Control method for combustion system of direct injection methanol generator based on multi-objective optimization

Abstract

The invention relates to the technical field of process control, in particular to an in-cylinder direct injection methanol generator combustion system control method based on multi-objective optimization, which comprises the following steps: acquiring multi-source sensor data of a methanol generator, performing feature extraction and evaluation on the multi-source sensor data, and generating a control weight distribution matrix; performing weighted fusion by using the control weight distribution matrix to obtain a fusion control vector; the method comprises the following steps: constructing an in-cylinder process control model, adaptively adjusting model parameters in combination with physicochemical characteristics of methanol fuel and in-cylinder combustion characteristics, comprehensively evaluating energy conversion efficiency, emission level and operation stability by adopting a multi-objective optimization algorithm based on the process control model and in combination with real-time working condition data, and carrying out iterative search on a Pareto optimal solution set. Dynamically generating adaptive control parameters; and generating a control instruction according to the self-adaptive control parameters, and adjusting the operation parameters of the methanol generator in real time. Accurate control over the combustion system is achieved by dynamically adjusting the control parameters of the methanol generator.

Description

Technical Field

[0001] This invention relates to the field of process control technology, specifically to a control method for the combustion system of an in-cylinder direct injection methanol generator based on multi-objective optimization. Background Technology

[0002] In existing technologies, the process control of methanol generators mainly relies on traditional PID controllers or static model control strategies based on fixed parameters. These strategies adjust system operating parameters through preset control logic to maintain stable operation. However, due to the high nonlinearity of operating conditions and the complexity of multi-source inputs, existing process control methods struggle to achieve real-time adjustment. The control system's response speed to dynamic conditions is slow, parameter adjustment accuracy is insufficient, and it cannot effectively balance multi-objective optimization requirements, resulting in inadequate control efficiency and accuracy for current methanol generator combustion systems.

[0003] To address this, a control method for the combustion system of an in-cylinder direct injection methanol generator based on multi-objective optimization is proposed. Summary of the Invention

[0004] The purpose of this invention is to provide a control method for the combustion system of an in-cylinder direct injection methanol generator based on multi-objective optimization. This method achieves precise control of the combustion system by dynamically adjusting the control parameters of the methanol generator. First, multi-source sensor data from the methanol generator is collected. Features are extracted and evaluated from the multi-source sensor data to generate a control weight allocation matrix. This matrix is ​​then used for weighted fusion to obtain a fused control vector. Next, an in-cylinder process control model is constructed. Combining the physicochemical properties of methanol fuel with the in-cylinder combustion characteristics, the model parameters are adaptively adjusted. Finally, based on the process control model and real-time operating data, a multi-objective optimization algorithm is used to comprehensively evaluate energy conversion efficiency, emission levels, and operational stability. Adaptive control parameters are dynamically generated by iteratively searching for the Pareto optimal solution set. Based on these adaptive control parameters, control commands are generated to adjust the operating parameters of the methanol generator in real time.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A control method for the combustion system of a methanol generator with in-cylinder direct injection based on multi-objective optimization includes:

[0007] Real-time acquisition and processing of multi-source sensor data from methanol generator; feature extraction of multi-source sensor data to obtain control vectors; evaluation of multi-source sensor data to generate control weight allocation matrix; weighted fusion of control vectors using control weight allocation matrix to obtain fused control vectors.

[0008] An in-cylinder process control model is constructed, and the model parameters are adaptively adjusted by combining the physicochemical properties of methanol fuel with the in-cylinder combustion characteristics to generate a mapping relationship between adjustable control parameters and energy conversion efficiency, emission levels and operational stability.

[0009] Based on the in-cylinder process control model and combined with real-time operating data, a multi-objective optimization algorithm is used to comprehensively evaluate energy conversion efficiency, emission levels and operational stability. Adaptive control parameters are dynamically generated by iteratively searching the Pareto optimal solution set.

[0010] Based on adaptive control parameters, control commands are generated to adjust the operating parameters of the methanol generator in real time, and to perform multi-objective collaborative optimization and adaptive energy consumption control of the in-cylinder energy conversion process.

[0011] Preferably, the method of extracting features from multi-source sensor data to obtain the control vector adopts a combination of multi-scale time-frequency analysis and sparse coding. Multi-scale time-frequency analysis decomposes the multi-source sensor data into multi-scale time-frequency components, and sparse coding is applied to process the multi-scale time-frequency components to obtain the control vector. A dynamic weighted fusion mechanism is used to evaluate the correlation between the confidence level of multi-source sensor data and the system state in real time. Based on the preset priority corresponding to the real-time identified operating conditions, the weight allocation of each data source in the fusion process is adaptively adjusted to construct a control weight allocation matrix.

[0012] Preferably, the adaptive adjustment of model parameters includes: concatenating the physicochemical properties of methanol fuel with the dynamic characteristics of in-cylinder combustion and the fused control vector to form a comprehensive input feature set; constructing an in-cylinder process control model using the comprehensive input feature set; continuously updating the weights and bias parameters of the in-cylinder process control model using an online incremental learning mechanism based on the real-time fused control vector; and generating a mapping relationship between adjustable control parameters and energy conversion efficiency, emission levels, and operational stability through the in-cylinder process control model.

[0013] Preferably, the construction process of the in-cylinder process control model includes: adopting a parameter initialization strategy based on transfer learning, using model weights pre-trained based on multiple representative historical operating condition data as initial parameters, and fine-tuning only the weights of the fully connected layers of the in-cylinder process control model in subsequent training; applying adaptive regularization technology to establish a mapping relationship between regularization strength and real-time prediction error and operating condition complexity of the model, and adjusting the weight matrix of the in-cylinder process control model according to the mapping relationship.

