An engine cold start combustion state diagnosis and optimization method based on speed fluctuation

Abstract

The application discloses an engine cold start combustion state diagnosis and optimization method based on rotating speed fluctuation, collects rotating speed and cylinder pressure data of each cycle of the engine in a conventional environment and extreme working conditions as a data training set, artificially marks a combustion state, learns and trains a combustion state diagnosis model based on past experimental data, identifies abnormal rotating speed fluctuation when abnormal combustion occurs, uses another set of extreme environment engine cold start experimental data as a model verification set, compares with signals collected by a cylinder pressure sensor in real time, and judges accuracy and real-time performance of the model. The method judges the combustion state of each cylinder of the engine through abnormal rotating speed fluctuation of the engine, identifies cylinders, phases, rotating speeds and torques and other parameters when abnormal combustion occurs in a transient working condition process of the engine, and serves as a judgment basis for a control parameter MAP and a correction parameter of the transient process of the cold start, and can further evaluate the effect after the correction of the parameters.

Description

Technical Field

[0001] This invention relates to engine technology and automatic control methods, specifically to a method for diagnosing and optimizing engine cold start combustion state based on speed fluctuations. Background Technology

[0002] Engines assembled in vehicles are generally not equipped with cylinder pressure sensors due to space and cost constraints. When developing engine cold start performance, the focus is often on optimizing and adjusting engine cold start calibration control parameters based on long-term engineering experience. However, there is a lack of effective theoretical basis and data support, and conventional methods cannot identify and accurately find the optimization direction of calibration parameter MAP.

[0003] There is an urgent need for a method that can accurately, efficiently, and easily diagnose the combustion status of engines during cold starts. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to address the difficulty in starting existing engines due to their relatively slow speed and tendency for abnormal combustion during cold starts. Furthermore, conventional methods for transient conditions like cold starts are insufficient to accurately identify the causes of performance degradation and the inadequacy of control parameter correction. This invention provides a method for diagnosing and optimizing the combustion state of engines during cold starts based on speed fluctuations.

[0005] During cold starts, engine speeds are relatively slow and combustion abnormalities are prone to occur, leading to starting difficulties. Conventional methods struggle to accurately identify the causes of performance degradation and correct control parameters for such transient conditions. This invention uses abnormal engine speed fluctuations to determine the combustion state of each cylinder, identifying parameters such as the cylinder, phase, speed, and torque where combustion abnormalities occur during cold starts. This provides a basis for calibrating and judging the MAP (Modular Analog Parameter) and its correction parameters for the cold start transient process, and also allows for further evaluation of the effectiveness of parameter corrections.

[0006] Technical Solution: This invention provides a method for diagnosing and optimizing engine cold start combustion state based on engine speed fluctuations. It collects engine speed and cylinder pressure data for each cycle of the cold start process under normal and extreme conditions as a data training set, manually labels the combustion state, and trains a combustion state diagnosis model based on past experimental data to identify abnormal speed fluctuations when abnormal combustion occurs. Another set of engine cold start experimental data under extreme conditions is used as a model validation set and compared with the signals collected in real time by the cylinder pressure sensor to determine the accuracy and real-time performance of the model.

[0007] Furthermore, the activation functions were chosen as the Sigmoid function and the tanh function, and a GRU algorithm prediction model was constructed by combining a multi-head attention mechanism.

[0008] Furthermore, a GRU model with a multi-head attention mechanism trained on a 120°CA(-50~70) dataset is adopted.

[0009] Furthermore, an LSTM model based on the 120°CA(-60~60) dataset is adopted.

[0010] Furthermore, it includes the following steps:

[0011] (1) Conduct bench tests on the cold start performance of the engine with cylinder pressure sensor under different conditions to obtain the speed fluctuation during normal combustion and the speed fluctuation during abnormal combustion.

[0012] (2) For the speed fluctuation during normal combustion, record the parameter data under normal combustion conditions in different environments. For the speed fluctuation during abnormal combustion, locate and mark the data information of the abnormal combustion state, and at the same time perform data verification and proceed to step (5).

[0013] (3) Train a large-scale combustion state recognition model using intelligent algorithms and the data results from step (2);

[0014] (4) Optimize relevant parameters;

[0015] (5) Conduct accuracy verification of the combustion state diagnostic model. If the result is accurate, proceed to step (6). If the result is incorrect, return to step (4) or increase the data sample and start over.

