A hydrogen internal combustion engine early ignition rapid detection and identification method

Abstract

The application discloses a hydrogen internal combustion engine early combustion rapid detection and identification method, and the method comprises the following steps: after the collected original cylinder pressure signal to be detected is pretreated, the original cylinder pressure sequence to be detected is input into a convolutional autoencoder model to obtain a reconstructed cylinder pressure sequence to be detected, and the reconstruction error between the original cylinder pressure sequence to be detected and the reconstructed cylinder pressure sequence to be detected is calculated in real time; if the reconstruction error is within the initial screening threshold range of abnormality judgment, the current cycle is directly determined as normal combustion, and the detection is ended; otherwise, the current cycle is determined as a suspected abnormal cycle of suspected abnormal combustion, and early combustion accurate judgment and grading based on double-index Mahalanobis distance are triggered; the early combustion accurate judgment and grading based on double-index Mahalanobis distance comprises grading the suspected abnormal cycle into normal combustion, strong early combustion and weak early combustion according to the size of the Mahalanobis distance. Through the construction of a two-stage detection mechanism, the machine learning is used for rapid initial screening of the cylinder pressure signal, and then the double-index fusion mutual judgment method is combined, so that the detection accuracy is ensured, and the calculation cost is significantly reduced.

Description

Technical Field

[0001] This invention relates to the field of combustion diagnostic technology for hydrogen internal combustion engines, and in particular to a rapid detection method for pre-ignition in hydrogen internal combustion engines, which can be used for the rapid and accurate identification of abnormal combustion phenomena such as pre-ignition. Background Technology

[0002] Hydrogen energy has become the most promising fossil fuel alternative in internal combustion engine applications due to its outstanding environmental advantages. However, hydrogen internal combustion engines are prone to pre-ignition under high-load conditions due to the wide flammability limit, low ignition energy, and high flame velocity of hydrogen fuel. Pre-ignition can induce knocking, leading to extremely high pressure oscillations, severe thermal loads, damage to internal combustion engine components, and seriously affecting the engine's lifespan and stability.

[0003] Traditional pre-ignition detection methods rely on independent judgments based on cylinder pressure signal thresholds or combustion thermodynamic characteristic parameters. However, threshold-based methods are significantly affected by noise interference and transient changes in operating conditions, resulting in poor detection robustness. Furthermore, the limited information dimension of a single indicator makes it difficult to accurately capture the subtle features of weak pre-ignition stages, leading to insufficient detection sensitivity. While combustion thermodynamic parameter-based methods can reflect the combustion state, they require complex integration and thermodynamic calculations on cylinder pressure signals, resulting in high computational load and long cycles, which cannot meet the real-time control requirements of onboard ECUs, exhibiting significant computational delays and response lags. Introducing supervised learning methods such as convolutional neural networks or recurrent neural networks has improved recognition capabilities to some extent, but these methods are highly dependent on high-quality labeled data. Given the sporadic and highly destructive nature of pre-ignition in hydrogen internal combustion engines, constructing a complete fault dataset covering multiple operating conditions and intensities is extremely difficult. Therefore, with the increasing real-time requirements for pre-ignition detection in hydrogen internal combustion engines, a fast and accurate pre-ignition detection and identification method is urgently needed. Summary of the Invention

[0004] In response to existing technologies, this invention provides a rapid detection and identification method for pre-ignition in hydrogen internal combustion engines. It uses an unsupervised learning model to quickly screen cylinder pressure signals, and then combines a dual-index fusion and cross-judgment method to significantly reduce computational costs and time delays while ensuring detection accuracy.

[0005] To address the aforementioned technical issues, this invention proposes a rapid detection and identification method for pre-ignition in hydrogen internal combustion engines. This method performs a rapid coarse judgment of pre-ignition based on the combustion state of the original cylinder pressure signal under test using a convolutional autoencoder (CAE). If the result of the rapid coarse judgment of pre-ignition is suspected abnormal combustion, a further fine judgment and classification of pre-ignition based on the squared Mahalanobis distance of the dual indicators is then performed.