[0014] Preferably, the process of the multi-objective optimization algorithm using a parallel evolutionary computation method includes: constructing a multivariate regression model, fitting the nonlinear mapping relationship between adjustable control parameters and energy conversion efficiency, emission levels, and operational stability, and obtaining a multi-objective optimization function based on dynamic adjustment of weights under operating conditions; using a parallel evolutionary computation method based on a non-dominated sorting genetic algorithm, employing the multi-objective optimization function as a fitness function to perform iterative search for the Pareto optimal solution set; combining the optimization process with a dynamic constraint adaptive adjustment mechanism, dynamically adjusting the constraints in the optimization process according to the real-time operating condition complexity through an adaptive penalty function; and applying an elite solution sharing mechanism and congestion distance sorting to maintain the diversity of the Pareto optimal solution set.

[0015] Preferably, the adaptive control parameters are generated by a hybrid decision-making method combining fuzzy logic and reinforcement learning, dynamically adjusting the weights of energy conversion efficiency, emission levels, and operational stability. The fuzzy logic constructs a rule set based on real-time operating data and fuel characteristics to generate initial weight allocations, while the reinforcement learning performs online optimization of the fuzzy logic rule set based on long-term performance feedback. Bayesian inference is used to calculate the posterior probability distribution of the predicted performance of each set of control parameters, and the optimal combination of control parameters is selected from the Pareto optimal solution set based on operating condition priorities.

[0016] Preferably, the generation process of the control command combines a closed-loop control mechanism that integrates real-time feedback correction and predictive control. By comparing the actual system response with the prediction results of the in-cylinder process control model, the control parameters are dynamically corrected. Based on the prediction results of the in-cylinder process control model, the control parameter trajectory for multiple future time steps is generated and updated through a rolling optimization method. The system disturbance caused by changes in operating conditions is predicted through a convolutional neural network, generating a feedforward compensation signal. The control parameters corrected by the feedback correction stage and the compensation signal generated by the feedforward compensation stage are weighted and fused to generate and execute the control command.

[0017] 1. Based on data acquisition from multiple sensors, a corresponding fused control vector is generated simultaneously with the acquired data. Dynamic weighted fusion technology is used to optimize feature representation, and statistical characteristics and machine learning methods are employed to detect outliers in the data. Then, context-aware imputation is used to fill in missing data values, effectively solving the problem of low control accuracy caused by anomalies in multi-source data or sensor anomalies. First, data from different sources are weighted and fused according to certain rules to form a new dataset, and different weight coefficients are assigned to different input signals to enhance data completeness. Second, when there are missing values ​​in the original dataset, context-aware imputation is used to ensure consistency between components, better cope with complex operating conditions, and improve the system's control robustness.

[0018] 2. An adaptive network topology is proposed, using a hybrid neural network model composed of multiple sub-networks for combustion process modeling. Based on this, a multi-objective evolutionary algorithm and reinforcement learning decision-making method are employed to achieve the final optimal control output. This overcomes the shortcomings of existing fixed control strategies, enabling optimal performance under numerous constraints. Specifically, a multi-objective evolutionary algorithm is used to obtain several possible feasible paths as the initial search space, followed by reinforcement learning training to derive the corresponding optimal action sequence. This allows for the selection of the most suitable control command from a large number of predictive performance indicators, ensuring the entire control system operates in a relatively optimal state. This effectively avoids failures caused by uncertainties and reduces the additional burden of manual intervention.

[0019] 3. The optimal control command is subsequently tracked and corrected through a closed-loop circuit. Furthermore, the feedforward compensation principle is used to mitigate the effects of various unknown disturbances. This novel control approach inherits the advantages of previous methods while also offering the benefits of later development, effectively reducing the lag effect and difficulty in precise tracking inherent in conventional controllers. In addition, continuous online updates of the error term minimize the probability of steady-state deviation and overshoot, helping to reduce the overall system fluctuation. Moreover, in the event of sudden situations or load changes, pre-emptive compensation measures can be implemented for rapid response. The closed-loop control mode not only improves the safety and reliability of unit operation but also significantly expands the controllable range and reduces operational complexity. Attached Figure Description

[0020] Figure 1 A schematic flowchart of a combustion system control method for an in-cylinder direct injection methanol generator based on multi-objective optimization provided in an embodiment of the present invention;

[0021] Figure 2 This is a schematic diagram of the process for acquiring multi-source sensor data provided in an embodiment of the present invention;

[0022] Figure 3 A schematic diagram of the operation flow of the in-cylinder process control model provided in an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Please see Figures 1 to 3 This invention provides a control method for a combustion system of an in-cylinder direct injection methanol generator based on multi-objective optimization, the technical solution of which is as follows:

[0025] Example 1:

[0026] To improve the combustion efficiency of methanol generators used as engine power sources, this invention introduces a multi-objective optimization-based control method for the combustion system of in-cylinder direct injection methanol generators. The specific process is as follows: Figure 1 As shown.

[0027] Real-time acquisition of multi-source sensor data on operating condition fluctuations and fuel characteristics of methanol generators, followed by data processing through feature extraction and dynamic weighted fusion, as detailed in the following process. Figure 2 As shown.

