Engine power compensation control method, electronic equipment and vehicle

Abstract

The invention relates to the technical field of vehicle control, in particular to an engine power compensation control method, electronic equipment and a vehicle. The method comprises the steps that current environment data of the vehicle and engine running state data are obtained; inputting the environment data into an environment prediction model, and extracting current environment features; the current environment data and the engine operation state data are jointly input into an engine torque prediction model, and the maximum engine torque value in the current environment is obtained; inputting the current environment characteristics and the maximum torque value into a torque compensation control instruction generation model to generate a corresponding torque compensation control instruction; and finally, performing power compensation control on the engine according to the instruction. According to the method, a cooperative control framework integrating environment prediction, torque estimation and intelligent decision making is constructed, dynamic precise compensation of engine dynamic response in a complex environment is achieved, and the torque output precision and the self-adaptive control capacity of the vehicle under special working conditions are effectively improved.

Description

Technical Field

[0001] This application relates to the field of vehicle control technology, and in particular to an engine power compensation control method, electronic equipment, and vehicle. Background Technology

[0002] In existing technologies, when vehicles are driving in complex scenarios, they rely on static compensation strategies or driver experience, which leads to sluggish engine torque response, discontinuous power output, and insufficient control stability. This not only affects the vehicle's power and passability but also increases the driving burden and safety risks. Summary of the Invention

[0003] In view of this, the purpose of this application is to provide an engine power compensation control method, electronic equipment and vehicle to solve at least some of the technical problems in the related art.

[0004] To achieve the above objectives, this application provides an engine power compensation control method, which includes: Acquire the vehicle's current environmental data and engine operating status data; The current environmental data is input into the environmental prediction model to obtain the current environmental features. The environmental prediction model is a pre-trained neural network model that can determine the current environmental features based on the current environmental data. The current environmental data and the engine operating status data are input into the engine torque prediction model to obtain the maximum torque value of the engine under the current environment. The engine torque prediction model is a pre-trained neural network model that can determine the maximum torque value of the engine under the current environment based on the current environmental data and the engine operating status data. The current environmental characteristics and the maximum torque value are input into the torque compensation control command generation model to obtain the engine torque compensation control command. The torque compensation control command generation model is a neural network model that can determine the engine torque compensation control command based on the current environmental characteristics and the maximum torque value. The engine is subjected to power compensation control according to the torque compensation control command.

[0005] Based on the same inventive concept, this application also provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable by the processor, wherein the processor implements the method described above when executing the computer program.

[0006] Based on the same inventive concept, this application also provides a vehicle including the above-mentioned electronic devices.

[0007] As can be seen from the above, the engine power compensation control method, electronic equipment, and vehicle provided in this application achieve joint perception of the vehicle's external environment and internal power system operating conditions by simultaneously acquiring current environmental data and engine operating status data. This overcomes the limitations of traditional methods that rely solely on a single data source for static compensation, providing an information foundation for subsequent dynamic and precise compensation. By processing the aforementioned data using pre-trained, structurally adapted neural network models, key features and torque values ​​can be efficiently extracted from multiple dimensions such as temporal environmental fluctuations and engine nonlinear response characteristics, providing a core decision-making basis for generating precise compensation control commands. By fusing the current environmental features and the maximum torque value and inputting them into the torque compensation control command generation model, the future environmental change trend and the theoretical output limit of the current system are integrated. By adjusting the engine in real time according to the generated torque compensation control commands, the power system can achieve precise torque compensation under complex operating conditions, thereby significantly improving the engine's power response speed and output continuity, and effectively enhancing the vehicle's passability, control stability, and driving safety in special environments. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in this application or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a schematic flowchart of the engine power compensation control method according to an embodiment of this application; Figure 2 This is a schematic diagram of the engine power compensation device according to an embodiment of this application; Figure 3 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with specific embodiments and the accompanying drawings.

[0011] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this application should have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," and similar terms used in the embodiments of this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are only used to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0012] LSTM: Long Short-Term Memory. The paper mentions its use in environmental prediction models.

[0013] RBF: Radial Basis Function. The text mentions its use in engine torque prediction models (RBF-NN).

[0014] MSE: Mean Squared Error. It is mentioned in the text as being used as a loss function.

[0015] MAE: Mean Absolute Error. It is mentioned in the text as a metric used for model evaluation.

[0016] RMSE: Root Mean Square Error. It is mentioned in the text as a metric used for model evaluation.

[0017] ANFIS: Adaptive Neuro-Fuzzy Inference System. In this paper, it is used as a concrete implementation of the "torque compensation control command generation model".

[0018] DeepAR: A probabilistic temporal prediction model. In this paper, it is presented as a specific implementation of an "environmental prediction model".

[0019] ARIMA: Autoregressive Integrated Moving Average. It is mentioned in the text as an alternative environmental prediction model.

[0020] PID stands for Proportional-Integral-Derivative. It's mentioned in the text as an alternative control logic.

[0021] RBF-NN: Radial Basis Function Neural Network. In this paper, it is used as a specific implementation of an "engine torque prediction model".

[0022] OTA: Over-The-Air, meaning over-the-air downloading / remote diagnostics. The text mentions its use for data acquisition.

[0023] CAN: Controller Area Network. The text mentions its use for in-vehicle data communication.

[0024] ECU: Electronic Control Unit. In this text, it specifically refers to the engine control unit.

[0025] CPU: Central Processing Unit. Mentioned in the text as being used in electronic devices.

[0026] RAM: Random Access Memory. Mentioned in the text as being used in electronic devices.

[0027] ROM: Read Only Memory. Mentioned in the text as being used in electronic devices.

[0028] GPS: Global Positioning System. Mentioned in the text as being used for positioning.

[0029] USB: Universal Serial Bus. It is mentioned in the text as a communication interface.

[0030] WIFI: Wireless Fidelity. Mentioned in the text as a communication method.

[0031] ASIC: Application Specific Integrated Circuit. Mentioned in the text as a form of processor.

[0032] In related technologies, when vehicles operate in complex environments such as high altitudes or other conditions that cause drastic changes in intake conditions, their engine power compensation control mainly relies on static lookup table methods calibrated for plain environments or driver experience to achieve passive and lagging torque adjustments. These methods are ill-suited to dynamically changing low-pressure, low-oxygen environments, leading to inaccurate engine intake volume prediction, decreased combustion efficiency, and consequently, problems such as sluggish torque response, discontinuous and fluctuating power output, severely impacting vehicle performance, passability, and driving safety.

[0033] To improve environmental adaptability, some existing solutions introduce single environmental sensing devices such as barometric pressure sensors and altitude sensors to monitor external intake conditions in real time and compensate parameters such as fuel injection quantity and ignition angle according to preset linear or piecewise correction coefficients. However, such methods only react to current transient environmental data and lack the ability to predict short-term environmental changes, thus failing to achieve feedforward adjustment of the control strategy. Furthermore, the engine's torque response to thin air is highly nonlinear, and simple linear correction models cannot accurately characterize its dynamic characteristics, resulting in insufficient compensation accuracy. Other solutions attempt to introduce more engine state parameters for comprehensive judgment, but these typically employ control logic based on fixed rules or simple mapping tables. Their rule sets cannot cover all complex operating conditions, have poor generalization ability, and cannot quantitatively assess the engine's real-time output capability limits under current conditions.

[0034] In practical applications, the limitations of a single data source and the static nature of control logic lead to unstable system performance. Environmental sensor data itself contains noise and short-term disturbances, which can easily cause command oscillations if used directly for control. Engine status data only reflects the system's operating results under the current control command and does not reveal the deep mapping relationship between the environment and output capability, resulting in poor compensation effects. This may cause the engine to operate in a high-load, low-efficiency range or fail to meet the driver's transient torque requirements under conditions such as climbing hills and overtaking.

[0035] One intuitive improvement strategy is to simply weight or logically combine multiple signals such as environmental data and engine data. For example, when the air pressure is below a certain value and the throttle opening is above a certain value, a fixed amount of fuel injection compensation is added. However, due to the variability of the high-altitude environment and the strong coupling of the engine system, this static combination rule is difficult to achieve fine and adaptive control, and is prone to undercompensation or overcompensation. Another approach is to try to establish a complex intake and combustion model based on physical mechanisms for real-time calculation. However, such models have a high computational load, complex parameter calibration, and are extremely sensitive to sensor accuracy and individual vehicle differences, making them difficult to reliably deploy and apply in actual in-vehicle embedded systems.

[0036] Based on this, the applicant discovered that it is necessary to design a collaborative compensation scheme based on deep fusion of multi-source information and intelligent time-series prediction in the engine power control system to solve the problem of how to comprehensively utilize time-series environmental data and real-time engine status data in complex driving scenarios such as high altitudes, to proactively predict environmental change trends and accurately estimate the engine's current maximum torque output potential through machine learning models, and to generate dynamic and precise torque compensation control commands accordingly, thereby maximizing power output performance while ensuring engine operating stability.

[0037] The embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0038] The engine power compensation control method proposed in the embodiments of this application, such as Figure 1 As shown, it includes: S101. Obtain the vehicle's current environmental data and engine operating status data.

