A vehicle-mounted virtual engine monitoring method based on logic isolation window sampling and mixed time sequence neural network

Abstract

The present application belongs to the technical field of vehicle engine state monitoring and virtual sensor modeling, and discloses a vehicle virtual engine monitoring method based on logical isolation window sampling and hybrid time sequence neural network. The time sequence files of multi-source heterogeneous engines are recursively discovered, column name mapped and physically cleaned. A logical isolation band with a length not less than the length of the observation window minus one is constructed between the training set, the validation set and the test set, and the sliding sampling step is adaptively set according to the working condition type, so that the original row sequence is reconstructed into a three-dimensional time sequence tensor. The tensor is input into the hybrid neural network, and the prediction results of the DOC front temperature, the intake volume and the NOx concentration are output. Finally, the global row number mapping is used to realize the cross-file backtracking of the prediction residual, and the monitoring results are generated in combination with the segmented compliance rate evaluation. The method improves the time sequence reliability, dynamic feature extraction capability and industrial traceability of the vehicle edge side virtual sensor modeling.

Description

Technical Field

[0001] This invention relates to the field of vehicle engine condition monitoring and virtual sensor modeling technology, specifically to a vehicle virtual engine monitoring method based on logically isolated window sampling and hybrid temporal neural networks. Background Technology

[0002] With the rapid development of the automotive industry, the real-time monitoring and accurate prediction of engine status by the engine control unit (ECU) has become increasingly important. Under complex and variable operating conditions, physical sensors often face problems such as measurement drift and accuracy fluctuations in high and low temperature environments. This has spurred the widespread application of virtual sensor technology. As a software algorithm, virtual sensors can deduce key state quantities that are difficult to measure directly from easily measurable parameters, providing a new technical approach for engine monitoring.

[0003] Currently, virtual sensor technology in the field of vehicle engine monitoring mainly focuses on data-driven model building. Existing technologies have adopted convolutional neural networks, long short-term memory networks, graph structure temporal networks, or spatiotemporal feature fusion methods to model engine life, operating status, or irregular multivariate time-series signals.

[0004] However, existing technologies still have the following technical shortcomings in virtual sensor modeling of vehicle engines: First, existing time-series data processing methods often overlook the overlap effect of sliding windows at time boundaries when dividing the training and test sets. This allows information from the training set to seep into the test set through overlapping rows, causing time-series leakage. For example, while existing methods for modeling irregular time series focus on the modeling accuracy of asynchronously sampled data, they do not adequately address the physical isolation between the training and test windows under sliding window partitioning conditions. This problem causes the model to tend to learn numerical memory rather than real physical laws, resulting in overestimation of evaluation results.

[0005] Secondly, engine operating conditions are typically highly unevenly distributed, with steady-state data volume far exceeding that of key transient data. Existing solutions based on spatiotemporal feature fusion or regression prediction focus more on the network structure itself, paying insufficient attention to the imbalance of sample density under different operating conditions. Conventional uniform sampling methods are prone to causing the model to overfit steady-state logic, thereby weakening its ability to capture highly dynamic load changes.

[0006] Third, engine monitoring data is typically stored in fragmented, multi-file format. While existing technologies can model multivariate time-series data, they generally lack a unified cross-file physical address mapping mechanism. When the model exhibits prediction bias, it is difficult to accurately trace the output residuals back to the specific original file and its corresponding physical line number or operating condition location, affecting the traceability of the virtual sensor system and the efficiency of engineering maintenance.

[0007] In addition, existing technologies often focus on data preprocessing or single temporal network structures, lacking a hybrid modeling chain that integrates input feature saliency enhancement, local dynamic feature extraction, medium- and long-term time dependency modeling, and key time location enhancement. This results in insufficient ability to jointly predict multiple key engine state variables under complex dynamic conditions. Summary of the Invention

[0008] To address the issues of temporal leakage risk, uneven operating condition distribution, difficulty in unifying the organization of fragmented multi-file data, and insufficient extraction of local dynamic features and long-term dependencies in existing vehicle-mounted engine virtual sensor sliding window modeling, this invention provides a vehicle-mounted virtual engine monitoring method based on logically isolated window sampling and hybrid temporal neural networks.

