Engine cold start combustion state diagnosis and optimization method based on rotating speed fluctuation

Abstract

The invention discloses an engine cold start combustion state diagnosis and optimization method based on rotation speed fluctuation, which comprises the following steps: acquiring rotation speed and cylinder pressure data of each cycle in a cold start process of an engine under a conventional environment and an extreme working condition as a data training set, manually marking a combustion state, learning and training a combustion state diagnosis model based on previous experimental data, and optimizing the combustion state. The abnormal fluctuation of the rotating speed when abnormal combustion occurs is recognized, another group of engine cold start experimental data in the extreme environment is used as a model verification set and is compared with signals collected by a cylinder pressure sensor in real time, and the accuracy and the real-time performance of a model are judged. According to the method, the combustion state of each cylinder of the engine is judged according to the abnormal fluctuation condition of the rotating speed of the engine, and parameters such as the cylinders, the phase, the rotating speed and the torque of abnormal combustion in the cold start transient working condition process of the engine are recognized and serve as a control parameter MAP and correction parameter calibration judgment basis of the cold start transient process. And meanwhile, the effect after parameter correction can be further evaluated.

Description

Technical Field

[0001] The invention relates to engine technology and automatic control methods, and in particular to a method for diagnosing and optimizing engine cold start combustion states based on rotation speed fluctuations. Background Art

[0002] Engines installed on complete vehicles are generally not equipped with cylinder pressure sensors due to space, cost and other limitations. When developing engine cold start performance, the engine cold start performance and long-term engineering experience are often used to optimize and adjust the engine cold start calibration control parameters. There is a lack of effective theoretical basis and data support, and conventional methods cannot identify and accurately find the optimization direction of the calibration parameter MAP.

[0003] There is an urgent need for an accurate, efficient and simple method for diagnosing the combustion state of an engine during cold start. Summary of the invention

[0004] Purpose of the invention: The purpose of the present invention is to address the situation in which the speed of existing engines during the cold start phase is relatively slow and abnormal combustion is prone to occur, resulting in starting difficulties. At the same time, conventional methods for transient conditions such as cold starts are difficult to accurately determine the causes of performance degradation and the shortcomings of control parameter correction. The present invention provides an engine cold start combustion state diagnosis and optimization method based on speed fluctuations.

[0005] The engine speed is relatively slow during the cold start phase and is prone to abnormal combustion, which makes starting difficult. Conventional methods are difficult to accurately determine the causes of performance degradation and control parameter correction methods for transient conditions such as cold start. The present invention determines the combustion state of each cylinder of the engine by abnormal fluctuations in engine speed, identifies the cylinder, phase, speed, torque and other parameters where abnormal combustion occurs during the transient cold start process of the engine, and provides a basis for calibrating and determining the control parameter MAP and its correction parameters during the transient cold start process. It can also further evaluate the effect of parameter correction.

[0006] Technical solution: The present invention provides an engine cold start combustion state diagnosis and optimization method based on speed fluctuation. The speed and cylinder pressure data of each cycle of the engine cold start process under normal environment and extreme working conditions are collected as a data training set, and the combustion state is manually marked. The combustion state diagnosis model is learned and trained based on past experimental data to identify abnormal speed fluctuations when abnormal combustion occurs. Another set of extreme environment engine cold start experimental data is used as a model verification set, which is compared with the signal 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 are selected as Sigmoid function and tanh function, and the GRU algorithm prediction model is constructed in combination with the multi-head attention mechanism.

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

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

[0010] Further, the following steps are included:

[0011] (1) Carry out cold start performance bench tests under different conditions on engines equipped with cylinder pressure sensors to obtain speed fluctuations during normal combustion and speed fluctuations during abnormal combustion;

[0012] (2) For speed fluctuations during normal combustion, record parameter data under normal combustion conditions in different environments; for speed fluctuations during abnormal combustion, locate and mark data information of abnormal combustion conditions, perform data verification, and proceed to step (5);

[0013] (3) training a large combustion state recognition model using an intelligent algorithm and the data results of step (2);

[0014] (4) Optimize relevant parameters;

[0015] (5) Verify the accuracy of the combustion state diagnosis model. If the result is accurate, proceed to step (6). If the result is wrong, return to step (4) or increase the data sample and start again.

