Detonation Initiation Prediction Method and System Based on Two-Stream Neural Network Model

Abstract

This invention provides a detonation initiation prediction method and system based on a two-stream neural network model, relating to the fields of industrial artificial intelligence and combustion process monitoring and prediction. The method includes: acquiring and preprocessing the input feature variables of the target detonation process; using a two-stream neural network model to predict the preprocessed time-series data block to obtain predicted values ​​of key detonation initiation parameters of the target detonation process; the two-stream neural network model includes: a two-stream feature extraction module for parallel extraction of global spatial features and temporal dynamic features; a multilayer perceptron module for predicting the two types of features separately, obtaining predicted values ​​corresponding to the two branches, and using the predicted value corresponding to the global feature extraction branch as the final output; while the predicted value corresponding to the temporal feature extraction branch is only used to assist in training the global feature extraction branch. This invention overcomes the deficiency of existing detonation initiation prediction models in simultaneously and effectively capturing the global spatial correlation and temporal dynamic evolution patterns of data.

Description

Technical Field

[0001] This invention relates to the fields of industrial artificial intelligence, combustion process monitoring and prediction technology, and in particular to a detonation initiation prediction method and system based on a dual-flow neural network model. Background Technology

[0002] The deflagration-to-detonation transition (DDT) is a key physical phenomenon in combustion and explosion processes, with significant research value in propulsion systems, detonation engines, and safety protection. Accurately predicting the critical conditions and dynamic evolution of detonation initiation is crucial for optimizing combustion design, preventing accidents, and improving energy release efficiency. However, the detonation initiation process involves complex chemical reaction kinetics, turbulence, and shock wave interactions, exhibiting strong nonlinearity, multi-scale characteristics, and high noise. Traditional simulation models based on physical mechanisms are computationally expensive and have poor adaptability, making them unsuitable for real-time prediction and engineering applications.

[0003] Existing methods for predicting detonation processes can be mainly classified into the following categories:

[0004] (1) Simulation model based on physical mechanism: The detonation process is simulated by solving the equations of combustion chemical reaction kinetics and fluid dynamics. This type of method has clear physical meaning, but the modeling is complex, it depends on accurate physical property parameters, the calculation is expensive, and it is difficult to apply in real time.

[0005] (2) Regression methods based on shallow machine learning, including linear regression, support vector regression, decision trees and their ensemble methods. These methods are computationally efficient, but have limited ability to capture multivariate, strongly nonlinear, and time-series dynamics, and their prediction accuracy is prone to bottlenecks.

[0006] (3) Deep learning-based methods include: Multilayer perceptrons: capable of handling nonlinear relationships, but with limited ability to model high-dimensional features and temporal dependencies; Recurrent neural networks: such as Long Short-Term Memory (LSTM) networks, designed specifically for sequence data, but with problems such as training not being parallelizable and weak ability to capture long-range dependencies; Transformer models: self-attention mechanisms can compute and model global dependencies in parallel, but for industrial time-series data with strong local correlations and noise, there is still room for improvement in the ability to extract local details and temporal dynamics, and a single model is difficult to simultaneously take into account deep patterns in both spatial and temporal dimensions.

[0007] Therefore, existing technical solutions generally suffer from one or more of the following problems: difficulty in simultaneously and effectively extracting global spatial features and temporal dynamic features of data; sensitivity to noise and outliers in the data; need to improve the generalization ability of the model; and failure to design differentiated optimization strategies for the extraction process of different features. Summary of the Invention

[0008] To address the shortcomings of the existing technologies, this invention proposes a detonation initiation prediction method and system based on a dual-flow neural network architecture. By fusing global spatial features and time series features, it aims to achieve higher accuracy and stronger robustness in detonation initiation prediction by collaboratively mining global and temporal information from the data through two parallel feature extraction paths. This overcomes the shortcomings of existing detonation initiation prediction models that cannot effectively capture both global spatial correlation and temporal dynamic evolution patterns of the data simultaneously, thereby enabling high-precision, real-time prediction of key detonation initiation parameters.

[0009] On the one hand, this invention proposes a detonation initiation prediction method based on a dual-stream neural network model, which includes the following process:

[0010] The input characteristic variables of the target detonation process are acquired in real time, a time series of the input characteristic variables is formed and preprocessed to obtain a preprocessed time series data block;

[0011] A two-stream neural network model is used to predict the preprocessed time series data blocks to obtain the predicted values ​​of the key parameters for detonation initiation in the target detonation process.

[0012] The dual-stream neural network model includes:

[0013] The dual-stream feature extraction module includes a global feature extraction branch and a temporal feature extraction branch, which are used to perform global feature extraction and temporal feature extraction on the preprocessed time-series data blocks in parallel to obtain global spatial features and temporal dynamic features;

[0014] The multilayer perceptron module is used to perform nonlinear mapping on the global spatial features and temporal dynamic features respectively to obtain the predicted values ​​corresponding to the global feature extraction branch and the temporal feature extraction branch. The predicted values ​​corresponding to the global feature extraction branch are used as the predicted values ​​of the key parameters for detonation initiation. The predicted values ​​corresponding to the temporal feature extraction branch are only used during the training phase to assist in training the global feature extraction branch.

[0015] Furthermore, the input characteristic variables include: thermodynamic parameters, chemical kinetic parameters, and flow parameters; wherein the thermodynamic parameters include: combustion zone pressure and combustion zone temperature; the chemical kinetic parameters include: heat release rate and key component concentration; and the flow parameters include: flow velocity and turbulence intensity.

[0016] Furthermore, the method for preprocessing the time series of the input feature variables is as follows:

[0017] For any input feature variable, perform Hampel filtering on the time series of that input feature variable to obtain the filtered time series; the specific process is as follows:

[0018] For any data point in the time series Define a Centered on, window size is A sliding window, representing the data within the sliding window as And set a threshold constant. ;

[0019] Calculate the median of the data within the sliding window. ;

[0020] Calculate the sum of each data point and the median within the sliding window. The absolute median deviation is obtained by calculating the median of the absolute value sequence of the differences between the two values. ;

[0021] The absolute median deviation Convert to standard deviation estimate ;

[0022] Based on the standard deviation estimation According to the set threshold constant Determine data points Are there any anomalies? If so, change the data points. Replace with median If no anomalies are found, the data points are retained. This leads to the filtered time series.

[0023] Batch normalization is performed on the filtered time series to obtain the batch normalized time series, and then the batch normalized time series of all input feature variables are obtained.

[0024] Use a length of A window is used to slide and truncate the time series of all input feature variables after batch normalization, with each window trunculating a fixed time step. The data points of all input feature variables within each consecutive time step form a group of size . The feature variable matrix; where The number of input feature variables;

[0025] All feature variable matrices are used as preprocessed time series data blocks.