[0009] The purpose of this invention is to eliminate temporal overlap and contamination between training data and test data by establishing a logical isolation band, to adaptively adjust the sample distribution of steady-state data and transient data by using the sampling density of condition perception, and to achieve joint prediction of multiple key engine state variables by using a hybrid neural network composed of an input feature attention module, a convolutional temporal feature extraction module, a long short-term memory module and a temporal self-attention module.

[0010] The technical solution adopted by this invention to solve its technical problem is: a vehicle-mounted virtual engine monitoring method based on logical isolation window sampling and hybrid temporal neural network. Based on operating status data collected from bench tests, vehicle tests, or vehicle engine control units, a vehicle-mounted virtual engine monitoring model is constructed to replace or assist actual sensors in online estimation of multiple physical quantities. The physical quantities to be monitored include one or more of the following: engine intake air volume, DOC pre-temperature, and original NOx sensor values ​​from the engine exhaust. Specifically, the method includes the following steps: S1. Recursively scan the on-board engine operation data storage directory, read multi-source heterogeneous time series files formed by bench tests, vehicle tests or on-board engine control units, and convert data fields of different formats into standard input features and standard output targets through preset column name mapping rules. S2. Perform physical criterion cleaning on the data obtained in step S1; S3. Divide the data within a file according to the time sequence of the individual file and perform logical isolation window sampling processing; determine the data before a preset proportion in a single file as the training and validation candidate segment, and the data after the preset proportion as the test candidate segment; set a test isolation band with a length not less than the observation window length minus one between the training and validation candidate segments and the test candidate segments; further divide the training and validation candidate segments into training sub-segments and validation segments, and set a validation isolation band with a length not less than the observation window length minus one between the training sub-segments and the validation segments; the data rows within the test isolation band and the validation isolation band are not used as prediction target points for the training, validation, or test windows, nor are they entered into the corresponding window input sequence; S4. Set the adaptive sliding sampling step size according to the working condition type of the file, perform window sampling on the cleaned time series file data, and reconstruct the original row sequence into a three-dimensional input tensor containing the sample dimension, window length dimension and feature dimension. S5. Fit an input normalizer and an output normalizer based on the original row data of the training segments of all files, and use the input normalizer and the output normalizer to transform the training tensor, the validation tensor and the test tensor. S6. Input the training tensor, validation tensor and test tensor obtained in step S5 into the hybrid temporal neural network to output the prediction results of multiple physical quantities to be monitored. S7. Construct an adaptively adjusted Huber loss function for one or more output targets among engine intake air volume, DOC inlet temperature and engine exhaust NOx sensor values, and train the hybrid temporal neural network. S8. Perform inverse standardization on the model prediction results, and establish the correspondence between the prediction results and the original physical storage line numbers through cross-file global line number mapping. Output the prediction results, residual tracking results and segmented compliance evaluation results for offline evaluation or online deployment verification of the vehicle virtual engine monitoring model.

[0011] The standard input features are the engine operating status parameters that the vehicle engine control unit can collect or calculate, and the standard output target is the physical quantity to be monitored that the vehicle virtual engine monitoring model needs to estimate online.

[0012] The physical criteria cleaning includes null value filling, outlier removal, negative value truncation of non-temperature features, and rotation speed validity screening. The physical criterion cleaning includes retaining temperature-related features that are allowed to have negative environmental values, truncating negative values ​​for flow, emission, and pressure-related features, and deleting sample rows with any feature value exceeding a preset abnormal threshold.

[0013] The logical isolation band in step S3 includes a verification isolation band and a test isolation band. The length of both the verification isolation band and the test isolation band is L, and L≥W-1, where W is the observation window length. The verification isolation band is used to ensure that the input window corresponding to any predicted target point of the verification segment does not contain the data line of the training sub-segment. The test isolation band is used to ensure that the input window corresponding to any predicted target point of the test segment does not contain the data line of the training verification candidate segment.

[0014] In step S4, a sliding sampling step size greater than 1 is used for the steady-state characteristic file, and a sliding sampling method with a step size of 1 is used for the transient operating condition file, in order to balance the sample density distribution under different operating conditions.

[0015] When the file name contains a preset steady-state condition label, the system uses a preset increasing step size to perform full loading sampling without physically truncating the number of rows of the original data.

[0016] The input normalizer and output normalizer in step S5 are fitted based on the splicing result of the original line data of the training segment of all files, and do not use the data statistics of the validation segment or test segment.