[0016] (6) Obtain a combustion state diagnostic model based on engine speed fluctuations;

[0017] (7) Determine if an abnormal combustion condition has occurred: If yes, determine and return the speed and torque of each cylinder where the abnormal combustion state occurred, correct the engine control parameters in the cold start response MAP online or offline, record the instantaneous speed fluctuation signal of the engine each time the engine is cold started, and repeat step (6). If no, determine that the engine cold start performance is better.

[0018] Furthermore, the parameter data in step (2) under normal combustion conditions in different environments include fluctuation characteristics, speed increase rate, and cylinder pressure curve.

[0019] Furthermore, the abnormal status data in step (2) includes cylinder status, cylinder pressure curve, and engine speed.

[0020] Engines assembled in vehicles are generally not equipped with cylinder pressure sensors due to space and cost constraints. When developing engine cold start performance, the focus is often on optimizing and adjusting engine cold start calibration control parameters based on long-term engineering experience. However, there is a lack of effective theoretical basis and data support, and conventional methods cannot identify and accurately find the optimization direction of calibration parameter MAP.

[0021] This invention utilizes the engine's built-in speed sensor to fully leverage the relatively low engine speed during the cold start phase. It collects the engine speed signal in real time during the cold start process, combines it with fuel injection control parameters and the combustion phase of each cylinder, and obtains the speed fluctuation rise curve when the engine has good cold start performance. It filters and analyzes the correspondence between each speed fluctuation parameter and the combustion state of each cylinder, and records the speed curve and fluctuation range of each cylinder during the engine's cold start process for efficient combustion.

[0022] When the engine malfunctions or the environmental conditions are harsh during a cold start, combustion instability can occur. By collecting speed fluctuations during the cold start process and comparing them with a previously trained cold start speed fluctuation model, and by comparing the deviations of each speed fluctuation during the cold start process with the speed fluctuations under normal combustion conditions, it is possible to determine whether abnormal combustion or even misfire occurs in the response speed and the combustion state of the ignition cylinder. This provides a basis for MAP calibration of fuel injection parameters during the cold start process. After calibration, this method can also be used to determine whether the optimized parameters have achieved good results during subsequent starts.

[0023] Unlike engine bench tests, engines mounted on a vehicle do not have expensive and unreliable cylinder pressure sensors, but they will definitely have speed sensors and timing chains to control the engine's fuel injection and combustion phases. During a vehicle's cold start, cylinder pressure cannot be obtained, making it impossible to determine the in-cylinder combustion situation. This poses a challenge to the accurate assessment and optimization calibration of the vehicle's cold start performance.

[0024] Considering that the in-cylinder combustion conditions in each cycle during the cold start acceleration phase affect the engine speed increase rate, there is a complex relationship between in-cylinder pressure and engine speed, which serve as the basis for judging combustion status. This invention proposes an engine cold start speed fluctuation-cylinder pressure judgment method based on intelligent algorithms. By collecting engine speed and cylinder pressure data for each cycle of the cold start process under normal and extreme conditions as a data training set, and manually marking the combustion status as normal, abnormal, or misfire, a combustion status diagnosis model is trained based on past experimental data to accurately identify abnormal speed fluctuations when abnormal combustion occurs. Then, another set of extreme environment engine cold start experimental data is used as the model validation set and compared with the signals collected in real time by the cylinder pressure sensor to judge the accuracy and real-time performance of the model.

[0025] Combustion diagnostic algorithms can utilize various artificial intelligence algorithms such as neural networks, machine learning, and reinforcement learning. They can be adjusted and improved for different application environments and engine characteristics. Through training with a large amount of engine cold start experimental data, a suitable activation function can be selected, and the diagnostic phase and judgment cycle caused by torque transmission path and engine structural parameters can be finely adjusted. Based on the model verification results, the penalty function, weight factor, and other model parameters can be optimized. After a certain number of iterations and optimizations, a relatively accurate engine cold start combustion state diagnostic model can be obtained. The logic and method of constructing the diagnostic model are largely the same.

[0026] By diagnosing the engine's cold start process, we can determine the frequency, speed, degree of degradation, and corresponding cylinders of abnormal combustion during the transient cold start operation. Based on this, we can accurately locate the control parameters of the air circuit, fuel circuit, and electrical circuit related to engine cold start, as well as the corrected MAPs under different conditions. These parameters can then be adjusted based on engineering experience. This method also allows for a quick and effective assessment of whether the engine's cold start performance has improved after the relevant control parameters have been corrected.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0028] 1. This method judges the combustion state of each cylinder of the engine by abnormal fluctuations in engine speed, and identifies the cylinder, phase, speed and torque parameters of the cylinder where the combustion abnormality occurs during the cold start transient process. It provides a basis for the calibration and judgment of the control parameter MAP and its correction parameters during the cold start transient process, and can also further evaluate the effect of parameter correction.