[0006] Rapid early combustion coarse judgment is performed based on the combustion state of the original cylinder pressure signal under test using a convolutional autoencoder (CAE). This includes: preprocessing the acquired original cylinder pressure signal under test, inputting the original cylinder pressure sequence under test into the convolutional autoencoder model to obtain the reconstructed cylinder pressure sequence under test, and calculating the reconstruction error between the original cylinder pressure sequence under test and the reconstructed cylinder pressure sequence under test in real time; if the reconstruction error is within the initial screening threshold range for anomaly judgment, it is directly judged as normal combustion, and the detection ends; otherwise, the current cycle is judged as a suspected abnormal cycle with suspected abnormal combustion, triggering early combustion fine judgment and classification based on the dual-index squared Mahalanobis distance.

[0007] Pre-ignition precision judgment and classification based on dual-index squared Marsh distance, including classifying the suspected abnormal cycle as normal combustion, strong pre-ignition, or weak pre-ignition according to the magnitude of the squared Marsh distance.

[0008] Furthermore, in the rapid detection and identification method for pre-ignition in hydrogen internal combustion engines described in this invention, the construction and training of the convolutional autoencoder model and the determination of the pre-ignition coarse judgment threshold include:

[0009] S1-1. Construct an unsupervised learning sample library: When the hydrogen internal combustion engine is working normally, collect the original cylinder pressure signals of no less than 500 normal combustion working cycles, and construct a training set and a validation set containing normal combustion samples.

[0010] S1-2. Data Standardization Preprocessing: After filtering, normalizing, and phase alignment of the original cylinder pressure signals in the training set, the original cylinder pressure sequence is obtained. Filtering removes high-frequency noise while preserving signal characteristics; phase alignment is performed on all sample cycles to ensure that the combustion characteristics between different cycles are strictly consistent in time; the normalization process maps the original cylinder pressure signals to the [0, 1] interval to eliminate differences in absolute pressure values ​​under different load conditions. The calculation expression is as follows:

[0011] (1);

[0012] In equation (1), The original cylinder pressure value before normalization. This is the normalized cylinder pressure value. and These represent the minimum and maximum cylinder pressure values ​​in this working cycle, respectively.

[0013] S1-3. Construction and Training of the Convolutional Autoencoder Model: First, a convolutional autoencoder model is constructed, comprising an encoder, a decoder, and a loss function. The encoder extracts cylinder pressure temporal features representing the combustion state from the original cylinder pressure sequence and compresses these features into a latent feature vector. The decoder generates a reconstructed cylinder pressure sequence based on the latent feature vector. The loss function measures the difference between the original and reconstructed cylinder pressure sequences. Then, the model is trained unsupervised using a normal sample training set. The optimization objective is to minimize the mean squared error (MSE) between the original and reconstructed cylinder pressure sequences. The formula for calculating the mean squared error is as follows:

[0014] (2);

[0015] In equation (2), Let N be the reconstruction error of the i-th cycle, and N be the number of sampling points of the original cylinder pressure sequence in a single working cycle. This represents the original cylinder pressure sequence value at the j-th sampling point in the i-th work cycle. This represents the reconstructed cylinder pressure sequence value at the j-th sampling point in the i-th working cycle.

[0016] S1-4. Statistical Threshold Definition: Based on the mean and standard deviation of the reconstruction error of the validation set samples, the coarse threshold for premature combustion is determined; the calculation expressions for the mean and standard deviation of the reconstruction error are as follows:

[0017] (3);

[0018] (4);

[0019] In equations (3) and (4), To reconstruct the mean error, M is the total number of samples in the validation set. To reconstruct the standard deviation of the error; based on the calculation results, the range of the coarse threshold for pre-ignition is determined as follows: .

[0020] S1-5. Online Real-Time Coarse Judgment: The preprocessed single-cycle raw cylinder pressure signal acquired in real time is input into the model to obtain the reconstructed cylinder pressure sequence, and its reconstruction error MSE_test is calculated in real time. If MSE_test is within the judgment threshold range, it is directly judged as normal combustion, and the detection ends; otherwise, it is marked as suspected abnormal combustion, triggering the fine judgment process.