[0016] (6) Obtaining a combustion state diagnosis model based on engine speed fluctuation;

[0017] (7) Determine whether an abnormal combustion condition occurs: If so, determine and return the speed and torque of each cylinder at which the abnormal combustion condition occurs, 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 loop step (6); if not, determine that the engine cold start performance is better.

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

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

[0020] Engines installed on complete vehicles are generally not equipped with cylinder pressure sensors due to space, cost and other limitations. When developing engine cold start performance, the engine cold start performance and long-term engineering experience are often used to optimize and adjust the engine cold start calibration control parameters. There is a lack of effective theoretical basis and data support, and conventional methods cannot identify and accurately find the optimization direction of the calibration parameter MAP.

[0021] The present invention makes full use of the relatively low engine speed during the cold start phase through the speed sensor already provided on the engine, collects the speed signal of the engine during the cold start process in real time, combines the injection control parameters and the combustion phase of each cylinder, obtains the speed fluctuation increase curve when the engine has better cold start performance, filters and analyzes the corresponding relationship between each speed fluctuation parameter and the combustion state of each cylinder, and records the speed curve and fluctuation range of efficient combustion of each cylinder during the cold start process of the engine.

[0022] When the engine fails or the environment conditions during cold start are bad, the cold start process will have unstable combustion. By collecting the speed fluctuations during the cold start process and comparing them with the previously trained cold start speed fluctuation state model, and comparing the speed fluctuation deviations of each speed fluctuation during the cold start process with the speed fluctuations of the normal combustion state, it is determined whether the combustion state under the response speed and the ignition cylinder has abnormal combustion or even misfire, thereby providing a basis for the cold start process injection parameter MAP calibration. After calibration, this method can also be used to determine whether the optimized parameters have achieved good results when starting again.

[0023] Unlike engine bench tests, the engine mounted on the vehicle will not be equipped with expensive and unreliable cylinder pressure sensors, but will definitely be equipped with speed sensors and timing chains to control the injection and combustion phase of the engine. During the cold start of the vehicle, it is impossible to obtain cylinder pressure and thus impossible to determine the combustion conditions in the cylinder during the cold start process, which makes it difficult to accurately determine and optimize the cold start performance of the vehicle.

[0024] Taking into account that the in-cylinder combustion conditions of each cycle in the cold start acceleration stage will affect the engine speed rise rate, there is a certain complex relationship between the in-cylinder pressure and the engine speed, which are the basis for judging the combustion state. The present invention proposes an engine cold start speed fluctuation-cylinder pressure judgment method based on an intelligent algorithm. By collecting the speed and cylinder pressure data of each cycle of the cold start process of the engine under normal environments and extreme working conditions as a data training set, the normal, abnormal, misfire and other combustion states are artificially marked, and the combustion state diagnosis model is learned and trained based on past experimental data to accurately identify abnormal speed fluctuations when abnormal combustion occurs. Then, another group of extreme environment engine cold start experimental data is used as a model verification set, which is compared with the signal collected in real time by the cylinder pressure sensor to judge the accuracy and real-time performance of the model.

[0025] Various artificial intelligence algorithms such as neural networks, machine learning and reinforcement learning can be applied to the combustion diagnosis algorithm. It can be adjusted and improved according to different application environments and engine characteristics. Through training with a large amount of engine cold start experimental data, the appropriate activation function can be selected, and the diagnostic phase and judgment period caused by the torque transmission path and engine structural parameters can be properly fine-tuned. According to the model verification results, the model parameters such as the penalty function and weight factor are optimized. After a certain number of iterations and optimizations, an engine cold start combustion state diagnosis model with good accuracy can be obtained. The logic and method of constructing the diagnostic model are basically the same.