[0029] 2. This method can approximate the combustion state of each cylinder during the cold start process by relying on the existing sensors of the engine. It does not require the installation of an expensive cylinder pressure sensor. It can be calculated by processing and analyzing the transient signals collected by the speed sensor. It is low in cost and highly versatile.

[0030] 3. This method can accurately identify abnormalities in the combustion state of each cylinder through algorithm optimization. Traditionally, a single cylinder pressure sensor can only observe the combustion state of one cylinder. At the same time, it can also perform more refined diagnosis based on the current environmental conditions, engine parameters, etc.

[0031] 4. This combustion diagnostic and identification algorithm can be used in a standalone analysis device. By inputting engine speed sensor signals and basic engine structural parameters, it analyzes the combustion state of each cylinder. Analysis can be performed online in real-time or offline based on data. Simultaneously, the algorithm can be directly programmed into the engine controller for online or offline application, or embedded in the cloud for remote data analysis and diagnostics. The application is flexible and can be freely configured according to the controller's computing power and application scenarios. Attached Figure Description

[0032] Figure 1 Training data was collected for the cold start process at an altitude of 4000m and an ambient temperature of -10℃.

[0033] Figure 2 This is a schematic diagram of the combustion process.

[0034] Figure 3 Constructing LSTM layers and fully connected layers;

[0035] Figure 4 Use the Sigmoid activation function;

[0036] Figure 5 The tanh activation function;

[0037] Figure 6 For the Softmax function;

[0038] Figure 7 The comparison of model test set validation results is shown in the following figures: (a) cold start cylinder pressure curve and the LSTM model judgment result based on the 720°CA dataset; (b) cold start speed curve and the LSTM model judgment result based on the 720°CA dataset; (c) cold start cylinder pressure curve and the LSTM model judgment result based on the 120°CA (-60~60) dataset; and (d) cold start speed curve and the LSTM model judgment result based on the 120°CA (-60~60) dataset.

[0039] Figure 8 For the comparison of model test set validation results, (a) cold start cylinder pressure curve and the judgment result of LSTM model based on 120°CA (-50~70) dataset, (b) cold start speed curve and the judgment result of LSTM model based on 120°CA (-50~70) dataset, (c) cold start cylinder pressure curve and the judgment result of LSTM model based on 120°CA (-40~80) dataset, (d) cold start speed curve and the judgment result of LSTM model based on 120°CA (-40~80) dataset;

[0040] Figure 9 This is a GRU algorithm model that incorporates a multi-head attention mechanism;

[0041] Figure 10The comparison of the model test set validation results includes: (a) cylinder pressure curve and the judgment result of GRU model based on 720°CA dataset; (b) speed curve and the judgment result of GRU model based on 720°CA dataset; (c) cold start cylinder pressure curve and the judgment result of GRU model based on 120°CA (-60~60) dataset; and (d) cold start speed curve and the judgment result of GRU model based on 120°CA (-60~60) dataset.

[0042] Figure 11 For the comparison of model test set validation results, (a) cold start cylinder pressure curve and GRU model judgment result based on 120°CA(-50~70) dataset, (b) cold start speed curve and GRU model judgment result based on 120°CA(-50~70) dataset, (c) cold start cylinder pressure curve and GRU model judgment result based on 120°CA(-40~80) dataset, (d) cold start speed curve and GRU model judgment result based on 120°CA(-40~80) dataset;

[0043] Figure 12 Optimization and comparison of engine cold start combustion state diagnostic models based on speed fluctuations;

[0044] Figure 13 The results of the cold start speed fluctuation-cylinder pressure judgment algorithm model at an altitude of 3000m and an ambient temperature of -20℃ are shown in (a) speed curve and model judgment result, and (b) cylinder pressure curve and model judgment result.

[0045] Figure 14 The IMEP graphs for the first 15 cycles of cold start acceleration of the test engine at an altitude of 3000m and an ambient temperature of -20℃ are shown.

[0046] Figure 15 This is a flowchart of the diagnostic and optimization method of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be further described below.

