In-cylinder flame multi-characteristic parameter extraction and instantaneous evaluation method based on machine learning

Abstract

The invention discloses an in-cylinder flame multi-characteristic parameter extraction and instantaneous evaluation method based on machine learning, and belongs to the technical field of power energy. Aiming at the problem that a traditional method is difficult to accurately extract flame transient characteristic parameters in a dynamic combustion environment, the method comprises the following steps: acquiring an in-cylinder flame image sequence through an endoscopic high-speed camera, and extracting an event driving key frame and a uniform time key frame by adopting optical flow analysis and fixed interval sampling; utilizing a double-flow feature extraction network to respectively obtain flame spatial form features and radiation features; capturing a flame parameter time evolution law through a time sequence convolutional network, and dynamically fusing key frame features in combination with a self-attention mechanism; and finally outputting flame temperature field and soot distribution prediction results. According to the method, instantaneous evaluation of the combustion process is realized, and a basis is provided for engine performance optimization and combustion fault diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of power energy technology, and in particular relates to a method for extracting and instantaneously evaluating multiple feature parameters of in-cylinder flames based on machine learning. Background Technology

[0002] In diesel engine combustion, key characteristic parameters such as flame temperature and soot formation directly affect combustion efficiency and emission levels. Traditional flame parameter monitoring mainly relies on physical models and sensor-based real-time monitoring methods, which have significant limitations in dynamic and complex combustion environments. Physical models are typically based on numerous assumptions, making it difficult to accurately describe the transient characteristics of the actual combustion process, especially under high-frequency transient changes in each combustion cycle, where model prediction accuracy drops significantly. While sensor-based measurement techniques can acquire some combustion parameters, they are limited by sensor response speed and installation location, failing to comprehensively reflect all information about in-cylinder flame development. Furthermore, sensors are easily damaged under high-temperature and high-pressure environments, resulting in high maintenance costs.

[0003] In practical applications, traditional methods struggle to simultaneously meet the real-time and accuracy requirements of flame parameter monitoring. The rapidly changing combustion process necessitates sub-millisecond monitoring frequencies, while traditional physical models suffer from high computational complexity, hindering real-time prediction. While sensor measurements provide real-time data, the limited number of measurement points fails to comprehensively reflect the in-cylinder flame temperature distribution and soot formation. Particularly during critical combustion stages such as the initial flame propagation phase, peak flame stage, and the main soot formation period, traditional methods often fail to accurately capture the rapid changes in flame characteristics. These issues severely restrict the real-time optimization and fault diagnosis capabilities of engine combustion processes.

[0004] The root cause of these technical shortcomings lies in the complexity and dynamic nature of the combustion process. Flame development involves multi-physics coupling processes such as turbulence, chemical reactions, and heat and mass transfer, and its parameter changes are highly nonlinear and uncertain. Traditional methods either oversimplify the physical processes or have limited measurement means, making it difficult to fully characterize this complexity. Furthermore, while new measurement methods such as high-speed cameras can acquire rich flame image information, the real-time processing and analysis of massive amounts of data faces enormous challenges, lacking effective feature extraction and modeling methods. These difficulties render existing technologies significantly inadequate for monitoring and evaluating engine combustion processes, necessitating the development of new technical solutions to address these issues. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a machine learning-based method for extracting and instantaneously evaluating multiple feature parameters of in-cylinder flames, thereby resolving the issues present in the prior art.

[0006] In a first aspect, to achieve the above objectives, the present invention provides a machine learning-based method for extracting and instantaneously evaluating multiple feature parameters of in-cylinder flames, comprising the following steps:

[0007] Acquire a sequence of flame images inside the engine cylinder using an endoscopic high-speed camera system;

[0008] Extract event-driven keyframes and uniform time keyframes from the flame image sequence;

[0009] A dual-stream feature extraction network was used to extract the spatial morphological features and radiation features of the flame, respectively.