[0017] The hybrid temporal neural network comprises a sequentially coupled input feature attention module, a convolutional temporal feature extraction module, a long short-term memory module, a temporal self-attention module, and a multi-output fully connected regression module. It is used to simultaneously extract feature dimension importance, local dynamic features, medium- to long-term dependencies, and key time-position responses from the time-series operation data of the vehicle engine. The input feature attention module generates weight coefficients for each input feature through a multilayer perceptron and a sigmoid mapping, performing dimension-wise weighting on the original input features before feeding them into subsequent networks. The convolutional temporal feature extraction module includes at least two one-dimensional convolutional layers to extract local dynamic features within the observation window. The long short-term memory module is used to model medium- to long-term dependencies. The temporal self-attention module is used to enhance the feature responses at key time positions.

[0018] In step S7, the adaptive adjustment of the Huber loss function sets target loss weights for each output target and dynamically updates the corresponding target loss weights during model training to improve the optimization priority of key target quantities in joint training.

[0019] In step S8, the cross-file global line number mapping establishes a global offset by accumulating the original length of the preceding file, and maps each predicted target point to the absolute physical line number in the corresponding original file.

[0020] Compared with the prior art, the present invention has the following advantages: Improve the reliability of time series modeling evaluation: By setting a logical isolation band based on window length between the training set, validation set and test set, the time series leakage problem caused by sliding window overlap is effectively avoided, and the evaluation results more realistically reflect the model's generalization ability to unknown operating conditions.

[0021] Improving the balance of sample distribution under different working conditions: By dynamically setting the sliding sampling step size according to the working condition type, it is beneficial to alleviate the problem of coexistence of redundancy in steady-state working condition samples and scarcity in transient working condition samples, and improve the model's adaptability to complex dynamic working conditions.

[0022] Enhanced ability to jointly extract local transient features and long-term dependencies: By introducing a hybrid temporal neural network consisting of an input feature attention module, a convolution module, a long short-term memory module, and a temporal self-attention module, it is possible to simultaneously characterize the local dynamic changes and cross-time step dependencies of engine signals.

[0023] Improve the robustness of multi-target joint prediction: By constructing a weighted Huber loss function for multiple output targets, it is possible to balance the importance of different target quantities and the need for stable training under large error samples.

[0024] Enhance industrial traceability: By using a cross-file global line number mapping mechanism, the model output is associated with the original physical line address, which is beneficial for cross-file tracking of prediction residuals and rapid location of abnormal operating conditions.

[0025] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0026] Figure 1 This is an overall flowchart of the vehicle virtual engine monitoring method based on logical isolation window sampling and hybrid temporal neural network described in this invention.

[0027] Figure 2 This is a schematic diagram of the hybrid temporal neural network structure used in this invention.

[0028] Figure 3 This is a schematic diagram of the logic isolation band and window sampling described in this invention.

[0029] Figure 4The following are time-series line graphs comparing the predicted and actual values ​​of the virtual sensor under test conditions: (a) Pre-DOC temperature prediction results on the first 80% of the training set data; (b) Pre-DOC temperature prediction results on the last 20% of the test set data; (c) Intake volume prediction results on the first 80% of the training set data; (d) Intake volume prediction results on the last 20% of the test set data; (e) Engine exhaust NOx sensor value prediction results on the first 80% of the training set data; (f) Engine exhaust NOx sensor value prediction results on the last 20% of the test set data.

[0030] Figure 5 This is a schematic diagram of the cross-file global line number mapping and residual tracking result output in step S8 of the present invention. Detailed Implementation

[0031] This invention provides a vehicle-mounted virtual engine monitoring method based on logical isolation window sampling and hybrid temporal neural network, including heterogeneous data source recursive discovery and parsing, physical criterion cleaning, logical isolation window partitioning, condition perception sampling step size adjustment, three-dimensional temporal tensor reconstruction, global standardization based on all training segment data, hybrid temporal neural network modeling, multi-output weighted robust training, and cross-file global line number mapping and segmented evaluation.

[0032] Preferably, the heterogeneous data source recursive discovery and parsing refers to automatically traversing engine experimental data files in multi-level nested directories through recursive scanning logic, and uniformly converting raw fields of different formats into standard input features and standard output targets through column name mapping rules.

[0033] Preferably, the physical criterion cleaning refers to performing null value filling, outlier removal, non-temperature feature negative value truncation, and rotation speed validity screening on the original data to improve the data quality and physical rationality entering the modeling process.

