A performance testing system and method for automotive hybrid engines

Abstract

This invention discloses a performance testing system and method for hybrid vehicle engines, belonging to the field of automotive powertrain testing technology. It includes: a data acquisition module, a data preprocessing module, a deep feature extraction module, a model training module, a real-time performance testing module, and a result output and early warning module. The data acquisition module is used to collect historical and real-time operating data of the hybrid engine; the historical operating dataset is labeled with corresponding performance tags. Through a three-level progressive extraction architecture of "single-domain basic features – inter-domain spatiotemporal correlation features – supervised learning to refine features," combined with masked autoencoders to filter redundant noise, supervised contrastive learning to enhance performance discriminative power, and gated linear units to complete precise feature selection, it achieves layer-by-layer feature refinement and high-order latent feature mining.

Description

Technical Field

[0001] This invention relates to the field of automotive powertrain testing technology, and in particular to a performance testing system and method for automotive hybrid engines. Background Technology

[0002] With the widespread adoption of hybrid vehicle technology, hybrid-specific engines, as the core power source of hybrid systems, have fundamentally different operating characteristics from traditional fuel engines: hybrid engines need to be deeply coupled with drive motors, power batteries, and vehicle controllers, and frequently switch between pure electric, series hybrid, parallel hybrid, direct drive, and energy recovery operating conditions during operation. They have core characteristics such as strong coupling of the five physical domains of heat-engine-fuel-electricity-vehicle, large dynamic range of operating parameters, complex transient characteristics of operating condition switching, and severe fluctuations in heat load caused by intermittent operation.

[0003] Existing engine performance testing technologies have the following core technical shortcomings that make them unsuitable for hybrid engine characteristics: Multi-domain coupling interference cannot be resolved, and the purity of basic feature extraction is insufficient: Existing technologies mostly directly extract overall features from all-dimensional parameters without considering the characteristic differences and coupling interference of the five physical domain parameters of the hybrid engine. This results in the extracted features containing a large amount of cross-domain redundant noise, making it impossible to accurately capture early abnormal features in a single domain, and easily leading to false alarms and missed detections. The lack of cross-domain dynamic correlation feature capture capability and extremely poor adaptability to operating conditions: The performance abnormalities of hybrid engines are mostly manifested as abnormalities in the coupling relationship between multiple domains, rather than the threshold exceeding of a single parameter. Existing technologies can only extract single-parameter time-series features or fixed parameter correlation features, and cannot capture the dynamic changes in inter-domain correlations during operating condition switching. They can only achieve detection under steady-state operating conditions, and the detection accuracy for transient operating conditions where hybrid engines account for more than 60% is less than 50%. The feature extraction level is too simple, and the ability to identify early performance degradation is seriously insufficient: Existing deep learning-based detection schemes mostly use a single convolutional or recurrent neural network structure, which cannot achieve progressive feature extraction and purification, cannot mine hidden high-order features that are strongly related to engine performance, and have an accuracy rate of less than 30% in identifying early engine performance degradation, and cannot achieve early warning of faults. The generalization and real-time performance of the model cannot be balanced: existing high-precision detection models are large in size and require high computing power, making them unsuitable for deployment in vehicle-mounted embedded environments; while lightweight models have insufficient feature extraction capabilities and poor generalization, making them unsuitable for hybrid engines of different models and driving mileages, requiring extensive recalibration and posing significant challenges for engineering applications. The detection dimensions are limited and cannot meet the needs of full life cycle health management: the existing solution can only determine whether a fault exists, and cannot simultaneously complete multi-task detection such as performance level assessment, quantitative index regression, anomaly type location, and degradation trend prediction, and cannot support the full life cycle health management and predictive maintenance of hybrid engines.

[0004] In response to the aforementioned core deficiencies of existing technologies, there is an urgent need to develop a hybrid engine performance testing solution that can adapt to the characteristics of multi-domain strong coupling and frequent switching of multiple operating conditions of hybrid engines, with sufficient feature extraction, high detection accuracy, and a balance between generalization and real-time performance. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide a performance testing system and method for hybrid engines in automobiles; it can solve the problems of large multi-domain coupling interference, insufficient ability to capture cross-domain dynamic correlation features, weak early performance degradation identification ability, and poor adaptability to operating conditions in the existing technology.

[0006] Technical solution: To solve the above-mentioned technical problems, according to one aspect of the present invention, more specifically, a performance testing system and method for a hybrid vehicle engine, comprising: a data acquisition module, a data preprocessing module, a deep feature extraction module, a model training module, a real-time performance testing module, and a result output and early warning module; Data acquisition module: used to collect historical and real-time operating data of hybrid engines. The historical operating data is labeled with corresponding performance tags. The data preprocessing module is used to perform data cleaning, missing value imputation, outlier removal, data standardization, time-series synchronization alignment, and working condition segmentation labeling on the collected historical and real-time running datasets in sequence, so as to obtain the preprocessed historical datasets and preprocessed real-time data. The deep feature extraction module is used to extract deep features from the preprocessed input data. The first-level features are extracted based on the preprocessed input data through a parallel branch structure with multi-domain divide-and-conquer. The second-level features are extracted based on the first-level feature set through a dual-branch spatiotemporal graph convolutional network. The third-level features are extracted based on the second-level features through a masked autoencoder combined with supervised contrastive learning and gated linear units. Model training module: used to train a multi-task performance detection model for hybrid engines based on the extracted deep features and corresponding performance labels; Real-time performance detection module: This module is used to input the three-level fusion deep features corresponding to real-time data into the trained performance detection model and output the real-time performance detection results of the hybrid engine. The results output and early warning module is used to visualize the performance test results and trigger graded early warnings when abnormal performance or excessive attenuation is detected.

[0007] Furthermore, the data acquisition module is divided into a cloud-based historical data acquisition unit and an on-board real-time data acquisition unit, and the acquisition range covers the full-dimensional operating parameters of the five physical domains of the hybrid engine: thermal management domain, mechanical domain, fuel combustion domain, electric drive coupling domain, and vehicle operating condition domain. Thermal management parameters include, but are not limited to, coolant inlet and outlet temperatures, cylinder block temperature, cylinder head temperature, oil cooler inlet and outlet temperatures, EGR cooler temperature, and thermostat opening. Mechanical parameters include, but are not limited to, engine speed, output torque, cylinder pressure, valve lift, crankshaft position, oil pressure, oil temperature, and crankcase pressure. Fuel combustion parameters include, but are not limited to, injection quantity, injection advance angle, injection pulse width, intake pressure, intake temperature, throttle opening, exhaust temperature, front and rear oxygen sensor voltages, air-fuel ratio, and EGR valve opening. Electric drive coupling domain parameters include, but are not limited to, drive motor speed, torque, output power, charging and discharging current, charging and discharging voltage, stator winding temperature, rotor temperature, power battery SOC, power battery SOH, cell maximum / minimum temperature, and charge / discharge rate; The vehicle operating condition parameters include, but are not limited to, the current drive mode, vehicle speed, throttle opening, brake opening, transmission gear, vehicle acceleration, ambient temperature, atmospheric pressure, and altitude output by the VCU.