[0010] The spatial morphological features and radiation features are input into a temporal modeling network to capture the temporal evolution of flame parameters;

[0011] Features of different types of keyframes are dynamically fused using a self-attention mechanism;

[0012] Output the predicted results of flame temperature field and soot distribution.

[0013] Optionally, the process of extracting event-driven keyframes includes:

[0014] Detecting moments of drastic changes in flames based on optical flow analysis algorithms;

[0015] When the optical flow velocity exceeds a set threshold, it is determined to be a critical event frame.

[0016] Optionally, the dual-stream feature extraction network includes:

[0017] Spatial feature flow: a CNN network is used to extract flame area, symmetry, and brightness gradient distribution features.

[0018] Radiation feature flow: Extract the light intensity ratio from the dual-channel image to generate a radiation feature vector.

[0019] Optionally, the temporal modeling network employs a temporal convolutional network, the structure of which includes:

[0020] Multi-layer dilated causal convolution stacking;

[0021] Each layer's output is superimposed on the original input;

[0022] Insert Dropout and BatchNorm between layers.

[0023] Optionally, the self-attention mechanism is a multi-head self-attention mechanism, and its working process includes:

[0024] The temporal features are decomposed into Q, K, and V matrices;

[0025] Calculate the scaling dot product attention per head;

[0026] Dynamically assign weights to different types of keyframes.

[0027] Optional, also includes:

[0028] A multi-task loss function is used to jointly optimize the prediction of temperature field and soot distribution;

[0029] The phased training strategy involves first freezing the parameters of the image feature extraction network to train the temporal model, and then unfreezing all parameters for end-to-end fine-tuning.

[0030] Secondly, the present invention also provides a machine learning-based system for extracting and instantaneously evaluating multiple feature parameters of in-cylinder flames, used to implement a machine learning-based method for extracting and instantaneously evaluating multiple feature parameters of in-cylinder flames, the system comprising:

[0031] The image acquisition module is used to acquire a sequence of images of flames inside the engine cylinder via an endoscopic high-speed camera system;

[0032] A keyframe extraction module is used to extract event-driven keyframes and uniform time keyframes from the flame image sequence.

[0033] The feature extraction module is used to extract the spatial morphological features and radiation features of the flame using a dual-stream feature extraction network.

[0034] The temporal modeling module is used to input the spatial morphological features and radiation features into the temporal modeling network to capture the temporal evolution of flame parameters;

[0035] The feature fusion module is used to dynamically fuse features of different types of keyframes through a self-attention mechanism;

[0036] The results output module is used to output the predicted results of the flame temperature field and soot distribution.

[0037] Thirdly, the present invention also provides a computer terminal device, comprising:

[0038] One or more processors;

[0039] A memory, coupled to the processor, for storing one or more programs;

[0040] When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the machine learning-based in-cylinder flame multi-feature parameter extraction and instantaneous evaluation method in the first aspect described above.

[0041] Fourthly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the steps of the machine learning-based in-cylinder flame multi-feature parameter extraction and instantaneous evaluation method described in the first aspect above.

[0042] Fifthly, the present invention also provides a computer program product, including a computer program, characterized in that, when the computer program is executed by a processor, it implements the steps of the machine learning-based in-cylinder flame multi-feature parameter extraction and instantaneous evaluation method described in the first aspect above.

[0043] Compared with the prior art, the present invention has the following advantages and technical effects:

[0044] This invention provides a machine learning-based method for extracting and instantaneously evaluating multiple feature parameters of in-cylinder flames. By combining CNN spatial feature extraction with TCN temporal modeling, this invention achieves accurate prediction of the in-cylinder flame temperature field and soot distribution. A hybrid keyframe extraction algorithm effectively captures transient features and gradual changes in the flame, improving the model's generalization ability. A self-attention mechanism dynamically fuses different types of keyframe features, optimizing the understanding of the combustion process. Lightweight network design and parallel computing architecture ensure real-time performance. Transfer learning supports combustion state recognition under different operating conditions, expanding the application scope. A dual-stream feature network and a multi-task loss function collaboratively optimize the accuracy of temperature and soot prediction. A staged training strategy reduces model variance and enhances stability. Attached Figure Description