[0021] Furthermore, in the rapid detection and identification method for pre-ignition in hydrogen internal combustion engines described in this invention, the calculation based on the dual-index squared Mahalanobis distance includes:

[0022] S2-1. Constructing the baseline feature space: Calculate the dual-feature indices of all normal samples in the training and validation sets, and construct a two-dimensional feature vector set composed of the maximum pressure rise rate and the CA5 pressure rise rate. Simultaneously, calculate the mean vector of this two-dimensional feature vector set. Covariance Matrix A statistical baseline distribution for normal combustion is established, which is composed as follows:

[0023] (5);

[0024] (6);

[0025] In equations (5) and (6), Let i be the two-dimensional feature vector of the i-th cycle. The maximum pressure rise rate in the i-th cycle. Let CA5 be the pressure rise rate in the i-th cycle. This represents the total number of samples in the training and validation sets.

[0026] The expression for calculating the mean vector is as follows:

[0027] (7);

[0028] The expression for calculating the covariance matrix is ​​as follows:

[0029] (8);

[0030] S2-2, Mahalanobis Distance Statistical Determination: For suspected abnormal cycles, calculate the maximum pressure rise rate and CA5 pressure rise rate of the original cylinder pressure signal to be tested, constructing a two-dimensional feature vector of the sample. The statistical distance between this two-dimensional feature vector and the baseline distribution center is calculated using the Mahalanobis distance algorithm. Mahalanobis distance can eliminate dimensional differences and correlation interference between two indicators. The expression for calculating the squared Mahalanobis distance is as follows:

[0031] (9);

[0032] In equation (9), This is the two-dimensional feature vector of the sample to be tested;

[0033] S2-3, Fault Classification and Identification: Based on the magnitude of the squared Mahalanobis distance, the suspected abnormal cycle is classified as normal combustion, strong pre-ignition, or weak pre-ignition. If the squared Mahalanobis distance is less than the first threshold, it is judged as normal combustion; if the squared Mahalanobis distance is between the first and second thresholds, it is judged as weak pre-ignition; if the squared Mahalanobis distance is greater than the second threshold, it is judged as strong pre-ignition. The two-dimensional feature vector composed of the maximum pressure rise rate and the CA5 pressure rise rate approximately follows a two-dimensional normal distribution; therefore, the squared Mahalanobis distance of the test sample... It approximately follows a chi-square distribution with 2 degrees of freedom. and adopt Quantiles determine the judgment threshold; In this text, the subscript 2 represents the degree of freedom. It is the quantile probability, determined. The first threshold, The second threshold is used, therefore the Mahalanobis distance method is used to determine the threshold as follows: when When the cycle is considered normal combustion, it is determined that the cycle is normal combustion; when If the cycle is determined to be a strong pre-ignition, it is determined to be a weak pre-ignition.

[0034] Compared with the prior art, the beneficial effects of the present invention are:

[0035] (1) The present invention is based on an unsupervised learning model of convolutional autoencoder, which can achieve coarse judgment of combustion state by relying only on the original cylinder pressure signal. It does not require complex integral, differential and heat release rate calculation for each cycle. The detection process has low computational load and fast response speed, which significantly reduces the overall computational load and resource consumption.

[0036] (2) This invention employs a two-stage detection mechanism of coarse judgment using machine learning and fine judgment using dual indicators. After quickly screening out suspected abnormal combustion, it uses the dual-indicator square Mahalanobis distance method to achieve refined identification. This method can effectively identify mild pre-ignition, strong pre-ignition, and normal combustion, with high identification accuracy, significantly better than the single threshold method.

[0037] (3) The convolutional autoencoder model used in this invention has a small number of parameters and can run in real time on an embedded hardware platform without the need for a high-performance computing unit. It is particularly suitable for hydrogen internal combustion engine vehicle control systems and bench real-time monitoring scenarios. This model only requires a large number of normal cylinder pressure signals and does not rely on the massive amount of pre-ignition fault data that is difficult to obtain. It can quickly adapt to different hydrogen internal combustion engine models or different operating conditions of the same internal combustion engine. It has good engineering applicability and ease of promotion, and solves the model generalization problem caused by insufficient fault samples in supervised learning methods. Attached Figure Description

[0038] Figure 1 This is a logic diagram of the rapid detection and identification method for pre-ignition in hydrogen internal combustion engines according to the present invention;

[0039] Figure 2 This is a flowchart of the method for rapid detection and identification of pre-ignition in hydrogen internal combustion engines according to the present invention. Detailed Implementation

[0040] This invention discloses a rapid detection and identification method for pre-ignition in hydrogen internal combustion engines, belonging to the field of hydrogen internal combustion engine combustion diagnosis technology. The method mainly includes: rapid coarse pre-ignition judgment based on the combustion state of the original cylinder pressure signal under test using a convolutional autoencoder (CAE); if the result of the rapid coarse pre-ignition judgment is suspected abnormal combustion, further fine judgment and classification of pre-ignition based on the squared Mahalanobis distance of two indices are performed. By constructing a two-level detection mechanism, machine learning performs rapid initial screening of the cylinder pressure signal, and then combined with a dual-indicator fusion and mutual judgment method, the computational cost is significantly reduced while ensuring detection accuracy.

