). The three-dimensional tensor 3D-DEST is decomposed into three factor matrices (combustion factor matrix, emission factor matrix, and component factor matrix) and a core tensor. The decomposition rank R is set to 3 (empirical value), which extracts three potential characteristic patterns. The decomposition process forces all factor elements to be ≥ 0, consistent with the physical meaning of engineering parameters (e.g., concentration and wear loss cannot have negative values). The optimization algorithm uses the alternating least squares (ALS) method, with an upper limit of 100 iterations and a convergence threshold of 1e-6.

[0056] Low-rank feature subspace extraction focuses on the analysis of the core tensor. The core tensor has a size of 3×3×3, and its element values ​​represent the coupling strength between various factors. The Emission Degradation Indicating Subspace (EDIS) is defined as the subspace consisting of the second column of the combustion factor matrix (associated with unstable combustion), the first column of the emission factor matrix (associated with NOx generation), and the third column of the component factor matrix (associated with wear consistency). The energy ratio (EDIS energy ratio, EER) of this subspace is calculated as (the sum of squares of the elements of the subspace core tensor) / (the sum of squares of the elements of the entire core tensor). A significant degradation pattern is considered present when the EER is greater than 0.35.

[0057] The output Dynamic Emission State Feature Vector Set (DESFVS) contains three vectors: Vector 1 (combustion factor vector): The five elements in the second column of the combustion factor matrix represent the projection weights of the unstable combustion mode on each pressure parameter; Vector 2 (emission factor vector): the six elements in the first column of the emission factor matrix, representing the distribution of the NOx dominant mode on the chemical parameters; Vector 3 (component factor vector): The four elements in the third column of the component factor matrix represent the manifestation of wear inconsistency in mechanical parameters.

[0058] For example, an emission factor vector might be [0.9, 0.1, 0.7, 0.3, 0.2, 0.05], indicating that NO concentration (weight 0.9) and NOx generation index (weight 0.7) are the primary contributors to the degradation pattern. This set of vectors serves as input to the Emission Health Map model, transforming raw data into interpretable features.

[0059] Advanced signal processing methods such as variational modal decomposition and harmonic wavelet packet entropy are used to extract characteristic indicators with clear physical meaning from the raw data. These include combustion stability reflected by the inflection point of the pressure rise rate, the degree of piston ring wear characterized by the attenuation of vibration energy in a specific frequency band, and the changing trend of spectral absorption peaks related to the NOx generation path. The raw data is converted into a quantitative feature set that can be directly used for state assessment. Feature engineering is used to reveal the coupling relationship between combustion efficiency, mechanical wear and emission generation hidden behind the data, laying the foundation for establishing an accurate emission health model.

[0060] S203, mapping the feature vectors to map nodes using a pre-built emission health map model based on the dynamic emission state feature vector set, and constructing and updating an emission health state map reflecting the current comprehensive emission health status of the engine in real time; Specifically, we can perform graph attention network matching based on the dynamic emission state feature vector set, map the nearest nodes in the pre-built emission health knowledge graph, and obtain the initial node mapping relationship; Structure and initialization of pre-built emission health knowledge graph The Emission Health Knowledge Graph (EHKG) is a pre-generated structured knowledge base whose nodes represent engine emissions-related entities (such as the combustion chamber, piston ring, three-way catalytic converter, and NOx generation pathway), and whose edges represent causal relationships between entities (e.g., "piston ring wear → cylinder pressure leakage → elevated unburned hydrocarbons"). Each node stores a multidimensional attribute vector, including historical health status baseline values ​​(e.g., the normal cylinder pressure fluctuation range is 0.5-1.2 MPa), fault mode association weights (e.g., the correlation coefficient between piston ring wear and hydrocarbon emissions is 0.85), and physical constraints (e.g., coefficients of the temperature-pressure coupling equation). The graph is constructed by combining an expert knowledge base with a historical fault database. For example, the OBD fault code P0420 (low catalyst efficiency) is mapped to the "catalyst aging" node, which is then connected to the "oxygen sensor failure" and "air-fuel ratio imbalance" nodes by bidirectional edges. During graph initialization, each node's health baseline vector is trained using tens of thousands of sets of normal operating condition data, serving as a reference for subsequent matching.

[0061] The core operation process of graph attention network matching A dynamic emission state feature vector (e.g., a 50-dimensional vector containing combustion stability, piston ring wear, and NOx path characteristics) is input to a graph attention network (GAT). The GAT first calculates the similarity score (SS) between the input feature vector and each node attribute in the EHKG. The equivalent operation is as follows: For each node, a linear transformation is performed on the input and node features using a learnable weight matrix. The attention mechanism is then used to calculate the cosine similarity between the two in the latent space, and the influence weights of adjacent nodes are added. For example, for the "cylinder pressure leakage" node, the GAT simultaneously calculates the strength of its association with the adjacent "piston ring wear" and "valve seal failure" nodes, ultimately outputting a matching confidence (MC) (ranging from 0 to 1). The system sets the MC threshold to 0.7, and a match is considered valid only when the MC is ≥ 0.7. After the matching is completed, the node with the highest confidence score is selected as the nearest neighbor node (NNN). For example, if abnormal cylinder pressure fluctuation is detected (the characteristic value deviates by 30% from the baseline) and is accompanied by an increase in crankcase blowby flow, the GAT may map the input vector to the "increase in piston ring-cylinder liner clearance" node with MC=0.92.

[0062] Generation and verification of initial node mapping relationship The matching results generate an Initial Node Mapping Table (INMT), recording the correspondence between dynamic feature vectors and EHKG nodes (for example, feature vector ID-7 maps to the node "Catalyst Oxygen Storage Capacity Decreased"). To ensure mapping reliability, the system activates a cross-modal validation mechanism. For example, when the GAT maps a feature vector to the "Injector Clogged" node, the validation module compares the soot absorption peak intensity (absorbance > 0.8 in the 2.5-micron band) in the real-time exhaust spectrum with the clogged feature threshold stored in the node. If there is a mismatch, a secondary match is triggered. The final output is a confidence-labeled mapping relationship for subsequent processing.

[0063] According to the initial node mapping relationship, the node state diffusion algorithm is used to calculate the semantic similarity weight between the dynamic emission state feature vector and the node, and the weighted edge connection relationship is obtained; Physical Basis and Algorithm Initialization of Node State Diffusion The Node State Diffusion Algorithm (NSDA) is designed based on the fault propagation characteristics within engine systems. For example, piston ring wear not only directly affects blowby but also indirectly increases NOx production by reducing the compression ratio. Starting from the initial mapping node, the algorithm performs multi-hop diffusion (MHD) along the edges of the EHKG. The diffusion depth is controlled by the preset propagation order (PO), typically set to 3, meaning that the algorithm only computes information up to the third level of neighbors of the target node. Diffusion weights are initialized according to the thermodynamic decay principle: edges directly connected to the fault source are assigned a base weight of 0.9, with a weight decay factor of 0.3 for each additional hop (for example, a second-order edge weight = 0.9 × 0.3 = 0.27). Edge type weights are also considered (e.g., a weight of 1.0 for mechanical wear edges and 0.8 for chemical reaction edges).