[0026] By diagnosing the cold start process of the engine, the number of abnormal combustion states, speed, degree of degradation, and corresponding cylinder information during the cold start transient working process can be determined. Based on this, the control parameters of the gas circuit, oil circuit, and circuit related to the cold start of the engine and the corrected MAP under different conditions can be accurately located, and the relevant parameters can be corrected according to engineering experience. This method can also quickly and effectively determine whether the cold start performance of the engine is improved after the relevant control parameters are corrected.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] 1. This method determines the combustion state of each cylinder of the engine through the abnormal fluctuation of the engine speed, identifies the cylinder, phase, speed and torque parameters where the abnormal combustion occurs during the transient working condition of the engine cold start, and provides a basis for calibrating and determining the control parameter MAP and its correction parameters during the transient process of the cold start. At the same time, it can also further evaluate the effect of the parameter correction.

[0029] 2. This method can rely on the existing sensors of the engine to perform an approximate analysis of the combustion state of each cylinder during the cold start process. There is no need to install an expensive cylinder pressure sensor. It only needs to process and analyze the transient signal collected by the speed sensor. It is low-cost and highly versatile.

[0030] 3. This method can accurately identify abnormalities in the combustion state of each cylinder through algorithm optimization. Traditionally, a 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. The combustion diagnosis and recognition algorithm can be used in an independent analysis device. The combustion status of each cylinder can be analyzed by inputting the engine speed sensor signal, basic engine structural parameters, etc. The analysis can be performed online in real time or offline based on data. At the same time, the algorithm can also be directly written to the engine controller for online or offline application, or implanted in the cloud for remote data analysis and diagnosis. The application method is flexible and can be freely configured according to the controller computing power and application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 The training data is sampled for the cold start process at an altitude of 4000m and an ambient temperature of -10℃;

[0033] Figure 2 It is a schematic diagram of the combustion state;

[0034] Figure 3 Built for LSTM layers and fully connected layers;

[0035] Figure 4 is the Sigmoid activation function;

[0036] Figure 5 is the tanh activation function;

[0037] Figure 6 is the Soffmax function;

[0038] Figure 7 Comparison of the model test set verification results, including: (a) cold start cylinder pressure curve and LSTM model judgment result based on 720°CA data set, (b) cold start speed curve and LSTM model judgment result based on 720°CA data set, (c) cold start cylinder pressure curve and LSTM model judgment result based on 120°CA (-60~60) data set, (d) cold start speed curve and LSTM model judgment result based on 120°CA (-60~60) data set;

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

[0040] Fig. 9 It is a GRU algorithm model combined with a multi-head attention mechanism;

[0041] Fig.10Comparison of model test set validation results, including (a) cylinder pressure curve and GRU model judgment result based on 720°CA dataset, (b) speed curve and GRU model judgment result based on 720°CA dataset, (c) cold start cylinder pressure curve and GRU model judgment result based on 120°CA (-60~60) dataset, (d) cold start speed curve and GRU model judgment result based on 120°CA (-60~60) dataset;

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

[0043] Fig.12 To compare and optimize the engine cold start combustion state diagnosis model based on speed fluctuation;

[0044] Fig.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) the speed curve and the model judgment result, and (b) the cylinder pressure curve and the model judgment result.

[0045] Fig.14 It is the IMEP diagram of the first 15 cycles of the cold start-up period of the test engine at an altitude of 3000m and an ambient temperature of -20℃. Fig.15 The figure is a flow chart of the diagnosis and optimization method of the present invention. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be further described below.

[0047] This embodiment uses the in-cylinder combustion and speed fluctuation data of a straight-six-cylinder diesel engine during the cold start process under different ambient temperatures and altitudes as a data set to train a combustion diagnosis model and verify the feasibility of the diagnosis method.

[0048] The cold start bench tests of diesel engines were conducted multiple times under high-altitude low-temperature conditions at an altitude of 0m to 4000m and an ambient temperature of 20 to -30℃. The cylinder pressure fluctuation and speed fluctuation data of each working cycle of the entire cold start process were obtained, and the data were checked and verified cycle by cycle. High-quality data sets can provide accurate, consistent and sufficient signals, so that the neural network can effectively learn the key features and laws of the target task during the training process, thereby improving the prediction accuracy and generalization ability of the model. When constructing and preparing data sets, great attention should be paid to their quality control and optimization. Figure 1 .