[0039] In practice, the current environmental data is a multi-dimensional time-series information set characterizing the external atmospheric environment of the vehicle, including instantaneous altitude, barometric pressure sensor output, ambient temperature, and derived features used to capture temporal changes (such as barometric pressure change rate and altitude change rate). This current environmental data can be collected and acquired through the vehicle's environmental perception sensors, over-the-air (OTA) diagnostic system, and onboard CAN bus. For example, real-time geographical location and altitude information can be obtained using a GPS / BeiDou module installed on the vehicle body; ambient barometric pressure values ​​can be collected at a frequency of 1Hz to 10Hz using a barometric pressure sensor installed near the vehicle's grille; and ambient temperature can be collected using a temperature sensor located outside the vehicle. Furthermore, combining dynamic information such as vehicle acceleration and speed change rate can more accurately explain nonlinear fluctuations in air pressure. All collected environmental data undergoes standardized cleaning and timestamp alignment on a cloud platform to remove noise and outliers, and is clustered based on geographic coordinates in typical plateau areas (such as altitudes of 2500-4000 meters) to form structured time-series data with environmental level labels. This data reflects the low oxygen levels and potential trends in the vehicle's environment, serving as a fundamental input for assessing engine intake conditions and predicting future environmental fluctuations.

[0040] In specific implementation, the engine operating status data is a multi-dimensional set of parameters characterizing the engine's real-time operating status and powertrain response, including: engine speed, throttle opening, instantaneous fuel injection quantity, air-fuel ratio, intake pressure, intake temperature, coolant temperature, deviation between actual torque output and commanded torque, and turbocharger pressure (if applicable). This engine operating status data can be collected and acquired through the vehicle's onboard network (such as the CAN bus) and corresponding sensors or controllers. Specifically: engine speed, throttle opening, fuel injection pulse width, air-fuel ratio, intake pressure and temperature, and coolant temperature are acquired through the engine ECU; the actual torque output is acquired through the vehicle powertrain controller or a dedicated torque sensor and compared with the commanded torque issued by the ECU to calculate the torque deviation; and boost pressure is acquired through a turbocharger pressure sensor. The above data is sampled at a frequency of 1Hz to 10Hz, and after timestamp alignment and normalization at the vehicle end or cloud, it is fused with current environmental data to form a structured input for model training and real-time inference. This data comprehensively reflects the engine's real-time combustion status, thermal load, and torque response capability under the current environment and control commands. It is the core basis for accurately estimating the engine's maximum available torque and formulating precise compensation strategies.

[0041] S102. Input the current environmental data into the environmental prediction model to obtain the current environmental features, wherein the environmental prediction model is a pre-trained neural network model that can determine the current environmental features based on the current environmental data.

[0042] In practice, the time-series prediction module integrated on the vehicle or cloud receives and processes current environmental data in real time. This current environmental data undergoes standardization and time window alignment consistent with the training phase (e.g., forming a time-series tensor containing 48 hours of historical data, one sampling point per minute, and a total of 2880 steps, covering variables such as historical altitude, air pressure, ambient temperature, vehicle acceleration, and throttle opening rate of change). The environmental prediction model is a DeepAR probabilistic time-series prediction model based on a Long Short-Term Memory (LSTM) network, pre-trained using massive amounts of high-altitude vehicle operation data. The environmental prediction model performs feature extraction and joint probabilistic modeling on this input time-series tensor. Its output includes not only predicted air pressure and altitude values ​​for the next 5 to 30 minutes but also confidence intervals for these predictions, collectively forming a current environmental feature vector. This vector encodes the short-term environmental change trends, fluctuation amplitudes, and uncertainties, providing crucial forward-looking input for subsequent control decisions.

[0043] S103. Input the current environmental data and the engine operating status data into the engine torque prediction model to obtain the maximum torque value of the engine under the current environment. The engine torque prediction model is a neural network model that is pre-trained and can determine the maximum torque value of the engine under the current environment based on the current environmental data and the engine operating status data.

[0044] In practice, the power system state estimation module receives and integrates current environmental data and engine operating status data in real time. These multidimensional data (such as real-time air pressure, altitude, engine speed, throttle opening, and intake pressure) undergo normalization processing consistent with the model training phase, forming a multidimensional feature vector. The engine torque prediction model is a radial basis function (RBF) based neural network model (RBF-NN) pre-trained using historical data containing engine operating conditions and measured torque labels under different altitudes and air pressures. The engine torque prediction model processes the input feature vector through its nonlinear mapping layer (Gaussian kernel function) and outputs a scalar value, namely the maximum torque value. This value represents the theoretically maximum effective torque that the engine hardware system can output under the current thin-oxygen environment and specific operating parameters, setting a theoretical upper limit for power compensation in the current environment for the controller.

[0045] S104. Input the current environmental characteristics and the maximum torque value into the torque compensation control command generation model to obtain the engine torque compensation control command. The torque compensation control command generation model is a neural network model that can determine the engine torque compensation control command based on the current environmental characteristics and the maximum torque value.

[0046] In practical implementation, the control strategy decision module receives the current environmental feature vector and the maximum torque value. The torque compensation control command generation model is an Adaptive Neural Fuzzy Inference System (ANFIS), whose internal structure includes a fuzzification layer, a rule base, and a defuzzification layer. The torque compensation control command generation model first fuzzifies the input variables (such as predicted air pressure, maximum torque value, current throttle opening, etc.) according to a preset Gaussian or triangular membership function. Then, based on a fuzzy rule base learned from high-altitude road condition data (e.g., "if the predicted air pressure is 'extremely low' and the maximum torque value is 'insufficient,' then output 'significantly increase' the throttle opening"), fuzzy inference is performed. Finally, the torque compensation control command generation model, through defuzzification calculation, outputs specific, executable torque compensation control commands, typically in the form of a correction amount for the target throttle opening, an increment value for the injection pulse width, or a bias value for the target air-fuel ratio.

[0047] S105. Perform power compensation control on the engine according to the torque compensation control command.

[0048] In practice, the vehicle's control execution unit (usually the engine electronic control unit (ECU) or its extended functional modules) receives torque compensation control commands in real time via the vehicle's bus (such as the CAN bus). The ECU parses these commands and converts them into precise control signals for the underlying actuators. For example, it drives the electronic throttle motor based on the throttle opening correction; adjusts the injector energizing time of the high-pressure fuel rail based on the injection pulse width correction; or comprehensively adjusts fuel injection and intake based on the air-fuel ratio bias. These adjustments aim to dynamically compensate for insufficient intake and reduced combustion efficiency caused by low air pressure and lean air, thereby ensuring that the engine's actual torque output is as close as possible to the driver's needs or the system's optimal target. The entire closed-loop chain from data perception to control execution is optimized, with the total latency controlled within 200 milliseconds, ensuring the real-time performance of power compensation and the vehicle's power responsiveness. Through this process, the system achieves feedforward, adaptive active power compensation in complex environments such as high altitudes.

[0049] By simultaneously acquiring current environmental data and engine operating status data, the above scheme achieves joint perception of the vehicle's external environment and internal power system operating conditions. This overcomes the limitations of traditional methods that rely solely on a single data source for static compensation, providing an information foundation for subsequent dynamic and precise compensation. By processing the data using pre-trained, structurally adapted neural network models, key features and torque values ​​can be efficiently extracted from multiple dimensions, such as temporal environmental fluctuations and engine nonlinear response characteristics, providing a core decision-making basis for generating precise compensation control commands. By fusing the current environmental features and the maximum torque value and inputting them into the torque compensation control command generation model, the future environmental change trend and the theoretical output limit of the current system are integrated. By adjusting the engine in real time according to the generated torque compensation control commands, the power system can achieve precise torque compensation under complex operating conditions, thereby significantly improving the engine's power response speed and output continuity, and effectively enhancing the vehicle's passability, control stability, and driving safety in special environments.

[0050] In some embodiments, in S102, the current environmental data is input into an environmental prediction model to obtain current environmental features. The terrain structure recognition model is a pre-trained neural network model that can determine the current terrain structure features based on the current terrain data.

[0051] In practical implementation, the corresponding environmental prediction models include: the Autoregressive Integral Moving Average (ARIMA) model based on classical time series analysis and the probabilistic time series prediction model based on deep learning (such as DeepAR). Among them, the DeepAR model based on Long Short-Term Memory (LSTM) network is preferred. The DeepAR model can jointly model and output the probability distribution of multiple time series variables, and is particularly suitable for forward-looking prediction and uncertainty quantification of variables such as air pressure and altitude in non-stationary environments at high altitudes.

[0052] The system collects massive amounts of historical multi-source time-series data (including historical altitude, air pressure, ambient temperature, and vehicle dynamics) gathered from vehicles operating in high-altitude regions. This data is then cleaned, standardized, and timestamped to construct a structured training dataset. During training, high-altitude ranges are embedded as auxiliary labels into the model input, and negative log-likelihood is used as the loss function. The initial DeepAR model parameters are iteratively optimized using the Adam optimizer. This training process is repeated, and cross-validation is performed on independent high-altitude validation sets (such as the Qinghai-Tibet Highway and Sichuan-Tibet Highway data) to ultimately obtain a stable environmental prediction model with good generalization capabilities.

[0053] Thus, after acquiring the current environmental data, the data is first standardized and aligned with the time window, consistent with the training phase (e.g., forming a time-series tensor containing 48 hours of historical data, with one sampling point per minute). The environmental prediction model receives this preprocessed time-series data. Internally, it first constructs a time-related sequence feature representation, and then uses multi-layer LSTM units to extract time-series patterns from the sequence, capturing temporal context features such as diurnal cycle variations and short-term fluctuations. Finally, the environmental prediction model outputs predicted air pressure and altitude values ​​within a specified future time window (e.g., 5-30 minutes), along with their probability distribution and confidence interval, collectively forming a current environmental feature vector. This vector encodes the future short-term environmental change trend, fluctuation amplitude, and uncertainty, providing crucial forward-looking input for subsequent compensation control.

[0054] S201. Standardize and align the current environmental data with the time window to obtain preprocessed current environmental data.