[0064] Dynamic calculation of semantic similarity weights For each diffusion path, the semantic similarity weight (SSW) between the dynamic feature vector and the path end node is calculated. This weight is achieved through bidirectional feature alignment: Forward matching: Extract key parameters from the feature vector (such as the cylinder pressure increase rate of 0.15 MPa / degree) and calculate the overlap ratio (Overlap Ratio, OR) with the fault parameter range stored in the node (such as the normal range of 0.18-0.22).

[0065] Backward deduction: Based on the physical constraint equations defined by the node (such as the modified ideal gas law), reverse deduction is made to determine whether the current eigenvector satisfies the node causal relationship (for example, if the monitored exhaust back pressure exceeds the limit by 1.8 bar, the SSW of the "catalyst plugging" node increases by 0.2).

[0066] The final SSW is equivalent to the OR and reverse verification scores according to the formula. For example, the SSW of a certain feature vector and the "oxygen sensor aging" node is 0.6 (OR contribution) + 0.3 (reverse verification) = 0.9.

[0067] Generation of weighted edge connections The algorithm outputs a weighted edge connection matrix (WECM). The rows of the matrix represent the initial mapping nodes, the columns represent the diffusion coverage nodes, and the element values ​​are SSW. For example: The initial node "Injector Carbon Deposits" (A) spreads to the immediately adjacent node "Air-Fuel Ratio Imbalance" (B), with an SSW of 0.85; the second-order node "Decreased Combustion Efficiency" (C), with an SSW of 0.72; and the third-order node "Excessive Nitrogen Oxides" (D), with an SSW of 0.53.

[0068] The matrix also labels the edge types (solid lines indicate strong causal relationships, and dotted lines indicate weak associations), providing a topological basis for subsequent graph updates.

[0069] Based on the weighted edge connection relationship, a temporal graph convolutional network is used to fuse the historical health status sequence and update the key node scores to obtain a real-time health score matrix; Structural Design of Temporal Graph Convolutional Network The Temporal Graph Convolutional Network (TGCN) consists of three layers of processing units: Spatial convolutional layer: Based on the topological relationships in the WECM, it aggregates the health status of adjacent nodes. For example, for the "three-way catalytic converter" node, it aggregates the features of its upstream "air-fuel ratio" node and downstream "tailpipe emissions" node. The aggregation weight is determined by the SSW value in the WECM (for example, the weight of the air-fuel ratio node is 0.7, and the weight of the tailpipe node is 0.6).

[0070] Temporal convolution layer: A one-dimensional convolution kernel (kernel length K = 5, covering the past five operating cycles) is used to scan the node's historical state sequence. For example, for the cylinder pressure stability node, the standard deviation sequence of fluctuations over the last five cycles ([0.08, 0.12, 0.15, 0.18, 0.22]) is extracted and convolved to output trend features.

[0071] Gated fusion unit: The forget gate (FG) and input gate (IG) are used to control the ratio of historical information to current updates. For example, when a sudden change in operating conditions (such as rapid acceleration) is detected, the FG reduces the historical weight to 0.3 and prioritizes current data.

[0072] Key node scoring update mechanism Node scores are calculated based on a multi-factor decay model: Base Health Score (BHS): Derived from the initial EHKG baseline value (e.g. piston ring node BHS = 95 / 100).

[0073] Real-time Attenuation (RTA): Calculated from the characteristic deviation output by TGCN. For example, if the current cylinder wall vibration energy is 20% higher than the baseline, then RTA = 15 minutes.

[0074] Propagation Attenuation Factor (PAF): Calculated based on the weighted SSW of the associated edges in the WECM. For example, the PAF of the "Piston Ring Wear" node increases by 0.2 times due to the associated "Increased Blowby Volume" node (SSW=0.8).

[0075] The final node score update formula is equivalent to: New Score = BHS - RTA × (1 + PAF). For example, a node with BHS = 90, RTA = 10, and PAF = 0.3 → New Score = 90 - 10 × 1.3 = 77.

[0076] Output of real-time health score matrix The updated scores of all nodes form the Real-time Health Score Matrix (RHSM). The matrix is ​​partitioned by engine subsystem: Combustion domain: includes nodes such as cylinder pressure stability (score 83) and combustion efficiency (76); Mechanical domain: piston ring wear (68), valve guide clearance (72), etc. Emission domain: hydrocarbon generation (79), nitrogen oxide pathway (81), etc.

[0077] The matrix is ​​labeled with health levels (e.g., >80 is green, 60-80 is yellow, <60 is red) for display in the visualization interface.

[0078] Based on the real-time health score matrix and the abnormal path backtracking mechanism, the causal chain of combustion domain, component domain and emission domain is reconstructed to obtain an emission health status map that reflects the current comprehensive emission health status of the engine.

[0079] Trigger conditions for abnormal path backtracking The Anomaly Path Backtracking Mechanism (APBM) is triggered in two scenarios: Explicit trigger: When the score of any node in the RHSM is lower than the threshold (e.g., the emission domain node is less than 75 points).

[0080] Implicit trigger: When the difference between cross-domain correlation scores exceeds the limit (for example, the combustion domain score is 85 but the emission domain score is only 70, and the difference is 15> the preset threshold of 10).

[0081] After the trigger, the system takes the low-scoring node as the end point and traces back along the edge in the WECM to the source node of the highest SSW. For example, the backtracking path from "Nitrogen Oxide Exceeding the Standard" (score 65) is: Excessive nitrogen oxides ← Combustion temperature too high (SSW=0.8) ← Air-fuel ratio too lean (SSW=0.9) ← Oxygen sensor drift (SSW=0.95) Multi-domain causal chain reconstruction technology The reconstruction process follows the three-domain coupling rule: From the combustion domain to the component domain: Verify the consistency of thermodynamic parameters (e.g., a peak in-cylinder temperature of 1800°C requires the piston top surface coating to be intact).

[0082] From component domain to emission domain: Check the compatibility of mechanical state and chemical reaction chain (e.g. when piston ring clearance is >0.2 mm, unburned hydrocarbons should be >200 ppm).

[0083] The bidirectional constraint solver eliminates contradictory paths: For example, a backtracking path claims that "injector blockage causes the air-fuel ratio to be too rich", but the exhaust oxygen content in the real-time data (Lambda value = 1.05>1.0) shows that it is actually too lean. This path is marked as illegal.