[0041] To make the objectives and technical solutions of this invention clearer, the invention will be described in detail with reference to the accompanying drawings. However, the following embodiments are not intended to limit the invention in any way.

[0042] This embodiment provides a rapid detection method for pre-ignition in hydrogen internal combustion engines, the core logic diagram of which is as follows: Figure 1 As shown, the main steps include: acquiring the original cylinder pressure signal to be measured; after data preprocessing, inputting the original cylinder pressure sequence to be measured into a convolutional autoencoder CAE model for pre-ignition detection. If the reconstruction error is within the threshold range, it is determined to be normal combustion; otherwise, it is marked as suspected abnormal combustion. Next, the Mahalanobis distance method is used to accurately detect pre-ignition in suspected abnormal combustion, and the cycle is classified as normal combustion, strong pre-ignition, and weak pre-ignition according to the magnitude of the Mahalanobis distance.

[0043] like Figure 2 The flowchart shown below illustrates a rapid detection method for pre-ignition in hydrogen internal combustion engines, including the following steps:

[0044] Step 1: Select raw cylinder pressure signals that are manually labeled or confirmed as normal combustion cycles using a traditional combustion analyzer. Randomly divide the collected M normal cycle samples into a training set and a validation set. M is an integer greater than 500, with the training set comprising 80% and the validation set comprising 20%.

[0045] It should be specifically noted that the purpose of the training sample set is to allow the model to learn and fit the intrinsic structure and temporal dependence of the cylinder pressure signal during normal combustion, and to determine the model parameters; while the purpose of the validation set is to objectively evaluate the model's fitting effect and determine the model's threshold. Generally, the larger the number M of normal combustion cycle samples, the more fully the samples cover the normal combustion state, which is more conducive to improving the stability of model training and the reliability of determining the pre-ignition coarse threshold.

[0046] Step 2: Preprocess the original cylinder pressure signals in the training sample set, including filtering and denoising, normalization, and phase alignment. In this embodiment, Savitzky-Golay smoothing filtering is used; the crankshaft angle corresponding to the ignition moment is selected as the zero-phase reference point, and all collected cyclic data are time-domain shifted to significantly reduce the baseline value of reconstruction error.

[0047] Step 3: Using the preprocessed original cylinder pressure sequence as input, perform unsupervised training on the CAE model to minimize the reconstruction loss. Simultaneously, based on the mean reconstruction error of the validation set samples... and standard deviation Determine an appropriate judgment threshold.

[0048] The CAE model structure used in this embodiment is as follows: The encoder includes two one-dimensional convolutional layers and two max-pooling layers, with ReLU activation function. Since the cylinder pressure signal is a one-dimensional time-series data that varies with the crankshaft angle, the encoder uses one-dimensional convolutional layers; each convolutional layer is followed by a max-pooling layer; ReLU introduces a nonlinear mapping, enabling the model to learn and fit the complex nonlinear relationship between cylinder pressure and combustion heat release. The decoder includes two upsampling layers and two deconvolutional layers. The upsampling layers correspond to the max-pooling layers in the encoder and are mainly used to recover the spatial dimension of the feature map; the deconvolutional layers are used to reconstruct the signal. The loss function used is MSE.

[0049] After unsupervised training, the CAE model generates reconstruction errors for all normal combustion cycle samples. It should follow or approximately follow a normal distribution. Since premature combustion is an abnormal event, its reconstruction error MSE_test will deviate significantly from the mean of normal samples. Therefore, the Three Sigma principle is used to ensure that the reconstruction error of the vast majority (over 99.8%) of normal samples falls within the range of... Inside.