[0055] In practice, the data processing module integrated in the vehicle or cloud receives raw environmental data streams from multiple sources of sensors and the bus. This data includes instantaneous altitude sequences obtained via GPS / BeiDou modules, air pressure values ​​collected by atmospheric pressure sensors, ambient temperature values ​​collected by external temperature sensors, and vehicle dynamic data (such as acceleration and rate of change of speed) obtained from the CAN bus to interpret environmental fluctuations. The preprocessing unit first standardizes all numerical variables, scaling them to zero mean and unit variance based on the mean and standard deviation of each variable calculated offline under typical high-altitude conditions, using Z-score normalization to eliminate the influence of different physical dimensions on the training and inference of the environmental prediction model. Subsequently, a time window alignment operation is performed: considering the different sampling frequencies (e.g., 1Hz for altitude, 10Hz for air pressure) and transmission delays that may exist in the data streams, an interpolation algorithm (e.g., linear interpolation) is used to uniformly resample all variables to a preset reference frequency (e.g., 1Hz), and, based on the timestamps provided by the high-precision time synchronization module, strictly aligns the data from different sources to the same time axis. Finally, following the input format required by the DeepAR environmental prediction model, a fixed length (e.g., 2880 points, corresponding to 48 hours) of historical data is extracted from the current time backward to construct a regular, multivariate preprocessed time-series data tensor, which serves as the direct input to the environmental prediction model. This process ensures the quality and consistency of the input data, which is a prerequisite for the environmental prediction model to make accurate time-series predictions.

[0056] S202. Input the preprocessed current environmental data into the environmental prediction model. The processing procedure in the environmental prediction model is as follows: S211. Based on the preprocessed current environmental data, construct a time-related sequence feature representation.

[0057] In practice, the input layer of the environmental prediction model receives a preprocessed, regularized time-series data tensor. This tensor has dimensions of [time step, feature dimension], for example, [2880, N], where N represents the number of features (including core variables such as historical altitude, air pressure, ambient temperature, and vehicle acceleration). The environmental prediction model first concatenates the numerical features of each time step with a series of auxiliary time features (periodic timestamp information processed by the embedding layer, such as minutes of the day, day of the week, and whether it is a holiday). Simultaneously, based on the vehicle's current GPS coordinates or the most recent historical average altitude, it determines the plateau altitude range label (e.g., "3000-3500m") and converts this category label into an embedding vector. Finally, the environmental prediction model constructs an enhanced feature vector for each time step, integrating the original observations, temporal context, and geographic environmental context, forming a time-related sequence feature representation, providing a rich information foundation for subsequent deep time-series modeling.

[0058] S212. Extract the temporal pattern from the sequence features to obtain temporal context features.

[0059] In practice, an encoder composed of multiple layers of Long Short-Term Memory (LSTM) network units processes the generated sequence feature representation. Each LSTM layer sequentially reads the input sequence, and its internal gating mechanisms (input gate, forget gate, output gate) selectively remember long-term trends and cyclical patterns (such as daily cyclical air pressure changes) while capturing short-term fluctuations and abrupt changes (such as a sharp drop in air pressure caused by entering a tunnel). Through multi-layer stacking, the lower-level LSTM captures fine-grained local patterns, while the upper-level LSTM aggregates this information to form an understanding of longer-term and more complex dependencies. Finally, the hidden state output by the last LSTM layer at each time step encodes all historical information up to that moment and its contextual relationships; these states together constitute the temporal contextual features of the environmental prediction model.

[0060] S213. Output the current environment features based on the temporal context features.

[0061] In practice, the output layer of the environmental prediction model decodes based on the temporal context features of the last time step (or multiple context features weighted by an attention mechanism). The environmental prediction model is a probabilistic prediction model, and its output layer typically consists of two parallel fully connected layers: one for predicting the conditional distribution parameters of the future target variable (air pressure, altitude), such as the mean (μ) of a Gaussian distribution; and another for predicting the scale parameter of this distribution, such as the standard deviation (σ), to characterize the uncertainty of the prediction. The observed values ​​(or predicted sample values) of the previous time step are used as part of the autoregressive input to iteratively predict multiple future time steps (e.g., 5 to 30 minutes). Therefore, for each future prediction time point, a probability distribution is output. Ultimately, the current environmental features output by the environmental prediction model are a set of predicted values ​​(usually the mean of the distribution) of air pressure and altitude for a specified future time period, along with their corresponding confidence intervals (e.g., μ±kσ). This feature vector provides the subsequent controller with a clear and quantifiable trend of future environmental changes and its reliability, serving as the core basis for implementing feedforward compensation.

[0062] The above approach enables accurate dynamic trend modeling and forward-looking analysis of current environmental time-series data using environmental prediction models. It extracts current environmental characteristics that precisely reflect short-term trends and uncertainties in air pressure and altitude. These characteristics not only include specific predicted values ​​but also provide confidence intervals through probabilistic modeling, allowing the system to quantify and predict risks. This provides a forward-looking perception capability for impending environmental deterioration (such as a sudden drop in air pressure) or improvement, offering crucial and reliable prior information on the future environment for subsequent torque estimation and compensation control.

[0063] As a preferred embodiment, the training process of this environmental prediction model is as follows: (1) Construct the initial environmental prediction model. The model is based on a multilayer long short-term memory network (LSTM) and is constructed as a probabilistic time series prediction architecture (DeepAR). Its output layer is designed with two branches to simultaneously predict the conditional distribution parameters of future sequences (such as the mean and standard deviation of the Gaussian distribution).

[0064] (2) An initial training dataset was constructed based on the actual driving data of a large number of vehicles operated by automakers in plateau areas. Key historical multi-source time-series data were extracted through the OTA remote diagnostic system and the vehicle CAN bus, including instantaneous altitude values, barometric pressure sensor output values, ambient temperature, vehicle acceleration, vehicle speed change rate, and throttle opening changes. These data samples were labeled with their respective plateau altitude ranges (e.g., "2500-3000m") based on their geographical coordinates at the time of collection, and their corresponding real future environmental sequences were determined based on the actual data in subsequent time windows. The labeled samples constituted the basic training sample set.

[0065] (3) For each time series data sample, perform standardization and time window alignment preprocessing consistent with the inference stage. Standardize numerical variables using Z-score based on overall data statistics. Then, using timestamp embedding technology, generate auxiliary feature vectors containing periodic (e.g., minutes, hours, days of the week) information for each time step. Concatenate the standardized time series variables, timestamp embedding vectors, and altitude segment label embedding vectors to form an enhanced time series tensor used as input for the initial environmental prediction model.

[0066] (4) Input the constructed enhanced time series tensor into the initial environment prediction model. The input sequence is processed sequentially by its multi-layer LSTM encoder to learn the long-term dependencies and periodic patterns in historical information and output the context code for each time step. Based on the context code of the last time step, the decoder iteratively predicts the probability distribution parameters of the target variables (air pressure, altitude) for multiple future time steps in an autoregressive manner.

[0067] (5) Negative log-likelihood loss is used as the objective function for optimization. This loss function directly measures the difference between the probability distribution predicted by the model and the actual observed future sequence. The loss value is calculated, and all weight parameters of the environment prediction model are updated by combining the backpropagation algorithm with the Adam optimizer.

[0068] (6) Repeat steps (4) and (5) to perform multiple rounds of iterative training on the environmental prediction model using the basic training sample set. After each round of training, use an independent validation set containing data from typical plateau routes such as the Qinghai-Tibet Highway and the Sichuan-Tibet Highway to evaluate the performance of the environmental prediction model and monitor changes in negative log-likelihood loss and prediction error (such as MAE). When the performance of the environmental prediction model on the validation set tends to stabilize, end this stage of training to obtain the pre-trained environmental prediction model.

[0069] (7) To improve the accuracy and robustness of the environmental prediction model under extreme dynamic conditions (such as continuous uphill climbing, entering and exiting long tunnels, encountering strong crosswinds, etc.), an additional plateau challenging condition sample set is constructed. This sample set focuses on the above-mentioned difficult scenarios and performs more refined future sequence labeling.

[0070] (8) Adjust the pre-trained environment prediction model using a plateau challenging working condition sample set. In this stage, a significantly reduced learning rate is used for fine-tuning training, so that the parameters of the environment prediction model can be further refined based on the learned plateau general pattern to improve its ability to model severe and rapid environmental fluctuations.

[0071] (9) The adjusted environmental prediction model was comprehensively evaluated on a completely independent plateau test set covering multiple regions and seasons. Evaluation metrics included the mean absolute error (MAE) and root mean square error (RMSE) of the predicted values, as well as the calibration of probabilistic predictions (such as interval coverage probability). Finally, environmental prediction models whose performance met the preset thresholds were determined as deployable environmental prediction models.

[0072] In some embodiments, in S103, the current environmental data and the engine operating status data are input into the engine torque prediction model to obtain the maximum torque value of the engine under the current environment. The engine torque prediction model is a neural network model that is pre-trained and can determine the maximum torque value of the engine under the current environment based on the current environmental data and the engine operating status data.

[0073] The corresponding engine torque prediction models include: lookup table interpolation based on traditional calibration data, mean model based on physical mechanisms, and data-driven radial basis function neural network (RBF-NN). Among these, the radial basis function neural network (RBF-NN) is preferred. The RBF-NN possesses strong nonlinear fitting capabilities and good local response characteristics, enabling it to accurately learn and map the complex coupling relationship between engine torque output and environmental and operating condition parameters from high-dimensional multi-source inputs. It is particularly suitable for rapid and high-precision real-time estimation of the engine's theoretical maximum output potential under current conditions in dynamically changing low-pressure, low-oxygen environments.