[0008] Furthermore, the data preprocessing module includes the following processing steps: Data cleaning: Remove duplicate data, empty data segments caused by communication interruptions, and invalid data with fixed values ​​caused by sensor failures; Missing value imputation: For a small number of missing values ​​with consecutive missing sampling points ≤ the preset value, linear interpolation is used to imput them; for data segments with consecutive missing sampling points exceeding the preset value, they are directly discarded to avoid interference from invalid data. Outlier removal: A two-layer removal mechanism combining the 3σ criterion and the isolated forest algorithm is adopted. First, extreme outliers exceeding 3 times the standard deviation are removed by the 3σ criterion. Then, the isolated forest algorithm is used to identify and remove hidden outliers during the operation condition switching process. Data standardization: The Z-Score standardization method is adopted, which maps all parameter values ​​to a standard normal distribution with a mean of 0 and a variance of 1 based on the mean and standard deviation of the historical dataset, thus completely eliminating the dimensional differences of different parameters. Timing synchronization alignment: Based on the system clock of the engine ECU, asynchronous data from different controllers and sensors are resampled and timestamped, and a unified resampling frequency is set; Operating condition segmentation and labeling: Based on the drive mode signal output by the VCU, combined with vehicle speed, throttle opening, motor power, and engine start / stop status, the operating data is divided into operating condition segments including pure electric operating condition, series hybrid operating condition, parallel hybrid operating condition, engine direct drive operating condition, and energy recovery operating condition. Each operating condition segment is labeled with a corresponding operating condition label for operating condition adaptation during feature extraction.

[0009] Furthermore, the extraction method for the first-level features is as follows: Each of the five physical domains is set up as a parallel and independent extraction branch, with each branch corresponding to the preprocessed data of one physical domain. There is no parameter sharing or cross interference between the branches, and they run completely independently. To address the characteristic of early abnormalities in hybrid engines manifested as transient nonlinear signal changes, the db4 mother wavelet was used to decompose the time-series data in the corresponding domain into 5-layer wavelet packets to obtain wavelet coefficients. Statistical features, including coefficient energy, entropy, standard deviation, skewness, and kurtosis for each frequency band, were extracted, and finally, the transient features in the frequency domain were output. To address the signal scale inconsistency caused by frequent switching of hybrid engine operating conditions, a three-layer cascaded deformable one-dimensional convolutional network is set up. Each convolutional layer is followed by a batch normalization layer, a LeakyReLU activation layer, and a max pooling layer. Through the adaptive receptive field of deformable convolution, the long-term trend and dynamic change characteristics of time-series data in the domain under different operating conditions are accurately captured, and finally the dynamic features of the time domain in the domain are output. Each branch concatenates the frequency-domain transient features and the time-domain dynamic features within the domain to obtain the basic features within the domain; the single-domain features of each branch are concatenated in a fixed order to obtain the first-level features.

[0010] Furthermore, the extraction process for the second-level features is as follows: Using the five physical domains corresponding to the first-level features as graph nodes, an undirected weighted graph is constructed corresponding to five nodes: mechanical domain, fuel combustion domain, thermal management domain, electric drive coupling domain, and vehicle operating condition domain. The initial value of the adjacency matrix is ​​obtained by calculating the Pearson correlation coefficient between each pair of domains using historical datasets. The value range is [0, 1], representing the initial correlation strength between the two domains. During training, it is adaptively updated with backpropagation of the model to achieve dynamic learning of the correlation strength between domains. The spatial convolution branch and the temporal convolution branch are set to input in parallel, with both inputs being the first-level feature set and the inter-domain association graph. The two branches run independently without parameter sharing. Spatial convolution branch: A two-layer cascaded graph convolutional network is used with LeakyReLU activation function. The feature information of adjacent nodes is aggregated through graph convolution to extract the static coupling correlation features between different domains. Temporal convolution branch: A 3-layer cascaded causal one-dimensional convolutional network is used with a causal filling mode to completely avoid the leakage of future information. The activation function is LeakyReLU to extract the dynamic change features of inter-domain correlation in the temporal dimension. The dual-branch spatiotemporal graph convolutional network submodule concatenates the static coupling features output by the spatial convolutional branch with the dynamic change features output by the temporal convolutional branch to obtain the second-level features.

[0011] Furthermore, the extraction process of the third-level features is as follows: An encoder-decoder structure is adopted, with the encoder being a 3-layer Transformer Encoder and the decoder being a 2-layer Transformer Decoder. The second-level features are input, and the masking rate is set to randomly mask some feature fragments. The encoder only encodes the unmasked features, and the decoder reconstructs the complete input features based on the encoded features. By minimizing the reconstruction loss, random noise and redundant features unrelated to engine performance are filtered out, and the preliminary purified features are output. A positive and negative sample set is constructed based on engine performance labels. Features within the same batch, with the same performance level and the same anomaly type are used as positive samples, while features with different performance levels and different anomaly types are used as negative samples. Normalized temperature scaling cross-entropy loss is used as the contrast loss function. A gated convolutional structure is adopted to perform the final precise screening of features enhanced by contrastive learning. Through the gating mechanism, a weight value between 0 and 1 is automatically assigned to each feature dimension. The weight value is positively correlated with the correlation between the feature and engine performance, and finally the third-level feature is output. Finally, the deep feature extraction module concatenates the first-level, second-level, and third-level features into deep features, which are then output to the model training module and the real-time performance detection module.

[0012] Furthermore, the model training module adopts a multi-branch, multi-task learning architecture. The backbone network is a 4-layer cascaded fully connected network, with a batch normalization layer, a Dropout layer, and a ReLU activation layer sequentially following each fully connected layer. The backbone network output has four parallel, independent task branches, namely: Performance level classification branch: The output layer adopts the Softmax activation function, and the output corresponds to four dimensions: power, economy, emissions, and reliability. Each dimension is divided into four levels: excellent, good, qualified, and unqualified. Performance index regression branch: The output layer adopts a linear activation function, and the output dimension corresponds to four core quantitative performance indicators, including the maximum output power and torque reserve coefficient of power performance, the effective thermal efficiency and fuel consumption per 100 kilometers of economy performance, the CO and NOx emission concentrations of emissions performance, and the cylinder leakage and oil consumption of reliability performance. Anomaly type identification branch: The output layer uses the Softmax activation function, and the output dimensions correspond to the classification results of normal, fuel injection system anomaly, ignition system anomaly, crankshaft and connecting rod mechanism anomaly, valve train mechanism anomaly, cooling system anomaly, lubrication system anomaly, intake system anomaly, and exhaust system anomaly. Attenuation trend prediction branch: The output layer adopts a linear activation function, and the output dimension is the predicted value of the comprehensive performance attenuation rate of the engine in the next 5,000 kilometers and 10,000 kilometers. The preprocessed historical dataset is divided into training, validation, and test sets using stratified sampling according to a preset ratio. A multi-task joint loss function is used. The AdamW optimizer is employed, and the learning rate adopts a cosine annealing decay strategy. During training, accuracy verification is performed on the validation set after each training epoch. If the validation set loss does not decrease for a preset number of consecutive epochs, an early stopping mechanism is triggered, and the weight parameter with the highest accuracy on the validation set is retained. After training, the model is lightweighted using INT8 quantization and structured pruning methods.