[0026] Furthermore, the global feature extraction branch is used to extract spatial features from the input data step by step to obtain global spatial features; wherein, the multivariate feature vector composed of data points of all input feature variables at the same time step in the preprocessed time series data block is used as the input data of the global feature extraction branch.

[0027] The time feature extraction branch is used to extract time features from the input data to obtain time dynamic features; wherein, each feature variable matrix in the preprocessed time series data block is used as the input data of the time feature extraction branch.

[0028] The multilayer perceptron module includes two identically structured multilayer perceptrons: a first multilayer perceptron and a second multilayer perceptron; wherein the first multilayer perceptron is used to perform nonlinear mapping on the global spatial features and output the predicted value corresponding to the global feature extraction branch; the second multilayer perceptron is used to perform nonlinear mapping on the temporal dynamic features and output the predicted value corresponding to the temporal feature extraction branch.

[0029] Furthermore, the global feature extraction branch includes: a linear embedding layer and an improved Transformer encoding module consisting of several stacked layers;

[0030] The linear embedding layer is used to perform linear projection on the multivariate feature vector corresponding to each time step to obtain an initial embedding vector, and the initial embedding vector is used as the input vector of the first layer improved Transformer coding module.

[0031] The improved Transformer encoding module is as follows:

[0032] The input vector of the current improved Transformer coding module is layer normalized to obtain the embedding vector, which is then input into the multi-head self-attention layer to obtain the output vector of the multi-head self-attention layer.

[0033] The output vector of the multi-head self-attention layer is residually concatenated with the embedding vector to obtain the attention feature vector;

[0034] The attention feature vector is input into the causal convolutional enhancement unit, and the output vector of the causal convolutional enhancement unit is normalized by the layer. The output vector of the causal convolutional enhancement unit after the layer normalization is input into the feedforward neural network to obtain the output vector of the feedforward neural network.

[0035] The output vector of the feedforward neural network is added to the output vector of the causal convolutional enhancement unit to obtain the output vector of the current improved Transformer coding module, which is then used as the input vector of the next layer of improved Transformer coding module.

[0036] The output vector of the last improved Transformer coding module is used as the global spatial feature. .

[0037] Furthermore, the causal convolution enhancement unit performs two parallel processing steps on the attention feature vector. One step involves performing a one-dimensional causal convolution operation on the attention feature vector and then passing it through a Tanh activation function to obtain a first output vector. The other step involves performing a one-dimensional causal convolution operation on the attention feature vector and then passing it through a Sigmoid activation function to obtain a second output vector. The first and second output vectors are then multiplied element-wise to obtain a gated enhanced feature vector. The gated enhanced feature vector is then subjected to a one-dimensional causal convolution operation, and after passing it through a ReLU activation function, it is multiplied element-wise with the attention feature vector to obtain the output vector of the causal convolution enhancement unit.

[0038] Furthermore, the temporal feature extraction branch includes: an adaptive emission attention module and several stacked iTransformer encoders;

[0039] The adaptive attention module is used to perform global average pooling and global max pooling on each feature variable matrix to obtain a global average pooling vector and a global max pooling vector, respectively; and then perform convolution operations on the global average pooling vector and the global max pooling vector to obtain a vector. and Then, for each vector and Applying the Sigmoid activation function and inputting the result into a linear mapping layer yields the temporal attention vector. Then, the time attention vector The time attention weight vector is obtained by element-wise multiplication with the input feature variable matrix. ;

[0040] Time attention weight vector Inputting a multi-layer stacked iTransformer encoder yields temporal dynamic features. ;

[0041] The iTransformer encoder comprises, in series, a multi-head self-attention sub-layer, a first normalization layer, a feedforward neural network, and a second normalization layer, and establishes residual connections between the input and output of the multi-head self-attention sub-layer and between the input and output of the feedforward neural network.

[0042] Furthermore, the specific method for using a two-stream neural network model to predict the preprocessed time-series data blocks and obtain the predicted values ​​of the key parameters for detonation initiation in the target detonation process is as follows:

[0043] Obtain the historical operational dataset of the target detonation process, including: time series of several sets of input feature variables, and measured values ​​of key detonation initiation parameters corresponding to each set of time series;

[0044] The key parameters for detonation initiation include: time domain parameters and intensity domain parameters;

[0045] For any set of input feature variables in the historical operation dataset, the time series is preprocessed to obtain the preprocessed time series data block, and together with the measured values ​​of the detonation initiation key parameters corresponding to the time series, a set of samples is obtained, thus obtaining several sets of samples.

[0046] All samples are divided into training and test sets according to a preset ratio;

[0047] Based on the adaptive hybrid loss function, the training set is input into the two-stream neural network model for iterative training, and the following operations are performed in each iteration:

[0048] For the samples in the training set, take a sample that contains... The training batches of samples are input into a two-stream neural network model to obtain the predicted values ​​corresponding to the global feature extraction branches of each sample. Predicted values ​​corresponding to the time feature extraction branch ;

[0049] Predicted values ​​based on the global feature extraction branch for each sample Predicted values ​​corresponding to the time feature extraction branch The adaptive mixture loss is calculated using the adaptive mixture loss function. ;

[0050] An optimization algorithm is used for backpropagation to minimize the adaptive mixing loss. To achieve this, update all learnable parameters in the two-stream neural network model;

[0051] When the model converges or reaches the preset number of training rounds, training is stopped, and a trained two-stream neural network model is obtained.

[0052] The trained two-stream neural network model is evaluated using a test set to obtain the final two-stream neural network model.

[0053] The preprocessed time-series data blocks are input into the final two-stream neural network model for prediction, to obtain the predicted values ​​of the key parameters for detonation initiation in the target detonation process.

[0054] Furthermore, the adaptive hybrid loss function is:

[0055] The Huber loss function is used to extract the predicted value corresponding to each branch based on the global features of each sample. Calculate the global characteristic flow loss using measured values ​​of key detonation initiation parameters. ;

[0056] The Log-Cosh loss function is used to extract the predicted value corresponding to each branch based on the temporal characteristics of each sample. Calculate the time characteristic flow loss based on the measured values ​​of key detonation initiation parameters. ;

[0057] The cosine similarity loss function is used to extract the predicted value corresponding to each branch based on the global features of each sample. Predicted values ​​corresponding to the time feature extraction branch Calculate consistency loss ;

[0058] Global Feature Flow Loss Temporal characteristic flow loss and consistency loss We perform weighted summation to obtain the adaptive hybrid loss. .