[0074] The system collects massive amounts of historical multi-source synchronous data (including environmental data such as air pressure, altitude, and ambient temperature; and engine operating status data such as engine speed, throttle opening, intake pressure, intake temperature, air-fuel ratio, and coolant temperature) from vehicles operating in high-altitude areas. This data is then correlated with the engine's actual maximum output torque at the same moment, obtained from feedback via a high-precision torque sensor or a rigorously calibrated ECU, as labels. All collected data undergoes cleaning, standardization, and timestamp alignment, eliminating invalid and idling segments to construct a structured training dataset. During training, high-altitude segments are used as auxiliary information to guide the model's sensitivity to environments with varying oxygen content. Mean squared error (MSE) is used as the loss function, and the initial RBF-NN model parameters (including the center and width of the radial basis functions and output layer weights) are iteratively optimized using backpropagation and optimizers (such as gradient descent or Adam). This training process is repeated, and cross-validation is performed on independent high-altitude validation sets (such as typical operating condition data from the Qinghai-Tibet Highway and Sichuan-Tibet Highway) to ultimately obtain a stable engine torque prediction model with good generalization capabilities.

[0075] Thus, after acquiring current environmental data and engine operating status data, the data is first normalized in the same way as during the training phase to eliminate the influence of dimensions and align with the training data distribution. The engine torque prediction model receives this preprocessed multidimensional feature vector. Internally, it first performs a nonlinear mapping of the input space through a radial basis function (such as a Gaussian function) hidden layer to extract torque response features that can characterize complex coupling relationships, and then performs comprehensive weighting through a linear output layer. Finally, the engine torque prediction model outputs a scalar value, namely the maximum torque value of the engine under the current environment. This value represents the theoretically maximum effective torque that the engine hardware system can stably output under the current scarce oxygen environment and real-time operating condition constraints, providing crucial, quantified power system capability boundary information for subsequent compensation control.

[0076] S301. Normalize the current environmental data and the engine operating status data to obtain preprocessed model input data.

[0077] In practice, the feature processing unit on the vehicle or in the cloud simultaneously receives current environmental data (such as real-time air pressure, altitude, and ambient temperature) and engine operating status data (such as engine speed, throttle opening, intake pressure, intake temperature, and air-fuel ratio). First, each feature dimension is normalized. Based on offline statistics and calculations of the mean and standard deviation of each feature under various high-altitude operating conditions, the Z-score method is used to scale all input values ​​to near zero mean and unit variance. This step aims to eliminate the influence of different units of measurement and ensure that the distribution of input data aligns with the distribution of training data, thereby guaranteeing the stability and accuracy of inference. After processing, all normalized features are concatenated to form the preprocessed model input data, i.e., a fixed-dimensional numerical vector.

[0078] S302. Input the preprocessed model input data into the engine torque prediction model. The processing procedure in the engine torque prediction model is as follows: S311. Construct a multidimensional feature vector based on the preprocessed model input data.

[0079] In practice, the input layer of the engine torque prediction model (RBF-NN) directly receives the preprocessed model input data. Since the model's input dimensions are determined during training, the current input numerical vector requires no additional transformation and is directly treated as the multi-dimensional feature vector to be processed by the engine torque prediction model. This vector integrates the current environmental constraints (such as low air pressure and high altitude) with the engine's real-time operating point (such as speed and load), forming the basis for the nonlinear mapping of the engine torque prediction model.

[0080] S312. Perform nonlinear mapping on the multidimensional feature vector to extract torque response features.

[0081] In practice, the multidimensional feature vector is fed into the second layer of the engine torque prediction model: the radial basis function hidden layer. This layer contains multiple radial basis function neurons (or "nuclei"), each corresponding to a high-dimensional spatial center point learned during training. The engine torque prediction model calculates the Euclidean distance between the input vector and all center points, and nonlinearly maps these distances to "activation values" using radial basis functions such as Gaussian functions. The closer the input vector is to a certain center point, the higher the activation value of the corresponding neuron. Through this transformation, the high-dimensional linear space of the input is mapped to another feature space, in which complex nonlinear relationships (e.g., the torque saturation effect under the combined effects of air pressure, engine speed, and throttle opening) can be described by simpler functions (subsequent linear layers). The output of this hidden layer, i.e., the set of activation values ​​of all neurons, constitutes the torque response features represented internally by the engine torque prediction model, encoding the engine's response characteristics under the current input conditions.

[0082] S313. Based on the torque response characteristics, output the maximum torque value of the engine under the current environment.

[0083] In practice, the third layer of the engine torque prediction model, the linear output layer, receives torque response features from the hidden layers. This layer performs a linear weighted sum of the input features, ultimately outputting a scalar value. This scalar value is then denormalized (using the mean and standard deviation corresponding to the input normalization) to be converted back into physically meaningful torque units. This is the maximum torque value predicted by the engine torque prediction model under the current environment. This value represents the theoretical estimate of the maximum effective torque that the engine can stably output, learned from historical data, under the combined constraints of current environmental pressure, intake air density, and engine operating temperature.

[0084] The above scheme, through unified normalization preprocessing, ensures that the feature data input to the engine torque prediction model is consistent with the training distribution, laying a solid foundation for stable inference of the engine torque prediction model. By directly constructing a multi-dimensional feature vector representing the current comprehensive state, complete information required for accurate mapping is obtained. The core radial basis function nonlinear mapping layer, utilizing its local response characteristics, transforms the complex input space into a feature space that can clearly characterize the engine's response characteristics under the current specific environment and operating point, thereby capturing the complex coupling relationships that are difficult to describe by traditional methods. The final linear weighted output and inverse normalization transform the internal features into torque estimates with clear physical meaning.

[0085] As a preferred embodiment, the training process of the engine torque prediction model is as follows: (1) Constructing the initial engine torque prediction model. The model is based on a radial basis function neural network (RBF-NN) and adopts a standard three-layer structure. The first layer is the input layer, and its dimension is consistent with the number of selected features. The second layer is the radial basis function (RBF) hidden layer, and the number of nodes (center points) is predetermined through cross-validation, and a Gaussian function is used as the activation function. The third layer is the linear output layer, which outputs a single-value torque estimate. The initialization of model parameters (especially the position and width of the RBF center points) is performed based on the training sample input using methods such as K-Means clustering.

[0086] (2) Based on massive amounts of high-altitude vehicle operation data, an initial training dataset was constructed. Through OTA and the vehicle bus, historical model input features (including ambient air pressure, altitude, engine speed, throttle opening, intake pressure, intake temperature, air-fuel ratio, etc.) and corresponding label data (i.e., the actual maximum output torque measured by a high-precision torque sensor at the same time, or the strictly calibrated actual torque feedback value of the ECU) were collected. Simultaneously, each sample was labeled with its corresponding high-altitude range (e.g., "3000-3500m"). The labeled samples constituted the basic training sample set.

[0087] (3) Preprocess the training samples. For each dimension of the input feature in the sample, calculate its mean and standard deviation over the entire training set, and perform Z-score normalization to form a normalized input feature vector. The label torque value is also scaled accordingly to fit the model output range.

[0088] (4) Perform model forward propagation. Input the preprocessed sample input feature vector into the initial RBF-NN model. The model first calculates the distance between the input and all center points of the RBF layer and then activates the model.

[0089] (5) Define and calculate the loss function. The mean squared error (MSE) is used as the loss function to calculate the difference between the torque predicted by the model and the true labeled torque of the sample, and the loss value is obtained.

[0090] (6) Iterative optimization of model parameters. Supervised learning is adopted, and the model parameters are updated based on the loss value through backpropagation algorithm and optimizer (such as gradient descent or its variant). This process includes two main parts: optimizing the weights of the linear output layer using gradient descent; and optimizing the center point position and width parameters of the RBF layer using clustering or gradient-based methods (or phased fixed center points). Steps (4) and (5) are iteratively executed until the model loss on the training set converges.

[0091] (7) To improve the estimation accuracy of the model under extreme low pressure (such as altitude above 4000 meters) or special working conditions (such as turbine hysteresis region, high temperature thermal decay), a plateau challenging working condition sample set is constructed. This sample set focuses on the above-mentioned data sparse or modeling difficult regions and performs more accurate torque labeling.

[0092] (8) The pre-trained RBF-NN model is fine-tuned using a plateau challenging working condition sample set. In this stage, a small learning rate is used to focus on adjusting the center point parameters and output layer weights in the RBF layer that are related to the challenging working conditions, so that the model can better fit the torque response surface under extreme conditions.

[0093] (9) The adjusted model was comprehensively evaluated on a completely independent high-altitude test set. Evaluation metrics included the mean absolute error (MAE) and root mean square error (RMSE) between predicted and actual torque, as well as the error distribution at key altitudes. Finally, the model whose performance met the preset accuracy and robustness thresholds was determined as the deployable engine torque prediction model.

[0094] In some embodiments, in S104, the current environmental characteristics and the maximum torque value are input into the torque compensation control command generation model to obtain the engine torque compensation control command. The torque compensation control command generation model is a neural network model that can determine the engine torque compensation control command based on the current environmental characteristics and the maximum torque value.

[0095] In specific implementation, the corresponding torque compensation control command generation model includes: control logic based on classical proportional-integral-derivative (PID), control logic based on heuristic rules, and a data-driven adaptive neural fuzzy inference system (ANFIS). Preferably, the ANFIS model integrates the intuitive rule-based expression capability of fuzzy logic systems with the adaptive learning capability of neural networks. It can autonomously learn and optimize control strategies from data without completely relying on precise physical models. This makes it particularly suitable for generating dynamic, accurate, and robust torque compensation control commands based on forward-looking environmental predictions and engine capability boundaries in complex and variable high-altitude environments.