[0013] Furthermore, the real-time performance detection module obtains the corresponding deep features in real time, then inputs them into the lightweight multi-task performance detection model to complete forward inference and output real-time performance detection results, including engine comprehensive performance score, performance level and quantitative indicators in four dimensions, anomaly type identification results, performance degradation trend prediction results, and remaining service life assessment results. Simultaneously, real-time collected, maintenance-calibrated, and manually labeled valid operational data are added to the incremental training set. By freezing the backbone network and only fine-tuning the top-level parameters of the task branches and feature extraction modules, the performance detection model is updated online incrementally. Without full retraining, it can adapt to the performance changes throughout the vehicle's entire life cycle and continuously optimize the model's detection accuracy and vehicle adaptability.

[0014] Furthermore, the result output and early warning module sets up a three-level early warning mechanism based on the corresponding detection results, and sets corresponding processing suggestions for each level of early warning mechanism.

[0015] According to another aspect of the present invention, a method for testing the performance of a hybrid vehicle engine is provided. This method is implemented based on the above-described hybrid vehicle engine performance testing system and specifically includes the following steps: S1. Historical Data Collection and Labeling: Collect historical operating datasets of hybrid engines covering all operating conditions, the entire life cycle, and all performance states. Through engine bench calibration tests and vehicle drum tests, label each set of historical data with corresponding performance tags, including performance level tags, quantitative performance index tags, anomaly type tags, and degradation trend tags. S2. Data standardization preprocessing: The historical operational dataset is subjected to a standardization preprocessing process in sequence, which includes data cleaning, missing value imputation, outlier removal, data standardization, time series synchronization and alignment, and working condition segmentation labeling, to obtain the preprocessed historical dataset. S3. Deep Feature Extraction: Perform deep feature extraction on the preprocessed historical dataset. The specific steps are as follows: S31, First-level feature extraction: Based on the preprocessed historical dataset, the domain is divided according to the five physical domains. An independent parallel extraction branch is set for each domain. The transient features in the frequency domain are extracted by wavelet packet transform, and the dynamic features in the time domain are extracted by deformable one-dimensional convolution. The basic features within each domain are concatenated to obtain the first-level features. S32. Second-level feature extraction: Based on the first-level features, construct the inter-domain association graphs corresponding to the five domains. Extract the static inter-domain association features through the spatial convolution branch of the dual-branch spatiotemporal graph convolutional network, and extract the temporal dynamic association features through the temporal convolution branch. Then, concatenate the features to obtain the second-level features. S33, Third-level feature extraction: Based on the second-level features, feature pre-purification is completed through masked autoencoder, the performance and discriminativeness of features are enhanced through supervised contrastive learning, and the final feature selection is completed through gated linear units to obtain the third-level features; S34. Feature Fusion: Channel concatenation of the first-level features, second-level features, and third-level features is performed to obtain a three-level fused deep feature vector; S4. Multi-task model training and optimization: Divide the training set, validation set and test set into an 8:1:1 ratio, construct a multi-branch multi-task learning model architecture, set the multi-task joint loss function and training hyperparameters, complete the model training, validation and optimization, and perform lightweight processing after reaching the preset accuracy to generate a multi-task performance detection model that can be deployed on the vehicle.

[0016] S5. Real-time data acquisition: After the vehicle is powered on, the system collects all-dimensional operating data of the five physical domains of the hybrid engine in real time through the vehicle CAN bus at a fixed sampling frequency of 100Hz, and generates a set of data to be tested every 10 seconds. S6. Real-time data preprocessing: Perform the same preprocessing operation as step S2 in the training phase on the real-time collected data to obtain the preprocessed real-time data, ensuring that the data format and distribution are completely consistent with the training set. S7. Real-time feature extraction: Perform deep feature extraction on the preprocessed real-time data in the same way as step S3 in the training phase to obtain a 1280-dimensional three-level fusion deep feature vector corresponding to the real-time data. S8. Real-time performance inference: Input the real-time feature vector into the trained lightweight multi-task performance detection model, complete the forward inference, and output the engine real-time performance detection result. S9. Results Display and Tiered Warning: The test results are displayed visually through the vehicle's central control screen and instrument panel. When an abnormal performance or attenuation rate is detected that exceeds a preset threshold, a corresponding tiered warning is triggered, and maintenance suggestions are pushed. S10. Online Incremental Model Update: After each preset mileage or after a maintenance check, real-time running data that has been validated and manually labeled is added to the incremental training set to perform online incremental updates to the model and continuously optimize the model's detection accuracy.

[0017] Beneficial effects: The parallel branch feature extraction architecture with multi-domain divide-and-conquer in five physical domains is adopted. There is no parameter sharing or cross-interference between branches, which avoids the interference of cross-domain redundant noise on feature extraction from the root. At the same time, it combines wavelet packet transform to extract frequency domain transient features and deformable one-dimensional convolution to extract time domain dynamic features, accurately capturing the transient nonlinear signal changes of early anomalies in hybrid engines. This solves the pain points of existing technologies where single-domain early anomaly features are easily submerged and the false alarm and false negative rates are high, and significantly improves the sensitivity of early performance degradation identification.

[0018] An undirected weighted inter-domain association graph of five physical domains is constructed, and the adjacency matrix can be adaptively updated during model training. The static coupling association features and temporal dynamic change features between domains are extracted simultaneously through a two-branch spatiotemporal graph convolutional network, accurately capturing the implicit anomalies of inter-domain coupling relationships during frequent switching of operating conditions. It completely solves the problem that the existing technology is only suitable for steady-state operating conditions and has insufficient detection accuracy for transient operating conditions, and comprehensively covers all operating conditions of pure electric, series hybrid, parallel hybrid, engine direct drive and energy recovery, greatly improving the detection accuracy of hybrid engines under high-frequency transient operating conditions.

[0019] By employing a three-level progressive extraction architecture of "single-domain basic features - inter-domain spatiotemporal correlation features - supervised learning to purify features," combined with masked autoencoders to filter redundant noise, supervised contrastive learning to enhance performance discriminativeness, and gated linear units to complete precise feature selection, the system achieves layer-by-layer feature purification and high-order latent feature mining. This addresses the problems of existing technologies having a single feature extraction level and insufficient latent feature mining that is strongly correlated with performance, significantly improving the accuracy of identifying early performance degradation in engines and enabling ultra-early warning of faults.

[0020] Employing a multi-task joint learning architecture, combined with lightweight processing of INT8 quantization and structured pruning, the model size is significantly compressed with minimal precision loss, and the inference latency of a single set of data meets the computing power requirements of in-vehicle embedded environments. At the same time, through an online incremental update mechanism that freezes the backbone network and only fine-tunes the top-level parameters, it can adapt to the performance changes of the entire vehicle life cycle and the hybrid engines of different models without full retraining. This solves the contradiction of existing technologies that "high-precision models cannot be deployed in vehicles and lightweight models have poor generalization," and significantly reduces the calibration cost and adaptation difficulty for engineering applications.