[0045] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0046] Figure 1 This is a schematic diagram of the experimental apparatus according to an embodiment of the present invention;

[0047] Figure 2 Images of flame temperature and KL factor from an embodiment of the present invention;

[0048] Figure 3 This is a flowchart of a method according to an embodiment of the present invention. Detailed Implementation

[0049] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0050] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0051] This invention provides a machine learning-based algorithm for extracting and evaluating multiple feature parameters of in-cylinder flames in real time. It can be measured using existing mature engine endoscopic visualization systems, overcoming the problems of slow identification speed and difficulty in accurately extracting transient flame change features in dynamic and complex combustion environments by traditional flame recognition methods. It enables the prediction of temperature field and soot distribution throughout the combustion cycle and achieves real-time evaluation of the combustion process.

[0052] Example 1

[0053] like Figure 1 As shown, this embodiment provides a machine learning-based method for extracting and instantaneously evaluating multiple feature parameters of in-cylinder flames, including:

[0054] Acquire a sequence of flame images inside the engine cylinder using an endoscopic high-speed camera system;

[0055] Extract event-driven keyframes and uniform time keyframes from the flame image sequence;

[0056] A dual-stream feature extraction network was used to extract the spatial morphological features and radiation features of the flame, respectively.

[0057] The spatial morphological features and radiation features are input into a temporal modeling network to capture the temporal evolution of flame parameters;

[0058] Features of different types of keyframes are dynamically fused using a self-attention mechanism;

[0059] Output the predicted results of flame temperature field and soot distribution.

[0060] Experimental setup as follows Figure 1 As shown, an endoscopic high-speed camera system is installed through an angled opening in the engine cover to obtain real-time images of the flames inside the engine cylinders. Figure 1 The experimental setup of the present invention is provided, which consists of three parts: a fuel supply system, including a gas rail, a rail pressure control device, etc.; an injector and an inlet pressure sensor; and a signal acquisition module, in which the inlet pressure signal is amplified by an amplifier and sent to an acquisition card to obtain the inlet pressure signal.

[0061] Flame data calibration:

[0062] like Figure 2 As shown, flame feature recognition mainly focuses on key parameters such as flame morphology, temperature, and soot coefficient. Based on the original two-color method algorithm, the surface temperature of channel G, the surface temperature of channel B, and the flame temperature are compared using comparative data. A machine learning model is then trained. The two-color method is used to solve for the temperature corresponding to each pixel based on the following formula:

[0063]

[0064] In the formula, T a1 T a2 C1 and C2 are the brightness temperatures corresponding to the radiation at wavelengths λ1 and λ2, respectively, obtained beforehand using a blackbody furnace; the actual temperature T is the unknown quantity to be determined. C1 and C2 are Planck's first and second constants, respectively, with values ​​of 3.7419 × 10⁻⁶. -16 W / m 2 and 1.4388×10 -2 m·K. The physical quantity KL characterizing the soot concentration in a flame can be obtained from the following formula:

[0065]

[0066] Where λ is the wavelength, and the constant a is generally taken as 1.38 in the visible light range, T a This refers to the brightness temperature.

[0067] As one implementation method in this embodiment, the process of extracting event-driven keyframes includes:

[0068] Detecting moments of drastic changes in flames based on optical flow analysis algorithms;

[0069] When the optical flow velocity exceeds a set threshold, it is determined to be a critical event frame.

[0070] Specifically, the flame keyframe extraction process includes:

[0071] To capture the transient characteristics of flame temperature and soot generation during combustion, keyframes need to be extracted from a continuous sequence of images captured by a high-speed camera. Keyframes are divided into two categories:

[0072] ①Event-driven keyframes:

[0073] The algorithm detects moments of dramatic flame changes (such as ignition, peak, and extinction) based on optical flow analysis. The Horn-Schunck optical flow equation calculates the motion vector field by minimizing the brightness gradient error between adjacent frames.