[0096] The system collects massive amounts of historical closed-loop control data from vehicles operating in high-altitude areas. This data includes: model input features, i.e., historical data corresponding to the current environmental characteristics and maximum torque value (such as historical environmental predictions, historical torque estimates, historical throttle openings, intake pressures, etc.); and model label data, i.e., ideal control commands obtained through actual vehicle operation verification or offline optimization under the input features, which best match the actual engine torque output with driver needs or the system's optimal goal (such as throttle opening correction and injection pulse width correction coefficient). All collected data undergoes cleaning, standardization, and alignment to remove unreasonable control segments, constructing a structured supervised learning training dataset. During the training phase, high-altitude segments are embedded as auxiliary labels to improve the model's discriminative power of control strategies under different oxygen content environments. The torque compensation control command generation model training employs a hybrid learning algorithm, where the membership function parameters (preconditions) are optimized using gradient descent (or the Adam optimizer) combined with backpropagation, while the rule consequent parameters are directly estimated using the least squares method. The initial ANFIS model parameters are iteratively optimized by minimizing the error between the model's output control command and the ideal control command in the label (using mean squared error, MSE, as the loss function). This training process is repeated, and cross-validation and policy evaluation are performed on an independent plateau validation set. Finally, a torque compensation control command generation model with stable generalization ability and excellent control performance is obtained.

[0097] Thus, after acquiring the current environmental characteristics and maximum torque value, this information is first fused and standardized with some real-time control variables to generate an input vector consistent with the training phase format. The torque compensation control command generation model receives this input vector. Internally, it first transforms the precise input into fuzzy semantic information through a fuzzification layer, then performs inference based on a pre-set fuzzy rule base through a rule layer to determine the activation strength of each rule, and finally aggregates and converts the fuzzy inference results into a precise numerical value through a defuzzification layer. Ultimately, it outputs one or more scalar values, namely the engine's torque compensation control command (such as the throttle target opening increment and the injection pulse width correction coefficient). This command aims to dynamically adjust engine control parameters based on predictions of the future environment and the current system's capability limits to compensate for torque loss caused by thin air, thereby enabling timely and smooth response of actual power output to driving demands. This command provides the core decision-making basis for subsequent direct execution control.

[0098] S401. Perform feature fusion on the current environmental features and the maximum torque value to obtain a fuzzy control input vector.

[0099] In practice, the control decision module receives current environmental characteristics (including predicted future air pressure / altitude values ​​and their confidence intervals) and the maximum torque value. First, feature fusion is performed: key elements from the current environmental characteristics (e.g., the predicted average air pressure value for the next 5 minutes, predicted air pressure change trend, and predicted altitude value) are concatenated with the maximum torque value and some key control variables collected in real time (e.g., current throttle opening and intake pressure). Then, this concatenated multidimensional vector is standardized to ensure that each component is within a preset universe of discourse. Finally, a unified and well-defined fuzzy control input vector is generated. This vector integrates future environmental trends, system capability limits, and the current control state, providing comprehensive input information for fuzzy inference.

[0100] S402. The fuzzy control input vector is input into the torque compensation control command generation model. The processing procedure in the torque compensation control command generation model is as follows: S411. The fuzzy control input vector is fuzzified according to the preset membership function to obtain the fuzzification result.

[0101] In practice, the fuzzification layer of the ANFIS model receives a fuzzy control input vector. For each input variable in this vector (such as "predicted air pressure", "maximum torque value", "current throttle opening"), the ANFIS model predefines 3 to 5 fuzzy subsets (such as "extremely low", "low", "normal"), each defined by a membership function (usually Gaussian or triangular). This layer calculates the membership degree (a value between 0 and 1) of the precise value of each input variable relative to the membership function of each fuzzy subset. For example, "predicted air pressure value" may belong to both the "low" fuzzy set (membership degree 0.7) and the "normal" fuzzy set (membership degree 0.3). The membership degrees of all input variables to all fuzzy subsets together constitute the fuzzification result, transforming precise numerical inputs into fuzzy semantic descriptions.

[0102] S412. Based on the fuzzification result and the preset fuzzy rule library, determine the corresponding fuzzy control rules.

[0103] In practice, the rule layer of the ANFIS model receives the fuzzification results. Internally, the ANFIS model stores a fuzzy rule base generated from historical plateau operational data. Each rule's antecedent defines a combination of multiple fuzzy subsets of input variables, while the consequent corresponds to a preliminary conclusion of an output control command. Based on the current fuzzification results, the rule layer calculates the activation strength of the antecedent for each rule (usually using a minimum or product operation). For example, if a rule states "if the predicted air pressure is 'low' and the maximum torque value is 'insufficient'", its activation strength is determined by the minimum membership degree of these two conditions. All rules are triggered to varying degrees according to their activation strength, thus determining the corresponding fuzzy control rules and their strengths.

[0104] S413. Based on the fuzzy control rules, determine the torque compensation control command.

[0105] In practice, subsequent layers of the ANFIS model (normalization layer, defuzzification layer) process the activated rules. First, the output of each rule is weighted (the weight is the rule activation strength). Then, the weighted outputs of all activated rules are aggregated into a precise value using a defuzzification method (such as weighted average or centering). This value is then scaled and transformed as necessary to generate the final torque compensation control command. This command is typically a specific actuator correction, such as the increment (%) of the target throttle opening, the correction coefficient of the injection pulse width, or the bias value of the target air-fuel ratio, which can be directly sent to the engine ECU for execution.

[0106] The above scheme enables the intelligent reasoning of features that integrate predictive information, capability boundaries, and real-time states using an Adaptive Neural Fuzzy Inference System (ANFIS), achieving the generation of torque compensation control commands under complex high-altitude conditions. This scheme combines the ability of fuzzy logic to handle expert knowledge and uncertainty with the ability of neural networks to automatically learn and optimize parameters from data.

[0107] As a preferred embodiment, the training process of the torque compensation control command generation model is as follows: (1) Constructing an initial torque compensation control command generation model. The five-layer network structure of the Adaptive Neural Fuzzy Inference System (ANFIS) is used as the initial model. Based on the number of input variables (such as predicted air pressure, maximum torque value, throttle opening, etc.) and the expected granularity, a set of Gaussian membership function parameters (center and width) are initialized for each input variable, and an initial fuzzy rule base consisting of the combination of all input fuzzy subsets with the number of rules being a Cartesian product is initialized. The consequent parameters of the rules are randomly initialized.

[0108] (2) A training dataset is constructed based on massive closed-loop operation data of vehicles operating on plateaus. Model input features at historical moments (i.e., historical environmental predictions, historical torque estimates, and historical operating condition data corresponding to the inference time step) and corresponding label data (i.e., the ideal control command obtained after verification or optimization under the given input conditions, which best matches the actual engine torque with the target torque, such as the optimal throttle correction) are collected. Simultaneously, the samples are labeled with their plateau altitude range. The labeled samples constitute the supervised learning training set.

[0109] (3) Preprocess the training samples. Standardize and fuse the input features in the same way as in the inference stage to generate fuzzy control input vectors. Standardize the output labels (control commands) as well.

[0110] (4) Perform model forward propagation. Input the preprocessed input vector into the initial ANFIS model, perform the complete fuzzification, rule reasoning, and defuzzification process, and obtain the control commands predicted by the model.

[0111] (5) Define and calculate the loss function. The mean squared error (MSE) is used as the loss function to calculate the difference between the control command predicted by the model and the ideal control command in the sample labels, and the loss value is obtained.

[0112] (6) Iterative optimization of model parameters is achieved using a hybrid learning algorithm. The training process employs a hybrid algorithm combining backpropagation and least squares estimation. In each training cycle or batch: Forward pass: With the membership function parameters fixed, the least squares method is used to directly calculate and update the consequent parameters of the rules in order to minimize the current loss.

[0113] Backpropagation: Fix the updated rule consequent parameters, backpropagate the error using gradient descent (or optimizers such as Adam), and update the membership function parameters (center and width) of the premise part (i.e., input variables). Iterate through steps (4) to (6) to gradually adjust the fuzzy partition and rule weights of the model to better fit the mapping relationship between the input and the ideal control output.

[0114] (7) To improve the smoothness and safety of the model's decision-making under extreme conditions or near the control boundary, a sample set of challenging control scenarios is constructed. This sample set focuses on scenarios such as steep altitude rise areas, sudden pressure change areas, and torque demand limit areas, and includes more refined and optimized control command labels.

[0115] (8) Fine-tune the pre-trained ANFIS model using a set of challenging scenario samples. Use a lower learning rate to further refine the model parameters, optimize its control strategy under critical conditions, avoid sudden instruction changes, and improve robustness.

[0116] (9) The adjusted model was evaluated using closed-loop or simulation tests on a completely independent high-altitude test dataset. Evaluation metrics included the accuracy of the control command in tracking the target torque, response speed, control stability at different altitudes, and the smoothness of the strategy. Finally, the model that met the performance requirements was selected as the deployable torque compensation control command generation model.

[0117] In some embodiments, S401 involves feature fusion of the current environmental features and the maximum torque value to obtain a fuzzy control input vector, including: S4011. Perform heterogeneous data alignment and standardization processing on the current environmental features to obtain the first structured features.