[0049] The expected effect based on the model is to judge whether the cylinder pressure value is normal according to the speed fluctuation during the cold start process, that is, whether abnormal combustion has occurred. When constructing the data set, the speed and cylinder pressure data of the cycle are read through the data of the Kibox combustion analyzer, and the combustion data is used to judge whether it is normal combustion, so as to annotate the results of this data set. Figure 2 The figure shows the combustion state curve of diesel ignition. When the combustion is normal in the cylinder, the average cylinder pressure of the engine is large, the work capacity is strong, and the speed will increase quickly. On the contrary, when abnormal combustion occurs in the cylinder, the average cylinder pressure of the engine is small, the work capacity is weak or even negative work is done, and the speed increase rate is slow or even drops. Figure 2 .

[0050] In order to accurately determine at which speed and in which cylinder the abnormal combustion occurred during the high transient process of the engine cold start, the corresponding scale is defined according to the number and position of the diesel engine cylinder pressure sensor installed when constructing the data set, so as to improve the effect of data training and combustion diagnosis and recognition. For example, in this embodiment, the diesel engine cylinder pressure sensor is arranged in the fifth cylinder. In order to improve the accuracy of combustion state recognition, two judgment scales are used:

[0051] (1) The first data division scale is to describe the combustion conditions of all cylinders based on the cylinder pressure of the fifth cylinder. Under this data division scale, a single data set covers the transient process of a complete cycle of 720° crankshaft rotation angle, and the average combustion state of the six cylinders of the engine is determined by the speed fluctuations of this cycle.

[0052] (2) Another case is when the number of engine cylinders N is such that the ignition interval of each cylinder is 720° / N. Considering the uniformity of the ignition intervals of each cylinder in an 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.

[0053] The specific judgment cycle to be adopted mainly depends on hardware factors such as the controller operation speed and the engine speed. Considering that the energy generated by the power stroke needs to pass through a transmission chain composed of a series of mechanical components to be transmitted to the engine crankshaft and cause the speed fluctuation, there may be a certain delay between the generation of speed fluctuation and the change of combustion state. Therefore, a study on the impact of different moderate delays was carried out with the ignition interval as the center to further improve the accuracy of combustion state diagnosis. This embodiment is a six-cylinder diesel engine, so 120°CA is selected as the speed fluctuation range, and delays of 10°CA and 20°CA are respectively performed to verify the evaluation results. Table 1 shows the data set scheme divided by 120°CA.

[0054] Table 1 Dataset scheme for 120°CA partitioning

[0055]

[0056] (1) Construction of the speed fluctuation-combustion state judgment model integrating LSTM

[0057] The LSTM training process is mainly divided into 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. Secondly, the forward propagation is calculated. Finally, the gradient is calculated by the gradient descent method and the optimizer is used to update the model parameters. Figure 3 .

[0058] Figure 3 The construction of LSTM layer and fully connected layer is shown. The input of the model is only the speed sequence data, so the input feature dimension input_size is 1. The expected effect of model training is to give a judgment result for the current cycle, so the output dimension class_num is 2, that is, to give a judgment on whether the current cycle is a misfire cycle. The length of the input sequence data sequence_size of LSTM is set to 720 and 120 according to the different scales of the data set division. In order to map the hidden state of the last layer of LSTM to the output dimension class_num, a fully connected layer is defined. The weights of the fully connected layer need to be properly initialized. Improperly initialized weights will lead to gradient explosion or gradient disappearance problems, and the final model cannot be used. Considering the high transient state of the diesel engine cold start process and the activation function used inside the LSTM unit, Xavier initialization is selected to initialize the weights of the fully connected layer. For each layer, Xavier can ensure the consistency of variance.

[0059] Activation function plays a vital role in neural networks. It can help neural networks overcome linear limitations and enable them to capture nonlinear relationships in data. At the same time, the activation function can map the input of the previous layer to more advanced features, which is conducive to capturing the intrinsic structure and discriminative information of the data. For classification tasks, the activation function can generate nonlinear decision boundaries. The existence of activation functions greatly improves the learning ability and application scope of neural networks.