[0059] On the other hand, this invention proposes a detonation initiation prediction system based on a dual-stream neural network model, the system comprising:

[0060] The data preprocessing module is used to acquire the input characteristic variables of the target detonation process in real time, form the time series of the input characteristic variables, and perform preprocessing to obtain the preprocessed time series data block;

[0061] The data preprocessing module includes:

[0062] The data acquisition unit is used to acquire several input characteristic variables of the target detonation process in real time and form a time series of each input characteristic variable;

[0063] The filtering unit is used to filter the time series of each input feature variable using the Hampel filtering method based on the median and absolute median deviation to obtain the filtered time series.

[0064] The batch normalization unit is used to perform batch normalization on the filtered time series to obtain the batch normalized time series of all input feature variables;

[0065] Sliding cut-off unit, used to extract segments of length... A window is used to slide and truncate the time series of all input feature variables after batch normalization, with each window trunculating a fixed time step. All data points of the input feature variables within a continuous time step are used to form a feature variable matrix; all feature variable matrices are used as preprocessed time series data blocks.

[0066] The model building module is used to build and train a two-stream neural network model to obtain a trained two-stream neural network model.

[0067] The parameter prediction module is used to extract features and perform regression prediction on the preprocessed time series data blocks using a dual-stream neural network model to obtain the predicted values ​​of the key parameters for detonation initiation in the target detonation process.

[0068] The beneficial effects of adopting the above technical solution are as follows:

[0069] The method of this invention achieves high-precision prediction of key parameters in the detonation initiation process by collaboratively extracting global spatial features and temporal dynamic features through a dual-stream architecture.

[0070] In the two-stream architecture, the improved Transformer encoding module in the first branch enhances global spatial feature extraction and noise robustness through causal convolution enhancement; the adaptive attention module in the second branch, combined with iTransformer, accurately captures the evolution patterns and key dependencies of the time series. This invention's method comprehensively enhances feature extraction capabilities through dual-stream complementarity, thereby enabling in-depth mining of multi-dimensional information from the data.

[0071] The method of this invention is specifically designed with an adaptive hybrid loss function H-LLoss, which can dynamically adjust the loss weights according to the real-time training status of the two branches. This effectively solves the common problem of loss scale imbalance in two-stream network training, thereby accelerating model convergence and improving training efficiency and training stability.

[0072] The method of this invention also includes a robust data preprocessing workflow based on Hampel filters, which can effectively handle common noise and outliers in experimental or industrial data. The model structure is clear and highly modular, facilitating integration into existing detonation process monitoring and diagnostic systems, and possesses strong engineering practicality. Attached Figure Description

[0073] Figure 1 This is a flowchart of the detonation initiation prediction method based on a dual-stream neural network model in this embodiment;

[0074] Figure 2 This is a schematic diagram of the detonation initiation prediction method based on a dual-stream neural network model in this embodiment;

[0075] Figure 3 This is a schematic diagram of the structure of the improved Transformer encoding module in this embodiment;

[0076] Figure 4 This is a schematic diagram of the adaptive attention module in this embodiment;

[0077] Figure 5 This is a structural diagram of the detonation initiation prediction system based on a dual-stream neural network model in this embodiment. Detailed Implementation

[0078] To facilitate understanding of this application, specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and embodiments. The following embodiments are illustrative of the invention but are not intended to limit its scope. Rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of this application.

[0079] Example 1:

[0080] This embodiment presents a detonation initiation prediction method based on a two-stream neural network model, such as... Figure 1 As shown, the method includes the following procedures:

[0081] The input characteristic variables of the target detonation process are acquired in real time, and the time series of the input characteristic variables are formed and preprocessed to obtain the preprocessed time series data block.

[0082] The input characteristic variables include: thermodynamic parameters, chemical kinetic parameters, and flow parameters; wherein the thermodynamic parameters include: combustion zone pressure and combustion zone temperature; the chemical kinetic parameters include: heat release rate and key component concentration; and the flow parameters include: flow velocity and turbulence intensity.

[0083] In this embodiment, a filtering method based on median and absolute median deviation is used to filter the time series of each input feature independently to remove noise and outliers. Then, batch normalization is performed on all input feature variables after Hampel filtering to accelerate network convergence and improve training stability.

[0084] Specifically, the preprocessing method is as follows:

[0085] For any input feature variable, perform Hampel filtering on the time series of that input feature variable to obtain the filtered time series.

[0086] The method for performing Hampel filtering on the time series of the input feature variable to obtain the filtered time series is as follows:

[0087] For any data point in the time series Define a Centered on, window size is A sliding window, representing the data within the sliding window as And set a threshold constant. .

[0088] Calculate the median of the data within the sliding window.

[0089] (1);

[0090] in, Indicates The median of the data within the center sliding window; This indicates the calculation of the median of the data within the sliding window; Represents the first in a time series One data point; Represents the first in a time series One data point; Represents the first in a time series Data points.

[0091] Calculate the absolute value sequence of the differences between each data point within the sliding window and the median, and obtain the absolute median deviation by calculating the median of this absolute value sequence. (Median Absolute Deviation, MAD).

[0092] (2);

[0093] in, Indicates The absolute median deviation of the data within the centered sliding window.

[0094] The absolute median deviation Convert to standard deviation estimate .

[0095] In this embodiment, a constant 1.4826 is used as the conversion factor to convert the absolute median deviation. Converting to an estimate comparable to the standard deviation, i.e., standard deviation estimation. :

[0096] (3);

[0097] Based on the standard deviation estimation According to the set threshold constant Determine data points Are there any anomalies? If so, change the data points. Replace with median If no anomalies are found, the data points are retained. This leads to the filtered time series.

[0098] (4);

[0099] in, This represents a new data point obtained based on the anomaly detection result.

[0100] In this embodiment, the window size is set. With threshold constant Both are 3, for the th in the time series Data points ,calculate Median of data within the sliding window and absolute median deviation This leads to the standard deviation estimate. .like Then it is believed It is an outlier. Replace with Otherwise, retain the original value. This process significantly smooths the data curve and removes obvious outliers.

[0101] Batch normalization is performed on the filtered time series to obtain the batch-normalized time series, and then the batch-normalized time series of all input feature variables are obtained.

[0102] In this embodiment, the batch normalization process is represented as follows:

[0103] (5);

[0104] in, For any data value in the filtered time series; Indicates to The data values ​​obtained after normalization; and These are the mean and variance of the current batch of data, respectively. Global statistics are used during testing. It is a very small constant added to prevent the denominator from being zero, usually 1. ; and These are the learnable scaling parameter and offset parameter, respectively. In this embodiment, they will be... The initial value is set to 1. The initial value is set to 0.001. After normalization, the distribution of each feature data is more regular, which is beneficial to model convergence.