[0074]

[0075] When the optical flow velocity v > 5 pixels / frame, it is determined to be a critical event frame number, where I t Let I be the partial derivative of brightness over time, and let I be the brightness.

[0076] ② Uniform time keyframes:

[0077] Sampling is performed at fixed intervals (1ms) according to the length of the combustion cycle (e.g., 10ms) to ensure coverage of the gradual process of flame development.

[0078] As one implementation method in this embodiment, the dual-stream feature extraction network includes:

[0079] Spatial feature flow: a CNN network is used to extract flame area, symmetry, and brightness gradient distribution features.

[0080] Radiation feature flow: Extract the light intensity ratio from the dual-channel image to generate a radiation feature vector.

[0081] Specifically, the two-stream feature extraction network includes:

[0082] ① Spatial Feature Flow (CNN):

[0083] Flame morphology features are extracted using EfficientNet-B3, and its composite scaling formula is as follows:

[0084] Depth: d=1.2, Width: ω=1.1, Resolution: r=1.15 (4)

[0085] The output is a feature vector representing the flame area, symmetry, and brightness gradient distribution. Here, Depth represents the network depth factor, controlling the degree of increase in the number of layers; Width represents the network width factor, corresponding to the number of channels per layer; and Resolution represents the input image resolution factor, improving feature extraction quality by enlarging the input image size.

[0086] ②Radiative characteristic flow:

[0087] The intensity ratio is directly extracted from the dual-channel image and input into the fully connected layer to generate a 64-dimensional radiation feature vector.

[0088] As one implementation method in this embodiment, the temporal modeling network adopts a temporal convolutional network, the structure of which includes:

[0089] Multi-layer dilated causal convolution stacking;

[0090] Each layer's output is superimposed on the original input;

[0091] Insert Dropout and BatchNorm between layers.

[0092] As one implementation method in this embodiment, the self-attention mechanism is a multi-head self-attention mechanism, and its working process includes:

[0093] The temporal features are decomposed into Q, K, and V matrices;

[0094] Calculate the scaling dot product attention per head;

[0095] Dynamically assign weights to different types of keyframes.

[0096] Specifically, temporal modeling and attention mechanisms include:

[0097] This invention processes the temporal dependencies between keyframes using a time series model, constructs a temporal prediction model based on a temporal convolutional network (TCN), captures the temporal dependencies between keyframes, and achieves anomaly detection and abnormal development prediction. It can use a self-attention mechanism to assign weights to different keyframes, making full use of the information provided by these two types of keyframes to improve the understanding and prediction accuracy of the combustion process.

[0098] ① Temporal Convolutional Network (TCN) Architecture Design:

[0099] To capture the evolution of flame parameters over time, this invention employs a stack of four layers of dilated causal convolutions. Each layer has the following structure: the dilation factors are d = {1, 2, 4, 8}, ensuring the receptive field fully covers the entire combustion cycle; each layer contains 64 convolutional kernels (kernel_size = 3), and the activation function is GELU (Gaussian error linear unit), as shown in the following formula, where x is the input feature tensor of the TCN module, and Φ(x) is the standard normal cumulative function.

[0100] GELU(x)=x·Φ(x) (5)

[0101] The original input is superimposed on the output of each layer to avoid gradient vanishing; Dropout (probability 0.2) and BatchNorm are inserted between layers to prevent overfitting.

[0102] Output = TCN_Block(x) + x (6)

[0103] Where TCN_Block(x) is the feature obtained after applying a set of temporal convolution modules to x, and x is the input feature tensor of the TCN module.

[0104] ② Multi-head self-attention mechanism:

[0105] To dynamically fuse event-driven frames and uniform time frames, a 4-head self-attention mechanism is designed. First, the temporal characteristics H output by the TCN are decomposed into Q, K, and V matrices, with each head having a dimension d. k =d v =16;

[0106]

[0107] Where Q is the query matrix, K is the key matrix, V is the value matrix, and W is the value matrix. Q W K W V All of them are trainable parameter matrices.