[0060] Table 2 shows the activation functions and expressions used in this method, where Sigmoid and tanh activation functions are used in the LSTM forward propagation process, and Softmax activation function is used to give the model judgment result.

[0061] Table 2 Activation functions and expressions involved in the model

[0062]

[0063] Figure 4 is the Sigmoid activation function. Since the output range of the Sigmoid function is 0 to 1, it is often used as the activation function of the model that uses the predicted probability as the output. The Sigmoid function can map the received input to the interval (0, 1). This feature can well simulate the gating mechanism in LSTM and control the ratio of information flow to be retained and discarded. Although the Sigmoid function will have the problem of gradient saturation due to the derivative approaching 0 at both ends, the gating signals in LSTM are independent of each other, and other gates can still be calculated effectively, which can well avoid this shortcoming of the Sigmoid function. Figure 4 .

[0064] Figure 5 The tanh function is shown, which can map the received input to the range of (-1, 1), and the output has the characteristics of symmetry and centralization. Compared with the Sigmoid function, the output of the tanh function has a larger dynamic range and can provide a larger information storage space. At the same time, its 0-centered feature can maintain numerical stability. Although the tanh function also has the phenomenon that the gradient is close to 0 at both ends, the bilateral saturation feature enables it to have learning capabilities in both positive and negative directions. Therefore, it is often used as an activation function in LSTM to generate candidate cell states and update actual cell states.

[0065] Figure 6The Softmax function is shown. It is a normalized exponential function that can normalize a set of real values ​​into a set of probability distributions. Therefore, it is often used in multi-class classification problems. In multi-class classification problems, the Softmax function amplifies the gap between the original scores through exponential operations, so that the probability of the category with a higher original score is significantly higher than that of other categories. Therefore, through the probability distribution output by the Softmax function, the category with the highest probability is selected as the prediction result.

[0066] Finally, the learning rate was selected as 0.001, and the activation functions were Sigmoid and tanh to build the LSTM prediction model, and 1500 iterations of training were performed. The model with the best indicators was selected as the final prediction model. The relevant indicators are shown in Table 3.

[0067] Table 3 LSTM model evaluation indicators

[0068]

[0069] Figure 7 The comparison between the LSTM model judgment results and the experimental results of the data set divided by 720°CA and 120°CA is shown. Figure 7 (a)(b) It can be seen that during the cold start acceleration period, the LSTM model based on the 720°CA dataset can well capture the misfire cycle through speed fluctuations. There were 6 large speed fluctuations during the acceleration period, corresponding to 6 misfire cycles. The LSTM model based on the 720°CA dataset gave accurate results, but there were also 3 wrong judgments. When the speed just reached the target idle speed, the ECU would reduce the cyclic injection amount, which usually caused a short-term deterioration of the in-cylinder combustion in the low temperature environment of the plateau. The misfire cycle occurred at this operating point in the test set, and the LSTM model based on the 720°CA dataset accurately judged this situation. However, during the idle stage, due to the low temperature external environment of the plateau and the cold start condition, the diesel engine intake conditions and the formation of the combustible mixture in the cylinder will be greatly affected, thereby affecting the cylinder pressure peak of normal combustion in the cylinder, and finally causing fluctuations in speed. Although the model successfully judged 2 misfire cycles in this stage, it produced an additional 4 wrong judgments.

[0070] analyze Figure 7(c)(d), during the cold start acceleration period, the LSTM model based on the 120°CA (-60~60) dataset only made one misjudgment of the misfire cycle at the 6496°CA crankshaft angle. The remaining misfire cycles were able to give correct judgment results through speed fluctuations and no misfire cycles were missed. At the same time, the model also accurately identified the misfire cycle caused by the ECU starting to reduce fuel injection when the speed just reached idle. In the subsequent idle stage, the LSTM model based on the 120°CA (-60~60) dataset performed better, successfully judging two cylinder misfires from speed fluctuations, and the number of misjudgments was 2 times, which was 2 times less than the model of the 720°CA dataset.