[0105] Use a length of A window is used to slide and truncate the time series of all input feature variables after batch normalization, with each window trunculating a fixed time step. The data points of all input feature variables within each consecutive time step form a group of size . The feature variable matrix; where This represents the number of input feature variables.

[0106] All feature variable matrices are used as preprocessed time series data blocks.

[0107] In this embodiment, the preprocessed continuous time series is divided into units of length 1. The data is divided into W consecutive time steps using a sliding window with a step size of 1. Each window contains all data points corresponding to W consecutive time steps, forming a feature variable matrix. .

[0108] A two-stream neural network model is used to predict the preprocessed time-series data blocks to obtain the predicted values ​​of key detonation initiation parameters for the target detonation process.

[0109] The dual-stream neural network model includes:

[0110] The dual-stream feature extraction module is used to extract global and temporal features in parallel from the preprocessed time-series data blocks, obtaining global spatial features and temporal dynamic features.

[0111] The dual-stream feature extraction module includes a global feature extraction branch and a temporal feature extraction branch.

[0112] In this embodiment, the dual-stream feature extraction module is the core of the method, consisting of two parallel and complementary feature extraction branches.

[0113] The global feature extraction branch is used to extract spatial features from the input data step by step to obtain global spatial features; wherein, the multivariate feature vector composed of data points of all input feature variables at the same time step in the preprocessed time series data block is used as the input data of the global feature extraction branch.

[0114] The time feature extraction branch is used to extract time features from the input data to obtain time dynamic features; wherein, each feature variable matrix in the preprocessed time series data block is used as the input data of the time feature extraction branch.

[0115] In this embodiment, the global feature extraction branch and the temporal feature extraction branch share the same preprocessed time-series data block as input, but they analyze the data from different dimensions. Specifically, the global feature extraction branch treats each time step as an independent unit, focusing on the spatial relationships between the input feature variables at that time; the temporal feature extraction branch treats the entire time window as a whole, capturing the dynamic patterns of the variables' evolution over time. (Using a feature variable matrix...) For example, in the global feature extraction branch, the feature variable matrix... The multivariate feature vector corresponding to each time step As an independent input, This allows for the extraction of spatial dependencies between input feature variables step-by-step, ultimately yielding... The global feature sequence corresponding to each time step is used as the global spatial feature. The time feature extraction branch extracts the feature variable matrix. As input, capture The evolution pattern between time steps. During training, the sample data used in the global feature extraction branch is denoted as... ,in This represents all input feature variable data corresponding to the same time step; express The corresponding measured values ​​of key detonation initiation parameters. Similarly, the sample data used in the time feature extraction branch can be obtained, such as... Figure 2 As shown.

[0116] The global feature extraction branch includes: a linear embedding layer and an improved Transformer encoding module consisting of several stacked layers.

[0117] In this embodiment, as Figure 2 As shown, the global feature extraction branch uses an improved Transformer encoding module as its core, aiming to mine global dependencies and complex nonlinear patterns among multiple feature variables from multivariate feature data at the same time step. All stacks use the same improved Transformer encoding module structure, and the feasible stacking range is generally selected according to the amount of data, usually between 3 and 6 layers. In this embodiment, 3 layers are used.

[0118] The linear embedding layer is used to perform linear projection on the multivariate feature vector corresponding to each time step to obtain an initial embedding vector, and the initial embedding vector is used as the input vector of the first layer improved Transformer coding module.

[0119] In this embodiment, for each time step , and its corresponding multivariate feature vector The input data for the global feature extraction branch is mapped to a higher-dimensional representation through a linear embedding layer, i.e. Figure 3 The positional encoding process shown yields the initial embedding vector. ,and , For the embedded dimension.

[0120] (6);

[0121] in, This represents a linear mapping layer.

[0122] The improved Transformer encoding module is as follows:

[0123] The input vector of the current improved Transformer coding module is layer normalized to obtain the embedding vector, which is then input into the multi-head self-attention layer to obtain the output vector of the multi-head self-attention layer.

[0124] In this embodiment, the improved Transformer encoding module of the first layer is taken as an example, such as... Figure 3 As shown, the input vector of the improved Transformer coding module for this layer is the initial embedding vector. Perform layer normalization to obtain the embedding vector. This is to improve training stability.

[0125] (7);

[0126] in, Presentation layer normalization processing.

[0127] In this embodiment, as Figure 3 As shown, the input vector of the current improved Transformer coding module is input into a multi-head self-attention sub-layer, which has multiple attention heads. Specifically, for the first layer of the improved Transformer coding module, the input vector is input into a multi-head self-attention sub-layer. Input a multi-head self-attention layer, and use a learnable weight matrix to... Convert to query matrix Key matrix Sum matrix The formula for calculating the attention of each single head is:

[0128] (8);

[0129] Among them, the query matrix The key matrix is ​​obtained by linearly transforming the input features through a learnable weight matrix. The value matrix is ​​obtained by linearly transforming the input features through a learnable weight matrix. It is obtained by linear transformation of input features through a learnable weight matrix; This represents the dimension of the key vector, used to scale the dot product result to prevent gradient explosion or vanishing. This indicates that the attention score matrix is ​​obtained by performing matrix multiplication on the transpose of the query matrix and the key matrix; This represents the Softmax activation function, used to normalize the attention score matrix into a probability distribution; This represents the output of the single-head self-attention mechanism, which is the result of weighted summation of the value matrix based on the attention weights.

[0130] To capture information more comprehensively, a multi-head self-attention layer performs multiple self-attention calculations in parallel. Each head independently learns a different subspace, and then the outputs of these heads are concatenated and subjected to a linear transformation to obtain the output vector. :

[0131] (9);

[0132] Where h represents the total number of attention heads; Indicates the first The output of the first self-attention head; the first The calculation method for each attention point is as follows: ,in They represent the first The query can learn the parameter matrix for each head, the key, and the value. This indicates that the outputs of multiple attention heads are concatenated along the feature dimension. The learnable parameter matrix for the output projection is used to map the concatenated multi-head features to the target dimension.

[0133] The output vector of the multi-head self-attention layer is residually concatenated with the embedding vector to obtain the attention feature vector.

[0134] In this embodiment, as Figure 3 As shown, the output vector By performing a residual concatenation with the initial embedding vector, we obtain the attention feature vector, denoted as:

[0135] (10);

[0136] in, This represents the attention feature vector.

[0137] The attention feature vector is input into the causal convolutional augmentation unit, and the output vector of the causal convolutional augmentation unit is normalized by the layer. The normalized output vector of the causal convolutional augmentation unit is then input into the feedforward neural network to obtain the output vector of the feedforward neural network.