[0084] Generation and updating of emission health status maps The final output Emission Health State Graph (EHSG) contains three layers: Topology layer: Integrates updated node scores (e.g., reducing the oxygen sensor node score from 82 to 73) with the verified causal chain (highlighted in red the "oxygen sensor drift → air-fuel ratio imbalance → excessive nitrogen oxides" path).

[0085] Spatiotemporal layer: embeds time series trends in node attributes (e.g., the slope of the piston ring wear score is -0.8 / minute over the past 10 minutes).

[0086] Decision-making layer: Attach maintenance recommendations to key nodes (e.g., associate the "oxygen sensor drift" node with the "recommend cleaning oxygen sensor" instruction).

[0087] The graph uses the Incremental Update Protocol (IUP) to ensure real-time performance: a global update is triggered only when the new score changes by more than Δ=5% compared to the previous one, otherwise a local refresh is performed.

[0088] The emission health model built based on knowledge graph technology performs graph attention matching on the extracted feature vectors with pre-established knowledge nodes such as combustion theory, failure mode and emission standards. The health score of each node is dynamically updated through the node state diffusion algorithm to form a visual three-dimensional health status topology map, realizing intelligent mapping from discrete features to system-level health status, intuitively displaying the fault transmission paths between the engine subsystems, providing a structured reasoning framework for subsequent root cause analysis, and significantly improving the diagnostic efficiency of complex emission problems.

[0089] S204: Based on the emission health state map, perform map node state deviation analysis and path anomaly detection processing to identify the root cause components or processes that lead to potential or actual emission violations, and generate an emission anomaly diagnosis report containing specific failure modes; Specifically, the node deviation can be calculated based on the emission health status map, the KL divergence between the key node status and the health benchmark can be quantified, and the node deviation vector can be obtained; Node health benchmark library construction The system pre-stores a baseline model of emissions health profiles for the entire engine lifecycle. This model is generated based on big data training from tens of thousands of engines of the same model. Each profile node (such as "combustion efficiency," "piston ring sealing," and "NOx generation path") is associated with a multi-dimensional feature distribution, including a mean vector (MV), covariance matrix (CM), and probability density function (PDF). For example, the health baseline for the "combustion efficiency" node is defined as: a cylinder pressure waveform peak of 12.5±0.3 megapascals (MPa) and a pressure rise rate inflection point at 8±1 crank angle degrees (CAD) after top dead center. The baseline database is partitioned by engine operating condition (e.g., idle, 2000 revolutions per minute (RPM), and full load) to ensure compatibility with operating conditions.

[0090] Real-time node status probabilistic modeling In the current real-time emission health state map, each node's state is quantified by a set of dynamic emission feature vectors (such as the pressure oscillation amplitude of the combustion stability feature vector and the soot concentration in the exhaust spectrum). The system extracts the corresponding feature dimension for each node and generates a probability distribution of the real-time state using a kernel density estimation (KDE) algorithm. For example, the "piston ring wear" node is characterized by the vibration energy attenuation gradient (in decibels per micrometer, dB / μm) in the 2-5 kHz frequency band. This is converted into a probability distribution curve using a Gaussian kernel function and aligned with the probability distribution of the same dimension in the healthy baseline.

[0091] KL divergence deviation quantification KL divergence (Kullback-Leibler Divergence, KLD) is used as a deviation measurement tool, and the calculation formula is: Here, P(x) is the real-time state probability distribution, and Q(x) is the healthy baseline probability distribution. For example, when the overlap between the real-time urea injection feedback concentration distribution P(x) and the baseline Q(x) for the "NOx Generation Path" node decreases, the KLD value increases. The system calculates KLD values ​​for all key nodes in the graph (preset 56 core nodes) in parallel to generate a Node Deviation Vector (NDV). This vector has the same dimensions as the number of nodes, with each element representing the degree of abnormality (in information bits) for the corresponding node. For example, a KLD greater than 3.0 bits for the piston ring wear node triggers the warning threshold.

[0092] Based on the node deviation vector, the minimum causal path search algorithm is used to locate the core fault propagation chain that causes excessive emissions. Graph Causal Topology Analysis The emissions health state graph is essentially a directed weighted graph, with nodes representing states (e.g., "carbon deposits in cylinder" or "oxygen sensor failure") and edges representing causal relationships (e.g., "piston ring wear → insufficient cylinder pressure → incomplete combustion"). Each edge is assigned a causal strength weight (CSW), derived from historical failure case statistics and an expert rule base. For example, the CSW from "injector blockage" to "excessive hydrocarbon emissions" is 0.92 (range 0-1), while the CSW from "excessive nitrogen oxide emissions" is only 0.15. The system loads the graph's adjacency matrix (AM) and weight matrix (WM) as the topological foundation for path search.

[0093] Minimum Causal Path Search Starting with the node where the KLD in the node deviation vector exceeds the threshold (e.g., "Excessive exhaust particulate matter") as the endpoint, a reverse search is conducted to identify possible fault sources. Using a modified Dijkstra algorithm, the optimization objective is to maximize the total causal strength of the path (higher strength indicates a more likely fault association), while limiting the path length to ≤ 5 hops to prevent excessive tracing. Starting from the endpoint, the algorithm iteratively traverses the incoming neighbor nodes, updates the path's accumulated weight (AW), and records the optimal predecessor node (PN). For example, the search revealed the optimal path for "Excessive particulate matter" to be: Crankcase ventilation valve blockage (CSW = 0.88) → Oil vapor backflow (CSW = 0.95) → Increased carbon deposits in the cylinder (CSW = 0.91) → Abnormal combustion chamber temperature (CSW = 0.89) → Excessive particulate matter.

[0094] Core fault propagation chain generation When multiple paths reach the same source, the one with the highest total path weight (PTW) is selected as the core fault propagation chain (CFPC). Each chain outputs a node sequence with local KLD values ​​and path weights. For example, a CFPC for locating excessive NOx might be: EGR valve sticking (KLD = 4.2 bits) → insufficient exhaust gas recirculation rate (KLD = 3.8 bits) → high cylinder temperature (KLD = 5.1 bits) → thermal NOx surge (KLD = 6.0 bits).

[0095] The system automatically marks critical path nodes and filters low-confidence chains with PTW < 2.5 (empirical threshold).

[0096] Based on the core fault propagation chain, thermodynamic constraint verification is performed to screen out false paths that violate the temperature-pressure-flow coupling equation and obtain the verified fault chain; Engine physics constraint library loading The system's preset thermodynamic constraint library contains three types of essential physical equations, whose parameters are bound to the engine's real-time sensor data: Temperature-pressure coupling equation: based on the actual working fluid state equation Where P is the combustion chamber pressure, in megapascals (MPa); V is the cylinder volume calculated from the piston position, in cubic centimeters (cm³); γ is the specific heat ratio of the working fluid, which is 1.35 for gasoline engines; n is the number of moles of gas in the cylinder, inferred from the intake oxygen concentration; R is the gas constant, 8.314; and T1 is the theoretical temperature, in Kelvin (K). This equation requires that an increase in the real-time cylinder pressure must be accompanied by a simultaneous increase in the theoretical temperature.