[0028] First, high-frequency sampling is used to acquire data from multiple sensor sources to capture fluctuations in operating conditions and fuel characteristics. This method samples 10 times per second, acquiring operating condition parameters, including load level, engine speed, and in-cylinder pressure, as well as fuel characteristic parameters, including fuel flow rate, oxygen concentration, and fuel composition. Timestamp alignment technology is used to synchronize the data streams from each sensor, generating a unified time-series dataset. A multi-channel parallel acquisition method is employed, utilizing a digital signal processor to perform Fast Fourier Transform and data buffering to optimize data throughput, ensuring real-time performance and data integrity during the acquisition process, and providing high-precision input for subsequent feature extraction and processing.

[0029] Furthermore, by combining multi-scale time-frequency analysis with sparse coding, control vectors are extracted from multi-source sensor data to capture the dynamic characteristics of the combustion system. Discrete wavelet transform is employed, using the Daubechies wavelet basis to decompose the time-series data into multi-scale time-frequency components, extracting dynamic fluctuation features of operating conditions and fuel characteristics, such as the transient correlation between pressure and flow rate or the time-series dependence of temperature and oxygen concentration. Subsequently, sparse coding is applied, and control vectors are generated through L1 norm regularization, preserving key dynamic features.

[0030] Furthermore, by real-time evaluation of the confidence level of multi-source sensor data and the correlation with system state, a control weight allocation matrix is ​​constructed to enhance the robustness of data processing and adaptability to nonlinear operating conditions. Based on historical data, the confidence level of each sensor data is calculated using standard deviation analysis to generate a confidence score, reflecting data stability. The fusion weights of multi-source sensor data are adaptively adjusted during the fusion process according to operating condition priorities. These operating condition priorities are obtained by consulting a pre-set priority configuration table within the system. This configuration table pre-stores different priority combinations corresponding to different operating conditions. The relationship between fusion weights and operating condition priorities is determined through a weight-operating condition correspondence table. For example, high-load driving pressure data has a higher priority, as does cold-start temperature data. The correlation between multi-source sensor data and system state is calculated using the Pearson correlation coefficient to generate a correlation coefficient matrix. Combining the confidence score and the correlation coefficient, a control weight allocation matrix is ​​generated using a weighted normalization method. The matrix dimension is the number of sensors, and it is iteratively updated every second to reflect changes in operating conditions. The control vector is then fused based on the weight matrix using a weighted averaging method to generate a fused control vector.

[0031] By generating control vectors based on multi-source sensor data fusion and incorporating key features such as in-cylinder pressure, injection quantity, and oxygen concentration, a dynamic weighted fusion mechanism is employed to optimize feature representation capabilities, significantly improving the accuracy and adaptability of combustion system control. Compared to traditional single-sensor data processing methods, this mechanism captures the complex dynamic characteristics of the methanol combustion process through comprehensive analysis of multi-source data, enhancing the information integrity of the feature vectors. Dynamic weighted fusion adaptively adjusts weight allocation according to real-time operating condition priorities, effectively balancing the impact of different features on model predictions, reducing feature redundancy and noise interference, thereby improving the modeling accuracy of the in-cylinder process control model for nonlinear operating conditions. This provides high-quality input for subsequent multi-objective optimization, optimizing energy conversion efficiency, emission levels, and operational stability.

[0032] Furthermore, based on the real-time detection of abnormal operating conditions in the combustion system using the fused control vector, an autoregressive model trained on historical data is employed to predict the statistical characteristics of the fused control vector under normal operating conditions, and this result is used as the baseline prediction value. Then, the system calculates the actual statistical characteristics of the current fused control vector through a sliding window. These statistical characteristics include mean, variance, skewness, and kurtosis, reflecting the central tendency, fluctuation amplitude, distribution asymmetry, and sharpness of the features, respectively. The actual statistical characteristics are compared with the model-predicted statistical characteristics; if the deviation exceeds two standard deviations, it is judged as an anomaly. Finally, to confirm the authenticity of the anomaly, the system calculates the Pearson correlation coefficient between the fused control vectors. If the correlation coefficient is lower than a threshold set experimentally by the expert team, the anomaly is finally confirmed.

[0033] Furthermore, a context-aware completion method is employed to generate high-precision backup data based on the fused control vector and control weight allocation matrix, filling in data gaps and maintaining control continuity. A lightweight convolutional long short-term memory network is used, utilizing one-dimensional convolutional layers to extract local features of the fused control vector and combining this with long short-term memory units to capture time-series dependencies, generating completed data. The fusion weights of the completed data are dynamically adjusted using the control weight allocation matrix to ensure consistency with real-time operating conditions.

[0034] Furthermore, through self-learning feedback optimization, the performance of the predictive sub-model is continuously improved using long-term anomaly detection data, enhancing its adaptability to future abnormal operating conditions. The k-means clustering method is employed to extract the distribution patterns of anomaly features, constructing a dynamic anomaly pattern library. Based on the latest anomaly data and operating condition characteristics, an online incremental learning algorithm and gradient descent method are used to update the weights and time windows of the predictive sub-model in real time. The anomaly pattern library is used to predict potential anomaly trends through pattern matching, pre-adjusting the model parameters.

[0035] An in-cylinder process control model is constructed, combining the physicochemical properties of methanol fuel with in-cylinder combustion characteristics. The model structure and parameters are adaptively adjusted to generate a mapping relationship between adjustable control parameters and energy conversion efficiency, emission levels, and operational stability. The operation flow of the in-cylinder process control model is as follows: Figure 3 As shown.