[0071] Figure 8 The comparison between the LSTM model judgment results and the experimental results of the data set divided by 120 degrees is shown. Figure 8 (a)(b) Analysis shows that the LSTM model based on the 120°CA (-50~70) data set did not miss any misfire cycles during the acceleration period. It judged 6 misfire cycles through 6 large speed fluctuations, but made 2 misjudgments, judging normal combustion cycles as misfire cycles. During the idling stage, the model judged 2 misfire cycles through speed fluctuations, but made 3 additional misjudgments of misfire cycles. Figure 8 (c) (d) Analysis shows that the LSTM model based on the 120°CA (-50~70) data set had 6 large speed fluctuations during the acceleration period, successfully realizing the identification of 6 misfire cycles, but additionally misjudged 3 normal combustion cycles. During the idling stage, the model realized the judgment of 2 misfire cycles through speed fluctuations, but additionally misjudged 3 misfire cycles.

[0072] From the analysis of the LSTM model performance evaluation indicators in Table 3, the LSTM model based on the 120°CA (-60~60) dataset has the best performance, with an accuracy of 94.20% and a precision of 94.76%, which shows that the model has sufficient credibility in judging the misfire cycle. Considering that the diesel engine cylinder pressure data is the fifth cylinder and its actual ignition range, the LSTM model based on the 120°CA dataset is closer to the actual situation in terms of the dataset. In principle, the LSTM model based on the 120°CA dataset should have better performance indicators. Comparing the LSTM model performance evaluation indicators of the 720°CA and 120°CA datasets, the LSTM model based on the 120°CA dataset does have better performance indicators than the LSTM model based on the 720°CA dataset.

[0073] (2) Construction of an improved GRU algorithm model combined with attention mechanism

[0074] The calculation of GRU depends on the update of the previous time step, while the calculation of the multi-head attention mechanism is independent of the sequence position. This feature can be combined with the parallel computing capabilities of modern hardware to improve the model training speed. Therefore, the GRU algorithm combined with the multi-head attention mechanism has advantages in parallel computing and higher overall efficiency. At the same time, the GRU algorithm combined with the multi-head attention mechanism recalculates the weight distribution of the past moment at each new time, allowing the model to flexibly focus on the most relevant input part according to the current task requirements, improving the adaptability and accuracy of the model. The GRU algorithm combined with the multi-head attention mechanism has many advantages, which enables the trained model to show stronger performance and generalization ability in sequence prediction tasks.

[0075] Fig. 9 The construction of the GRU algorithm combined with the multi-head attention mechanism is demonstrated. The input feature dimension and output feature dimension 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, the attention score is calculated by the dot product operation, and then the attention score is scaled to stabilize the training. The softmax function is applied to normalize the attention weight along a specific dimension, and then the weighted sum is calculated according to the attention weight. Finally, the weighted result of the attention is nonlinearly transformed through a fully connected layer and connected with the original signal residual to obtain the final attention output. The result after the attention output is passed to the GRU layer for sequence modeling, and the final hidden state is output. Finally, the final prediction result is given by the softmax function.

[0076] In this embodiment, the activation functions are selected as Sigmoid function and tanh function, and the GRU algorithm prediction model is constructed in combination with the multi-head attention mechanism. Similarly, according to the sequence data of different scales, the length of the input sequence data is adjusted, and 1500 iterations of training are performed, and the one with the best performance index is selected. Table 4 shows the training results of the GRU model combined with the multi-head attention mechanism.

[0077] Table 4 Evaluation indicators of the GRU model combined with multi-head attention mechanism

[0078]