[0118] In practice, the current environmental features are multi-dimensional vectors output by the DeepAR model, which may contain predicted air pressure and altitude values ​​and their corresponding confidence intervals at different future time points (e.g., 5, 10, and 20 minutes). These data differ in temporal granularity. First, heterogeneous data alignment is performed: a unified decision reference time point (e.g., 5 minutes in the future) is selected, and the predicted air pressure and altitude values ​​at this time point are used as core environmental data. Simultaneously, the mean or median of the prediction sequence near this time point is calculated as a trend feature, and the width of the corresponding confidence interval is extracted as a measure of prediction uncertainty. Next, standardization is performed: based on offline established statistical distributions (mean and standard deviation) of air pressure and altitude for different plateau altitude ranges, the Z-score method is used to standardize these aligned environmental prediction values ​​and their derived features (e.g., trend and uncertainty), transforming them to a unified, dimensionless numerical range. After the above processing, a set of comparable and scale-consistent features is generated, namely the first structured features.

[0119] S4012. Perform heterogeneous data alignment and standardization processing on the maximum torque value to obtain the second structured feature.

[0120] In practice, the maximum torque value is a single scalar output by the RBF-NN model, but its physical meaning and numerical range need to be integrated with features from other models. Since the torque estimate (i.e., the maximum torque value) is an absolute value, its "heterogeneity" lies in its difference from the expected torque benchmark under different environmental conditions (especially altitude). Therefore, heterogeneous data alignment is first performed: based on the current altitude range of the vehicle (e.g., "3500-4000m"), the engine's "calibrated rated torque" or "torque loss benchmark" at that altitude range is obtained as a reference value. The relative ratio or deviation between the maximum torque value and this reference value is calculated, transforming it into a relative index characterizing "torque potential" or "torque loss degree." Then, standardization is performed: based on the statistical distribution (mean and standard deviation) of this relative index under different operating conditions in historical data, it is Z-score standardized. Finally, the standardized "torque potential" index is used as the second structured feature.

[0121] S4013. Determine the confidence level of the first feature of the first structured feature and the confidence level of the second feature of the second structured feature.

[0122] In practice, it is necessary to assign weights to the structured features from different models that reflect the reliability of their current estimates. The confidence score of the first feature mainly originates from the uncertainty output of the DeepAR (Environmental Prediction Model). It can be calculated based on the "prediction uncertainty metric" (such as the confidence interval width) extracted from S4011; for example, the narrower the confidence interval, the higher the confidence score. It can also be quantified by combining the probability distribution parameters of the model output (such as the standard deviation σ of the prediction distribution). The confidence score can be set to the form 1 / (1+k×σ), where k is the adjustment coefficient. The confidence score of the second feature mainly evaluates the estimation reliability of the Engine Torque Prediction Model (RBF-NN) under the current input conditions. This can be judged based on the model's "familiarity" with the distribution of the training data for the input point. For example, the minimum Euclidean distance between the current input feature vector and all center points of the RBF-NN hidden layer can be calculated. The smaller the distance, the closer the input point is to the trained operating condition area, and the higher the confidence score; the larger the distance, the more it is in an extrapolation state, and the confidence score should be reduced. The confidence score can be designed as a function inversely proportional to the minimum distance.

[0123] S4014. Based on the first feature confidence level and the second feature confidence level, the first structured feature and the second structured feature are weighted and fused to obtain the fuzzy control input vector.

[0124] In practice, after obtaining the confidence scores of the two features (which can be normalized to weight coefficients, such as using the softmax function), a weighted fusion is performed. The first structured feature vector is multiplied by the first feature confidence weight, and the second structured feature scalar is multiplied by the second feature confidence weight. Then, the weighted features are concatenated. To provide the fuzzy controller with a more comprehensive decision-making context, current real-time, key engine control state variables (such as current throttle opening and current intake pressure) are also standardized and concatenated into the vector. Finally, the entire concatenated vector is standardized or scaled to ensure that the numerical range of each dimension fits the preset universe of discourse of the subsequent ANFIS model input layer (membership function). After this step, a unified fuzzy control input vector that integrates look-ahead environmental information, system capability assessment and reliability, and real-time state context is finally generated.

[0125] The above scheme enables effective alignment, standardization, and credibility-weighted fusion of predictive and estimated features from different sources and with varying properties. This not only solves the problems of scale inconsistencies and importance confusion that may arise from directly splicing multi-source heterogeneous data, but more importantly, by introducing a confidence mechanism, the fusion process can dynamically reflect the estimation quality of each preceding model at the current moment and under the current operating conditions. This endows the fuzzy controller with the ability to distinguish the reliability of information during decision-making, allowing it to rely more on engine torque estimation when environmental prediction uncertainty is high, or to focus more on environmental trend judgment when torque estimation reliability is low, thereby significantly improving the adaptability and robustness of the entire compensation control strategy in the complex dynamic environment of high-altitude areas.

[0126] In some embodiments, the method further includes: Adjustments are made to the environmental prediction model, the engine torque prediction model, and the torque compensation control command generation model, including: Historical vehicle operation data is acquired to form an adjustment dataset.

[0127] In practice, historical vehicle operation data is acquired to form an adjustment dataset. This adjustment dataset is not the raw, massive dataset, but rather carefully selected vehicle operation data specifically from typical high-altitude regions (e.g., the Tanggula Pass section of the Qinghai-Tibet Highway at altitudes above 3000 meters, the Litang-Batang section of the Sichuan-Tibet Highway, and the Baima Snow Mountain section of the Yunnan-Tibet Highway). This data is transmitted back via onboard terminals and includes multi-source time-series and state data collected in the high-altitude environment, isomorphic to the original training phase but with different distributions. The data undergoes the same cleaning, standardization, and timestamp alignment processes as the initial training, ensuring it contains the input features required by the model and corresponding ground truth labels (for supervised fine-tuning) or effective closed-loop feedback information (for self-supervised or reinforcement learning fine-tuning).

[0128] Based on the adjusted dataset, the environmental prediction model, the engine torque prediction model, and the torque compensation control command generation model are adjusted respectively; The adjustment to the environmental prediction model includes updating the weight parameters of the long short-term memory network of the environmental prediction model.

[0129] In practice, the DeepAR model is adjusted using a fine-tuning strategy from transfer learning. First, a pre-trained environment prediction model is loaded as a foundation. Then, the model is further trained using the aforementioned plateau adjustment dataset (containing time-series segments unique to the plateau, such as sudden pressure changes and rapid altitude variations). During this process, the original overall network structure of the model (including the number of LSTM layers and neurons) is preserved, but its weight parameters are allowed to be updated slightly based on the plateau data. A relatively low learning rate is typically used to make the LSTM network more sensitive to plateau-specific short-term pressure drops (e.g., entering a tunnel), sudden pressure increases (e.g., driving out of a valley), and fluctuations caused by winds in complex terrain, while retaining the general time-series patterns already learned by the model. By minimizing the negative log-likelihood loss on the plateau dataset, the weights of the LSTM are iteratively updated, making the model's output prediction distribution more closely match the actual observed sequences on the plateau.

[0130] The adjustment of the engine torque prediction model includes: refitting the center point and width parameters of the radial basis function neural network of the engine torque prediction model.

[0131] In practice, the adjustment of the RBF-NN model focuses on reconstructing the local response basis of its nonlinear mapping. Because the low air pressure and low oxygen environment of high-altitude areas alters the engine's torque response surface, the original RBF center points and widths determined based on broader data (including a large amount of plains data) may not optimally characterize high-altitude operating conditions. During adjustment, the engine operating conditions and measured torque data pairs from the high-altitude adjustment dataset are used to re-execute the center point selection process. For example, clustering algorithms such as K-Means are used to determine new center point locations only within the input feature space of the high-altitude data samples, making these center points more representative of typical high-altitude operating condition clusters. Simultaneously, based on the density of the high-altitude data distribution, the width (or variance) parameter of each radial basis function is recalculated or adjusted to control its local influence range in the high-altitude feature space. After this, the new RBF layer parameters can be fixed, and the weights of the linear output layer can be retrained (or fine-tuned) using high-altitude data, specifically improving the fitting accuracy of the entire model in the high-altitude region.

[0132] The adjustment of the torque compensation control command generation model includes: expanding the fuzzy rule base of the torque compensation control command generation model and adjusting the membership function parameters of the torque compensation control command generation model.

[0133] In practice, the adjustment of the ANFIS controller is divided into two parts: rule base enhancement and parameter refinement. Expanding the fuzzy rule base involves mining new and effective "if-then" control rules from the high-altitude adjustment dataset. New fuzzy rules are generated by analyzing the correspondence between successful control inputs (predictive features, torque estimates, etc.) and ideal control outputs (optimized compensation commands) under specific high-altitude scenarios (such as continuous climbing and extremely low air pressure). For example, "If the average predicted air pressure for the next 5 minutes is 'extremely low' and the current torque potential is 'severely insufficient,' then the injection pulse width correction should be 'significantly increased.'" These new rules are then added to the existing rule base. Adjusting membership function parameters: The parameters of the model's original membership functions (such as the membership function used to classify "air pressure") are initialized based on a broader range of data. In high-altitude environments, the typical range of variable values ​​changes (e.g., the "normal" air pressure interval shifts to the left overall). Therefore, it is necessary to utilize plateau data and fine-tune the center position and width of these Gaussian or triangular membership functions through supervised gradient descent, so that fuzzy partitions (such as "extremely low", "low", "normal") can more accurately cover and describe the actual distribution range of each input variable in the plateau environment, thereby making the premise part of fuzzy inference more accurate.