[0097] Flow conservation equation: expressed as in = ex + leak ,in, in is the intake air mass flow rate, in grams per second (g / s); ex is the exhaust mass flow rate; leak is the crankcase blowby flow rate. This equation requires that the total amount of gas entering the engine must equal the sum of the exhaust volume and the leakage volume. An error exceeding 5% triggers an abnormality.

[0098] Balanced chemical reaction equation: describing the rate of nitrogen oxide formation .in, is the activation temperature of nitrogen oxides generation, 38,000 K; T2 is the measured temperature in the cylinder; is the percentage of oxygen concentration in the cylinder; k is the fuel characteristic constant. This equation requires that a high-temperature, high-oxygen environment will inevitably lead to an increase in nitrogen oxide concentration.

[0099] For example, a CFPC claims that "turbocharger failure results in insufficient intake pressure, which reduces combustion temperature." However, if the boost pressure (BP) in the real-time data is 1.8 bar and the peak in-cylinder temperature reaches 1800 K, this violates the "low pressure corresponds to low temperature" constraint.

[0100] Real-time data-driven constraint validation For each causal node pair in the core fault propagation chain (e.g., "boost pressure ↓ → combustion temperature ↓"), extract the current operating data for verification: If there are direct measurements between nodes (such as boost pressure sensor values ​​and in-cylinder thermocouple readings), substitute them into the corresponding equations to calculate the theoretical values ​​and compare the residual error (RE) with the measured values.

[0101] If there is no direct data, it can be deduced through associated variables: for example, when verifying "EGR rate ↓→ cylinder temperature ↑", the actual EGR rate is inferred using the exhaust gas oxygen concentration (EGO) sensor, and then the theoretical cylinder temperature is calculated in combination with the intake temperature.

[0102] Set the residual tolerance threshold (e.g., temperature relative error > 15% is considered a violation) and automatically mark node pairs that violate the constraints.

[0103] False path screening and fault chain correction For node pairs that violate thermodynamic constraints, remove the causal edge or the entire path. For example: A certain CFPC path is fuel injector carbon deposits → insufficient fuel injection → reduced cylinder temperature → reduced nitrogen oxides. However, the cylinder temperature actually increased in real-time data (violating the constraint of "insufficient fuel injection should reduce the temperature"), so it is judged as a false path.

[0104] After correction, the effective sub-chain fuel injector carbon deposits are retained → poor atomization → increased soot generation.

[0105] The output of the verified path is a verified fault chain (VFC), and each chain is accompanied by a constraint verification pass flag (Verification Flag, VF=1) and a residual record.

[0106] According to the verified fault chain, the fault mode is decoupled to obtain a fault mode list; Component-level failure mode mapping The system maintains a Fault Mode Knowledge Base (FMKB) that maps graph nodes to specific physical component failure modes: Combustion-related nodes: For example, "combustion oscillation" corresponds to the following fault modes: injector needle valve sticking, spark plug carbon deposits, and cylinder head gasket leakage; Emission-related nodes: For example, "low nitrogen oxide conversion efficiency" corresponds to the following fault modes: SCR catalyst sulfur poisoning, urea nozzle blockage; Mechanical wear nodes: For example, "axial wear of the piston ring" corresponds to the failure mode: insufficient oil cleanliness and air filter failure leading to abrasive particle intrusion.

[0107] For example, the "crankcase blowby flow rate exceeds the standard" node in VFC is decoupled into three possible modes through FMKB: piston ring fracture (probability weight 40%), cylinder liner scratch (35%), and valve guide seal failure (25%).

[0108] Multiple fault chain competition decoupling When multiple VFCs point to the same component (for example, two chains indicate "piston ring wear" and "cylinder liner out-of-round," both leading to excessive blowby), the Dempster-Shafer Theory (DST) algorithm is used to calculate the belief probability (BP) of each failure mode. The input is the path weight (PTW) and node KLD value of each VFC, and the output is a normalized failure probability distribution. For example: Piston ring fracture: BP=62% (supporting evidence: blowby gas flow + vibration high-frequency attenuation); cylinder liner strain: BP=28% (supporting evidence: abnormal temperature gradient + low-frequency vibration); others: BP=10%.

[0109] Failure mode list generation Outputs a Fault Mode List (FML) sorted in descending order by confidence probability. Each record contains: faulty component (e.g., injector in cylinder 3); specific mode (e.g., needle valve dynamic response delay); confidence probability (e.g., BP=0.78); and associated emissions (e.g., excessive hydrocarbons (HC>50ppm)).

[0110] For example: 1. Faulty component: EGR cooler, mode: cooling line blockage, BP=0.85, impact: NOx exceeds the standard; 2. Faulty component: turbocharger, mode: wastegate valve sticking, BP=0.67, impact: particulate matter & CO exceeds the standard.

[0111] Based on the failure mode list, component failure probabilities are fused and combined with real-time data to generate an emission anomaly diagnosis report with confidence intervals.

[0112] Component-level failure probability calculation Based on the confidence probability (BP) in the failure mode list and combined with the real-time component working stress data, the final failure probability (FP) is calculated: Where S is the real-time stress (such as the offset of the piston ring temperature relative to the design value), S_{max} is the maximum allowable stress (such as the temperature limit of the material), and omega_t is the time decay factor (such as the average development rate of the failure mode). For example: The fault mode "injector nozzle carbon deposit" has a BP of 0.75. The real-time injector temperature is 130°C (design upper limit = 150°C), so S / S_{max} is 0.87. The carbon deposit growth model gives omega_t = 0.92 (based on the number of continuous operating hours). The final FP = 0.75 × 0.87 × 0.92 ≈ 0.60.

[0113] Confidence interval construction Considering data uncertainty, confidence intervals (CI) are introduced for key parameters: Measurement error: For example, the cylinder pressure sensor accuracy is ±0.5 MPa, resulting in CI = [3.2, 4.0] bits for the combustion efficiency node KLD; Model error: such as the standard deviation of the thermodynamic constraint verification residual (SD = 0.08); Probability propagation: Use Monte Carlo simulation (MCS) to randomly perturb the input parameters 1000 times and output a 90% confidence interval for FP (e.g., FP = 0.60 ± 0.07).