[0034] Preferably, the logical isolation window division refers to dividing a single file into training, validation, and testing segments in chronological order, and establishing a logical isolation band with a length not less than the observation window length minus one at the boundaries of each dataset, thereby avoiding cross-set overlap of the input observation intervals of the sliding window.

[0035] Preferably, the adjustment of the sampling step size for working condition perception refers to automatically setting the sliding sampling step size according to the working condition type to which the file belongs. For working conditions with high steady-state redundancy, a step size greater than 1 is used for sparse sampling, and for working conditions with significant dynamic fluctuations, a high-density sampling step size of 1 is used to balance the spatial distribution of samples.

[0036] Preferably, the three-dimensional temporal tensor reconstruction refers to mapping the cleaned multi-dimensional engine temporal signal into a three-dimensional input tensor containing sample dimension, window length dimension and feature dimension in a sliding window manner, so as to provide a unified temporal input format for deep learning models.

[0037] Preferably, the global standardization refers to fitting the input standardizer and the output standardizer based on the splicing result of the original line data of the training segments of all files, and applying them to the training set, validation set and test set, without using the information of the validation segment or test segment to participate in the fitting.

[0038] Preferably, the hybrid temporal neural network includes an input feature attention module, a convolutional temporal feature extraction module, a long short-term memory module, a temporal self-attention module, and a fully connected regression output module; wherein, the input feature attention module is used to enhance the expression of key input variables, the convolutional module is used to extract local transient features, the long short-term memory module is used to represent medium- and long-term temporal dependencies, and the temporal self-attention module is used to highlight the feature responses at key time positions.

[0039] Preferably, the multi-output weighted robust training refers to jointly optimizing the model by setting independent weights for multiple output targets using the Huber loss function, so as to balance the training priority of different target quantities and robustness under large error scenarios.

[0040] Preferably, the cross-file global line number mapping refers to establishing a global absolute line number by using the cumulative bias of the original length of the preceding file, and mapping each prediction result of the model back to the specific physical line address in the original test file, so as to support cross-file residual tracking and abnormal operating condition backtracking.

[0041] Preferably, the segmented evaluation refers to calculating the segmented compliance rate based on the numerical range of different target physical quantities using either an absolute error criterion or a relative error criterion, in order to reflect the industrial applicability of the virtual sensor in different measurement ranges.

[0042] This invention constructs a virtual sensor modeling chain suitable for vehicle monitoring scenarios through logical isolation window partitioning, standardized constraints, weighted robust training, and global mapping mechanisms.

[0043] Furthermore, the method supports outputting model prediction results, segmented compliance results, and time-series visualization results that correspond one-to-one with the physical line numbers of the original file after training is completed, so as to facilitate engineers to perform model evaluation and abnormal operating condition analysis.

[0044] Furthermore, the method can be deployed in an in-vehicle edge computing environment with limited computing resources, and achieves multi-target state prediction through a unified input tensor format and a compact hybrid neural network structure.

[0045] Furthermore, the method can be applied to the virtual sensing and long-term monitoring of key engine state parameters such as DOC inlet temperature, intake air volume, and NOx concentration.

[0046] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0047] An example of an on-board virtual engine monitoring method based on logical isolation window sampling and hybrid temporal neural networks is as follows: In the scenario of vehicle engine monitoring, the system needs to use easily measurable input signals to virtually predict key state variables such as DOC inlet temperature, intake air volume, and NOx concentration. To this end, this invention constructs a complete technical process from heterogeneous data discovery, physical criterion cleaning, logical isolation window sampling, temporal tensor quantization reconstruction, hybrid neural network modeling to cross-file result tracking.

[0048] Step 1: Recursive discovery of heterogeneous data sources and parsing of standard fields.

[0049] The system recursively scans industrial data directories containing multi-level nested folders, automatically reads CSV files in different encoding formats, and converts engine time series fields from different sources into standard input features and standard output targets through preset column name mapping rules, providing a consistent data interface for subsequent unified modeling.

[0050] Step 2: Physical criteria cleaning and logical isolation window division.

[0051] The system performs null value imputation, outlier removal, negative value truncation of non-temperature features, and rotation speed validity screening on the raw data. Then, it divides the data into training, validation, and test segments according to the time sequence of individual files, and establishes a logical isolation band with a length not less than one observation window length at the boundary of adjacent datasets to avoid including the original physical lines used in training in the test window input.

[0052] Step 3: Adaptive adjustment of operating condition sensing sampling density.