[0021] The single model simultaneously performs four parallel tasks: performance level classification, core performance index regression, accurate anomaly type identification, and performance degradation trend prediction. It can output comprehensive performance scores, remaining service life assessments, and other full-dimensional results, along with a three-level graded early warning mechanism. It solves the shortcomings of existing technologies that only have a single detection dimension and can only determine the presence or absence of faults. It comprehensively covers the full life cycle health management needs of hybrid engines, from early performance degradation and fault warning to life prediction, and provides full-dimensional data support for predictive maintenance and repair decisions. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the method. Detailed Implementation

[0023] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0024] Example 1 Detailed implementation of each module of the system Data acquisition module implementation (1) Historical Data Collection: Collect the operating data of this hybrid engine under different ambient temperatures, plains, plateaus, cities, highways, and mountain roads, the entire life cycle of 0-180,000 kilometers, normal working conditions, 4-level performance degradation level / 8 types of typical fault modes, and a total of 180,000 sets of valid data were collected. Each set of data lasted 10 seconds and the sampling frequency was 100Hz. (2) Performance labeling: Through engine bench testing, the performance labels for each set of data are completed. The specific labeling rules are as follows: Performance rating labels: four dimensions: power, economy, emissions, and reliability. Each dimension is divided into four levels according to the performance degradation rate: excellent (degradation rate < 5%), good (5% ≤ degradation rate < 10%), qualified (10% ≤ degradation rate < 15%), and unqualified (degradation rate ≥ 15%). Quantitative performance indicator labels include eight core indicators: maximum output power, torque reserve coefficient, effective thermal efficiency, combined fuel consumption per 100 kilometers, CO emission concentration, NOx emission concentration, cylinder leakage, and oil consumption. Abnormality type labels include 9 categories: normal, fuel injection system abnormality, ignition system abnormality, crankshaft and connecting rod mechanism abnormality, valve train abnormality, cooling system abnormality, lubrication system abnormality, intake system abnormality, and exhaust system abnormality. Attenuation trend label: Based on bench durability tests, the comprehensive performance degradation rate for each set of data over the next 5,000 kilometers and 10,000 kilometers is labeled.

[0025] Data preprocessing module implementation Standardization preprocessing was performed on the 180,000 sets of historical data collected, with the specific parameters set as follows: (1) Missing value filling: When there are ≤10 consecutive missing sampling points, linear interpolation is used to fill them; when there are >100 consecutive missing sampling points (10% of the length of a single data set), the data set is directly removed. (2) Outlier removal: First, extreme outliers exceeding 3 times the standard deviation are removed by the 3σ criterion, and then latent outliers are removed by the isolated forest algorithm (the number of trees is set to 100). Finally, 168,000 sets of valid preprocessed data are obtained. (3) Data standardization: Z-Score standardization was adopted, and based on the mean and standard deviation of the historical dataset, all 38 collected parameters were mapped to a standard normal distribution with a mean of 0 and a variance of 1. (4) Timing synchronization alignment: Based on the engine ECU clock, all parameters are resampled and timestamped to ensure that the timing of all parameters is fully synchronized; (5) Operating condition segment labeling: According to the VCU drive mode signal, the data is divided into 5 operating condition segments and labeled with the corresponding operating condition labels. Among them, the series hybrid operating condition accounts for 32%, the parallel hybrid operating condition accounts for 28%, the engine direct drive operating condition accounts for 22%, the pure electric operating condition accounts for 12%, and the energy recovery operating condition accounts for 6%.

[0026] Detailed implementation of the deep feature extraction module This module is the core of this embodiment, and its detailed parameter settings are as follows: (1) First-level feature extraction unit The 38 acquired parameters were divided into domains according to five physical domains, and the specific division results are as follows: Mechanical domain (7 parameters): engine speed, output torque, cylinder pressure, valve lift, crankshaft position, oil pressure, and oil temperature; Fuel combustion domain (9 parameters): fuel injection quantity, fuel injection advance angle, fuel injection pulse width, intake pressure, intake air temperature, throttle opening, exhaust temperature, oxygen sensor voltage, air-fuel ratio; Thermal management domain (5 parameters): coolant inlet and outlet temperatures, cylinder block temperature, cylinder head temperature, and thermostat opening. Electric drive coupling domain (8 parameters): drive motor speed, torque, power, charging and discharging current, charging and discharging voltage, winding temperature, power battery SOC, and cell temperature; Vehicle operating condition domain (9 parameters): drive mode, vehicle speed, throttle opening, brake opening, transmission gear, vehicle acceleration, ambient temperature, atmospheric pressure, and altitude.

[0027] Each domain corresponds to an independent parallel extraction branch. The wavelet packet transform subunit of each branch uses the db4 mother wavelet, with 5 layers of wavelet packet decomposition to extract 5 types of statistical features in 32 frequency bands, outputting 32-dimensional features. The deformable one-dimensional convolution subunit is set with 3 layers of deformable convolution, with convolution kernel sizes of 5, 9, and 13, dilation coefficients of 1, 2, and 4, and output channels of 128, 64, and 64, respectively. The pooling kernel size is 2, the stride is 2, and the final output is 64-dimensional features. Each branch is concatenated to obtain 96-dimensional features, and 5 branches are concatenated to obtain 480-dimensional first-level features.

[0028] (2) Second-level feature extraction unit The initial values ​​of the adjacency matrix of the inter-domain association graph are obtained by calculating the Pearson correlation coefficients between each pair of domains using historical datasets. The initial adjacency matrix is ​​as follows: Model training module implementation (1) Data set partitioning: 168,000 sets of preprocessed data were stratified and sampled in a ratio of 8:1:1 to obtain 134,400 sets of training set, 16,800 sets of validation set, and 16,800 sets of test set; (2) Model architecture: The backbone network has 4 fully connected layers with hidden layer dimensions of 1024, 512, 256 and 128, and Dropout coefficient of 0.3; the output has 4 parallel task branches with output dimensions of 16, 8, 9 and 2 respectively. (3) Loss function: The weight coefficients of the multi-task joint loss function are set as w1=0.4, w2=0.35, and w3=0.25; (4) Training hyperparameters: AdamW optimizer, initial learning rate 0.001, weight decay 0.01, cosine annealing decay, batch_size=32, maximum training epochs 300, early stopping patience=15; (5) Training results: After training, the model's performance on the test set was as follows: performance level classification accuracy 98.7%, F1 score 98.5%; anomaly type identification accuracy 99.2%, F1 score 99.0%; performance index regression determination coefficient R 2 ≥0.97, Mean Absolute Error (MAE) ≤1.8%; Decay Trend Prediction R 2 ≥0.95, MAE≤2.2%; (6) Lightweight processing: INT8 quantization and structured pruning are adopted, the model size is compressed from 128MB to 28MB, the compression rate is 78%, the accuracy loss is 0.32%, and the inference latency of a single set of data is 76ms, which meets the requirements of vehicle.