[0138] This embodiment introduces a causal convolutional enhancement unit after the multi-head self-attention layer, and then performs a residual connection between the output of the feedforward neural network sub-layer and the output of the causal convolutional enhancement unit. This design enables the model to better capture local details and resist noise while extracting global contextual information. Specifically, as... Figure 3As shown, the causal convolution enhancement unit uses one-dimensional causal convolution and adaptively fuses and enhances features through a parallel dual-branch gating mechanism (using Tanh and Sigmoid activations respectively). After fusion, it performs a residual connection with the original input of the unit, namely the attention feature vector.

[0139] The causal convolution enhancement unit performs two parallel processing steps on the attention feature vector. One step involves performing a one-dimensional causal convolution operation on the attention feature vector and then passing it through a Tanh activation function to obtain a first output vector. The other step involves performing a one-dimensional causal convolution operation on the attention feature vector and then passing it through a Sigmoid activation function to obtain a second output vector. The first and second output vectors are then multiplied element-wise to obtain a gated enhanced feature vector. Finally, the gated enhanced feature vector is subjected to a one-dimensional causal convolution operation, and after passing it through a ReLU activation function, it is multiplied element-wise with the attention feature vector to obtain the output vector of the causal convolution enhancement unit.

[0140] The causal convolution enhancement unit is one of the key improvements to the Transformer coding module in this embodiment. The input causal convolutional enhancement unit enhances local feature extraction and suppresses noise. The causal convolutional enhancement unit comprises two parallel one-dimensional causal convolution operations. This is followed by Tanh and Sigmoid activation functions, and then the outputs of the two are multiplied element-wise to construct a gating mechanism. The process is represented as follows:

[0141] (11);

[0142] in, This represents the feature vector after gating enhancement; This represents the Tanh activation function; This represents the Sigmoid activation function; This indicates element-wise multiplication.

[0143] Subsequently, the output vector of the causal convolution enhancement unit is obtained through the ReLU activation function and residual structure. , represented as:

[0144] (12);

[0145] in, This represents the ReLU activation function.

[0146] The output vector of the causal convolution enhancement unit Perform layer normalization again, as follows:

[0147] (13);

[0148] in, This represents the output vector of the causal convolutional enhancement unit after normalization of the representation layer.

[0149] Will Input a feedforward neural network (FFN) that contains two linear transformations and a ReLU activation function, expressed as follows:

[0150] (14);

[0151] in, The output vector of the feedforward neural network, such as Figure 3 As shown, this embodiment uses a multilayer perceptron as a feedforward neural network; and All are learnable weight matrices; and These are all bias vectors; during training, , , and The initial values ​​are all obtained through random initialization.

[0152] The output vector of the feedforward neural network is added to the output vector of the causal convolutional enhancement unit to obtain the output vector of the current improved Transformer coding module, which is then used as the input vector of the next layer of improved Transformer coding module.

[0153] The output vector of the last improved Transformer coding module is used as the global spatial feature. .

[0154] In this embodiment, the output vector of the feedforward neural network is compared with... Adding them together yields the output vector of the improved Transformer encoding module. .

[0155] (15);

[0156] Because the improved Transformer encoding module has a multi-layer stacked structure, only the output vector of the last layer is used as the global spatial feature. .

[0157] In this embodiment, the time feature extraction branch is used to extract dynamic patterns within a continuous time window. This branch aims to capture the evolution trend and periodic patterns of emission data from the continuous time window.

[0158] The temporal feature extraction branch includes: an adaptive attention module and several stacked iTransformer encoder layers.

[0159] The adaptive attention module is used to perform global average pooling and global max pooling on each feature variable matrix to obtain a global average pooling vector and a global max pooling vector, respectively; and then perform convolution operations on the global average pooling vector and the global max pooling vector to obtain a vector. and Then, for each vector and Applying the Sigmoid activation function and inputting the result into a linear mapping layer yields the temporal attention vector. Then, the time attention vector The time attention weight vector is obtained by element-wise multiplication with the input feature variable matrix. .

[0160] In this embodiment, the feature variable matrix Input an adaptive attention module to adaptively assign importance weights at different time steps. For example... Figure 4 As shown, the data processing procedure of the adaptive attention module is as follows:

[0161] For the characteristic variable matrix respectively Global average pooling and global max pooling are performed to obtain two pooled vectors that converge different statistical information. These two pooled vectors are then passed through an independent 1×1 convolutional layer for feature interaction, enabling information exchange between different data sets.

[0162] (16);

[0163] in, This indicates a global average pooling operation; This represents the global max pooling operation; express The convolution operation.

[0164] right and The sigmoid activation function is applied to each element, and the results are then input into a linear mapping layer for linear transformation to generate a temporal attention vector. .

[0165] (17);

[0166] in, Represents a linear mapping layer; temporal attention weight vector This refers to the weighted time series features, which automatically highlight historical moments that are more important for the current prediction.

[0167] Then convert the time attention vector The time attention weight vector is obtained by element-wise multiplication with the input feature variable matrix. .

[0168] (18);

[0169] Time attention weight vector Inputting a multi-layer stacked iTransformer encoder yields temporal dynamic features. .

[0170] The iTransformer encoder comprises, in series, a multi-head self-attention sub-layer, a first normalization layer, a feedforward neural network, and a second normalization layer, and establishes residual connections between the input and output of the multi-head self-attention sub-layer and between the input and output of the feedforward neural network.

[0171] In this embodiment, weighted temporal features are input into a multi-layer stacked iTransformer encoder. The iTransformer encoder has the same structure as the standard Transformer encoder, including multi-head self-attention, layer normalization, residual connections, and a feedforward neural network. The iTransformer model treats the sequence of each input feature variable over the entire time window as a token, effectively modeling long-term dependencies and dynamic patterns across time steps using its multi-head self-attention mechanism and feedforward neural network. The output of the last iTransformer encoder layer serves as a high-level temporal dynamic feature. .

[0172] The multilayer perceptron module is used to perform nonlinear mapping on the global spatial features and temporal dynamic features respectively to obtain the predicted values ​​corresponding to the global feature extraction branch and the temporal feature extraction branch. The predicted values ​​corresponding to the global feature extraction branch are used as the predicted values ​​of the key parameters for detonation initiation. The predicted values ​​corresponding to the temporal feature extraction branch are only used during the training phase to assist in training the global feature extraction branch.

[0173] The multilayer perceptron module includes two identically structured multilayer perceptrons: a first multilayer perceptron and a second multilayer perceptron; wherein the first multilayer perceptron is used to perform nonlinear mapping on the global spatial features and output the predicted value corresponding to the global feature extraction branch; the second multilayer perceptron is used to perform nonlinear mapping on the temporal dynamic features and output the predicted value corresponding to the temporal feature extraction branch.