[0053] The system dynamically configures the sliding sampling step size based on the file's operating condition label. For operating condition files with high steady-state redundancy, sparse sampling is performed with a step size greater than 1; for transient operating condition files with significant dynamic fluctuations, high-density sampling with a step size of 1 is used, thereby balancing the proportion of steady-state samples and transient samples in the training space.

[0054] Step 4: High-dimensional temporal feature tensor quantization and global standardization.

[0055] The cleaned multidimensional time series signal is reconstructed into a three-dimensional input tensor containing sample dimension, window length dimension and feature dimension through a sliding window operator; and the input normalizer and output normalizer are fitted based on the splicing result of the original line data of all file training segments, and the training set, validation set and test set are uniformly transformed without using the data of the validation segment or test segment to participate in the fitting.

[0056] Step 5: Modeling and training of hybrid temporal neural networks.

[0057] The standardized temporal tensor obtained in step four is input into a hybrid temporal neural network. Preferably, the network first performs saliency modulation on the feature dimension of each time step through an input feature attention module; then, it extracts local dynamic features through a multi-layer one-dimensional convolutional structure; next, it models medium- to long-term dependencies through a long short-term memory module; then, it uses a temporal self-attention module to enhance the response at key time locations; and finally, it generates predicted values ​​for multiple target physical quantities through a fully connected regression output module. During training, a Huber loss function with independent weights for multiple output targets is used for joint optimization, and the optimal model parameters are selected by combining the validation set loss.

[0058] Step 6: Cross-file global physical line number mapping and segmentation evaluation.

[0059] The system establishes a global offset by accumulating the original length of the preceding file, and maps each predicted target point back to the absolute physical line number in the corresponding original file. After outputting the prediction results, the system calculates the segmented achievement rate according to the absolute error criterion or the relative error criterion based on the interval where different target physical quantities are located, and draws a long-term time series comparison chart of the true value and the predicted value with the global physical line number as the horizontal axis to support the backtracking of abnormal working conditions and the analysis of model performance.

[0060] Through the above steps, this invention realizes a complete virtual sensor monitoring process, from automatic access to heterogeneous engine data, logical isolation window sampling, hybrid temporal neural network modeling to cross-file residual backtracking and segmented evaluation. Compared with methods that only focus on data segmentation or only use a single temporal network structure, this invention can better balance temporal reliability, dynamic feature extraction capability, and industrial traceability.

[0061] The above description is merely a preferred embodiment of the present invention, and the present invention is not limited to the above embodiments. It is understood that other improvements and variations that are directly derived or conceived by those skilled in the art without departing from the spirit and concept of the present invention should be considered to be included within the protection scope of the present invention.

Claims

1. A method for monitoring an on-board virtual engine based on logically isolated window sampling and a hybrid temporal neural network, characterized in that, Based on the operating status data collected from bench tests, vehicle tests, or on-board engine control units, an on-board virtual engine monitoring model is constructed to replace or assist actual sensors in online estimation of multiple physical quantities; the physical quantities to be monitored include one or more of the following: engine intake air volume, DOC inlet temperature, and original NOx sensor values ​​from the engine exhaust. Specifically, the following steps are included: S1. Recursively scan the on-board engine operation data storage directory, read multi-source heterogeneous time series files formed by bench tests, vehicle tests or on-board engine control units, and convert data fields of different formats into standard input features and standard output targets through preset column name mapping rules. S2. Perform physical criterion cleaning on the data obtained in step S1; S3. Divide the data within a file according to the time sequence of the individual file and perform logical isolation window sampling processing; determine the data before a preset proportion in a single file as the training and validation candidate segment, and the data after the preset proportion as the test candidate segment; set a test isolation band with a length not less than the observation window length minus one between the training and validation candidate segments and the test candidate segments; further divide the training and validation candidate segments into training sub-segments and validation segments, and set a validation isolation band with a length not less than the observation window length minus one between the training sub-segments and the validation segments; the data rows within the test isolation band and the validation isolation band are not used as prediction target points for the training, validation, or test windows, nor are they entered into the corresponding window input sequence; S4. Set the adaptive sliding sampling step size according to the working condition type of the file, perform window sampling on the cleaned time series file data, and reconstruct the original row sequence into a three-dimensional input tensor containing the sample dimension, window length dimension and feature dimension. S5. Fit an input normalizer and an output normalizer based on the original row data of the training segments of all files, and use the input normalizer and the output normalizer to transform the training tensor, the validation tensor and the test tensor. S6. Input the training tensor, validation tensor and test tensor obtained in step S5 into the hybrid temporal neural network to output the prediction results of multiple physical quantities to be monitored. S7. Construct an adaptively adjusted Huber loss function for one or more output targets among engine intake air volume, DOC inlet temperature and engine exhaust NOx sensor values, and train the hybrid temporal neural network. S8. Perform inverse standardization on the model prediction results, and establish the correspondence between the prediction results and the original physical storage line numbers through cross-file global line number mapping. Output the prediction results, residual tracking results and segmented compliance evaluation results for offline evaluation or online deployment verification of the vehicle virtual engine monitoring model.