[0029] Implementation of Real-time Performance Monitoring and Early Warning Module The vehicle-mounted real-time detection module completes data acquisition, preprocessing, feature extraction, and inference every 10 seconds, and outputs real-time performance detection results. The three-level warning thresholds are set as follows: Level 1 warning (attenuation rate ≥ 5%), Level 2 warning (attenuation rate ≥ 15%), and Level 3 warning (a clear abnormality type is detected). The model's online incremental update is performed every 5,000 kilometers. It adopts the method of freezing the backbone network and only fine-tuning the top-level parameters. The time for a single round of incremental training is ≤ 10 minutes, and the accuracy is improved by 0.5%~1.2%.

[0030] Example 2 Full-scene data collection Bench test data acquisition: Data acquisition was completed on the engine high and low temperature environment bench and the vehicle four-wheel drive drum test bench. The ambient temperature covered -30℃~45℃, the altitude covered 0~4500m, and the operating conditions covered the full operating scenarios of hybrid engines, including idle speed, partial load, full load, cold start, hot start, and start-stop cycle. The bench durability test simulated the performance degradation state of the entire life cycle from 0 to 200,000 kilometers. The simulation of 8 typical fault modes, such as fuel injection system blockage / leakage, ignition coil aging, and cylinder wear, was completed. Each fault mode was set with 3 degradation levels: mild, moderate, and severe. A total of 220,000 sets of valid bench data were collected, with each set of data lasting 10 seconds and a sampling frequency of 100Hz.

[0031] Real-vehicle drum data collection: Data collection was completed on the drum test bench for the three global standard cycles of NEDC, WLTC, and CLTC-P, as well as typical real-vehicle conditions such as urban congestion, highway cruising, mountain road climbing, low-temperature cold start, and high-altitude driving. The data covered the full range of power battery SOC from 5% to 95%, and a total of 60,000 sets of real-vehicle condition data were collected, ultimately forming a complete historical dataset of 280,000 sets.

[0032] Standardized performance label calibration In strict accordance with national standards such as GB18297-2001 "Test Methods for Performance of Automobile Engines" and GB18352.6-2016 "Limits and Measurement Methods for Pollutant Emissions from Light-Duty Vehicles", performance labeling for each set of data was completed through bench calibration tests. The labeling system fully matches the claims of this invention. The specific calibration rules are as follows: Performance rating labels are divided into four dimensions: power, economy, emissions, and reliability. Each dimension is divided into four levels based on the performance degradation rate: excellent (degradation rate < 5%), good (5% ≤ degradation rate < 10%), acceptable (10% ≤ degradation rate < 15%), and unacceptable (degradation rate ≥ 15%). Quantitative performance indicator labels include eight core quantitative indicators: maximum output power, torque reserve coefficient, effective thermal efficiency, fuel consumption per 100 kilometers, CO emission concentration, NOx emission concentration, cylinder leakage, and oil consumption. Abnormality type tags: Includes 9 categories, including normal, fuel injection system abnormality, ignition system abnormality, crankshaft and connecting rod mechanism abnormality, valve train abnormality, cooling system abnormality, lubrication system abnormality, intake system abnormality, and exhaust system abnormality; Attenuation trend label: Indicates the overall engine performance attenuation rate over the next 5,000 kilometers and 10,000 kilometers for each set of data.

[0033] Data standardization preprocessing implementation Data cleaning: Remove duplicate data, continuous empty data segments caused by communication interruptions, and invalid data with fixed values ​​caused by abnormal sensor power supply. A total of 12,000 sets of invalid data were removed. Missing value imputation: When there are ≤8 consecutive missing sampling points, linear interpolation is used for imputation; when there are >8 consecutive missing sampling points, the data set is directly removed, with a total of 7,000 sets of low-quality data removed. Outlier removal: A two-layer removal mechanism combining the 3σ criterion and the isolated forest algorithm is adopted. First, extreme outliers exceeding 3 times the standard deviation are removed by the 3σ criterion. Then, the isolated forest algorithm is used to identify latent outliers. The isolated forest algorithm is set with 120 trees and a contamination value of 0.02. Finally, 261,000 sets of valid preprocessed data are obtained. Data standardization: The Z-Score standardization method is adopted. Based on the mean and standard deviation of the full historical dataset, all 42 collected parameters are mapped to a standard normal distribution with a mean of 0 and a variance of 1, thus completely eliminating the dimensional differences of different parameters. Timing synchronization alignment: Based on the system clock of the engine ECU, asynchronous data from different controllers are resampled and timestamped, with a unified resampling frequency of 100Hz, to ensure that the timing synchronization error of all parameters is ≤1ms; Operating condition segmentation labeling: Based on the drive mode signal output by the VCU, combined with vehicle speed, throttle opening, motor output power, and engine start / stop status, the operating data is divided into 5 operating condition segments and labeled accordingly. The final operating condition distribution is as follows: series hybrid operating condition accounts for 34%, parallel hybrid operating condition accounts for 26%, engine direct drive operating condition accounts for 21%, pure electric operating condition accounts for 14%, and energy recovery operating condition accounts for 5%, which perfectly matches the actual operating condition distribution of this vehicle model.

[0034] Deep feature extraction implementation First-level multi-domain divide-and-conquer parallel feature extraction The data is divided into five independent parallel extraction branches based on five physical domains. There is no parameter sharing or cross-interference between the branches. Each branch corresponds to the preprocessed data of one physical domain and extracts transient features in the frequency domain and dynamic features in the time domain, respectively. Frequency domain feature extraction: Each branch uses the db4 mother wavelet to perform 5-level wavelet packet decomposition on the time series data in the domain, obtaining wavelet coefficients of 32 frequency bands. Five statistical features of coefficients in each frequency band are extracted, namely energy, entropy, standard deviation, skewness and kurtosis, and 32-dimensional frequency domain transient features are output. Temporal feature extraction: Each branch is set with a 3-layer cascaded deformable one-dimensional convolutional network with kernel sizes of 7, 11, and 15, and dilation coefficients of 1, 2, and 4, respectively. After each convolutional layer, a batch normalization layer, a LeakyReLU activation layer, and a max pooling layer are set in sequence. The pooling kernel size is 2 and the stride is 2. The number of output channels of the 3 convolutional layers are 128, 64, and 32, respectively, and the final output is 32-dimensional temporal dynamic features. Feature splicing: Each branch splices the frequency domain transient features with the time domain dynamic feature channels to obtain 64-dimensional basic features in the domain. The features of the 5 branches are spliced ​​in a fixed order to obtain 320-dimensional first-level features.

[0035] Second-level bi-branch spatiotemporal graph convolution feature extraction An undirected weighted inter-domain association graph is constructed using five physical domains as five graph nodes. The initial value of the adjacency matrix is ​​obtained by calculating the Pearson correlation coefficient between each pair of domains using historical datasets, with a value range of [0,1]. During training, it is adaptively updated during backpropagation of the model. A parallel structure of spatial convolutional branch and temporal convolutional branch is set up, with no parameter sharing between the two branches. Spatial convolution branch: A two-layer cascaded graph convolutional network is used with the activation function LeakyReLU. Each layer has 128 output channels. The feature information of adjacent nodes is aggregated through graph convolution to extract the static coupling correlation features between domains and output 128-dimensional static correlation features. Temporal convolution branch: A 3-layer cascaded causal one-dimensional convolutional network is adopted, and a causal padding mode is used to avoid the leakage of future information. The activation function is LeakyReLU, the convolutional kernel size is 5, 7, and 9, the dilation coefficient is 1, 2, and 3, and the number of output channels per layer is 64. The dynamic change features of inter-domain correlation in the temporal dimension are extracted, and 64-dimensional dynamic correlation features are output. Feature concatenation: The feature channels of the dual-branch output are concatenated to obtain 192-dimensional second-level features.