[0108] Calculate the scaled dot product attention by head:

[0109]

[0110] Among them, Q i K i V i This is the submatrix corresponding to the i-th attention head, used for independently calculating self-attention for each head. For K i The order of the order.

[0111] By setting four attention heads, the weight ratio of event-driven keyframes and uniform time keyframes can be reasonably allocated, enhancing the model's applicability to different combustion modes.

[0112] As one implementation method in this embodiment, it also includes:

[0113] A multi-task loss function is used to jointly optimize the prediction of temperature field and soot distribution;

[0114] The phased training strategy involves first freezing the parameters of the image feature extraction network to train the temporal model, and then unfreezing all parameters for end-to-end fine-tuning.

[0115] Specifically, the model training and evaluation process includes:

[0116] ① Multi-task loss function design:

[0117] To simultaneously optimize the prediction accuracy of temperature field and soot distribution, a joint loss function is designed, taking into account both global trends and local details:

[0118] Temperature field prediction: Given the strong correlation between the high-temperature region of the flame core and the high-smoke concentration region, higher weights are applied to the high-smoke region to suppress background interference. The model's sensitivity to key combustion regions is enhanced through weighted mean square error (MSE).

[0119] Carbon soot distribution prediction: Dice loss is used to solve the class imbalance problem caused by sparse carbon soot distribution, and an edge-sensitive term is introduced to constrain the smoothness of the prediction boundary to avoid abnormal abrupt changes in carbon soot diffusion patterns.

[0120] Joint optimization: Dynamically balance the contribution of temperature and carbon soot task to the loss, and determine the task weight ratio through experimental verification to ensure the joint optimization of the two objectives.

[0121] ② Phased training strategy:

[0122] Phase 1: Temporal-spatial feature decoupling pre-training:

[0123] The parameters of the image feature extraction network are frozen, and the temporal model and attention module are trained first. This stage focuses on establishing the temporal evolution patterns between key frames to avoid spatial feature noise interfering with the temporal dependency modeling.

[0124] Phase Two: End-to-End Joint Fine-Tuning

[0125] All network parameters are unfrozen, and an adaptive learning rate strategy is used for overall optimization. The learning rate is dynamically adjusted using the cosine annealing algorithm, combined with gradient clipping techniques to stabilize the training process. An early stopping mechanism is introduced to prevent overfitting and retain the model with the best performance on the validation set.

[0126] Automated parameter tuning and verification:

[0127] A Bayesian optimization framework is employed to search for key hyperparameters (such as learning rate, regularization strength, and task weight ratio) to balance training efficiency and model performance. Cross-validation is used to evaluate the model's generalization ability, with a focus on monitoring the volatility of temperature field errors and the consistency of soot distribution. A phased training strategy significantly reduces model variance and improves stability under different operating conditions.

[0128] The algorithm flowchart of this invention is as follows: Figure 3 As shown, achieving instantaneous evaluation of in-cylinder flame is one of the core features of this invention. By rationally selecting and matching key frames, it is possible to achieve accurate and rapid identification of in-cylinder flame status and prediction of flame characteristic parameters, providing a basis for performance optimization and realizing online combustion fault diagnosis of the whole engine.

[0129] Step 1: Based on the flame feature values ​​derived from the convolutional neural network, data support is provided for the subsequent algorithm model establishment, and it serves as the basis for judging the model's recognition performance.

[0130] Step 2: Based on the two keyframe iteration methods, perform model training and parameter optimization under different keyframe selection and combination conditions;

[0131] Step 3: Based on the parameters obtained in the previous step, distinguish between model verification and verification of recognition and prediction effects under actual machine operation.

[0132] Step 4: Repeat steps 1 to 3 to achieve flame anomaly detection and flame characteristic parameter prediction.