2. The vehicle-mounted virtual engine monitoring method based on logically isolated window sampling and hybrid temporal neural network according to claim 1, characterized in that, The standard input features are the engine operating status parameters that the vehicle engine control unit can collect or calculate, and the standard output target is the physical quantity to be monitored that the vehicle virtual engine monitoring model needs to estimate online.

3. The vehicle-mounted virtual engine monitoring method based on logically isolated window sampling and hybrid temporal neural network according to claim 1, characterized in that, The physical criteria cleaning includes null value filling, outlier removal, negative value truncation of non-temperature features, and rotation speed validity screening. The physical criterion cleaning includes retaining temperature-related features that are allowed to have negative environmental values, truncating negative values ​​for flow, emission, and pressure-related features, and deleting sample rows with any feature value exceeding a preset abnormal threshold.

4. The vehicle-mounted virtual engine monitoring method based on logically isolated window sampling and hybrid temporal neural networks according to claim 1, characterized in that, In step S3, the lengths of both the test isolation band and the verification isolation band are L, satisfying L≥W-1, where W is the length of the observation window, ensuring that the input window corresponding to any target point in the test set does not contain any data rows used in training.

5. The vehicle-mounted virtual engine monitoring method based on logically isolated window sampling and hybrid temporal neural networks according to claim 1, characterized in that, In step S4, a sliding sampling step size greater than 1 is used for the steady-state characteristic file, and a sliding sampling method with a step size of 1 is used for the transient operating condition file, in order to balance the sample density distribution under different operating conditions.

6. The vehicle-mounted virtual engine monitoring method based on logically isolated window sampling and hybrid temporal neural networks according to claim 5, characterized in that, When the file name contains a preset steady-state condition label, the system uses a preset increasing step size to perform full loading sampling without physically truncating the number of rows of the original data.

7. The vehicle-mounted virtual engine monitoring method based on logically isolated window sampling and hybrid temporal neural network according to claim 1, characterized in that, The input normalizer and output normalizer in step S5 are fitted based on the splicing result of the original line data of the training segment of all files, and do not use the data statistics of the validation segment or test segment.

8. The vehicle-mounted virtual engine monitoring method based on logically isolated window sampling and hybrid temporal neural network according to claim 1, characterized in that, The hybrid temporal neural network comprises a sequentially coupled input feature attention module, a convolutional temporal feature extraction module, a long short-term memory module, a temporal self-attention module, and a multi-output fully connected regression module. It is used to simultaneously extract feature dimension importance, local dynamic features, medium- to long-term dependencies, and key time-position responses from the time-series operation data of the vehicle engine. The input feature attention module generates weight coefficients for each input feature through a multilayer perceptron and a sigmoid mapping, performing dimension-wise weighting on the original input features before feeding them into subsequent networks. The convolutional temporal feature extraction module includes at least two one-dimensional convolutional layers to extract local dynamic features within the observation window. The long short-term memory module is used to model medium- to long-term dependencies. The temporal self-attention module is used to enhance the feature responses at key time positions.

9. The vehicle-mounted virtual engine monitoring method based on logically isolated window sampling and hybrid temporal neural network according to claim 1, characterized in that, In step S7, the adaptive adjustment of the Huber loss function sets target loss weights for each output target and dynamically updates the corresponding target loss weights during model training to improve the optimization priority of key target quantities in joint training.

10. The vehicle-mounted virtual engine monitoring method based on logically isolated window sampling and hybrid temporal neural network according to claim 1, characterized in that, In step S8, the cross-file global line number mapping establishes a global offset by accumulating the original length of the preceding file, and maps each predicted target point to the absolute physical line number in the corresponding original file.

Images