[0079] Fig.10 The comparison between the judgment results of the GRU model combined with the attention mechanism and the experimental results of the data set divided into 720°CA and 120°CA is shown. Fig.10(a)(b) It can be seen that in the acceleration stage, the judgment results of the GRU model based on the 720°CA dataset are very accurate. It can judge 5 misfire cycles through speed fluctuations without misjudgment. However, when the idle speed is just reached, the ECU reduces the cycle injection amount, which leads to misfire cycles. The GRU model based on the 720°CA dataset failed to make a correct judgment. In the idle stage, the GRU model combined with the attention mechanism correctly identified 2 misfire cycles. Although there were 3 wrong judgments of judging normal ignition cycles as misfire cycles, it was reduced by 1 error compared with the LSTM model. The diesel engine cold start speed fluctuation-cylinder pressure judgment model constructed by the GRU algorithm with 720°CA partitioned dataset and combined with multi-head attention mechanism has an accuracy of 95.26%, and the f1 score reaches 0.9532. It has excellent performance in judging the in-cylinder combustion conditions of the diesel engine during the cold start and acceleration period. Analysis Fig.10 (c)(d) It can be seen that the GRU model based on the 120°CA (-60~60) data set correctly judged 5 misfire cycles in the acceleration stage. Although there was 1 wrong judgment, it was able to identify the misfire cycle caused by the ECU reducing fuel injection when the speed just reached the idle stage. In the idle stage, the model still performed well in identifying the misfire cycle, successfully judging the occurrence of the misfire cycle through two speed fluctuations, but there were also two wrong judgments.

[0080] Fig.11 The comparison between the judgment results of the GRU model with a 120-degree partition of the data set and the multi-head attention mechanism and the experimental results is shown. Fig.11 During the cold start acceleration period, the GRU model based on the 120°CA (-50~70) dataset has the best performance, and can accurately judge 6 misfire cycles based on the speed fluctuation, without any misjudgment or omission. The GRU model of the 120°CA (-40~80) dataset identified the normal combustion cycle as a misfire cycle twice during the acceleration period, but did not miss any misfire cycles. At the moment of just reaching idle speed, the ECU reduced the cyclic injection amount, which led to a misfire cycle. The GRU model based on the 120°CA dataset can accurately capture this situation and make the correct judgment. During the idle phase, the GRU models based on the 120°CA (-50~70) dataset and the 120°CA (-40~80) dataset both identified the misfire cycle. The difference between them is that the model based on the 120°CA (-50~70) dataset misjudged the normal combustion cycle once, while the model based on the 120°CA (-40~80) dataset misjudged the normal combustion cycle as misfire cycle four times.

[0081] A comprehensive comparison of the three GRU models of the 120°CA dataset shows that the GRU model trained on the -50~70°CA dataset and based on the multi-head attention mechanism has the best performance. This is consistent with the performance evaluation indicators of the model. The GRU model of the 120°CA (-50~70) dataset has an accuracy of 97.01%, which means that it has a high reliability in judging a certain working cycle as a misfire cycle based on speed fluctuations; at the same time, the f1 score is as high as 0.9742, proving that the model has good performance in judging behavior. Comparing the GRU models of the 720°CA dataset and the 120°CA dataset, the GRU model of the 120°CA dataset is better than the GRU model of the 720°CA dataset in all three performance evaluation indicators, which is consistent with the expectations of scientifically dividing the dataset.

[0082] Table 5 shows the constructed cold start speed fluctuation-combustion state judgment model number and the corresponding algorithm and data set. Fig.12 The following is a comparison chart of the results of different cold start speed fluctuation-combustion state judgment models. Fig.12 It can be seen that all models basically achieve the correct recognition of misfire cycles during the acceleration period and the idling period. Only the LSTM algorithm model (model IV) based on the 120°CA (-40~80) dataset recognizes the misfire cycle as a normal combustion cycle once. Among all the models, the best performance is the GRU algorithm model (model VII) based on the 120°CA (-50~70) dataset combined with the attention mechanism. It accurately captures each misfire cycle and achieves the minimum number of judgments on normal combustion cycles. From the perspective of the composition of the dataset, the sequence data length of the 720°CA dataset is 6 times that of the 120°CA dataset, which is not conducive to model memory, resulting in a certain degree of forgetting of previous information in the later stage of training. From the algorithmic analysis, the attention mechanism is used to calculate the global attention, so that each part retains the global information related to itself, and then the gating mechanism of the GRU ensures the preservation of long-term sequence information, which will not be clear over time or prediction irrelevant, and finally achieves comprehensive consideration of the changes in the entire time period.