[0114] Diagnostic report generation Generate a structured emission anomaly diagnostic report, which consists of four parts: Root cause summary: For example, "EGR cooler blockage (FP = 0.85 ± 0.05) caused NOx to exceed the standard of 0.8g / kWh (limit 0.4g / kWh)" Fault propagation path: Visually display the verified fault chain (e.g. EGR rate ↓ → cylinder temperature ↑ → thermal NOx ↑); Maintenance priority matrix: sort by FP and emission exceedance (e.g., immediately address faults with FP > 0.7 and exceedance > 50%); Real-time data attachment: key sensor raw data fragments (e.g., coolant temperature difference ΔT = 2°C corresponding to a clogged EGR valve → healthy baseline ΔT = 10°C); The report is output in PDF / JSON dual formats, supporting vehicle terminal display and cloud synchronization.

[0115] The KL divergence is used to quantify the degree of deviation between key nodes and the baseline state, and the minimum causal path search algorithm is combined to locate the fault propagation chain. False correlations are eliminated through thermodynamic constraint verification. The final output is a detailed diagnostic report that includes the location of the faulty component, failure mechanism description and confidence assessment. This breaks through the limitations of traditional threshold alarms, realizes fault tracing from phenomenon to essence, and accurately identifies the root causes of emission degradation such as injector carbon deposits and EGR valve sticking, providing a direct basis for precise maintenance.

[0116] S205 , based on the emission abnormality diagnosis report and the historical emission health status map sequence, a prediction algorithm based on map evolution is applied to predict the deterioration trend of a specific emission path of the engine, and a targeted preventive engine maintenance strategy is generated.

[0117] Specifically, the graph evolution spatiotemporal tensor can be constructed based on the emission anomaly diagnosis report and the historical emission health status graph sequence, and the node state time slices can be stored to obtain the graph evolution spatiotemporal tensor; After the system generates an emissions anomaly diagnosis report (including a list of fault modes, confidence intervals, and the core fault propagation chain), it must be combined with the historical emission health state graph sequence (graph snapshots stored by timestamp) to build a prediction basis. First, the Graph Evolution Spatiotemporal Tensor (GEST) is a four-dimensional data structure whose dimensions are defined as: Dimension 1: Graph node (Node, N). Represents entities in the emission health graph (such as "piston ring wear", "three-way catalytic converter efficiency", and "NOx generation path").

[0118] Dimension 2: Node state feature (Feature, F). Stores the dynamic state vector of each node (such as wear score 0-100, catalytic efficiency percentage, emission concentration ppm).

[0119] Dimension 3: Spatial Hierarchy (Hierarchy, H). Layered by engine subsystems (combustion domain, exhaust domain, and after-treatment domain).

[0120] Dimension 4: Time series (Time, T). Snapshots of historical graphs collected at regular intervals (e.g., every 10 minutes).

[0121] The construction process begins with Node State Time Slice (NSTS) extraction: the state vector of each node is extracted from the historical graph sequence in chronological order (for example, the characteristic value of the node "piston ring wear" at time T1 is [wear score = 72, vibration energy entropy = 0.85]), and the time drift caused by sensor delay is eliminated through the Temporal Alignment Module (TAM).

[0122] To improve tensor completeness, key information from the emission anomaly diagnosis report needs to be integrated: Fault mode injection: Encode the fault mode in the diagnostic report (such as "piston ring sticking - moderate") as additional node features (such as adding "fault flag = 1" and "confidence = 0.92").

[0123] Causal chain enhancement: Based on the core fault propagation chain in the report (for example, "cylinder liner wear → increased blowby → incomplete combustion → excessive CO"), a virtual "causal edge weight" is added to the tensor, allowing subsequent prediction algorithms to focus on the critical path.

[0124] Temporal and spatial correlation filling: Bidirectional long short-term memory (Bi-LSTM) networks are used to complete missing time slices (such as sensor failure periods). For example, health graph data from one hour before and after can be used to predict the node status at the missing time.

[0125] The final generated GEST tensor example: Dimensions: N × F × H × T = 50 nodes × 8-dimensional features × 3 spatial levels × 240 time slices (48 hours of data).

[0126] Stored in a distributed time series database (such as InfluxDB), it supports millisecond-level query responses and provides structured input for subsequent predictions.

[0127] During the project implementation, GEST construction needs to solve two major challenges: Real-time performance: Data slicing and alignment are completed on the vehicle side through the Edge Computing Node (ECN), and only compressed tensors are uploaded to the cloud, reducing bandwidth requirements.

[0128] Scalability Design: Hierarchical Tensor Chunking (HTC) is used to partition large tensors into blocks by spatial hierarchy (e.g., independent storage of combustion domain sub-tensors), avoiding single-point resource bottlenecks. The output of this step is a GEST tensor with complete spatiotemporal correlation, which is the core input for predicting evolution.

[0129] According to the spatiotemporal tensor of the graph evolution, the spatiotemporal graph neural network is used to predict the state trajectory of key nodes in the future preset period to obtain the predicted state trajectory; The Spatio-Temporal Graph Neural Network (STGNN) is a prediction engine whose structure is customized for the GEST tensor: Spatial dependency modeling: Using the Graph Convolution Module (GCM), we aggregate features of adjacent nodes based on the connections between graph nodes (e.g., the "piston ring wear" node is strongly connected to the "cylinder pressure fluctuation" node). For example, we approximate the graph convolution kernel using Chebyshev polynomials to calculate the neighborhood weighted average of node features.

[0130] Temporal dependency modeling: Using a Gated Temporal Convolution Module (GTCM), a dilated causal convolution (DCC) is used to capture long-term patterns (such as the gradual deterioration of wear), and a gating mechanism (similar to a GRU) is used to filter noise.

[0131] Spatiotemporal coupling: The Attention Fusion Layer (AFL) dynamically adjusts spatiotemporal weights. For example, when a sudden emission change is detected, the attention weight in the temporal dimension is increased to respond quickly.

[0132] The forecasting process is performed in three stages: Key Node Screening: Based on the emission anomaly diagnosis report, identify high-impact nodes (HIN) that require prediction. For example, if the report indicates that "NOx excess is caused by EGR valve failure," EGR valve-related nodes (such as "EGR opening deviation" and "cooling efficiency") are marked as HIN.

[0133] Multi-step rolling forecast: Taking the GEST tensor as input, STGNN performs iterative forecasting according to a preset period (such as the next 24 hours): Input: GEST tensor for the past 48 hours.

[0134] Output: Node state at time t+1 in the future.

[0135] The prediction result at time t+1 is fed back to the input, and the prediction at time t+2 is continued to form a rolling closed loop.

[0136] Uncertainty Quantification: Monte Carlo Dropout (MCD) is used to generate probabilistic prediction intervals. For example, the prediction for "Piston Ring Wear Score" after 24 hours is [85, 92] (95% confidence interval).

[0137] The predicted state trajectory (PST) contains two core types of information: Node state sequence: characteristic values ​​of key nodes at future time points (e.g., predicted values ​​of "three-way catalytic converter efficiency" every 2 hours: [78%, 76%, 74%, ...]).