[0036] First, an in-cylinder process control model is constructed, employing a hybrid neural network that integrates convolutional neural networks and long short-term memory (LSTM) networks to process the fused feature vector. The convolutional neural network contains three convolutional layers to extract the local spatiotemporal patterns of the fused control vector, capturing the short-term impact of injection quantity changes on peak pressure. The LSTM network contains two memory layers to model the long-term time dependence of the combustion process, capturing the cumulative effect of oxygen concentration on combustion rate. Fully connected layers output adjustable control parameters, including injection timing, injection quantity, and ignition advance angle. Then, the physicochemical properties of methanol fuel, including calorific value, volatility, and chemical reactivity, as well as the dynamic characteristics of in-cylinder combustion, including combustion rate, peak pressure, and heat release rate, are concatenated with the fused feature vector to form a comprehensive input feature set. The physicochemical properties of methanol fuel are obtained through experimental measurements and database queries, and after standardization, are input into the in-cylinder process control model. In-cylinder combustion characteristics are derived by obtaining the statistical properties of the fused feature vector, including mean, variance, skewness, and kurtosis, deriving the peak pressure, and then analyzing the heat release rate through time-frequency component analysis to collaboratively model the physicochemical mechanism of methanol combustion.

[0037] The parameters of the hybrid neural network are continuously updated based on real-time sensor data through an online incremental learning mechanism. Specifically, a gradient descent method is used to progressively update the weights and bias parameters of the convolutional neural network and the long short-term memory network based on the real-time fused control vector data stream. Real-time sensor data includes load pressure, fuel flow rate, and oxygen concentration. The fused control vector is dynamically adjusted by combining historical parameters using a weighted averaging method. Incremental learning ensures that the model continuously adapts to changes in operating conditions, such as rapid parameter updates during cold starts to optimize emission control.

[0038] Hybrid neural networks combine one-dimensional convolutional neural networks and long short-term memory networks to enhance the model's ability to represent nonlinear dynamic characteristics and improve the modeling accuracy for complex operating conditions. An online incremental learning mechanism continuously updates parameters based on real-time sensor data, ensuring the model adapts to dynamic operating conditions, reducing control deviations, improving energy conversion efficiency, lowering emissions, and enhancing operational stability.

[0039] Furthermore, a transfer learning parameter initialization method is employed to accelerate the convergence speed of the in-cylinder process control model under different operating conditions by fusing the pre-trained model with real-time operating condition data. A pre-trained model is trained based on historical operating condition data, which includes data covering various scenarios such as high-load driving, cold start, and low-load cruising, generating initial weights and bias parameters for a one-dimensional convolutional neural network and a long short-term memory network. The in-cylinder process control model is then trained using transfer learning, with only the weights of the fully connected layers being fine-tuned. Real-time operating condition data and pre-trained weights are fused using a weighted average method, and the weights are dynamically allocated based on operating condition similarity to ensure the model quickly adapts to new operating conditions. Regularization techniques are used to suppress the overfitting risk of the in-cylinder process control model. An adaptive L2 regularization method is adopted, and the regularization strength is dynamically adjusted according to the model prediction error and the complexity of the operating conditions. The complexity of the operating conditions is quantified by calculating the information entropy of the recently fused control vector. The higher the information entropy, the more complex the operating conditions are considered. For example, the regularization strength is increased under low-load cruise conditions to enhance generalization ability, while the strength is reduced under complex operating conditions to retain model flexibility. The regularization term is applied to the initial weights of the one-dimensional convolutional neural network and the long short-term memory network. The parameters are optimized through gradient updates to ensure a balance between prediction accuracy and generalization ability.

[0040] Transfer learning parameter initialization significantly accelerates model convergence speed by fusing pre-trained models with real-time operating data, adapts to diverse operating conditions, and improves control efficiency. Regularization techniques adjust the regularization strength, effectively suppressing the risk of overfitting, enhancing the model's generalization ability, ensuring the stability and reliability of control parameter predictions, and optimizing energy conversion efficiency, emission levels, and operational stability.

[0041] Based on the in-cylinder process control model and combined with real-time operating data, a multi-objective optimization algorithm is used to comprehensively evaluate energy conversion efficiency, emission levels and operational stability. Adaptive control parameters are dynamically generated by iteratively searching for the Pareto optimal solution set.

[0042] A multivariate regression model is constructed to generate a mapping relationship between adjustable control parameters and energy conversion efficiency, emission levels, and operational stability. The injection timing, injection quantity, and ignition advance angle output by the hybrid neural network are fitted to the nonlinear relationship between the control parameters and multiple objectives through the multivariate regression model, generating a multi-objective optimization function. This optimization function employs a weighted approach, balancing efficiency, emissions, and stability through weight coefficients. The weights are dynamically adjusted according to operating conditions; for example, efficiency weights are increased during high-load driving. This mapping relationship guides the combustion system to adjust parameters in real time, achieving multi-objective optimal control.

[0043] The generated multi-objective optimization function is used as the fitness function. A parallel evolutionary computation method is employed, and a non-dominated sorting genetic algorithm is used to comprehensively evaluate the energy conversion efficiency, emission levels, and operational stability predicted by the in-cylinder process control model, generating a Pareto optimal solution set. First, the algorithm initializes a set of control parameters, each including injection timing, injection quantity, and ignition advance angle. Then, the algorithm calculates the predicted performance based on the load level, fuel flow rate, and oxygen concentration, along with the mean, variance, skewness, and kurtosis of the fused control vector. Energy conversion efficiency is evaluated using the fuel consumption to power output ratio, emission levels are evaluated using NOx and unburned hydrocarbon concentrations, and operational stability is evaluated using the standard deviation of pressure fluctuations. The non-dominated sorting genetic algorithm evaluates the fitness of the population, retaining parameter combinations superior to other solutions through non-dominated sorting. The crossover operation uses simulated binary crossover to generate new parameter combinations; the mutation operation introduces randomness through Gaussian perturbation, iteratively updating the population until convergence to the Pareto optimal solution set, which contains multiple sets of control parameters balancing multi-objective performance.