[0083] Table 5 Cold start speed fluctuation-combustion state judgment model number and corresponding algorithm and data set

[0084] 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)

[0085] The figure below shows the experimental verification of the algorithm on another 4L diesel engine. It is the output result of the speed fluctuation-cylinder pressure judgment model of the diesel engine cold start at an altitude of 3000m and an ambient temperature of -20℃. Fig.13(a) It can be seen that the algorithm model calculated that an abnormal combustion state occurred in the range of 2880-3240°CA based on the input speed signal. According to the number of cylinders of the test engine, using this combustion state diagnosis algorithm, it was found that misfire cycles occurred in two cylinders in the working cycle, which caused fluctuations in speed.

[0086] Through the combustion diagnosis program, it was found that the engine fuel control parameters where abnormal combustion occurred at a speed of 150r / min needed to be adjusted and optimized. In the embodiment, the time of the first pre-injection of the engine at a speed of 150r / min was advanced to 125°CA before the compression top dead center (a total of 2 pre-injections), thereby improving the combustion performance at the current speed and making the engine cold start process more stable and faster.

[0087] The above is only a preferred embodiment of the present invention and does not limit the present invention in any way. Any technician in the relevant technical field, without departing from the scope of the technical solution of the present invention, makes any form of equivalent replacement or modification to the technical solution and technical content disclosed in the present invention, which does not depart from the content of the technical solution of the present invention and still falls within the protection scope of the present invention.

Claims

1. A method for diagnosing and optimizing engine cold start combustion state based on speed fluctuation, characterized in that: The speed and cylinder pressure data of each cycle of the engine cold start process under normal environment and extreme working conditions are collected as the data training set, and the combustion state is manually marked. The combustion state diagnosis model is trained based on past experimental data to identify abnormal speed fluctuations when abnormal combustion occurs. Another set of engine cold start experimental data in extreme environment is used as the model verification set, which is compared with the signal collected in real time by the cylinder pressure sensor to determine the accuracy and real-time performance of the model.

2. The engine cold start combustion state diagnosis and optimization method based on speed fluctuation according to claim 1, characterized in that: The activation functions are selected as Sigmoid function and tanh function, and the GRU algorithm prediction model is constructed in combination with the multi-head attention mechanism.

3. The engine cold start combustion state diagnosis and optimization method based on speed fluctuation according to claim 1, characterized in that: A GRU model with a multi-head attention mechanism trained on the 120°CA (-50~70) dataset is used.

4. The engine cold start combustion state diagnosis and optimization method based on speed fluctuation according to claim 1, characterized in that: The LSTM model based on the 120°CA (-60~60) dataset is adopted.

5. The engine cold start combustion state diagnosis and optimization method based on speed fluctuation according to claim 1, characterized in that: The steps include: (1) Carry out cold start performance bench tests under different conditions on engines equipped with cylinder pressure sensors to obtain speed fluctuations during normal combustion and speed fluctuations during abnormal combustion; (2) For speed fluctuations during normal combustion, record parameter data under normal combustion conditions in different environments; for speed fluctuations during abnormal combustion, locate and mark data information of abnormal combustion conditions, perform data verification, and proceed to step (5); (3) training a large combustion state recognition model using an intelligent algorithm and the data results of step (2); (4) Optimize relevant parameters; (5) Verify the accuracy of the combustion state diagnosis model. If the result is accurate, proceed to step (6). If the result is wrong, return to step (4) or increase the data sample and start again. (6) Obtaining a combustion state diagnosis model based on engine speed fluctuation; (7) Determine whether an abnormal combustion condition occurs: If so, determine and return the speed and torque of each cylinder at which the abnormal combustion condition occurs, 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 loop step (6); if not, determine that the engine cold start performance is better.

6. The engine cold start combustion state diagnosis and optimization method based on speed fluctuation according to claim 5, characterized in that: The parameter data of step (2) under normal combustion conditions in different environments include fluctuation characteristics, speed increase rate, and cylinder pressure curve.

7. The engine cold start combustion state diagnosis and optimization method based on speed fluctuation according to claim 5, characterized in that: The data information of the abnormal state in step (2) includes cylinder state, cylinder pressure curve, and engine speed.

8. The engine cold start combustion state diagnosis and optimization method based on speed fluctuation according to claim 5, characterized in that: The parameters in step (4) include model parameters and characteristic functions.

Images