[0048] This embodiment uses data on in-cylinder combustion and speed fluctuation during cold start of an inline 6-cylinder diesel engine under different ambient temperatures and altitudes as a dataset to train a combustion diagnostic model and verify the feasibility of the diagnostic method.

[0049] Multiple cold-start bench tests of diesel engines under high-altitude, low-temperature conditions were conducted at altitudes ranging from 0m to 4000m and ambient temperatures from 20℃ to -30℃. Cylinder pressure and speed fluctuation data for each working cycle during the entire cold-start process were acquired, and the data was checked and verified cycle by cycle. High-quality datasets provide accurate, consistent, and sufficient signals, enabling neural networks to effectively learn the key features and patterns of the target task during training, thereby improving the model's prediction accuracy and generalization ability. Therefore, quality control and optimization should be given high priority when constructing and preparing datasets. Figure 1 .

[0050] The expected effect of the model is to determine whether the cylinder pressure value is normal based on the degree of speed fluctuation during cold starts, i.e., whether abnormal combustion has occurred. When constructing the dataset, the speed and cylinder pressure data of that cycle are read through the Kibox combustion analyzer, and the combustion data is used to determine whether normal combustion has occurred, thereby labeling the results of this dataset. Figure 2 The diagram shows the combustion state curve of diesel fuel. When combustion is normal in the cylinder, the engine's average cylinder pressure is high, its work capacity is strong, and the engine speed increases relatively quickly. Conversely, when abnormal combustion occurs in the cylinder, the engine's average cylinder pressure is low, its work capacity is weak or even negative, and the engine speed increases less or even decreases. Figure 2 .

[0051] To accurately determine at which engine speed and in which cylinder the combustion anomaly occurred during the high transient process of engine cold start, a response scale is defined when constructing the dataset based on the number and location of the diesel engine cylinder pressure sensors. This improves the effectiveness of data training and combustion diagnosis identification. For example, in this embodiment, the diesel engine cylinder pressure sensor is located in the fifth cylinder. To improve the accuracy of combustion state identification, two judgment scales are used:

[0052] (1) The first data segmentation scale is based on the cylinder pressure of the fifth cylinder to give a general description of the combustion of all cylinders. Under this data segmentation scale, a single dataset encompasses the transient process of a complete cycle of 720° crankshaft rotation angle, and the average combustion state of the engine's six cylinders is determined by the speed fluctuation during this cycle.

[0053] (2) Another approach is to consider the number of cylinders N in the engine, with the ignition interval of each cylinder being 720° / N. Considering the uniformity of the ignition interval of each cylinder in the actual engine, the speed fluctuation within the ignition interval angle can be used as the judgment period, thereby identifying the abnormality of the combustion state of each cylinder and the corresponding speed.

[0054] The specific judgment cycle used depends primarily on hardware factors such as the controller's processing speed and the engine's rotational speed. Considering that the energy generated during the power stroke must travel through a transmission chain composed of a series of mechanical components to propagate to the engine crankshaft and thus cause speed fluctuations, there may be a certain delay between the generation of speed fluctuations and changes in combustion state. Therefore, the impact of different appropriate delays was studied, focusing on the ignition interval, to further improve the accuracy of combustion state diagnosis. This embodiment uses a six-cylinder diesel engine; therefore, 120°CA was selected as the speed fluctuation range, and delays of 10°CA and 20°CA were applied to verify and evaluate the results. Table 1 shows the dataset scheme for the 120°CA division.

[0055] Table 1 Dataset Scheme for 120° CA Partitioning

[0056]

[0057] (1) Construction of a speed fluctuation-combustion state judgment model fused with LSTM

[0058] The training process of LSTM mainly consists of three steps. First, the input data is processed to adapt to the input requirements of LSTM and converted into the format of the LSTM model. Second, forward propagation is calculated. Finally, gradient descent is used to calculate the gradient, and the optimizer is used to update the model parameters. Figure 3 .

[0059] Figure 3 This demonstrates the construction of LSTM layers and fully connected layers. The model's input is only rotational speed sequence data, therefore the input feature dimension is... _ The size is set to 1. The expected outcome of model training is to provide a judgment result for the current loop; therefore, for the output dimension class... `num` is set to 2, indicating whether the current loop is a failed loop. The input sequence length of the LSTM is `sequence`. _ The size is set to 720 or 120 depending on the dataset partitioning scale. This is to map the hidden states of the last layer of the LSTM to the output dimension class. On num, a fully connected layer is defined. The weights of the fully connected layer need proper initialization; improper initialization can lead to gradient explosion or vanishing gradients, rendering the final model unusable. Considering the high transient response during diesel engine cold starts and the activation function used within the LSTM unit, Xavier initialization is chosen to initialize the weights of the fully connected layer. For each layer, Xavier ensures consistent variance.