[0044] Furthermore, a dynamic constraint adaptive adjustment mechanism is introduced to optimize the generation efficiency of the Pareto optimal solution set. This mechanism dynamically adjusts the optimization constraints based on the real-time operating condition complexity. For example, under high-load conditions, emission level constraints are relaxed to prioritize energy conversion efficiency, while under cold-start conditions, operational stability constraints are tightened to reduce pressure fluctuations. Constraint adjustment is achieved through an adaptive penalty function, which dynamically updates weights based on operating condition priorities, weakening the constraint influence of secondary objectives and strengthening the optimization direction of primary objectives. The constraint boundary is adaptively adjusted based on the statistical characteristics of the fused control vector. For example, stability constraints are relaxed when variance is large, and emission constraint boundaries are dynamically adjusted based on skewness and kurtosis analysis. This dynamic constraint adaptive adjustment mechanism ensures efficient convergence of the algorithm under nonlinear operating conditions, avoiding getting trapped in local optima.

[0045] Furthermore, a multi-objective cooperative search strategy enhances the diversity and coverage of the Pareto solution set. The strategy employs an elite solution-sharing mechanism, preserving high-quality control parameter combinations across objectives during population evolution and accelerating global search efficiency through cross-subpopulation sharing. Diversity maintenance is achieved using congestion distance sorting, prioritizing parameter combinations at solution set boundaries and in sparse regions to ensure the solution set covers a broad target space encompassing energy conversion efficiency, emission levels, and operational stability. Cooperative search combines the time-frequency components of the fused control vector to analyze dynamic changes in operating conditions, dynamically adjusting the search direction and prioritizing parameter combinations adapted to complex operating conditions, such as generating stability-oriented solutions during low-load cruise. The strategy balances the exploration and utilization of objectives through adaptive weight allocation, ensuring the solution set contains diverse control strategies.

[0046] Parallelized evolutionary computation methods efficiently generate Pareto optimal solution sets through non-dominated sorting genetic algorithms, significantly improving multi-objective optimization efficiency and adapting to the high dynamics of in-cylinder direct-injection methanol generator combustion systems. A dynamic constraint adaptive adjustment mechanism dynamically optimizes constraints based on operating condition priorities, enhancing the algorithm's adaptability to nonlinear operating conditions and convergence speed, avoiding local optimum traps. A multi-objective cooperative search strategy expands the diversity and coverage of the solution set through elite solution sharing and crowding distance sorting, ensuring that control parameter combinations adapt to complex operating conditions and significantly optimizing energy conversion efficiency, emission levels, and operational stability.

[0047] Based on adaptive control parameters, control commands are generated to adjust the operating parameters of the methanol generator in real time, and to perform multi-objective collaborative optimization and adaptive energy consumption control of the in-cylinder energy conversion process.

[0048] Furthermore, a hybrid decision-making method combining fuzzy logic and reinforcement learning is employed to dynamically adjust the weights of energy conversion efficiency, emission levels, and operational stability based on the real-time system state, generating the optimal combination of control parameters. The fuzzy logic method constructs a rule set, inputting real-time operating data, including the statistical characteristics of the fused control vector and fuel characteristics, such as calorific value, volatility, and chemical reactivity. Fuzzy rules are defined through membership functions; for example, a higher mean increases the efficiency weight, while a larger variance increases the stability weight. The fuzzy inference system defuzzifies using a weighted average to generate initial weight allocations; for example, under high load conditions, efficiency accounts for 60%, stability for 30%, and emissions for 10%. In addition, a deep deterministic policy gradient method is adopted to optimize the fuzzy rule set based on the deviation between long-term predicted performance and actual performance. The long-term predicted performance refers to the comprehensive performance composed of three indicators: energy conversion efficiency, emission level and operational stability. The performance deviation is the reward and penalty signal of reinforcement learning. If the actual comprehensive performance is better than the prediction, a positive reward is obtained, and otherwise a negative penalty is imposed. Then, the weight allocation strategy is dynamically updated by adjusting the parameters of the membership function to ensure that the weights accurately match the changes in operating conditions.

[0049] Confidence interval analysis of predicted performance is used to select parameter combinations with the highest control reliability. A Bayesian inference method is employed, calculating the posterior probability distribution based on the predicted performance distribution of each control parameter in the Pareto solution set. The posterior distribution is modeled by fusing statistical characteristics of the control vectors; for example, skewness reflects the asymmetry of performance deviation, and kurtosis reflects the probability of abnormal fluctuations. Confidence intervals are defined as a 95% probability range, prioritizing parameter combinations with narrower confidence intervals and superior predicted performance, such as solutions with high efficiency and low deviation. The analysis process incorporates operating condition priority; for example, in cold start conditions, parameter combinations with high reliability are prioritized to ensure high reliability of the selected parameters under real-time operating conditions.

[0050] Fuzzy logic, combined with reinforcement learning, ensures a high degree of alignment between decision-making and operational objectives by dynamically adjusting performance index weights to accurately respond to real-time operating condition changes. This significantly improves the relevance and adaptability of control parameters. Confidence interval analysis assesses the reliability of predicted performance through Bayesian inference, selecting high-confidence parameter combinations to effectively reduce control uncertainty. This ensures efficient operation of the combustion system under complex conditions, significantly optimizing energy conversion efficiency, emission levels, and operational stability.