[0138] Pathway Deterioration Index: Calculates the degradation rate (slope) of a specific emission pathway (e.g., the "Unburned Hydrocarbon Generation Pathway"). For example, based on the predicted values ​​of the "Air-Fuel Ratio Fluctuation" and "Tailpipe HC Concentration" nodes, a path deterioration coefficient β of 0.15 (unit: ppm / hour) is fitted.

[0139] The prediction engine is deployed in a cloud-edge collaborative architecture: the STGNN lightweight model runs on the on-board ECU, and complex retraining is performed in the cloud to ensure real-time performance.

[0140] Based on the predicted state trajectory, a three-objective Pareto optimization process is performed to find the optimal balance point among emission risk, maintenance cost, and downtime, and to obtain the optimal intervention plan. Three-Objective Pareto Optimization (TOPO) requires balancing the following conflicting objectives: Objective 1: Emission Risk (ER). Quantify the probability of exceeding the standard during the predicted trajectory (e.g., the probability of NOx concentration exceeding the National VI limit of 50 mg / km).

[0141] Objective 2: Maintenance Cost (MC). This includes parts costs (e.g., piston ring replacement quote: ￥1200) and labor costs (￥300 / hour).

[0142] Objective 3: Downtime Duration (DD). Operational losses caused by vehicle downtime (e.g., a logistics vehicle loses ￥200 per hour).

[0143] The optimization problem is defined as: minimizing [ER, MC, DD]; the constraint condition is the technical feasibility of the maintenance action (DD ≥ 4 hours if the cylinder block needs to be disassembled).

[0144] The solution adopts the improved NSGA-III algorithm (non-dominated sorting genetic algorithm with reference points): Gene encoding: Encode the maintenance strategy as a binary gene string. For example, the gene bit definition is: Bit 1: whether to replace the piston ring (0 / 1); Bit 2: whether to clean the EGR valve (0 / 1); Bit 3: whether to calibrate the oxygen sensor (0 / 1).

[0145] Fitness evaluation: ER calculation: If the gene contains "replace piston ring," the wear node prediction value is updated based on the PST, and the probability of exceeding emission standards is recalculated.

[0146] MC calculation: Add up the parts and labor costs for the selected action.

[0147] DD calculation: Take the maximum value when parallel maintenance actions can overlap (for example, replacing a piston ring takes 3 hours and cleaning an EGR valve takes 1 hour, and if they are performed in parallel, DD = 3 hours).

[0148] Balance point screening: Generates a Pareto front in the three-dimensional target space and selects the most balanced solution using the entropy-weighted TOPSIS decision technique (Technique for Order Preference by Similarity to Ideal Solution). For example, the solution with the shortest Euclidean distance to the ideal solution (ER=0, MC=0, DD=0) is screened.

[0149] The Optimal Intervention Scheme (OIS) is output as structured instructions: Action combination: such as "Replace piston ring (priority 1) + calibrate oxygen sensor (priority 2)".

[0150] Execution timing: Set a time window based on the deterioration inflection point of the predicted trajectory (e.g., "piston ring replacement must be performed within 48 hours").

[0151] Economic indicators: The total cost of the solution is ￥1500, the shutdown time is 2 hours, and the emission risk is reduced to 5%.

[0152] According to the optimal intervention plan, the vehicle operation plan is integrated and combined with the database to output a preventive maintenance instruction set.

[0153] The Vehicle Operation Plan (VOP) is obtained from the fleet management system and includes: Itinerary: Transportation route for the next week (e.g., "Shanghai → Beijing, mileage 1200km").

[0154] Load spectrum: cargo weight distribution (e.g. full load of 20 tons on the outbound trip and empty on the return trip).

[0155] Working condition prediction: engine load rate based on road conditions (e.g. 85% load rate on mountainous roads).

[0156] The fusion logic is to embed maintenance actions from the OIS into the clearance window of the VOP, while also considering the impact of operating conditions on maintenance urgency. For example, if the forecast indicates that "piston ring wear worsens faster on plateau roads," and the VOP includes transportation on the Qinghai-Tibet Highway, the priority of this action will be automatically increased.

[0157] The process of generating a Preventive Maintenance Instruction Set (PMIS) is as follows: Resource matching: Query the maintenance knowledge graph database: Action "Replace piston ring" → Associated with the required tool (cylinder liner remover) → Match the inventory of nearby repair shops.

[0158] Dynamic Scheduling: Using Constraint Programming (CP): Input: VOP time window, maintenance station working time pool, spare parts inventory; Output: optimal maintenance appointment time (such as "June 5, 10:00, Suzhou Maintenance Station").

[0159] Instruction hierarchical generation: The driver's terminal pushes the message "Please schedule maintenance within 48 hours, code P0300"; the maintenance station system sends the message "Work Order #202406001: Replace piston ring, 2-hour workstation reservation required"; the supply chain system triggers the message "Piston ring spare parts transfer, destination Suzhou Station."

[0160] The adaptive mechanism of the instruction set ensures robustness: Real-time feedback closed loop: If the repair station responds "no piston rings in stock", the system automatically triggers alternative solutions: Transfer from other sites (increase cost ￥200, delay 1 day); downgrade to "piston ring cleaning + adding oil additives" (temporary solution, valid for 72 hours).

[0161] Multi-objective rebalancing: When alternative plans change the MC and DD, TOPO is re-run to fine-tune the OIS. The final PMIS output includes: a maintenance action list (including spare part models and work hours); a spatiotemporal scheduling plan (location, time, and personnel); an emergency downgrade plan (when the primary plan is unfeasible); and an economic and environmental benefit assessment (e.g., "projected reduction in NOx emissions by 1.2 kg, avoiding a ¥5,000 fine").

[0162] Through the spatiotemporal graph neural network, the historical health status evolution law is learned, the change trajectory of key parameters in the future is predicted, and the Pareto optimization is combined to seek a balance between multiple objectives such as emission risk, maintenance cost and downtime. The personalized strategy containing the optimal intervention time, maintenance content and expected effect is output, realizing the transition from passive maintenance to active prevention. Through predictive maintenance, sudden emission exceeding standard events are avoided, and at the same time, the maintenance resource allocation is optimized, which significantly reduces the operation and maintenance cost of the whole life cycle.

[0163] It can be seen that according to the operating status of the automobile internal combustion engine, a multi-dimensional fusion sensor data set is synchronously collected in real time; for the multi-dimensional fusion sensor data set, a dynamic emission state feature vector set is generated to characterize the current combustion efficiency, wear status of key components and emission generation path; based on the dynamic emission state feature vector set, an emission health state map reflecting the current comprehensive emission health status of the engine is constructed and updated in real time; based on the emission health state map, an emission abnormality diagnosis report containing specific fault modes is generated; based on the emission abnormality diagnosis report and the historical emission health state map sequence, the deterioration trend of the engine's specific emission path is predicted, and a targeted preventive engine maintenance strategy is generated, so as to achieve real-time and accurate diagnosis and trend prediction of the engine's emission health status, and improve the initiative of emission abnormality handling and maintenance efficiency.