[0060] Activation functions play a crucial role in neural networks, helping them overcome linear limitations and enabling them to capture non-linear relationships in data. Simultaneously, activation functions map the input of previous layers to higher-level features, which is beneficial for capturing the inherent structure and discriminative information of the data. For classification tasks, activation functions can generate non-linear decision boundaries. The existence of activation functions greatly enhances the learning ability and application scope of neural networks.

[0061] Table 2 shows the activation functions and expressions used in this method. The Sigmoid and tanh activation functions are used in the LSTM forward propagation process, and the Softmax activation function is used to give the model judgment result.

[0062] Table 2 shows the activation functions and expressions involved in the model.

[0063]

[0064] Figure 4 The Sigmoid activation function is commonly used in models that output predicted probabilities, as its output range is from 0 to 1. The Sigmoid function maps the received input to the (0, 1) interval, a property that effectively simulates the gating mechanism in LSTM, controlling the ratio of information retained to discarded. Although the Sigmoid function can cause gradient saturation due to its derivative approaching zero at both ends, in LSTM, the gating signals are independent, and other gates can still be computed effectively, thus mitigating this drawback of the Sigmoid function. Figure 4 .

[0065] Figure 5 The tanh function is demonstrated, which maps the received input to the range (-1, 1), exhibiting symmetry and centralization in its output. Compared to the sigmoid function, the tanh function has a larger dynamic range, providing more information storage space, while its zero-centered characteristic ensures numerical stability. Although the tanh function also exhibits gradients close to zero at both ends, its two-sided saturation characteristic allows it to learn in both positive and negative directions. Therefore, in LSTM, it is frequently used as an activation function for generating candidate cell states and updating actual cell states.

[0066] Figure 6The text demonstrates the Softmax function, a normalized exponential function that normalizes a set of real values ​​into a probability distribution, making it commonly used in multi-class classification problems. In multi-class classification, the Softmax function amplifies the differences between original scores through exponential operations, resulting in a significantly higher probability for the class with the highest original score compared to other classes. Therefore, the class with the highest probability is selected as the prediction result based on the probability distribution output by the Softmax function.

[0067] Finally, an LSTM prediction model was constructed with a learning rate of 0.001 and activation functions of Sigmoid and tanh. The model was trained for 1500 iterations, and the model with the best performance was selected as the final prediction model. The relevant metrics are shown in Table 3.

[0068] Table 3 Evaluation Metrics for LSTM Models

[0069]

[0070] Figure 7 This paper presents a comparison between the judgment results of the LSTM model and the experimental results when the dataset is partitioned using 720°CA and 120°CA. (Comprehensive analysis) Figure 7 As shown in (a) and (b), during the cold start acceleration period, the LSTM model based on the 720°CA dataset can effectively capture the de-ignition cycle through speed fluctuations. Six significant speed fluctuations occurred during acceleration, corresponding to six de-ignition cycles. The LSTM model based on the 720°CA dataset provided accurate results for all six, but also exhibited three incorrect identifications. When the engine speed just reaches the target idle speed, the ECU reduces the cyclic fuel injection quantity. In the low-temperature environment of high altitude, this typically leads to a temporary deterioration in in-cylinder combustion. A de-ignition cycle occurred at this point in the test set, and the LSTM model based on the 720°CA dataset accurately identified this situation. However, during the idle stage, due to the high-temperature environment and cold start conditions, the diesel engine's intake conditions and the formation of the combustible mixture in the cylinder are significantly affected, thus impacting the peak cylinder pressure for normal combustion and ultimately causing speed fluctuations. While the model successfully identified two de-ignition cycles during this stage, it generated four additional incorrect identifications.

[0071] analyze Figure 7(c)(d) During the cold start acceleration phase, the LSTM model based on the 120°CA (-60~60) dataset only made one incorrect judgment about the misfire cycle at the crankshaft angle of 6496°CA. For the remaining misfire cycles, it correctly identified them based on engine speed fluctuations and did not miss any. The model also accurately identified the misfire cycle caused by the ECU reducing fuel injection just as the engine speed reached idle. In the subsequent idle phase, the LSTM model based on the 120°CA (-60~60) dataset performed even better, successfully identifying two in-cylinder misfire events based on engine speed fluctuations, with only two incorrect judgments, a reduction of two compared to the model on the 720°CA dataset.