[0051] Based on the optimal control parameters generated through multi-objective optimization and combined with real-time operating data, control commands are generated through a command mapping model. The command mapping model employs a feedforward neural network to map the optimal control parameters into executable commands for the methanol generator, including injector pulse width, ignition coil trigger signal, and throttle opening. Input data includes the statistical characteristics of the fused control vector and real-time operating data. The statistical characteristics include mean, variance, skewness, and kurtosis. The real-time operating data includes load level, fuel flow rate, and oxygen concentration. Precise control commands are generated through nonlinear transformation of the feedforward neural network. The command generation process incorporates operating condition priorities; for example, under high load conditions, injection quantity is adjusted first to optimize energy conversion efficiency, while under cold start conditions, ignition advance angle is adjusted first to reduce emissions.

[0052] Furthermore, a real-time feedback correction mechanism compares the actual system response with the prediction results of the in-cylinder process control model, dynamically correcting control parameters and enhancing the accuracy of control commands. The actual system response is collected using sensors, including in-cylinder pressure, exhaust emission concentration, and speed fluctuations, and compared with the energy conversion efficiency, emission levels, and operational stability predicted by the in-cylinder process control model. The mean square error method is used to calculate the deviation, and control parameters are dynamically adjusted based on the magnitude of the deviation. For example, when actual emissions are higher than predicted values, the injection quantity is reduced to lower the concentration of unburned hydrocarbons. Feedback correction utilizes a Kalman filter method for optimization, optimizing the statistical characteristics of the fused control vector, smoothing noise interference, and ensuring the stability and reliability of the correction parameters. The correction process is updated every second to adapt to changes in operating conditions. In addition, a model predictive control method is employed, combined with real-time feedback correction, to form a closed-loop control mechanism, optimizing the dynamics of control commands. Model predictive control (MMCC) generates control parameter trajectories for multiple time steps based on the prediction results of the in-cylinder process control model, predicting energy conversion efficiency, emission levels, and operational stability. Then, a rolling optimization method is used to generate and update control parameter trajectories for future time steps. The direction of the predicted trajectory is adjusted based on real-time operating data and fused control vectors; for example, the stability trajectory is optimized during low-load cruising. Closed-loop control corrects control parameters by comparing the predicted trajectory with the actual system response, and updates the predictive model parameters using feedback correction results, ensuring that control commands match operating conditions. Furthermore, predictive control incorporates operating condition priorities to further adjust and optimize target weights; for example, emission control weights are increased during cold starts.

[0053] Furthermore, a feedforward compensation model is constructed to predict system disturbances based on real-time operating data, thereby reducing the response delay of control commands. The feedforward compensation model is a one-dimensional convolutional neural network containing two convolutional layers and a fully connected output layer. Its input consists of the time-frequency components of the fused control vector and real-time operating data, and its output is a compensation signal vector. The feedforward compensation model is trained offline. The training dataset is generated by analyzing historical operating data. The input samples are feature data sequences before various typical operating condition switching moments (such as load mutations), while the target output label is the actual control compensation amount required to quickly restore system stability after the operating condition switch occurs. In real-time operation, the compensation signal predicted by the feedforward compensation model can be directly superimposed on the main control command. When working in conjunction with feedback correction, the compensation signal is combined with the feedback correction parameters using a weighted fusion method, with the weights dynamically adjusted based on the operating condition complexity. The operating condition complexity is quantified by calculating the variance of recent input data; for example, under high load or complex operating conditions with large variance, the weight of the compensation signal is increased to achieve a rapid response.

[0054] Furthermore, by executing the generated control commands, the operating parameters of the methanol generator are adjusted in real time to achieve multi-objective collaborative optimization and adaptive energy consumption control. The control commands drive the injectors, ignition coils, and throttle valve to adjust injection timing, injection quantity, and ignition advance angle, optimizing the in-cylinder combustion process. Multi-objective collaborative optimization is achieved by dynamically balancing energy conversion efficiency, emission levels, and operational stability; for example, prioritizing efficiency improvement while controlling emissions under high load conditions. Adaptive energy consumption control, based on real-time operating data and the statistical characteristics of the fused control vector, adjusts the fuel injection quantity and air-fuel ratio to optimize energy consumption allocation; for example, reducing fuel injection to lower energy consumption during low-load cruising. The control process is continuously optimized through closed-loop feedback, ensuring the system maintains high efficiency under complex operating conditions.

[0055] Real-time feedback correction dynamically adjusts control parameters by comparing the actual system response with the predicted results, significantly improving the accuracy and robustness of control commands. Model predictive control combined with closed-loop regulation optimizes the dynamics of control parameters, ensuring multi-objective collaborative optimization. Feedforward compensation strategy reduces system response delay by predicting operating condition disturbances, enhancing the system's rapid adaptability to complex operating conditions. The overall mechanism significantly improves the comprehensive performance of the methanol generator in terms of energy conversion efficiency, emission levels, and operational stability, ensuring a highly efficient and stable in-cylinder energy conversion process.

[0056] This paper presents a complete control method for a methanol generator combustion system. It achieves dynamic adjustment of methanol generator control parameters by collecting data from multiple sensors and constructing a control weight matrix, thus providing high-quality input for subsequent models. By building an in-cylinder process control model, control deviations are reduced, energy conversion efficiency is improved, emissions are lowered, and operational stability is enhanced. Parallel evolutionary computation methods are employed to improve the optimization efficiency of multiple objectives. By comparing real-time feedback correction with actual system response and prediction results, control parameters are dynamically corrected, significantly improving the accuracy and robustness of control commands. Finally, precise control of the combustion system is achieved through dynamic adjustment of methanol generator control parameters.