[0164] Another embodiment of the present invention provides an emission monitoring system for an automobile internal combustion engine. Figure 3 , the system may include: Acquisition module 301, for synchronously acquiring, in real time and based on the operating state of the automobile internal combustion engine, a multi-dimensional fusion sensor data set including exhaust gas component spectral data, combustion chamber pressure fluctuation waveform, in-cylinder temperature gradient, crankcase blowby flow rate, and exhaust back pressure; Extraction module 302, configured to perform spatiotemporal alignment and feature extraction on the multi-dimensional fusion sensor data set to generate a dynamic emission state feature vector set representing current combustion efficiency, wear state of key components, and emission generation path; A construction module 303 is configured to map the dynamic emission state feature vector set to a map node using a pre-built emission health map model, thereby constructing and updating an emission health map reflecting the current comprehensive emission health status of the engine in real time; Detection module 304 is configured to perform state deviation analysis of nodes in the emission health state map and detect path anomalies based on the map, identify the root cause components or processes that lead to potential or actual emission violations, and generate an emission anomaly diagnosis report containing specific failure modes; The prediction module 305 is used to apply a prediction algorithm based on the evolution of the map according to the emission abnormality diagnosis report and the historical emission health status map sequence to predict the deterioration trend of the specific emission path of the engine and generate a targeted preventive engine maintenance strategy.

[0165] An embodiment of the present invention further provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps of any one of the above method embodiments when running.

[0166] Specifically, in this embodiment, the above-mentioned storage medium may be configured to store a computer program for performing the following steps: S201, based on the operating state of the automobile internal combustion engine, real-time synchronous acquisition of a multi-dimensional fusion sensor data set including exhaust gas component spectral data, combustion chamber pressure fluctuation waveform, cylinder temperature gradient, crankcase blowby flow rate, and exhaust back pressure; S202, performing spatiotemporal alignment and feature extraction on the multi-dimensional fusion sensor data set to generate a dynamic emission state feature vector set representing current combustion efficiency, wear status of key components, and emission generation path; S203, mapping the feature vectors to map nodes using a pre-built emission health map model based on the dynamic emission state feature vector set, and constructing and updating an emission health state map reflecting the current comprehensive emission health status of the engine in real time; S204: Based on the emission health state map, perform map node state deviation analysis and path anomaly detection processing to identify the root cause components or processes that lead to potential or actual emission violations, and generate an emission anomaly diagnosis report containing specific failure modes; S205 , based on the emission abnormality diagnosis report and the historical emission health status map sequence, a prediction algorithm based on map evolution is applied to predict the deterioration trend of a specific emission path of the engine, and a targeted preventive engine maintenance strategy is generated.

[0167] An embodiment of the present invention further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any one of the above method embodiments.

[0168] Specifically, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0169] Specifically, in this embodiment, the processor may be configured to execute the following steps through a computer program: S201, based on the operating state of the automobile internal combustion engine, real-time synchronous acquisition of a multi-dimensional fusion sensor data set including exhaust gas component spectral data, combustion chamber pressure fluctuation waveform, cylinder temperature gradient, crankcase blowby flow rate, and exhaust back pressure; S202, performing spatiotemporal alignment and feature extraction on the multi-dimensional fusion sensor data set to generate a dynamic emission state feature vector set representing current combustion efficiency, wear status of key components, and emission generation path; S203, mapping the feature vectors to map nodes using a pre-built emission health map model based on the dynamic emission state feature vector set, and constructing and updating an emission health state map reflecting the current comprehensive emission health status of the engine in real time; S204: Based on the emission health state map, perform map node state deviation analysis and path anomaly detection processing to identify the root cause components or processes that lead to potential or actual emission violations, and generate an emission anomaly diagnosis report containing specific failure modes; S205 , based on the emission abnormality diagnosis report and the historical emission health status map sequence, a prediction algorithm based on map evolution is applied to predict the deterioration trend of a specific emission path of the engine, and a targeted preventive engine maintenance strategy is generated.

[0170] The above describes in detail the structure, features and effects of the present invention based on the embodiments shown in the drawings. The above is only a preferred embodiment of the present invention, but the scope of implementation of the present invention is not limited to what is shown in the drawings. Any changes made in accordance with the concept of the present invention, or modifications to equivalent embodiments with equivalent changes, which do not exceed the spirit covered by the description and drawings, should be within the scope of protection of the present invention.

Claims

1. A method for monitoring emissions of an automobile internal combustion engine, characterized in that: The method comprises: Based on the operating status of the vehicle's internal combustion engine, a multi-dimensional fusion sensor data set including exhaust gas composition spectral data, combustion chamber pressure fluctuation waveform, in-cylinder temperature gradient, crankcase blowby flow rate, and exhaust back pressure is collected in real time and synchronously; Performing spatiotemporal alignment and feature extraction on the multi-dimensional fusion sensor data set to generate a dynamic emission state feature vector set representing current combustion efficiency, wear status of key components, and emission generation path; Based on the dynamic emission status feature vector set, a pre-built emission health map model is used to map the feature vectors to map nodes, and an emission health state map reflecting the current comprehensive emission health status of the engine is constructed and updated in real time; Based on the emission health state map, perform map node state deviation analysis and path anomaly detection processing to identify the root cause components or processes that lead to potential or actual emission violations, and generate an emission anomaly diagnosis report containing specific failure modes; According to the emission abnormality diagnosis report and the historical emission health status map sequence, a prediction algorithm based on map evolution is applied to predict the deterioration trend of the engine's specific emission path and generate a targeted preventive engine maintenance strategy.

2. The method according to claim 1, characterized in that The multi-dimensional fusion sensor data set including exhaust gas component spectral data, combustion chamber pressure fluctuation waveform, cylinder temperature gradient, crankcase blowby flow rate and exhaust back pressure is collected in real time and synchronously according to the operating state of the automobile internal combustion engine, including: Based on the phase-synchronized pulse signal output by the crankshaft position sensor, the sampling clocks of the cylinder pressure sensor, exhaust gas spectrometer, temperature array, blowby gas flowmeter, and back pressure sensor are synchronized to obtain the original signal stream with time base alignment. Based on the original signal stream aligned with the time base, the combustion cycle is segmented, and the compression stroke and power stroke data slices are divided according to the crankshaft angle to obtain the combustion stage grouped signal set; According to the grouped signal sets of the combustion stage, the engine geometric space mapping is performed. Based on the piston kinematics, the cylinder wall vibration signal is mapped to the position of the piston ring-cylinder liner friction pair to obtain the spatial registration vibration feature matrix. According to the spatial registration vibration feature matrix, multi-source noise cancellation processing is performed, exhaust spectrum, temperature and flow data are fused, and ignition electromagnetic interference is suppressed to obtain a multi-dimensional fusion sensing data set synchronized in time and space.