[0134] Through the above schemes, the parameters of the DeepAR environmental prediction model are fine-tuned, enabling its LSTM network to retain its general time-series understanding capabilities while specifically enhancing its ability to capture plateau-specific fluctuation patterns (such as sudden changes in air pressure), thereby improving the local accuracy of predicting future environmental trends. The engine torque prediction model (RBF-NN) undergoes structural reconstruction. By repositioning the center of the radial basis functions and adjusting their influence range in a high-dimensional feature space, it essentially "redraws" the torque response characteristic surface of the engine in the thin-oxygen environment of the plateau, significantly improving the estimation accuracy of the maximum torque value. The torque compensation control command generation model (ANFIS) undergoes dual optimization of knowledge and parameters. On the one hand, it expands the rule base, injecting specialized knowledge extracted from successful plateau control experience; on the other hand, it adjusts the membership function to closely align its fuzzy semantic division with the actual distribution of variables in the plateau environment, thereby generating more refined and robust control commands that better meet the actual needs of the plateau. This hierarchical and differentiated adjustment method avoids the high computational cost and potential loss of valuable prior knowledge caused by simple global retraining. It enables the entire dynamic compensation system to quickly adapt to the data distribution migration from the standard plain environment to the extreme plateau environment, effectively overcoming the problem of model performance degradation caused by the mismatch between training data and the real application environment, namely the domain gap.

[0135] In some embodiments, the method further includes: During vehicle operation, the output data of the environmental prediction model and the engine torque prediction model, the torque compensation control command execution data, and the actual torque feedback data of the engine are continuously collected.

[0136] In practical implementation, a lightweight data logger is integrated into the vehicle-mounted embedded system. This logger collects and caches a closed-loop feedback sample at a fixed time granularity (e.g., every minute). This sample includes: the input (historical environmental time series) and output (current environmental characteristics, i.e., future air pressure / altitude prediction sequence and its confidence interval) of the current environmental prediction model; the input (current environmental and operating condition data) and output (maximum torque value) of the engine torque prediction model; the torque compensation control command issued at the current moment (such as throttle increment command); and the actual engine torque feedback data obtained through the ECU or sensors, corresponding to the execution of the control command. After compression and anonymization, the data is uploaded in batches to the cloud-based strategy learning platform via the vehicle network.

[0137] For the environmental prediction model, based on the rate of change of the vehicle's current environmental characteristics output by the environmental prediction model at the current time point relative to the rate of change of the vehicle's current environmental characteristics output by the environmental prediction model at the previous time point, the prediction time window or confidence weight parameter of the environmental prediction model is adjusted.

[0138] In practice, the cloud-based learning platform's optimization module for the DeepAR environmental prediction model analyzes the uploaded data. It calculates the rate of change of predicted values ​​for a specific target (e.g., air pressure in the next 5 minutes) over consecutive time points. If the rate of change remains consistently high (e.g., exceeding a threshold even after moving average filtering), it indicates that the current environment is in a period of rapid fluctuation. At this point, the optimization module may trigger two adjustments: Dynamically adjusting the prediction time window: For example, temporarily shortening the model's primary prediction window from "the next 5-30 minutes" to "the next 2-15 minutes," making the model more focused on recent dramatic changes and improving short-term prediction accuracy. Adjusting confidence weight parameters: Within the model, learnable confidence weights can be introduced for input features from different time points or sources (e.g., the contribution of vehicle acceleration to air pressure prediction). When environmental changes are drastic, these weights are fine-tuned through online learning, making the model rely more heavily on features that better explain the current dynamics during inference. Adjustments are made based on newly uploaded plateau time-series data, achieved through several rounds of gradient descent updates, forming a new snapshot of model parameters.

[0139] For the engine torque prediction model, the kernel function weight distribution of the neural network of the engine torque prediction model is adjusted according to the deviation between the maximum torque value output by the engine torque prediction model and the actual torque feedback data.

[0140] In practical implementation, for the engine torque prediction model (RBF-NN), the online learning module calculates the deviation between the maximum torque value output by the model and the actual torque feedback achievable under the same operating conditions (which needs to be collected under conditions where the throttle opening is large enough to reflect the maximum capacity). The statistical measure of this deviation (such as the mean absolute error within the sliding window) is used to evaluate the model's accuracy in the current operating range. If the deviation continues to exceed the preset tolerance, it indicates that there is an error in the model's mapping relationship for the current individual vehicle (potentially due to hardware aging or changes in the intake system state) or the current environmental range. The learning module will use newly collected (input features, actual torque) sample pairs to fine-tune the weights of the linear output layer of the RBF-NN online. Furthermore, a dynamic kernel function weight adjustment strategy can be adopted, that is, introducing a learnable weight factor into each radial basis function (RBF kernel). Through online learning, the contribution of each kernel function to the final output is dynamically adjusted based on new data, thereby more flexibly correcting the torque response surface without changing the center point position, making it better fit the actual characteristics of the current vehicle.

[0141] For the torque compensation control command generation model, the execution parameters of the torque compensation control command are determined based on the execution data of the torque compensation control command, and the membership function parameters of the torque compensation control command generation model are adjusted based on the execution parameters of the torque compensation control command. Determining the execution parameters of the torque compensation control command based on the execution data of the torque compensation control command includes: If the deviation between the actual torque feedback data and the torque corresponding to the torque compensation control command continues to be greater than a first threshold within a preset time period, the execution state of the torque compensation control command is determined to be an invalid execution state, and the invalid execution state is determined as the execution parameter of the torque compensation control command.

[0142] In practical implementation, for the Torque Compensation Control Command Generation Model (ANFIS), the core of online learning is to evaluate the control effect and optimize the rule premises. First, the execution parameters are determined: the control effect is judged by comparing the actual torque feedback of the engine after the torque compensation control command is executed with the expected target torque. For example, a key execution parameter can be defined as "torque response underperformance rate" or "command tracking error." If this error parameter continuously exceeds a threshold within a preset time (e.g., 5 minutes), the current control rule is deemed ineffective under this type of input condition. Based on this judgment, adjustments to the ANFIS model are triggered. These adjustments primarily target the fuzzification layer: using recently collected input-output data pairs that lead to poor control performance, the parameters of the relevant membership functions are fine-tuned using online gradient descent. For example, the center and width of the fuzzy set membership functions for "extremely low air pressure" or "insufficient torque potential" are adjusted. This allows the model to dynamically optimize the fuzzy partitioning of input variables based on actual control effect feedback, thus triggering more suitable fuzzy rules and outputting more effective compensation commands when similar situations arise in the future. For extremely ineffective rules, a more complex rule base reconstruction process can be triggered.

[0143] Through the above approach, the system can achieve continuous adaptive optimization and personalized adaptation throughout the vehicle's entire lifecycle. The environmental prediction model can dynamically adjust its prediction strategy based on the real-time environment, the engine torque prediction model can learn and track the performance degradation or changes of individual vehicles, and the torque compensation controller can optimize its decision logic based on feedback from actual control effects. This significantly improves the system's stability and reliability in long-term use, enabling it to cope with various uncertainties in complex high-altitude conditions that are difficult to fully cover through initial training. It is a key innovation for maintaining excellent performance and long-term competitiveness after large-scale deployment of the technical solution. All online optimizations are performed in a secure cloud environment and updates are only pushed after simulation verification, ensuring the safety and controllability of the learning process.

[0144] Based on the same inventive concept, and corresponding to any of the above embodiments, this application also provides an engine power compensation control device.

[0145] refer to Figure 2 The device includes: The multi-source information acquisition module 210 is configured to acquire the vehicle's current environmental data and engine operating status data; The environmental feature prediction module 220 is configured to input the current environmental data into the environmental prediction model to obtain the current environmental features, wherein the environmental prediction model is a pre-trained neural network model that can determine the current environmental features based on the current environmental data. The maximum torque prediction module 230 is configured to input the current environmental data and the engine operating status data into the engine torque prediction model to obtain the maximum torque value of the engine under the current environment. The engine torque prediction model is a pre-trained neural network model that can determine the maximum torque value of the engine under the current environment based on the current environmental data and the engine operating status data. The control command generation module 240 is configured to input the current environmental features and the maximum torque value into the torque compensation control command generation model to obtain the torque compensation control command of the engine. The torque compensation control command generation model is a neural network model that can determine the torque compensation control command of the engine based on the current environmental features and the maximum torque value. The power compensation control module 250 is configured to perform power compensation control on the engine according to the torque compensation control command.

[0146] In some embodiments, the environmental feature prediction module 220 includes: An environmental data preprocessing unit is configured to standardize and align the current environmental data with a time window to obtain preprocessed current environmental data. The time-series feature construction unit is configured to construct a time-related sequence feature representation based on the preprocessed current environment data; The temporal pattern extraction unit is configured to extract temporal patterns from the sequence features to obtain temporal context features; The environment feature determination unit is configured to output the current environment features based on the temporal context features.

[0147] In some embodiments, the maximum torque prediction module 230 includes: The torque prediction preprocessing unit is configured to normalize the current environmental data and the engine operating status data to obtain preprocessed model input data. The feature vector construction unit is configured to construct a multidimensional feature vector based on the preprocessed model input data; A nonlinear mapping unit is configured to perform nonlinear mapping on the multidimensional feature vector to extract torque response features; The maximum torque determination unit is configured to output the maximum torque value of the engine under the current environment based on the torque response characteristics.

[0148] In some embodiments, the control command generation module 240 includes: a feature fusion unit configured to perform feature fusion on the current environmental features and the maximum torque value to obtain a fuzzy control input vector; The input fuzzification unit is configured to perform fuzzification processing on the fuzzy control input vector according to a preset membership function to obtain the fuzzification result; The rule reasoning unit is configured to determine the corresponding fuzzy control rules based on the fuzzification result and a preset fuzzy rule library; The control command determination unit is configured to determine torque compensation control commands based on the fuzzy control rules.