[0057] Example 2:

[0058] To improve the overall performance and control intelligence of the generator set when it is used as a power source driven by the engine, this embodiment provides a control method for the combustion system of an in-cylinder direct injection methanol generator based on multi-objective optimization.

[0059] First, sensors collect real-time operating data from the methanol generator at a frequency of 10 times per second, including in-cylinder pressure, engine speed, methanol fuel flow rate, and intake oxygen concentration. A variational autoencoder (VAE) composed of a long short-term memory (LSTM) network is employed to learn the latent distribution of each sequence of data and compress it into a control vector that captures its core dynamics. To enhance feature extraction accuracy, a graph neural network is introduced to model the spatial-temporal correlations of multi-sensor data. First, the system constructs a graph where each node represents a sensor. The initial feature of each node is the control vector generated for that sensor by the VAE. Edges between nodes are determined based on the pre-calculated Pearson correlation coefficient between sensor data; connections are established only when the absolute value of the correlation coefficient is greater than a threshold of 0.6. Subsequently, the constructed graph is input into a GAT model containing two graph attention layers. Through a message passing mechanism, each node can selectively aggregate information from its neighbors based on attention weights, thereby learning the dynamic influence relationships between different sensors under the current operating conditions. The GAT model ultimately outputs an updated fused control vector containing correlation information for each sensor node.

[0060] To address abnormal operating conditions, an anomaly detection model based on historical normal operating conditions is used, calculating the anomaly probability of the fused feature vector through a sliding window. If a sudden change in fuel flow is detected, the distribution of abnormal features is extracted using k-means clustering, the dynamic anomaly pattern library is updated, potential disturbances are predicted, and control parameters are pre-adjusted. This mechanism ensures data quality and the system's rapid response to anomalies. An in-cylinder process control model for the combustion process is constructed using a recurrent neural network (RNN) and a variational autoencoder (VAE). The RNN is used to process time-series data, and the latent features extracted by the VAE and graph neural network are used to generate the correlation between adjustable control parameters and energy conversion efficiency, emission levels, and operational stability. Through an online incremental learning mechanism, the model updates weights based on real-time sensor data to adapt to changes in operating conditions. To prevent overfitting, a Dropout strategy is employed, with the Dropout ratio dynamically adjusted according to the complexity of the operating conditions; for example, the ratio is reduced under high-load conditions to preserve model capacity. Transfer learning is used to initialize parameters, fuse pre-trained weights with real-time data, fine-tune the output layer, and accelerate model convergence.

[0061] Based on the in-cylinder process control model and combined with real-time operating data, a multi-objective optimization method is employed to evaluate energy conversion efficiency using the fuel consumption-to-power output ratio, emission levels using NOx and unburned hydrocarbon concentrations, and operational stability using the standard deviation of pressure fluctuations. Furthermore, a non-dominated sorting genetic algorithm is used to generate a Pareto optimal solution set to optimize the combination of control parameters, including injection timing, injection quantity, and ignition advance angle. A dynamic priority adjustment mechanism is introduced into the optimization process, adjusting the target weights according to operating conditions. Weight adjustment is automatically completed through operating complexity assessment, ensuring that the solution set adapts to different operating objectives. The decision-making process uses a proximal policy optimization reinforcement learning algorithm to select the optimal combination of control parameters from the Pareto solution set. The proximal policy optimization reinforcement learning algorithm is a policy gradient-based reinforcement learning method that improves training stability by limiting the policy update amplitude, making it suitable for real-time combustion control. Generator load, fuel flow rate, and oxygen concentration are considered as environmental states, control parameters as actions, and a weighted sum of comprehensive objectives is used as the reward. The probability distribution of actions is generated using a neural network. The input is a fused control vector, and the output is the mean and variance of the continuous action space. Specifically, the neural network adopts an actor-critic architecture, with both its policy and value networks consisting of multiple fully connected layers. The proximal policy optimization reinforcement learning algorithm collects real-time data and rewards, calculates the policy gradient, and uses a pruned objective function to limit the update amplitude and avoid policy oscillations. Each iteration updates the policy and value networks and optimizes parameter selection. To optimize multi-objective balance, the reward function is dynamically adjusted according to the operating conditions. Efficiency reward weights are increased under high load conditions, and the stability reward ratio is increased during cold starts. Entropy regularization encourages policy exploration, ensuring that the proximal policy optimization reinforcement learning algorithm generates balanced control parameters under complex operating conditions, guaranteeing the stability of the reinforcement algorithm and its suitability for complex dynamic environments, and ensuring control accuracy.

[0062] The posterior probability distribution is modeled by fusing the mean, variance, and skewness of the feature vectors to calculate the confidence interval. Parameter combinations with small confidence intervals and excellent performance are prioritized to ensure high reliability under real-time operating conditions and meet priority requirements. Optimal control parameters are converted into actuator control commands via a feedforward neural network to drive the injectors, ignition coils, and throttle valve, adjusting injection timing, injection quantity, and ignition advance angle to optimize the combustion process. Additionally, a recurrent neural network is used to predict operating condition disturbances (such as the impact of sudden load changes on cylinder pressure), generating a feedforward compensation signal that is superimposed on the control commands to improve system response speed. Real-time feedback correction dynamically optimizes control parameter accuracy by comparing actual responses with predicted results. Feedback correction and feedforward compensation are combined through weighted fusion, with weights adjusted according to operating condition complexity; for example, compensation weights are increased under high loads. Control commands continuously adjust fuel injection quantity and air-fuel ratio to optimize energy distribution; for example, fuel injection is reduced to lower energy consumption under low loads.