3. The method according to claim 2, characterized in that The method of performing spatiotemporal alignment and feature extraction on the multi-dimensional fusion sensor data set to generate a dynamic emission state feature vector set representing the current combustion efficiency, wear state of key components, and emission generation path includes: Based on the multi-dimensional fusion sensor data set, the cylinder pressure waveform variational mode decomposition is performed to extract the pressure rise rate inflection point and the indicated mean effective pressure energy distribution characteristics, and the combustion stability feature vector is obtained; Based on the exhaust spectral data, the chemical bond absorption peaks are correlated, and the emission concentration ratio and transformation path are analyzed to obtain the pollutant generation path characteristic vector; Based on the spatial registration vibration characteristic matrix, the harmonic wavelet packet entropy algorithm is used to quantify the energy attenuation gradient in the 2-5kHz frequency band to obtain the piston ring wear state characteristic vector; The combustion stability feature vector, pollutant generation path feature vector and piston ring wear state feature vector are tensor-fused to construct a three-dimensional dynamic emission state tensor. According to the three-dimensional dynamic emission state tensor, non-negative constrained tensor decomposition is performed to extract the low-rank feature subspace that characterizes emission degradation and obtain the dynamic emission state feature vector set.

4. The method according to claim 3, characterized in that The method includes mapping the feature vectors to map nodes using a pre-built emission health map model based on the dynamic emission state feature vector set, and constructing and updating an emission health state map that reflects the current comprehensive emission health status of the engine in real time, including: Based on the dynamic emission state feature vector set, graph attention network matching is performed to map the nearest nodes in the pre-built emission health knowledge graph to obtain the initial node mapping relationship; According to the initial node mapping relationship, the node state diffusion algorithm is used to calculate the semantic similarity weight between the dynamic emission state feature vector and the node, and the weighted edge connection relationship is obtained; Based on the weighted edge connection relationship, a temporal graph convolutional network is used to fuse the historical health status sequence and update the key node scores to obtain a real-time health score matrix; Based on the real-time health score matrix and the abnormal path backtracking mechanism, the causal chain of combustion domain, component domain and emission domain is reconstructed to obtain an emission health status map that reflects the current comprehensive emission health status of the engine.

5. The method according to claim 4, characterized in that Based on the emission health state map, the map node state deviation analysis and path anomaly detection processing are performed to identify the root cause components or processes that lead to potential or actual emission violations, and generate an emission anomaly diagnosis report containing specific failure modes, including: According to the emission health status map, the node deviation is calculated to quantify the KL divergence between the key node status and the health benchmark, and the node deviation vector is obtained; Based on the node deviation vector, the minimum causal path search algorithm is used to locate the core fault propagation chain that causes excessive emissions. Based on the core fault propagation chain, thermodynamic constraint verification is performed to screen out false paths that violate the temperature-pressure-flow coupling equation and obtain the verified fault chain; According to the verified fault chain, the fault mode is decoupled to obtain a fault mode list; Based on the failure mode list, component failure probabilities are fused and combined with real-time data to generate an emission anomaly diagnosis report with confidence intervals.

6. The method according to claim 5, characterized in that The method uses a prediction algorithm based on the evolution of the map based on the emission abnormality diagnosis report and the historical emission health status map sequence to predict the deterioration trend of the engine's specific emission path and generate a targeted preventive engine maintenance strategy, including: Based on the emission anomaly diagnosis report and the historical emission health status graph sequence, the graph evolution spatiotemporal tensor is constructed, and the node state time slices are stored to obtain the graph evolution spatiotemporal tensor; According to the spatiotemporal tensor of the graph evolution, the spatiotemporal graph neural network is used to predict the state trajectory of key nodes in the future preset period to obtain the predicted state trajectory; Based on the predicted state trajectory, a three-objective Pareto optimization process is performed to find the optimal balance point among emission risk, maintenance cost, and downtime, and to obtain the optimal intervention plan. According to the optimal intervention plan, the vehicle operation plan is integrated and combined with the database to output a preventive maintenance instruction set.

7. An emission monitoring system for an automobile internal combustion engine, characterized in that: The system comprises: The acquisition module is used to synchronously collect multi-dimensional fusion sensor data sets including exhaust gas component spectral data, combustion chamber pressure fluctuation waveform, cylinder temperature gradient, crankcase blowby flow rate and exhaust back pressure in real time according to the operating status of the automobile internal combustion engine; an extraction module for performing spatiotemporal alignment and feature extraction on the multi-dimensional fusion sensor data set to generate a dynamic emission state feature vector set representing current combustion efficiency, wear status of key components, and emission generation path; A construction module is used to map the feature vectors to map nodes based on the dynamic emission state feature vector set using a pre-built emission health map model, and to construct and update the emission health state map reflecting the current comprehensive emission health status of the engine in real time; a detection module configured to perform, based on the emission health state map, state deviation analysis of map nodes and path anomaly detection processing, identify the root cause components or processes that lead to potential or actual emission violations, and generate an emission anomaly diagnosis report containing specific failure modes; The prediction module is used to apply a prediction algorithm based on the evolution of the map based on the emission abnormality diagnosis report and the historical emission health status map sequence to predict the deterioration trend of the specific emission path of the engine and generate a targeted preventive engine maintenance strategy.

8. The system according to claim 7, characterized in that The acquisition module is specifically used to: Based on the phase-synchronized pulse signal output by the crankshaft position sensor, the sampling clocks of the cylinder pressure sensor, exhaust gas spectrometer, temperature array, blowby gas flowmeter, and back pressure sensor are synchronized to obtain the original signal stream with time base alignment. Based on the original signal stream aligned with the time base, the combustion cycle is segmented, and the compression stroke and power stroke data slices are divided according to the crankshaft angle to obtain the combustion stage grouped signal set; According to the grouped signal sets of the combustion stage, the engine geometric space mapping is performed. Based on the piston kinematics, the cylinder wall vibration signal is mapped to the position of the piston ring-cylinder liner friction pair to obtain the spatial registration vibration feature matrix. According to the spatial registration vibration feature matrix, multi-source noise cancellation processing is performed, exhaust spectrum, temperature and flow data are fused, and ignition electromagnetic interference is suppressed to obtain a multi-dimensional fusion sensing data set synchronized in time and space.

9. A storage medium, characterized in that: The storage medium stores a computer program, wherein the computer program is configured to execute the method according to any one of claims 1 to 6 when executed.

10. An electronic device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to perform the method according to any one of claims 1 to 6.

Images