[0036] Third-level feature purification and final fusion An encoder-decoder structure is used, combining supervised contrastive learning with gated linear units to complete the final feature extraction. Masked autoencoder pre-purification: The encoder is a 3-layer TransformerEncoder, the decoder is a 2-layer TransformerDecoder, the hidden layer dimension is 192, the number of attention heads is 8, the masking rate is set to 40%, and random noise and redundant features are filtered by minimizing the reconstruction loss, and the output is initially purified features. Supervised contrastive learning reinforcement: Positive and negative sample sets are constructed based on performance labels. Features with the same performance level and the same anomaly type within the same batch are positive samples, and vice versa. Normalized temperature-scaled cross-entropy loss is used as the contrastive loss function, and the temperature coefficient is set to 0.07 to enhance the distinguishability of features on engine performance status. Gated linear unit filtering: A gated convolutional structure is adopted to automatically assign a weight value of 0 to 1 to each feature dimension. The weights are positively correlated with the features and engine performance, and finally output 128-dimensional third-level features. Final feature fusion: The first-level 320-dimensional features, the second-level 192-dimensional features, and the third-level 128-dimensional features are concatenated to obtain a 640-dimensional three-level fused deep feature vector, which is then output to the model training module and the real-time performance detection module.

[0037] Multi-task performance testing model training and lightweight implementation The specific implementation process is as follows: Dataset partitioning: 261,000 sets of preprocessed data were stratified and sampled in a ratio of 8:1:1 to ensure that the distribution of working conditions, performance levels and fault types of each dataset was consistent, resulting in a training set of 208,800 sets, a validation set of 26,100 sets and a test set of 26,100 sets. Multi-task model architecture: A multi-branch, multi-task learning architecture is adopted. The backbone network is a 4-layer cascaded fully connected network. Each layer is followed by a batch normalization layer, a Dropout layer, and a ReLU activation layer. The hidden layer dimensions of the 4 fully connected layers are 512, 256, 128, and 64, respectively, and the Dropout coefficient is set to 0.25. The backbone network output has 4 parallel independent task branches. Performance level classification branch: The output layer uses the Softmax activation function, with 16 output dimensions, corresponding to 4 dimensions and 4 performance levels for each dimension; Performance Metrics Regression Branch: The output layer uses a linear activation function with an output dimension of 8, corresponding to 8 core quantitative performance metrics; Anomaly type identification branch: The output layer uses the Softmax activation function, with an output dimension of 9, corresponding to 9 categories of anomaly / normal classification results; Degradation trend prediction branch: The output layer uses a linear activation function with an output dimension of 2, corresponding to the predicted performance degradation rate over the next 5000 km and 10000 km. Training hyperparameter settings: A multi-task joint loss function is used, with total loss Loss = 0.35 × classification loss + 0.4 × regression loss + 0.25 × contrastive loss, where the classification loss uses cross-entropy loss and the regression loss uses mean squared error loss; the optimizer uses the AdamW optimizer, with an initial learning rate of 0.0008, weight decay of 0.01, and a cosine annealing decay strategy for the learning rate. The batch size is 64, the maximum number of training epochs is 350, and the early stopping mechanism patience is set to 20. Model training results: After training, the model's overall performance on the test set is as follows: Classification accuracy for performance level is 99.1%, F1 score is 98.8%; Anomaly type identification accuracy is 99.5%, F1 score is 99.3%; Performance index: Regression coefficient of determination R0 2 ≥0.98, Mean Absolute Error (MAE) ≤1.5%; Decay Trend Prediction R 2 ≥0.96, MAE≤2.0%, fully meeting the mass production precision requirements of automakers; Lightweight model processing: A lightweight method combining INT8 quantization and structured pruning is adopted to retain the core weights of the backbone network and feature extraction module, and only redundant channels and neurons are pruned. The final model size is compressed from 96MB to 19MB, with a compression rate of 80.2% and an accuracy loss of only 0.28%. The forward inference latency of a single 10s data set is ≤65ms, which is far lower than the 100ms latency threshold of the vehicle end, and fully meets the requirements of automotive-grade real-time detection.

[0038] Real-time vehicle-mounted detection and graded early warning system implemented The lightweight model is then deployed to the vehicle domain controller to complete real-vehicle adaptation and functional implementation. The specific implementation is as follows: Real-time detection process: After the vehicle is powered on, the system collects operating parameters in real time at a frequency of 100Hz via the CANFD bus. A set of data to be detected is generated every 10 seconds. The system performs the same preprocessing and three-level feature extraction process as the training phase to obtain a real-time 640-dimensional deep feature vector. This vector is then input into the model to complete forward inference and outputs the engine comprehensive performance score, 4-dimensional performance level, quantitative indicators, anomaly type, decay trend, and remaining service life full-dimensional detection results. A three-tiered early warning mechanism: This mechanism combines vehicle warranty requirements with maintenance and repair standards to set three levels of early warning thresholds and corresponding handling strategies. Level 1 Warning (Prompt Level): Triggered when the performance degradation rate in any dimension is ≥5% and <10%, or the overall performance score is <90 points and ≥80 points. The vehicle health interface on the central control screen displays a yellow prompt and pushes the suggestion "Engine performance is slightly degraded, and a full inspection is recommended at the next maintenance". It is also synchronized to the car manufacturer's user APP and after-sales system. Level 2 Warning (Warning Level): Triggered when the performance degradation rate in any dimension is ≥10% and <15%, or the comprehensive performance score is <80 points and ≥70 points. The instrument panel and the central control screen will display an orange warning simultaneously and push the suggestion "The engine performance has moderately degraded. It is recommended to go to the authorized service center for inspection within 1 month" and simultaneously send the nearest authorized service station. Level 3 Warning (Emergency Level): Triggered when the performance degradation rate in any dimension is ≥15%, or a clear abnormality type is identified, or the comprehensive performance score is <70 points. The instrument panel issues a red emergency warning and audible and visual alert, and pushes a message that "Engine performance is abnormal, there is a safety risk, please slow down and stop immediately and contact roadside assistance." The fault data is uploaded to the cloud emergency response system in real time.

[0039] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A performance testing system for a hybrid automotive engine, characterized in that, include: The system includes a data acquisition module, a data preprocessing module, a deep feature extraction module, a model training module, a real-time performance testing module, and a result output and early warning module. Data acquisition module: used to collect historical and real-time operating data of hybrid engines. The historical operating data is labeled with corresponding performance tags. The data preprocessing module is used to perform data cleaning, missing value imputation, outlier removal, data standardization, time-series synchronization alignment, and working condition segmentation labeling on the collected historical and real-time running datasets in sequence, so as to obtain the preprocessed historical datasets and preprocessed real-time data. Deep feature extraction module: used to extract deep features from the preprocessed input data; The first-level features are extracted based on the preprocessed input data through a parallel branching structure of multi-domain divide-and-conquer. The second-level features are extracted from the first-level feature set through a dual-branch spatiotemporal graph convolutional network. The third-level features are based on the second-level features and are extracted by combining masked autoencoders with supervised contrastive learning and gated linear units. Model training module: used to train a multi-task performance detection model for hybrid engines based on the extracted deep features and corresponding performance labels; Real-time performance detection module: This module is used to input the three-level fusion deep features corresponding to real-time data into the trained performance detection model and output the real-time performance detection results of the hybrid engine. The results output and early warning module is used to visualize the performance test results and trigger graded early warnings when abnormal performance or excessive attenuation is detected.