[0063] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A control method for a combustion system of an in-cylinder direct injection methanol generator based on multi-objective optimization, characterized in that, include: Real-time acquisition and processing of multi-source sensor data from methanol generator; feature extraction of multi-source sensor data to obtain control vectors; evaluation of multi-source sensor data to generate control weight allocation matrix; weighted fusion of control vectors using control weight allocation matrix to obtain fused control vectors. An in-cylinder process control model is constructed, and the model parameters are adaptively adjusted by combining the physicochemical properties of methanol fuel with the in-cylinder combustion characteristics to generate a mapping relationship between adjustable control parameters and energy conversion efficiency, emission levels and operational stability. Based on the in-cylinder process control model and combined with real-time operating data, a multi-objective optimization algorithm is used to comprehensively evaluate energy conversion efficiency, emission levels and operational stability. Adaptive control parameters are dynamically generated by iteratively searching the Pareto optimal solution set. Based on adaptive control parameters, control commands are generated to adjust the operating parameters of the methanol generator in real time, and to perform multi-objective collaborative optimization and adaptive energy consumption control of the in-cylinder energy conversion process.

2. The control method for the combustion system of an in-cylinder direct injection methanol generator based on multi-objective optimization according to claim 1, characterized in that, The method for extracting control vectors from multi-source sensor data employs a combination of multi-scale time-frequency analysis and sparse coding. Multi-scale time-frequency analysis decomposes the multi-source sensor data into multi-scale time-frequency components, and sparse coding is applied to process these components to obtain the control vector. A dynamic weighted fusion mechanism is used to evaluate the correlation between the confidence level of the multi-source sensor data and the system state in real time. Based on the preset priority corresponding to the real-time identified operating conditions, the weight allocation of each data source in the fusion process is adaptively adjusted to construct a control weight allocation matrix.

3. The control method for the combustion system of an in-cylinder direct injection methanol generator based on multi-objective optimization according to claim 1, characterized in that, The adaptive adjustment of model parameters includes: concatenating the physicochemical properties of methanol fuel with the dynamic characteristics of in-cylinder combustion and the fused control vector to form a comprehensive input feature set; constructing an in-cylinder process control model using the comprehensive input feature set; continuously updating the weights and bias parameters of the in-cylinder process control model using an online incremental learning mechanism based on the real-time fused control vector; and generating a mapping relationship between adjustable control parameters and energy conversion efficiency, emission levels, and operational stability through the in-cylinder process control model.

4. The control method for the combustion system of an in-cylinder direct injection methanol generator based on multi-objective optimization according to claim 3, characterized in that, The construction process of the in-cylinder process control model includes: adopting a parameter initialization strategy based on transfer learning, using model weights pre-trained based on multiple representative historical operating condition data as initial parameters, and fine-tuning only the weights of the fully connected layers of the in-cylinder process control model in subsequent training; applying adaptive regularization technology to establish a mapping relationship between regularization strength and the model's real-time prediction error and operating condition complexity, and adjusting the weight matrix of the in-cylinder process control model according to the mapping relationship.

5. The control method for the combustion system of an in-cylinder direct injection methanol generator based on multi-objective optimization according to claim 1, characterized in that, The process of the multi-objective optimization algorithm using a parallel evolutionary computation method includes: constructing a multivariate regression model, fitting the nonlinear mapping relationship between adjustable control parameters and energy conversion efficiency, emission levels, and operational stability, and obtaining a multi-objective optimization function based on dynamic adjustment of weights under operating conditions; using a parallel evolutionary computation method based on a non-dominated sorting genetic algorithm, employing the multi-objective optimization function as the fitness function to iteratively search for the Pareto optimal solution set; combining the optimization process with a dynamic constraint adaptive adjustment mechanism, dynamically adjusting the constraints in the optimization process according to the real-time operating condition complexity through an adaptive penalty function; and applying an elite solution sharing mechanism and congestion distance sorting to maintain the diversity of the Pareto optimal solution set.

6. The control method for the combustion system of an in-cylinder direct injection methanol generator based on multi-objective optimization according to claim 1, characterized in that, The adaptive control parameters are generated by a hybrid decision-making method that combines fuzzy logic and reinforcement learning, dynamically adjusting the weights of energy conversion efficiency, emission levels, and operational stability. The fuzzy logic constructs a rule set based on real-time operating data and fuel characteristics to generate initial weight assignments, while the reinforcement learning optimizes the rule set of the fuzzy logic method online based on long-term performance feedback. Bayesian inference is used to calculate the posterior probability distribution of the predicted performance of each set of control parameters. Combined with the priority of operating conditions, the optimal combination of control parameters is selected from the Pareto optimal solution set.

7. The control method for the combustion system of an in-cylinder direct injection methanol generator based on multi-objective optimization according to claim 1, characterized in that, The generation process of the control command combines a closed-loop control mechanism that integrates real-time feedback correction and predictive control. By comparing the actual system response with the prediction results of the in-cylinder process control model, the control parameters are dynamically corrected. Based on the prediction results of the in-cylinder process control model, the control parameter trajectory for multiple future time steps is generated and updated through a rolling optimization method; the system disturbance caused by changes in operating conditions is predicted through a convolutional neural network, and a feedforward compensation signal is generated; the control parameters corrected by the feedback correction link and the compensation signal generated by the feedforward compensation link are weighted and fused to generate control commands and execute them.

Images