[0050] Step 4: After preprocessing the raw cylinder pressure signal of a single cycle acquired in real time, as in Step 2, the resulting test sequence is input into the CAE model trained in Step 4 to obtain the reconstructed cylinder pressure sequence. The reconstruction error MSE_test is calculated in real time, and a threshold is used to determine whether the current cycle is normal. If 0 ≤ MSE_test ≤ ... If the output result is normal combustion, otherwise it is judged as suspected abnormal combustion, and proceed to step 5;

[0051] Step 5: Calculate the maximum pressure rise rate and CA5 pressure rise rate of the training sample set and the validation set, construct the baseline distribution, and calculate the mean vector matrix. Covariance matrix .

[0052] CA5 is defined as the crankshaft angle corresponding to a combustion heat release rate of 5%, reflecting the pressure rise rate in the initial stage of combustion. If pre-ignition occurs, flame propagation is advanced, and the pressure rise rate at CA5 is often larger. On the other hand, weak pre-ignition occurs in the early stages of compression but has a smaller impact, while strong pre-ignition occurs near TDC and causes severe pressure rise and detonation. Therefore, the maximum pressure rise rate and CA5 pressure rise rate are used to distinguish between strong pre-ignition, weak pre-ignition, and normal combustion.

[0053] Step 6: Calculate the maximum pressure rise rate and CA5 pressure rise rate of the abnormal combustion cycle, construct a two-dimensional feature vector using dual indices, and calculate the squared Mahalanobis distance of the cycle. Statistical analysis is performed on the degree of deviation from the baseline distribution to achieve precise judgment and classification of pre-ignition. If... 5.99, this cycle is considered normal combustion; if 9.21, determine if the cycle is a strong pre-ignition; otherwise, determine if it is a weak pre-ignition.

[0054] Although the present invention has been described above in conjunction with the accompanying drawings, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are preferred application examples that demonstrate the core technical ideas of the present invention, and are merely illustrative and not restrictive. Those skilled in the art can make many improvements and changes under the guidance of the present invention without departing from the spirit of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for rapid detection and identification of pre-ignition in hydrogen internal combustion engines, characterized in that, Based on the combustion state of the original cylinder pressure signal to be measured, a rapid pre-ignition coarse judgment is made. If the result of the rapid pre-ignition coarse judgment is suspected abnormal combustion, the pre-ignition fine judgment and classification are further based on the squared Mahalanobis distance of the dual index. Rapid early combustion coarse judgment is performed based on the combustion state of the original cylinder pressure signal under test using a convolutional autoencoder (CAE). This includes: preprocessing the acquired original cylinder pressure signal under test, inputting the original cylinder pressure sequence under test into the convolutional autoencoder model to obtain the reconstructed cylinder pressure sequence under test, and calculating the reconstruction error between the original cylinder pressure sequence under test and the reconstructed cylinder pressure sequence under test in real time; if the reconstruction error is within the initial screening threshold range for anomaly judgment, it is directly judged as normal combustion, and the detection ends; otherwise, the current cycle is judged as a suspected abnormal cycle with suspected abnormal combustion, triggering early combustion fine judgment and classification based on the dual-index squared Mahalanobis distance. Pre-ignition precision judgment and classification based on dual-index squared Marshman distance, including classifying the suspected abnormal cycle as normal combustion, strong pre-ignition, or weak pre-ignition according to the magnitude of the squared Marshman distance.