[0174] In this embodiment, global spatial features and time dynamic characteristics Two identical multilayer perceptrons (MLPs) are input, each containing at least one hidden layer and using the ReLU activation function. The resulting output layers for the two branches of the MLP are as follows: and ,in The predicted value corresponding to the global feature extraction branch; Extract the predicted values ​​corresponding to the branches for time features; and These are the predicted values ​​for the key parameter 1 of detonation initiation from the two branches, respectively. and These are the predicted values ​​for the two branches of the key parameter 2 for detonation initiation.

[0175] It should be noted that in this embodiment, during the training of the two-stream neural network model, a temporal feature extraction branch is used to assist in training the global feature extraction branch, thereby supplementing the temporal features and improving prediction accuracy. However, during model testing or application, only the predicted value corresponding to the global feature extraction branch is used as the final result.

[0176] The specific method for using a two-stream neural network model to predict the preprocessed time-series data blocks and obtain the predicted values ​​of key detonation initiation parameters of the target detonation process is as follows:

[0177] Obtain the historical operational dataset of the target detonation process, including: time series of several sets of input feature variables, and measured values ​​of key detonation initiation parameters corresponding to each time series.

[0178] The key parameters for detonation initiation include: time domain parameters and intensity domain parameters.

[0179] In this embodiment, historical time-series data is obtained from experimental or simulation datasets of the detonation process. This dataset includes multiple input feature variables such as combustion zone pressure, combustion zone temperature, and heat release rate, as well as key detonation initiation parameters as model outputs. The key detonation initiation parameters include time-domain parameters and intensity-domain parameters. The time-domain parameters describe the temporal characteristics of detonation initiation; in this embodiment, the initiation delay time is selected. The intensity-domain parameters describe the intensity characteristics of the detonation wave; in this embodiment, the peak pressure of the detonation wave is selected. Robust cleaning, denoising, and standardization are then performed on the original temperature, pressure, and heat release rate data.

[0180] For any set of input feature variables in the historical operation dataset, the time series is preprocessed to obtain the preprocessed time series data block, which, together with the measured values ​​of the detonation initiation key parameters corresponding to the time series, is used as a sample to obtain several sets of samples.

[0181] All samples are divided into training and test sets according to a preset ratio.

[0182] In this embodiment, sliding is performed at a fixed time step, and the construction length is... The preprocessed time series dataset is processed using a window to generate several samples. In this embodiment... The optimal value is 20. The detonation parameter value of the last time step within the window is set as the predicted target label for that window. Subsequently, all samples are randomly shuffled and divided into training and test sets according to a preset ratio, such as 7:3.

[0183] Based on the adaptive hybrid loss function, the training set is input into the two-stream neural network model for iterative training, and the following operations are performed in each iteration:

[0184] For the samples in the training set, take a sample that contains... The training batches of samples are input into a two-stream neural network model to obtain the predicted values ​​corresponding to the global feature extraction branches of each sample. Predicted values ​​corresponding to the time feature extraction branch .

[0185] Predicted values ​​based on the global feature extraction branch for each sample Predicted values ​​corresponding to the time feature extraction branch The adaptive mixture loss is calculated using the adaptive mixture loss function. .

[0186] To train the aforementioned two-stream network efficiently and stably, this embodiment specifically designs an adaptive hybrid loss function. This function consists of a global feature stream loss. Temporal characteristic flow loss and consistency loss It consists of three parts and introduces a dynamic weight balancing mechanism. For the global feature branch, Huber loss, which is robust to outliers, is used; for the time feature branch, Log-Cosh loss, which has good smoothness, is used; and for the consistency loss, cosine similarity loss, which can remove redundant features, is used.

[0187] The global feature flow loss The Huber loss function is adopted because it is insensitive to outliers, which is beneficial for robust learning of global features.

[0188] Let the true label value corresponding to the global feature extraction branch be... The true label value is derived from the measured values ​​of key detonation initiation parameters contained in each sample, and the hyperparameter is... This is used to control the switching point between quadratic loss and linear loss. In this embodiment, we take... Global feature flow loss Represented as:

[0189] (19);

[0190] in, For the first Key parameters of detonation initiation corresponding to the group of samples The actual label value; For the first Key parameters for detonation initiation obtained from the global feature extraction branch corresponding to the group of samples The predicted value.

[0191] The time-featured flow loss The Log-Cosh loss function is adopted, which is differentiable everywhere and very smooth, making it particularly suitable for time series regression tasks. It can avoid the problems of MAE being nondifferentiable at zero and MSE being sensitive to outliers.

[0192] Let the true label value corresponding to the time feature extraction branch be... The true label value is derived from the measured values ​​of key detonation initiation parameters included in each sample. Temporal feature flow loss. Represented as:

[0193] (20);

[0194] in, Represents the natural logarithm function; Represents the hyperbolic cosine function; For the first Key parameters for detonation initiation obtained from the temporal feature extraction branch corresponding to the sample group The predicted value; For the first Key parameters of detonation initiation corresponding to the group of samples The actual label value.

[0195] The consistency loss A cosine similarity loss function is used, which optimizes the network by measuring the similarity between the outputs of two branches, ensuring that the feature representations learned by the two branches are consistent when processing the same input. (Consistency loss) Represented as:

[0196] (twenty one);

[0197] in, express Norm; For the first The predicted value corresponding to the global feature extraction branch of the group of samples; For the first The predicted value corresponding to the branch of time feature extraction for the group of samples.

[0198] To balance the optimization difficulty of the two branches and prevent one branch from dominating the training process, the dynamic weighted loss function is defined as a weighted sum of the two losses. Its weights are dynamically adjusted according to their respective loss values, thereby automatically balancing the training difficulty of the two branches and promoting the rapid and stable convergence of the network.

[0199] (twenty two);

[0200] in, This represents the value of the adaptive hybrid loss function; The dynamic weights of the global feature flow; The dynamic weights of the time feature stream; and Initialize all values ​​to 1; It is a very small positive number, used to prevent the denominator from being zero; This represents a hyperparameter with a value of 0.15.

[0201] An optimization algorithm is used for backpropagation to minimize the adaptive mixing loss. To achieve this, update all learnable parameters in the two-stream neural network model, including: the weights of the linear embedding layers in each branch, and the query / key / value parameter matrix in the multi-head self-attention mechanism. The parameters include the kernel parameters in the causal convolutional enhancement unit, the weight matrices W1, W2 and biases b1, b2 in the feedforward neural network, the convolutional layer parameters in the adaptive attention module, the parameters in the iTransformer encoder, the multilayer perceptron parameters in the feature fusion module, and the learnable scaling parameters in the batch normalization layer. and offset parameter β.