2. The performance testing system for a hybrid vehicle engine according to claim 1, characterized in that: The data acquisition module is divided into a cloud-based historical data acquisition unit and an on-board real-time data acquisition unit. The acquisition scope covers the full-dimensional operating parameters of five physical domains: hybrid engine thermal management domain, mechanical domain, fuel combustion domain, electric drive coupling domain, and vehicle operating condition domain. Thermal management parameters include, but are not limited to, coolant inlet and outlet temperatures, cylinder block temperature, cylinder head temperature, oil cooler inlet and outlet temperatures, EGR cooler temperature, and thermostat opening. Mechanical parameters include, but are not limited to, engine speed, output torque, cylinder pressure, valve lift, crankshaft position, oil pressure, oil temperature, and crankcase pressure. Fuel combustion parameters include, but are not limited to, injection quantity, injection advance angle, injection pulse width, intake pressure, intake temperature, throttle opening, exhaust temperature, front and rear oxygen sensor voltages, air-fuel ratio, and EGR valve opening. Electric drive coupling domain parameters include, but are not limited to, drive motor speed, torque, output power, charging and discharging current, charging and discharging voltage, stator winding temperature, rotor temperature, power battery SOC, power battery SOH, cell maximum / minimum temperature, and charge / discharge rate; The vehicle operating condition parameters include, but are not limited to, the current drive mode, vehicle speed, throttle opening, brake opening, transmission gear, vehicle acceleration, ambient temperature, atmospheric pressure, and altitude output by the VCU.

3. The performance testing system for a hybrid vehicle engine according to claim 1, characterized in that: The data preprocessing module includes the following processing steps: Data cleaning: Remove duplicate data, empty data segments caused by communication interruptions, and invalid data with fixed values ​​caused by sensor failures; Missing value imputation: For a small number of missing values ​​with consecutive missing sampling points ≤ the preset value, linear interpolation is used to imput them; for data segments with consecutive missing sampling points exceeding the preset value, they are directly discarded to avoid interference from invalid data. Outlier removal: A two-layer removal mechanism combining the 3σ criterion and the isolated forest algorithm is adopted. First, extreme outliers exceeding 3 times the standard deviation are removed by the 3σ criterion. Then, the isolated forest algorithm is used to identify and remove hidden outliers during the operation condition switching process. Data standardization: The Z-Score standardization method is adopted, which maps all parameter values ​​to a standard normal distribution with a mean of 0 and a variance of 1 based on the mean and standard deviation of the historical dataset, thus completely eliminating the dimensional differences of different parameters. Timing synchronization alignment: Based on the system clock of the engine ECU, asynchronous data from different controllers and sensors are resampled and timestamped, and a unified resampling frequency is set; Operating condition segmentation and labeling: Based on the drive mode signal output by the VCU, combined with vehicle speed, throttle opening, motor power, and engine start / stop status, the operating data is divided into operating condition segments including pure electric operating condition, series hybrid operating condition, parallel hybrid operating condition, engine direct drive operating condition, and energy recovery operating condition. Each operating condition segment is labeled with a corresponding operating condition label for operating condition adaptation during feature extraction.

4. The performance testing system for a hybrid vehicle engine according to claim 1, characterized in that: The extraction method for the first-level features is as follows: Each of the five physical domains is set up as a parallel and independent extraction branch, with each branch corresponding to the preprocessed data of one physical domain. There is no parameter sharing or cross interference between the branches, and they run completely independently. To address the characteristic of early abnormalities in hybrid engines manifested as transient nonlinear signal changes, the db4 mother wavelet was used to decompose the time-series data in the corresponding domain into 5-layer wavelet packets to obtain wavelet coefficients. Statistical features, including coefficient energy, entropy, standard deviation, skewness, and kurtosis for each frequency band, were extracted, and finally, the transient features in the frequency domain were output. To address the signal scale inconsistency caused by frequent switching of hybrid engine operating conditions, a three-layer cascaded deformable one-dimensional convolutional network is set up. Each convolutional layer is followed by a batch normalization layer, a LeakyReLU activation layer, and a max pooling layer. Through the adaptive receptive field of deformable convolution, the long-term trend and dynamic change characteristics of time-series data in the domain under different operating conditions are accurately captured, and finally the dynamic features of the time domain in the domain are output. Each branch concatenates the frequency domain transient features and the time domain dynamic features within the domain to obtain the basic features within the domain. The first-level features are obtained by splicing the single-domain features of each branch in a fixed order.

5. The performance testing system for a hybrid vehicle engine according to claim 1, characterized in that: The extraction process for the second-level features is as follows: Using the five physical domains corresponding to the first-level features as graph nodes, an undirected weighted graph is constructed corresponding to five nodes: mechanical domain, fuel combustion domain, thermal management domain, electric drive coupling domain, and vehicle operating condition domain. The initial value of the adjacency matrix is ​​obtained by calculating the Pearson correlation coefficient between each pair of domains using historical datasets. The value range is [0, 1], representing the initial correlation strength between the two domains. During training, it is adaptively updated with backpropagation of the model to achieve dynamic learning of the correlation strength between domains. The spatial convolution branch and the temporal convolution branch are set to input in parallel, with both inputs being the first-level feature set and the inter-domain association graph. The two branches run independently without parameter sharing. Spatial convolution branch: A two-layer cascaded graph convolutional network is used with LeakyReLU activation function. The feature information of adjacent nodes is aggregated through graph convolution to extract static coupling and correlation features between different domains. Temporal convolution branch: A 3-layer cascaded causal one-dimensional convolutional network is used with a causal filling mode to completely avoid the leakage of future information. The activation function is LeakyReLU to extract the dynamic change features of inter-domain correlation in the temporal dimension. The dual-branch spatiotemporal graph convolutional network submodule concatenates the static coupling features output by the spatial convolutional branch with the dynamic change features output by the temporal convolutional branch to obtain the second-level features.

6. The performance testing system for a hybrid vehicle engine according to claim 1, characterized in that: The extraction process of the third-level features is as follows: It adopts an encoder-decoder structure, with the encoder consisting of a 3-layer Transformer Encoder and the decoder consisting of a 2-layer Transformer Decoder; Input the second-level features, set the masking rate, randomly mask some feature fragments, the encoder only encodes the unmasked features, the decoder reconstructs the complete input features based on the encoded features, and filters out random noise and redundant features unrelated to engine performance by minimizing the reconstruction loss, and outputs the preliminary purified features; A positive and negative sample set is constructed based on engine performance labels. Features within the same batch, with the same performance level and the same anomaly type are used as positive samples, while features with different performance levels and different anomaly types are used as negative samples. Normalized temperature scaling cross-entropy loss is used as the contrast loss function. A gated convolutional structure is adopted to perform the final precise screening of features enhanced by contrastive learning. Through the gating mechanism, a weight value between 0 and 1 is automatically assigned to each feature dimension. The weight value is positively correlated with the correlation between the feature and engine performance, and finally the third-level feature is output. Finally, the deep feature extraction module concatenates the first-level, second-level, and third-level features into deep features, which are then output to the model training module and the real-time performance detection module.