[0149] In some embodiments, the feature fusion unit specifically includes: The first feature processing subunit is configured to perform heterogeneous data alignment and standardization processing on the current environmental features to obtain the first structured features; The second feature processing subunit is configured to perform heterogeneous data alignment and standardization processing on the maximum torque value to obtain a second structured feature. The confidence level determination subunit is configured to determine the first feature confidence level of the first structured feature and the second feature confidence level of the second structured feature; The weighted fusion subunit is configured to perform weighted fusion of the first structured feature and the second structured feature based on the first feature confidence and the second feature confidence to obtain the fuzzy control input vector.

[0150] In some embodiments, the apparatus further includes: The model adjustment module is configured to adjust the environmental prediction model, the engine torque prediction model, and the torque compensation control command generation model, specifically including: The adjusted dataset building unit is configured to acquire historical vehicle operation data to form the adjusted dataset. The model parameter optimization unit is configured to adjust the environmental prediction model, the engine torque prediction model, and the torque compensation control command generation model respectively based on the adjustment dataset; The adjustment to the environmental prediction model includes updating the weight parameters of its long short-term memory network. The adjustment to the engine torque prediction model includes: refitting the center point and width parameters of its radial basis function neural network; The adjustment of the torque compensation control command generation model includes: expanding its fuzzy rule base and adjusting its membership function parameters.

[0151] In some embodiments, the apparatus further includes: The online learning module is configured to continuously collect output data from the environmental prediction model and the engine torque prediction model, torque compensation control command execution data, and actual engine torque feedback data during vehicle operation, and dynamically optimize the model based on this data. Specifically, this includes: The data acquisition and feedback unit is configured to collect and upload closed-loop feedback samples to the cloud learning platform. The online optimization unit for the environmental prediction model is configured to adjust the prediction time window or confidence weight parameters of the environmental prediction model based on the rate of change of environmental characteristics at continuous time points. The online optimization unit for the torque prediction model is configured to adjust the kernel function weight distribution of the neural network of the engine torque prediction model based on the deviation between the maximum torque value and the actual torque feedback data. The online optimization unit for the control command generation model is configured to adjust the membership function parameters of the torque compensation control command generation model based on the execution parameters of the torque compensation control command.

[0152] For ease of description, the above devices are described in terms of function, divided into various modules. Of course, in implementing this application, the functions of each module can be implemented in one or more software and / or hardware.

[0153] The apparatus of the above embodiments is used to implement the corresponding method in any of the foregoing embodiments and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0154] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides an electronic device, the electronic device including a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the methods described in any of the above embodiments.

[0155] Figure 3 This embodiment illustrates a more specific hardware structure of an electronic device, which may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.

[0156] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0157] The memory 1020 can be implemented in the form of ROM (Read-Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0158] The input / output interface 1030 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components within the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touchscreens, microphones, various sensors, etc., while output devices may include displays, speakers, vibrators, indicator lights, etc.

[0159] The communication interface 1040 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0160] Bus 1050 includes a pathway for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.

[0161] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0162] The electronic devices described above are used to implement the corresponding methods in any of the foregoing embodiments and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0163] Based on the same inventive concept, this application also provides a vehicle including the device or electronic device described in the above embodiments. The beneficial effects of embodiments having corresponding devices or electronic devices will not be elaborated further here.

[0164] It is understood that before using the technical solutions of the various embodiments in this disclosure, users will be informed of the type, scope of use, and usage scenarios of the personal information involved in an appropriate manner, and user authorization will be obtained.

[0165] For example, upon receiving a user's active request, a prompt message is sent to the user to explicitly inform them that the requested operation will require the acquisition and use of the user's personal information. This allows the user to independently choose, based on the prompt message, whether to provide personal information to the software or hardware such as electronic devices, applications, servers, or storage media performing the operations of this disclosed technical solution.

[0166] As an optional but not limited implementation, in response to a user's active request, sending a prompt message to the user can be done via a pop-up window, where the prompt message can be presented in text format. Furthermore, the pop-up window can also include a selection control allowing the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0167] It is understood that the above notification and user authorization process are merely illustrative and do not constitute a limitation on the implementation of this disclosure. Other methods that comply with relevant laws and regulations may also be applied to the implementation of this disclosure.

[0168] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this application (including the claims) is limited to these examples; within the framework of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this application as described above, which are not provided in the details for the sake of brevity.

[0169] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of this application, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of this application, and this also takes into account the fact that the details of the implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of this application will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuits) have been set forth to describe exemplary embodiments of this application, it will be apparent to those skilled in the art that the embodiments of this application can be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0170] Although this application has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0171] The embodiments of this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this application should be included within the protection scope of this application.

Claims

1. An engine power compensation control method, characterized in that, include: Acquire the vehicle's current environmental data and engine operating status data; The current environmental data is input into the environmental prediction model to obtain the current environmental features. The environmental prediction model is a pre-trained neural network model that can determine the current environmental features based on the current environmental data. The current environmental data and the engine operating status data are input into the engine torque prediction model to obtain the maximum torque value of the engine under the current environment. The engine torque prediction model is a pre-trained neural network model that can determine the maximum torque value of the engine under the current environment based on the current environmental data and the engine operating status data. The current environmental characteristics and the maximum torque value are input into the torque compensation control command generation model to obtain the engine torque compensation control command. The torque compensation control command generation model is a neural network model that can determine the engine torque compensation control command based on the current environmental characteristics and the maximum torque value. The engine is subjected to power compensation control according to the torque compensation control command.

2. The method according to claim 1, characterized in that, The step of inputting the current environmental data into the environmental prediction model to obtain the current environmental characteristics includes: The current environmental data is standardized and aligned with the time window to obtain preprocessed current environmental data; The preprocessed current environmental data is input into the environmental prediction model, and the processing procedure in the environmental prediction model is as follows: Based on the preprocessed current environmental data, a time-related sequence feature representation is constructed. Temporal pattern extraction is performed on the sequence features to obtain temporal context features; Based on the temporal context features, output the current environment features.

3. The method according to claim 1, characterized in that, The step of inputting the current environmental data and the engine operating status data into the engine torque prediction model to obtain the maximum torque value of the engine under the current environment includes: The current environmental data and the engine operating status data are normalized to obtain preprocessed model input data; The preprocessed model input data is input into the engine torque prediction model, and the processing procedure in the engine torque prediction model is as follows: Based on the preprocessed model input data, a multidimensional feature vector is constructed. The torque response features are extracted by performing a nonlinear mapping on the multidimensional feature vector; Based on the torque response characteristics, the maximum torque value of the engine under the current environment is output.

4. The method according to claim 1, characterized in that, The step of inputting the current environmental characteristics and the maximum torque value into the torque compensation control command generation model to obtain the engine's torque compensation control command includes: The current environmental features and the maximum torque value are fused to obtain a fuzzy control input vector; The fuzzy control input vector is input into the torque compensation control command generation model, and the processing procedure in the torque compensation control command generation model is as follows: The fuzzy control input vector is fuzzified according to a preset membership function to obtain the fuzzification result; Based on the fuzzification result and the preset fuzzy rule library, the corresponding fuzzy control rules are determined; Based on the fuzzy control rules, torque compensation control commands are determined.

5. The method according to claim 4, characterized in that, The step of fusing the current environmental features and the maximum torque value to obtain a fuzzy control input vector includes: The current environmental features are subjected to heterogeneous data alignment and standardization processing to obtain the first structured features; The maximum torque value is subjected to heterogeneous data alignment and standardization processing to obtain the second structured feature; Determine the confidence level of the first feature of the first structured feature and the confidence level of the second feature of the second structured feature; Based on the confidence scores of the first and second features, the first and second structured features are weighted and fused to obtain the fuzzy control input vector.

6. The method according to claim 1, characterized in that, The method further includes: Adjustments are made to the environmental prediction model, the engine torque prediction model, and the torque compensation control command generation model, including: Historical vehicle operation data is acquired to form an adjustment dataset; Based on the adjusted dataset, the environmental prediction model, the engine torque prediction model, and the torque compensation control command generation model are adjusted respectively; The adjustment to the environmental prediction model includes updating the weight parameters of the long short-term memory network of the environmental prediction model; The adjustment of the engine torque prediction model includes: refitting the center point and width parameters of the radial basis function neural network of the engine torque prediction model; The adjustment of the torque compensation control command generation model includes: expanding the fuzzy rule base of the torque compensation control command generation model and adjusting the membership function parameters of the torque compensation control command generation model.

7. The method according to claim 1, characterized in that, The method further includes: During vehicle operation, the output data of the environmental prediction model and the engine torque prediction model, the torque compensation control command execution data, and the actual torque feedback data of the engine are continuously collected. For the environmental prediction model, based on the current environmental characteristics of the vehicle output by the environmental prediction model at the current time point and the rate of change of the current environmental characteristics of the vehicle output by the environmental prediction model at the previous time point, the prediction time window or confidence weight parameter of the environmental prediction model is adjusted. For the engine torque prediction model, the kernel function weight distribution of the neural network of the engine torque prediction model is adjusted according to the deviation between the maximum torque value output by the engine torque prediction model and the actual torque feedback data. For the torque compensation control command generation model, the execution parameters of the torque compensation control command are determined based on the execution data of the torque compensation control command, and the membership function parameters of the torque compensation control command generation model are adjusted based on the execution parameters of the torque compensation control command.

8. The method according to claim 7, characterized in that, Determining the execution parameters of the torque compensation control command based on the execution data of the torque compensation control command includes: If the deviation between the actual torque feedback data and the torque corresponding to the torque compensation control command continues to be greater than a first threshold within a preset time, the execution state of the torque compensation control command is determined to be an invalid execution state, and the invalid execution state is determined as the execution parameter of the torque compensation control command.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 8.

10. A vehicle, characterized in that, Includes the electronic device as described in claim 9.

Images