[0202] In this embodiment, the Adam optimizer is used for model training, with an appropriate learning rate and batch size configured. During training, the loss curves on the training and test sets are monitored to ensure good model convergence and prevent overfitting. This iterative process is repeated multiple times until the model converges or reaches the preset number of training epochs.

[0203] Training stops when the model converges or reaches the preset number of training rounds, resulting in a well-trained two-stream neural network model.

[0204] The trained two-stream neural network model is evaluated using a test set to obtain the final two-stream neural network model.

[0205] In this embodiment, the trained model is used to make predictions on an independent test set, and quantitative evaluation is performed using metrics such as RMSE, MAE, and R². Experiments show that the dual-stream neural network model proposed in this embodiment can accurately predict key parameters of detonation initiation. The root mean square error (RMSE) is:

[0206] (twenty three);

[0207] The mean absolute error (MAE) is:

[0208] (twenty four);

[0209] The coefficient of determination R² is:

[0210] (25);

[0211] in, y is the number of test samples; n For the first The true label value of each test sample; For the first Predicted values ​​of key detonation initiation parameters for the target detonation process corresponding to each test sample; This represents the mean of the true label values.

[0212] A dual-stream neural network model is used to predict the preprocessed time-series data blocks to obtain the predicted values ​​of key detonation initiation parameters for the target detonation process.

[0213] In this embodiment, the trained and validated model parameters, along with the Hampel filter parameters and batch normalized statistics required for the preprocessing stage, are permanently stored. The trained and evaluated model and preprocessing parameters are then deployed to the production environment. In actual deployment, the aforementioned dual-stream neural network model can be integrated into a detonation process experiment or monitoring system. The system receives data streams in real time and performs preprocessing identical to that used in the training stage on newly arriving data.

[0214] In the preprocessing stage, the Hampel filter needs to use the current data point x t A symmetrical window centered on the data source. For real-time data streams, future data is unknown; therefore, this invention employs a delayed prediction strategy: caching the most recently received data. Historical data points, when new data x t Upon arrival, the data is not immediately filtered and predicted; instead, it waits for the subsequent k data points x. t+1 ,…..xt+k After all have arrived, then use x t A complete window is constructed around the center for Hampel filtering. At this point, x... t The corresponding predicted output is actually delayed by k steps. For scenarios requiring real-time output, an approximate window can be used, that is, an asymmetric window [x] is constructed using only historical data. t-2k ,…. x t As the input to the Hampel filter, although there is a slight loss of accuracy, it can meet the real-time requirements.

[0215] After preprocessing, an input window of length W is constructed and input into the trained dual-stream neural network model, which outputs predicted values ​​of key parameters in real time. These predictions can be used for process status early warning, combustion efficiency optimization, or safety control.

[0216] Example 2:

[0217] This embodiment presents a detonation initiation prediction system based on a dual-stream neural network model, such as... Figure 5 As shown, the system includes: a data preprocessing module, a model building module, and a parameter prediction module.

[0218] The data preprocessing module is used to acquire the input feature variables of the target detonation process in real time, form a time series of the input feature variables, and perform preprocessing to obtain a preprocessed time series data block.

[0219] The data preprocessing module includes: a data acquisition unit, a filtering unit, a batch normalization unit, and a sliding truncation unit.

[0220] The data acquisition unit is used to acquire several input feature variables of the target detonation process in real time and form a time series of each input feature variable.

[0221] The filtering unit is used to filter the time series of each input feature variable using the Hampel filtering method based on the median and absolute median deviation to obtain the filtered time series.

[0222] The batch normalization unit is used to perform batch normalization processing on the filtered time series to obtain the batch normalized time series of all input feature variables, so as to accelerate network convergence and improve training stability.

[0223] The sliding interception unit is used to employ a length of A window is used to slide and truncate the time series of all input feature variables after batch normalization, with each window trunculating a fixed time step. Data points of all input feature variables within a continuous time step are used to form a feature variable matrix; all feature variable matrices are used as preprocessed time series data blocks.

[0224] The model building module is used to build and train a two-stream neural network model to obtain a trained two-stream neural network model.

[0225] The parameter prediction module is used to extract features and perform regression prediction on the preprocessed time-series data block using a dual-stream neural network model to obtain predicted values ​​of key detonation initiation parameters for the target detonation process.

[0226] Example 3:

[0227] This embodiment proposes an electronic device, including: one or more processors, and a memory, wherein the memory is used to store instructions, and when the instructions are executed by the one or more processors, the one or more processors execute the detonation initiation prediction method based on the dual-stream neural network model.

[0228] The electronic device may be a mobile phone, computer, or tablet computer, etc., and includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, implements the detonation initiation prediction method based on a dual-stream neural network model as described in the embodiments. It is understood that the electronic device may also include an input / output (I / O) interface and communication components.

[0229] The processor is used to execute all or part of the steps in the detonation initiation prediction method based on a two-stream neural network model as described in the above embodiments. The memory is used to store various types of data, which may include, for example, instructions for any application or method in the electronic device, as well as application-related data.

[0230] The processor can be implemented as an Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), controller, microcontroller, microprocessor, or other electronic components, and is used to execute the detonation initiation prediction method based on the dual-stream neural network model described in the above embodiments.

[0231] Example 4:

[0232] This embodiment proposes a computer-readable storage medium that stores executable instructions. When these instructions are executed, if they are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0233] The computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the detonation initiation prediction method based on the dual-stream neural network model described in the various embodiments of this application.

[0234] The aforementioned storage media include: flash memory, hard disks, multimedia cards, card-type memory (e.g., SD (Secure Digital Memory Card) or DX (Memory Data Register, MDR) memory), random access memory (RAM), static random-access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic storage, disks, optical discs, servers, APP (Application) application stores, and other media capable of storing program verification codes. These media store computer programs, which, when executed by a processor, can implement the various steps of the detonation initiation prediction method based on the dual-stream neural network model described above.

[0235] Example 5:

[0236] This embodiment proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the detonation initiation prediction method based on a dual-stream neural network model.

[0237] Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a computer program product.

[0238] The various embodiments in this application are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0239] The scope of protection of this application is not limited to the embodiments described above. Obviously, those skilled in the art can make various modifications and variations to this disclosure without departing from the scope and spirit of this disclosure. If such modifications and variations fall within the scope of this disclosure and its equivalents, then the intent of this disclosure also includes these modifications and variations.