[0133] Based on this, the present invention provides a machine learning-based method for extracting and instantaneously evaluating multiple feature parameters of in-cylinder flames, which has the following advantages:

[0134] (1) This invention extracts spatial features such as flame shape and texture through CNN, and captures temporal dynamics by combining LSTM / TCN, so as to achieve sub-millisecond-level prediction of temperature field and carbon smoke distribution, with accuracy exceeding that of traditional physical models.

[0135] (2) The hybrid keyframe extraction algorithm based on optical flow analysis and brightness change can capture event features such as flame peaks and retain the gradual details of time-uniform frames, thereby improving the model's generalization ability.

[0136] (3) The present invention uses a self-attention mechanism to dynamically weight two types of key frame features (event frames and time uniform frames), effectively integrating macro trends and micro changes, which can significantly reduce the recognition error rate.

[0137] (4) The keyframe extraction algorithm compresses the data volume to 5%-10% of the original sequence. Combined with the lightweight CNN and TCN parallel computing architecture, it can achieve a faster inference speed. By fine-tuning the pre-trained model through transfer learning, it supports the transfer of combustion state under different loads and injection strategies, thus covering a wider range of operating conditions.

[0138] Example 2

[0139] In this embodiment, a computer terminal device is provided, including:

[0140] One or more processors;

[0141] A memory, coupled to the processor, for storing one or more programs;

[0142] When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the above-described machine learning-based in-cylinder flame multi-feature parameter extraction and instantaneous evaluation method.

[0143] In this embodiment, a computer-readable storage medium is also provided, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the above-described machine learning-based method for extracting and instantaneously evaluating multiple feature parameters of in-cylinder flames.

[0144] In this embodiment, an electronic device is also provided, including a memory and a processor. The memory stores a computer program, and the processor is configured to run the computer program to perform the steps of the above-described machine learning-based in-cylinder flame multi-feature parameter extraction and instantaneous evaluation method.

[0145] In this embodiment, a computer program product is also provided, including a computer program that, when executed by a processor, implements the steps of the above-described machine learning-based in-cylinder flame multi-feature parameter extraction and instantaneous evaluation method.

[0146] The aforementioned program can run on a processor or be stored in memory (or a computer-readable medium). Computer-readable media include both permanent and non-permanent, removable and non-removable media, and information storage can be achieved by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0147] These computer programs may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes can be implemented using different modules, and different steps can be implemented using different modules.

[0148] This embodiment provides such an apparatus or system. The system, referred to as a machine learning-based in-cylinder flame multi-feature parameter extraction and instantaneous evaluation system, includes:

[0149] The image acquisition module is used to acquire a sequence of images of flames inside the engine cylinder via an endoscopic high-speed camera system;

[0150] A keyframe extraction module is used to extract event-driven keyframes and uniform time keyframes from the flame image sequence.

[0151] The feature extraction module is used to extract the spatial morphological features and radiation features of the flame using a dual-stream feature extraction network.

[0152] The temporal modeling module is used to input the spatial morphological features and radiation features into the temporal modeling network to capture the temporal evolution of flame parameters;

[0153] The feature fusion module is used to dynamically fuse features of different types of keyframes through a self-attention mechanism;

[0154] The results output module is used to output the predicted results of the flame temperature field and soot distribution.

[0155] As one implementation method in this embodiment, the keyframe extraction module includes:

[0156] The optical flow analysis unit is used to detect moments of drastic changes in the flame.

[0157] The event determination unit is used to determine a critical event frame when the optical flow velocity exceeds a set threshold.

[0158] The time sampling unit is used to extract uniform time keyframes at fixed time intervals.

[0159] As one implementation method in this embodiment, the feature extraction module includes:

[0160] The spatial feature extraction unit uses a CNN network to extract flame area, symmetry, and brightness gradient distribution features.

[0161] The radiation feature extraction unit is used to extract the light intensity ratio from the dual-channel image to generate a radiation feature vector.

[0162] As one implementation method in this embodiment, the time series modeling module includes:

[0163] Dilated convolutional units are used to implement multi-layer dilated causal convolution stacks;

[0164] Residual connection units are used to superimpose the output of each layer onto the original input;

[0165] Regularization units are used to implement Dropout and BatchNorm operations between layers.