7. The performance testing system for a hybrid vehicle engine according to claim 1, characterized in that: The model training module adopts a multi-branch, multi-task learning architecture. The backbone network is a 4-layer cascaded fully connected network. Each fully connected layer is followed by a batch normalization layer, a Dropout layer, and a ReLU activation layer. The backbone network output has four parallel, independent task branches, namely: Performance level classification branch: The output layer adopts the Softmax activation function, and the output corresponds to four dimensions: power, economy, emissions, and reliability. Each dimension is divided into four levels: excellent, good, qualified, and unqualified. Performance index regression branch: The output layer adopts a linear activation function, and the output dimension corresponds to four core quantitative performance indicators, including the maximum output power and torque reserve coefficient of power performance, the effective thermal efficiency and fuel consumption per 100 kilometers of economy performance, the CO and NOx emission concentrations of emissions performance, and the cylinder leakage and oil consumption of reliability performance. Anomaly type identification branch: The output layer uses the Softmax activation function, and the output dimensions correspond to the classification results of normal, fuel injection system anomaly, ignition system anomaly, crankshaft and connecting rod mechanism anomaly, valve train mechanism anomaly, cooling system anomaly, lubrication system anomaly, intake system anomaly, and exhaust system anomaly. Attenuation trend prediction branch: The output layer adopts a linear activation function, and the output dimension is the predicted value of the comprehensive performance attenuation rate of the engine in the next 5,000 kilometers and 10,000 kilometers. The preprocessed historical dataset is divided into training, validation, and test sets according to a preset ratio using stratified sampling; a multi-task joint loss function is used; the AdamW optimizer is employed, and a cosine annealing decay strategy is used for the learning rate. During training, accuracy verification is performed on the validation set after each training round. If the loss on the validation set does not decrease for a preset number of consecutive rounds, an early stopping mechanism is triggered, and the weight parameter with the highest accuracy on the validation set is retained. After training, the model was lightweighted using INT8 quantization and structured pruning methods.

8. The performance testing system for a hybrid vehicle engine according to claim 1, characterized in that: The real-time performance detection module obtains the corresponding deep features in real time, then inputs them into the lightweight multi-task performance detection model to complete forward inference and output the real-time performance detection results, including the engine comprehensive performance score, the performance level and quantitative indicators of four dimensions, the anomaly type identification results, the performance degradation trend prediction results, and the remaining service life assessment results. Simultaneously, real-time collected, maintenance-calibrated, and manually labeled valid operational data are added to the incremental training set. By freezing the backbone network and only fine-tuning the top-level parameters of the task branches and feature extraction modules, the performance detection model is updated online incrementally. Without full retraining, it can adapt to the performance changes throughout the vehicle's entire life cycle and continuously optimize the model's detection accuracy and vehicle adaptability.

9. The performance testing system for a hybrid vehicle engine according to claim 1, characterized in that: The result output and early warning module sets up a three-level early warning mechanism based on the corresponding detection results, and sets corresponding processing suggestions for each level of early warning mechanism.

10. A method for testing the performance of a hybrid vehicle engine, characterized in that: This method is implemented based on the performance testing system for a hybrid vehicle engine as described in any one of claims 1-9, and specifically includes the following steps: S1. Historical Data Collection and Labeling: Collect historical operating datasets of hybrid engines covering all operating conditions, the entire life cycle, and all performance states. Through engine bench calibration tests and vehicle drum tests, label each set of historical data with corresponding performance tags, including performance level tags, quantitative performance index tags, anomaly type tags, and degradation trend tags. S2. Data standardization preprocessing: The historical operational dataset is subjected to a standardization preprocessing process in sequence, which includes data cleaning, missing value imputation, outlier removal, data standardization, time series synchronization and alignment, and working condition segmentation labeling, to obtain the preprocessed historical dataset. S3. Deep Feature Extraction: Perform deep feature extraction on the preprocessed historical dataset. The specific steps are as follows: S31, First-level feature extraction: Based on the preprocessed historical dataset, the domain is divided according to the five physical domains. An independent parallel extraction branch is set for each domain. The transient features in the frequency domain are extracted by wavelet packet transform, and the dynamic features in the time domain are extracted by deformable one-dimensional convolution. The basic features within each domain are concatenated to obtain the first-level features. S32. Second-level feature extraction: Based on the first-level features, construct the inter-domain association graphs corresponding to the five domains. Extract the static inter-domain association features through the spatial convolution branch of the dual-branch spatiotemporal graph convolutional network, and extract the temporal dynamic association features through the temporal convolution branch. Then, concatenate the features to obtain the second-level features. S33, Third-level feature extraction: Based on the second-level features, feature pre-purification is completed through masked autoencoder, the performance and discriminativeness of features are enhanced through supervised contrastive learning, and the final feature selection is completed through gated linear units to obtain the third-level features; S34. Feature Fusion: Channel concatenation of the first-level features, second-level features, and third-level features is performed to obtain a three-level fused deep feature vector; S4. Multi-task model training and optimization: Divide the training set, validation set and test set into an 8:1:1 ratio, construct a multi-branch multi-task learning model architecture, set the multi-task joint loss function and training hyperparameters, complete the model training, validation and optimization, and perform lightweight processing after reaching the preset accuracy to generate a multi-task performance detection model that can be deployed on the vehicle. 11.S5 Real-time data acquisition: After the vehicle is powered on, the full-dimensional operation data of the five physical domains of the hybrid engine is acquired in real time through the vehicle CAN bus at a fixed sampling frequency of 100Hz, and a set of data to be tested is generated every 10 seconds. S6. Real-time data preprocessing: Perform the same preprocessing operation as step S2 in the training phase on the real-time collected data to obtain the preprocessed real-time data, ensuring that the data format and distribution are completely consistent with the training set. S7. Real-time feature extraction: Perform deep feature extraction on the preprocessed real-time data in the same way as step S3 in the training phase to obtain a 1280-dimensional three-level fusion deep feature vector corresponding to the real-time data. S8. Real-time performance inference: Input the real-time feature vector into the trained lightweight multi-task performance detection model, complete the forward inference, and output the engine real-time performance detection result. S9. Results Display and Tiered Warning: The test results are displayed visually through the vehicle's central control screen and instrument panel. When an abnormal performance or attenuation rate is detected that exceeds a preset threshold, a corresponding tiered warning is triggered, and maintenance suggestions are pushed. S10. Online Incremental Model Update: After each preset mileage or after a maintenance check, real-time running data that has been validated and manually labeled is added to the incremental training set to perform online incremental updates to the model and continuously optimize the model's detection accuracy.

Images