[0072] Figure 8 This demonstrates a comparison between the judgment results of the LSTM model when the dataset is divided into 120-degree segments and the experimental results. (By...) Figure 8 Analysis of (a) and (b) shows that the LSTM model based on the 120°CA(-50~70) dataset did not miss any misfire cycles during the acceleration phase, identifying six misfire cycles through six large speed fluctuations. However, it made two incorrect identifications, misclassifying normal combustion cycles as misfire cycles. During the idling phase, the model identified two misfire cycles through speed fluctuations, but additionally made three incorrect identifications of misfire cycles. Figure 8 Analysis of (c) and (d) shows that the LSTM model based on the 120°CA(-50~70) dataset successfully identified six misfire cycles during the acceleration phase due to six significant speed fluctuations, but also made three incorrect identifications of normal combustion cycles. During the idling phase, the model identified two misfire cycles through speed fluctuations, but also made three incorrect identifications of misfire cycles.

[0073] Analyzing the LSTM model performance metrics in Table 3, the LSTM model based on the 120°CA (-60~60) dataset exhibits the best performance, with an accuracy of 94.20% and a precision of 94.76%. This indicates that the model has sufficient reliability in identifying misfire cycles. Considering the diesel engine cylinder pressure data is for the fifth cylinder and its actual ignition range, the LSTM model based on the 120°CA dataset is closer to reality in terms of dataset representation. In principle, the LSTM model based on the 120°CA dataset should have better performance metrics. Comparing the performance metrics of the LSTM models on the 720°CA and 120°CA datasets, the LSTM model based on the 120°CA dataset does indeed show better performance than the LSTM model on the 720°CA dataset.

[0074] (2) Construction of an improved GRU algorithm model incorporating attention mechanism

[0075] The GRU algorithm relies on updates from the previous time step for computation, while the multi-head attention mechanism is computationally independent of sequence position. This characteristic allows it to leverage the parallel computing capabilities of modern hardware, improving model training speed. Therefore, the GRU algorithm with multi-head attention has an advantage in parallel computing, resulting in higher overall efficiency. Furthermore, by recalculating the weight distribution of past time steps at each new time step, the GRU algorithm with multi-head attention allows the model to flexibly focus on the most relevant input components based on the current task requirements, improving the model's adaptability and accuracy. These advantages combined with the multi-head attention mechanism enable the trained model to exhibit stronger performance and generalization ability in sequence prediction tasks.

[0076] Figure 9 This paper demonstrates the construction of a GRU algorithm incorporating a multi-head attention mechanism. The input and output feature dimensions are consistent with the LSTM model, and the number of heads in the attention mechanism is set to 2. In the forward propagation of the multi-head attention mechanism, attention scores are calculated through dot product operations, and then scaled to stabilize training. A softmax function is applied to normalize the attention weights along a specific dimension, followed by a weighted summation based on the attention weights. Finally, a fully connected layer performs a non-linear transformation on the attention-weighted result and connects it to the residual of the original signal to obtain the final attention output. The attention output is then fed into the GRU layer for sequence modeling, outputting the final hidden state. Finally, a softmax function is used to provide the final prediction result.

[0077] In this embodiment, the activation functions are selected as the Sigmoid and tanh functions, and a GRU algorithm prediction model is constructed by combining a multi-head attention mechanism. Similarly, the length of the input sequence data is adjusted according to sequence data of different scales, and 1500 iterations of training are performed, selecting the one with the best performance index. Table 4 shows the training results of the GRU model combined with the multi-head attention mechanism.

[0078] Table 4 Evaluation metrics of the GRU model incorporating multi-head attention mechanism

[0079]