[0166] As one implementation method in this embodiment, the feature fusion module includes:

[0167] The matrix splitting unit is used to split the time series features into Q, K, and V matrices;

[0168] Attention computation unit, used to compute scaled dot product attention head-to-head;

[0169] The weight allocation unit is used to dynamically allocate weights to different types of keyframes.

[0170] As one implementation method in this embodiment, it also includes:

[0171] The loss calculation module is used to jointly optimize the prediction of temperature field and soot distribution using a multi-task loss function;

[0172] The training control module is used to implement a phased training strategy. First, the parameters of the image feature extraction network are frozen to train the time-series model, and then all parameters are unfrozen for end-to-end fine-tuning.

[0173] The system or apparatus is used to implement the functions of the methods in the above embodiments. Each module in the system or apparatus corresponds to each step in the method, as has been described in the method and will not be repeated here.

[0174] The above implementation method solves the problem of extracting and instantaneously evaluating multiple feature parameters of in-cylinder flames based on machine learning in related technologies, thereby ensuring that the problems existing in the prior art are resolved.

[0175] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for extracting and instantaneously evaluating multiple feature parameters of in-cylinder flames based on machine learning, characterized in that, Includes the following steps: Acquire a sequence of flame images inside the engine cylinder using an endoscopic high-speed camera system; Extract event-driven keyframes and uniform time keyframes from the flame image sequence; A dual-stream feature extraction network was used to extract the spatial morphological features and radiation features of the flame, respectively. The spatial morphological features and radiation features are input into a temporal modeling network to capture the temporal evolution of flame parameters; Features of different types of keyframes are dynamically fused using a self-attention mechanism; Output the predicted results of flame temperature field and soot distribution.

2. The method according to claim 1, characterized in that, The process of extracting event-driven keyframes includes: Detecting moments of drastic changes in flames based on optical flow analysis algorithms; When the optical flow velocity exceeds a set threshold, it is determined to be a critical event frame.

3. The method according to claim 1, characterized in that, The dual-stream feature extraction network includes: Spatial feature flow: a CNN network is used to extract flame area, symmetry, and brightness gradient distribution features. Radiation feature flow: Extract the light intensity ratio from the dual-channel image to generate a radiation feature vector.

4. The method according to claim 1, characterized in that, The temporal modeling network employs a temporal convolutional network, and its structure includes: Multi-layer dilated causal convolution stacking; Each layer's output is superimposed on the original input; Insert Dropout and BatchNorm between layers.

5. The method according to claim 1, characterized in that, The self-attention mechanism is a multi-head self-attention mechanism, and its working process includes: The temporal features are decomposed into Q, K, and V matrices; Calculate the scaling dot product attention per head; Dynamically assign weights to different types of keyframes.

6. The method according to claim 1, characterized in that, Also includes: A multi-task loss function is used to jointly optimize the prediction of temperature field and soot distribution; The phased training strategy involves first freezing the parameters of the image feature extraction network to train the temporal model, and then unfreezing all parameters for end-to-end fine-tuning.

7. A machine learning-based system for extracting and instantaneously evaluating multiple feature parameters of in-cylinder flames, characterized in that, The system includes: The image acquisition module is used to acquire a sequence of images of flames inside the engine cylinder via an endoscopic high-speed camera system; A keyframe extraction module is used to extract event-driven keyframes and uniform time keyframes from the flame image sequence. The feature extraction module is used to extract the spatial morphological features and radiation features of the flame using a dual-stream feature extraction network. The temporal modeling module is used to input the spatial morphological features and radiation features into the temporal modeling network to capture the temporal evolution of flame parameters; The feature fusion module is used to dynamically fuse features of different types of keyframes through a self-attention mechanism; The results output module is used to output the predicted results of the flame temperature field and soot distribution.

8. A computer terminal device, characterized in that, include: One or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors perform the steps of the method as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-6.

Images