2. The method for rapid detection and identification of pre-ignition in hydrogen internal combustion engines according to claim 1, characterized in that, The construction and training of the convolutional autoencoder model and the determination of the early fire coarse judgment threshold include: S1-1. Construct an unsupervised learning sample library: When the hydrogen internal combustion engine is working normally, collect the original cylinder pressure signals of no less than 500 normal combustion working cycles, and construct a training set and a validation set containing normal combustion samples. S1-2, Data standardization preprocessing: After filtering, normalizing, and phase aligning the original cylinder pressure signals in the training set, the original cylinder pressure sequence is obtained. The normalization process maps the original cylinder pressure signal to the [0, 1] interval, and the calculation expression is as follows: （1）； In equation (1), The original cylinder pressure value before normalization. This is the normalized cylinder pressure value. and These represent the minimum and maximum cylinder pressure values ​​in this working cycle, respectively. S1-3, Construction and training of convolutional autoencoder models: First, a convolutional autoencoder model is constructed, comprising an encoder, a decoder, and a loss function. The encoder extracts cylinder pressure temporal features characterizing the combustion state from the original cylinder pressure sequence and compresses these features into a latent feature vector. The decoder generates a reconstructed cylinder pressure sequence based on the latent feature vector. The loss function measures the difference between the original cylinder pressure sequence and the reconstructed cylinder pressure sequence. Then, the model is trained unsupervised using a normal sample training set, with the optimization objective being to minimize the mean squared error (MSE) between the original cylinder pressure sequence and the reconstructed cylinder pressure sequence; the formula for calculating the mean squared error is as follows: （2）； In equation (2), Let N be the reconstruction error of the i-th cycle, and N be the number of sampling points of the original cylinder pressure sequence in a single working cycle. This represents the original cylinder pressure sequence value at the j-th sampling point in the i-th work cycle. The reconstructed cylinder pressure sequence value at the j-th sampling point in the i-th working cycle; S1-4. Determine the coarse threshold for premature combustion based on the mean and standard deviation of the reconstruction error of the validation set samples; The expressions for calculating the mean and standard deviation of the reconstruction error are as follows: （3）； （4）； In equations (3) and (4), To reconstruct the mean error, M is the total number of samples in the validation set. To reconstruct the standard deviation of the error; based on the calculation results, the range of the coarse threshold for pre-ignition is determined as follows: .

3. The method for rapid detection and identification of pre-ignition in hydrogen internal combustion engines according to claim 2, characterized in that, In steps S1-2, the filtering and denoising process is used to remove high-frequency noise and retain signal characteristics; the phase alignment process refers to performing phase alignment on all sample cycles to ensure that the combustion characteristics between different cycles are strictly consistent in time; the normalization process is used to eliminate the difference in absolute pressure values ​​under different load conditions.

4. The method for rapid detection and identification of pre-ignition in hydrogen internal combustion engines according to claim 1, characterized in that, The calculation based on the dual-index Mahalanobis distance includes: S2-1. The maximum pressure rise rate and CA5 pressure rise rate of the statistical training set and validation set constitute a two-dimensional feature vector set. Calculate the mean vector of this feature vector set. Covariance matrix A statistical baseline distribution for normal combustion is established, which is composed as follows: （5）； （6）； In equations (5) and (6), Let be the two-dimensional feature vector of the i-th cycle. The maximum pressure rise rate in the i-th cycle. Let CA5 be the pressure rise rate in the i-th cycle. This represents the total number of samples in the training and validation sets. The expression for calculating the mean vector is as follows: （7）； The expression for calculating the covariance matrix is ​​as follows: （8）； S2-2. For suspected abnormal cycles, calculate the maximum pressure rise rate and CA5 pressure rise rate of the original cylinder pressure signal to be tested, and construct a two-dimensional feature vector of the sample to be tested; use the Mahalanobis distance algorithm to calculate the statistical distance between the two-dimensional feature vector and the reference distribution center; The expression for calculating Mahalanobis distance is as follows: （9）； In equation (9), This represents the two-dimensional feature vector of the sample to be tested. S2-3. Based on the magnitude of the squared Mahalanobis distance, classify the suspected abnormal cycle as normal combustion, strong pre-ignition, or weak pre-ignition: The two-dimensional feature vector composed of the maximum pressure rise rate and CA5 pressure rise rate approximately follows a two-dimensional normal distribution. The squared Mahalanobis distance of the sample under test... It approximately follows a chi-square distribution with 2 degrees of freedom. ,use Quantiles determine the judgment threshold; In this text, the subscript 2 represents the degree of freedom. It is the quantile probability, determined. The first threshold, The second threshold is determined using the squared Mahalanobis distance method as follows: If the square Mahalanobis distance If the value is less than or equal to the first threshold, it is determined to be normal combustion. If the square Mahalanobis distance If the value falls between the first and second thresholds, it is classified as weak pre-ignition. If the square Mahalanobis distance If the value is greater than or equal to the second threshold, it is determined to be a strong pre-ignition.

5. The method for rapid detection and identification of pre-ignition in hydrogen internal combustion engines according to claim 4, characterized in that, In steps S2-3, when the squared Mahalanobis distance... When the time is right, the cycle is determined to be normal combustion; when the squared Mahalanobis distance is right... If the cycle is determined to be a strong pre-ignition, it is determined to be a weak pre-ignition.

Images