[0080] Figure 10 This paper presents a comparison between the judgment results of the GRU model, which incorporates an attention mechanism and partitions the dataset using 720°CA and 120°CA, and the experimental results. Analysis Figure 10As shown in (a) and (b), during the acceleration phase, the GRU model based on the 720°CA dataset accurately identified five misfire cycles based on engine speed fluctuations without any misjudgments. However, when idling is reached, the ECU reduces the cyclic fuel injection, leading to a misfire cycle, which the GRU model based on the 720°CA dataset failed to correctly identify. During the idling phase, the GRU model incorporating the attention mechanism correctly identified two misfire cycles. Although it made three incorrect judgments, mistaking normal ignition cycles for misfire cycles, this was one less error compared to the LSTM model. The diesel engine cold start speed fluctuation-cylinder pressure judgment model constructed using the 720°CA dataset and the GRU algorithm incorporating a multi-head attention mechanism achieves an accuracy of 95.26% and an f1 score of 0.9532, demonstrating excellent performance in judging in-cylinder combustion during the cold start acceleration phase of a diesel engine. Analysis Figure 10 (c) and (d) show that the GRU model based on the 120°CA(-60~60) dataset correctly identified 5 misfire cycles during the acceleration phase. Although one misidentification occurred, it was able to identify the misfire cycle caused by the ECU reducing fuel injection just as the engine speed reached idle. During the idle phase, the model still performed well in identifying misfire cycles, successfully detecting the occurrence of a misfire cycle through two speed fluctuations, but there were also two instances of incorrect identification.

[0081] Figure 11 This paper presents a comparison between the judgment results of the GRU model, which splits the dataset by 120 degrees and incorporates a multi-head attention mechanism, and the experimental results. Analysis Figure 11 During the cold start acceleration phase, the GRU model based on the 120°CA (-50~70) dataset performed best, accurately identifying six misfire cycles based on engine speed fluctuations without any misjudgments or omissions. The GRU model on the 120°CA (-40~80) dataset, however, correctly identified two normal combustion cycles as misfire cycles during the acceleration phase, but did not miss any misfire cycles. Furthermore, the GRU model based on the 120°CA dataset accurately detected and correctly identified the misfire cycle caused by the ECU reducing fuel injection volume just upon reaching idle speed. During the idling phase, both the GRU models based on the 120°CA(-50~70) and 120°CA(-40~80) datasets identified the misfire cycle. The difference was that the model based on the 120°CA(-50~70) dataset made one incorrect judgment of the normal combustion cycle, while the model based on the 120°CA(-40~80) dataset identified the normal combustion cycle as the misfire cycle four times.

[0082] Comparing the three GRU models on the 120°CA dataset, the GRU model trained on the -50–70°CA dataset and based on a multi-head attention mechanism shows the best performance. This is consistent with the model performance evaluation metrics. The GRU model on the 120°CA (-50–70°) dataset has an accuracy of 97.01%, meaning it has high reliability in judging a work cycle as a burnout cycle based on rotational speed fluctuations; at the same time, its F1 score is as high as 0.9742, proving that the model has good performance in making judgments. Comparing the GRU models on the 720°CA and 120°CA datasets, the GRU model on the 120°CA dataset outperforms the GRU model on the 720°CA dataset in all three performance evaluation metrics, which is consistent with the expectation of scientifically dividing the dataset.

[0083] Table 5 lists the constructed cold start speed fluctuation-combustion state judgment model numbers and their corresponding algorithms and datasets. Figure 12 A comparison chart of results from different cold start speed fluctuation-combustion state judgment models. Based on... Figure 12 It can be seen that all models basically achieved correct identification of the misfire cycle during the acceleration and idling periods. Only the LSTM algorithm model based on the 120°CA (-40~80) dataset (Model IV) once identified the misfire cycle as a normal combustion cycle. Among all models, the best performing model is the GRU algorithm model based on the 120°CA (-50~70) dataset combined with an attention mechanism (Model VII). It accurately captured each misfire cycle and achieved the fewest judgments of the normal combustion cycle. From the perspective of dataset composition, the sequence data length of the 720°CA dataset is 6 times that of the 120°CA dataset, which is not conducive to model memorization and leads to a certain degree of forgetting of earlier information in the later stages of training. From an algorithmic perspective, the attention mechanism performs global attention calculation, so that each part retains the global information related to itself. Then, the GRU gating mechanism ensures the preservation of long-term sequence information and does not become clear over time or due to irrelevance in prediction, ultimately achieving a comprehensive consideration of changes over the entire time period.

[0084] Table 5. Cold Start Speed ​​Fluctuation-Combustion State Judgment Model Numbers and Corresponding Algorithms and Datasets

[0085] Model Number algorithm Training dataset I LSTM 720°CA II LSTM 120°CA (-60~60) III LSTM 120°CA (-50~70) IV LSTM 120°CA (-40~80) V GRU with attention mechanism 720°CA VI GRU with attention mechanism 120°CA (-60~60) VII GRU with attention mechanism 120°CA (-50~70) VIII GRU with attention mechanism 120°CA (-40~80)

[0086] The figure below shows the experimental verification of this algorithm on another 4L diesel engine. It presents the output results of the engine speed fluctuation-cylinder pressure judgment model during cold start at an altitude of 3000m and an ambient temperature of -20℃. Analysis Figure 13(a) It can be seen that the algorithm model calculates that an abnormal combustion state occurred in the CA range of 2880 to 3240° based on the input speed signal. According to the number of cylinders of the test engine, the combustion state diagnosis algorithm is used to find that two cylinders in the working cycle have misfired, which leads to fluctuations in speed.

[0087] By using a combustion diagnostic program, it was found that abnormal combustion occurred at a speed of 150 r / min. The engine fuel control parameters needed to be adjusted and optimized. In this example, the combustion performance at the current speed was improved by advancing the time of the first pre-injection at the engine speed of 150 r / min to 125° CA before the top dead center of the compression (a total of 2 pre-injections). The engine cold start process was more stable and faster.

[0088] The above are merely preferred embodiments of the present invention and do not constitute any limitation on the present invention. Any equivalent substitutions or modifications made by those skilled in the art to the technical solutions and content disclosed in the present invention without departing from the scope of the present invention shall be deemed to have remained within the scope of protection of the present invention.

Claims

1. A method for diagnosing and optimizing the combustion state of an engine during cold start based on engine speed fluctuations, characterized in that, Engine speed and cylinder pressure data for each cycle of cold start under both normal and extreme conditions are collected as a training set. Combustion states are manually labeled, and a combustion state diagnostic model is trained based on past experimental data to identify abnormal speed fluctuations when abnormal combustion occurs. Another set of cold start experimental data from an engine under extreme conditions is used as a model validation set and compared with signals collected in real time by the cylinder pressure sensor to determine the model's accuracy and real-time performance. By selecting the Sigmoid and tanh activation functions and combining them with a multi-head attention mechanism, a GRU algorithm prediction model is constructed. A GRU algorithm prediction model combining multi-head attention mechanism is used. The input feature dimension `input_size` is set to 1, the output dimension `class_num` is set to 2, and the number of heads in the attention mechanism is set to 2. In the forward propagation of the multi-head attention mechanism, attention scores are calculated through dot product operations, and then scaled to stabilize training. A softmax function is applied to normalize the attention weights along a specific dimension, and then a weighted sum is calculated based on the attention weights. Finally, a fully connected layer performs a non-linear transformation on the attention-weighted result and connects it to the residual of the original signal to obtain the final attention output. The attention output is then fed into the GRU layer for sequence modeling and outputs the final hidden state. Finally, the softmax function provides the final prediction result. The aforementioned engine cold start combustion state diagnosis and optimization method based on engine speed fluctuation includes the following steps: (1) Conduct bench tests on the cold start performance of the engine with cylinder pressure sensor under different conditions to obtain the speed fluctuation during normal combustion and the speed fluctuation during abnormal combustion; (2) For the speed fluctuation during normal combustion, record the parameter data under normal combustion conditions in different environments. For the speed fluctuation during abnormal combustion, locate and mark the data information of the abnormal combustion state, and perform data verification at the same time. Then proceed to step (5). The parameter data under normal combustion conditions in different environments include fluctuation characteristics, speed increase rate, and cylinder pressure curve. The data information of the abnormal state includes cylinder state, cylinder pressure curve, and engine speed. (3) Train the combustion state diagnostic model using the data results from step (2); (4) Optimize relevant parameters; (5) Conduct accuracy verification of the combustion state diagnostic model. If the result is accurate, proceed to step (6). If the result is incorrect, return to step (4) or increase the data sample and start over. (6) Obtain a combustion state diagnostic model based on engine speed fluctuations; (7) Determine if an abnormal combustion condition has occurred: If yes, determine and return the speed and torque of each cylinder where the abnormal combustion state occurred, correct the engine control parameters in the cold start response MAP online or offline, record the engine instantaneous speed fluctuation signal each time the engine is cold started, and repeat step (6). If no, determine that the engine cold start performance is better.

2. The engine cold start combustion state diagnosis and optimization method based on speed fluctuation as described in claim 1, characterized in that, A multi-head attention mechanism GRU algorithm prediction model was trained on a dataset with engine crankshaft angles ranging from -50°CA to 70°CA, spanning 120°CA.

3. The engine cold start combustion state diagnosis and optimization method based on speed fluctuation as described in claim 1, characterized in that, Step (4) parameters include model parameters and feature functions.

Images