Claims

1. A method for predicting detonation initiation based on a double-flow neural network model, characterized in that, This method The process includes the following: The input characteristic variables of the target detonation process are acquired in real time, a time series of the input characteristic variables is formed and preprocessed to obtain a preprocessed time series data block; A two-stream neural network model is used to predict the preprocessed time series data blocks to obtain the predicted values ​​of the key parameters for detonation initiation in the target detonation process. The dual-stream neural network model includes: The dual-stream feature extraction module includes a global feature extraction branch and a temporal feature extraction branch, which are used to perform global feature extraction and temporal feature extraction on the preprocessed time-series data blocks in parallel to obtain global spatial features and temporal dynamic features; The multilayer perceptron module is used to perform nonlinear mapping on the global spatial features and temporal dynamic features respectively to obtain the predicted values ​​corresponding to the global feature extraction branch and the temporal feature extraction branch. The predicted values ​​corresponding to the global feature extraction branch are used as the predicted values ​​of the key parameters for detonation initiation. The predicted values ​​corresponding to the temporal feature extraction branch are only used during the training phase to assist in training the global feature extraction branch.

2. The detonation initiation prediction method based on a dual-stream neural network model according to claim 1, characterized in that, The input characteristic variables include: thermodynamic parameters, chemical kinetic parameters, and flow parameters; wherein the thermodynamic parameters include: combustion zone pressure and combustion zone temperature; the chemical kinetic parameters include: heat release rate and key component concentration; and the flow parameters include: flow velocity and turbulence intensity.

3. The detonation initiation prediction method based on a dual-stream neural network model according to claim 1, characterized in that, The method for preprocessing the time series of the input feature variable is as follows: For any input feature variable, perform Hampel filtering on the time series of the input feature variable to obtain the filtered time series; the specific process is as follows: For any data point in the time series Define a Centered on, window size is A sliding window, representing the data within the sliding window as And set a threshold constant. ; Calculate the median of the data within the sliding window. ; Calculate the sum of each data point and the median within the sliding window. The absolute median deviation is obtained by calculating the median of the absolute value sequence of the differences between the two values. ; The absolute median deviation Convert to standard deviation estimate ; Based on the standard deviation estimation According to the set threshold constant Determine data points Are there any anomalies? If so, change the data points. Replace with median ; If no anomalies are found, the data points are retained. This leads to the filtered time series. Batch normalization is performed on the filtered time series to obtain the batch normalized time series, and then the batch normalized time series of all input feature variables are obtained. Use a length of A window is used to slide and truncate the time series of all input feature variables after batch normalization, with each window trunculating a fixed time step. The data points of all input feature variables within each consecutive time step form a group of size . The feature variable matrix; where The number of input feature variables; All feature variable matrices are used as preprocessed time series data blocks.

4. The detonation initiation prediction method based on a dual-stream neural network model according to claim 3, characterized in that, The global feature extraction branch is used to extract spatial features from the input data step by step to obtain global spatial features; wherein, the multivariate feature vector composed of data points of all input feature variables at the same time step in the preprocessed time series data block is used as the input data of the global feature extraction branch. The time feature extraction branch is used to extract time features from the input data to obtain time dynamic features; wherein, each feature variable matrix in the preprocessed time series data block is used as the input data of the time feature extraction branch. The multilayer perceptron module includes two identically structured multilayer perceptrons: a first multilayer perceptron and a second multilayer perceptron; wherein the first multilayer perceptron is used to perform nonlinear mapping on the global spatial features and output the predicted value corresponding to the global feature extraction branch; the second multilayer perceptron is used to perform nonlinear mapping on the temporal dynamic features and output the predicted value corresponding to the temporal feature extraction branch.

5. The detonation initiation prediction method based on a dual-stream neural network model according to claim 4, characterized in that, The global feature extraction branch includes: a linear embedding layer and an improved Transformer coding module consisting of several stacked layers; The linear embedding layer is used to perform linear projection on the multivariate feature vector corresponding to each time step to obtain an initial embedding vector, and the initial embedding vector is used as the input vector of the first layer improved Transformer coding module. The improved Transformer encoding module is as follows: The input vector of the current improved Transformer coding module is layer normalized to obtain the embedding vector, which is then input into the multi-head self-attention layer to obtain the output vector of the multi-head self-attention layer. The output vector of the multi-head self-attention layer is residually concatenated with the embedding vector to obtain the attention feature vector; The attention feature vector is input into the causal convolutional enhancement unit, and the output vector of the causal convolutional enhancement unit is normalized by the layer. The output vector of the causal convolutional enhancement unit after the layer normalization is input into the feedforward neural network to obtain the output vector of the feedforward neural network. The output vector of the feedforward neural network is added to the output vector of the causal convolutional enhancement unit to obtain the output vector of the current improved Transformer coding module, which is then used as the input vector of the next layer of improved Transformer coding module. The output vector of the last improved Transformer coding module is used as the global spatial feature. .

6. The detonation initiation prediction method based on a dual-stream neural network model according to claim 5, characterized in that, The causal convolution enhancement unit performs two parallel processing steps on the attention feature vector. One step involves performing a one-dimensional causal convolution operation on the attention feature vector and then passing it through a Tanh activation function to obtain a first output vector. The other step involves performing a one-dimensional causal convolution operation on the attention feature vector and then passing it through a Sigmoid activation function to obtain a second output vector. The first and second output vectors are then multiplied element-wise to obtain a gated enhanced feature vector. Finally, the gated enhanced feature vector is subjected to a one-dimensional causal convolution operation, and after passing it through a ReLU activation function, it is multiplied element-wise with the attention feature vector to obtain the output vector of the causal convolution enhancement unit.

7. The detonation initiation prediction method based on a dual-stream neural network model according to claim 6, characterized in that, The temporal feature extraction branch includes: an adaptive emission attention module and several stacked iTransformer encoders; The adaptive attention module is used to perform global average pooling and global max pooling on each feature variable matrix to obtain a global average pooling vector and a global max pooling vector, respectively; and then perform convolution operations on the global average pooling vector and the global max pooling vector to obtain a vector. and Then, for each vector and Applying the Sigmoid activation function and inputting the result into a linear mapping layer yields the temporal attention vector. Then, the time attention vector The time attention weight vector is obtained by element-wise multiplication with the input feature variable matrix. ; Time attention weight vector Inputting a multi-layer stacked iTransformer encoder yields temporal dynamic features. ; The iTransformer encoder comprises, in series, a multi-head self-attention sub-layer, a first normalization layer, a feedforward neural network, and a second normalization layer, and establishes residual connections between the input and output of the multi-head self-attention sub-layer and between the input and output of the feedforward neural network.

8. The detonation initiation prediction method based on a dual-stream neural network model according to claim 7, characterized in that, The specific method for using a two-stream neural network model to predict the preprocessed time-series data blocks and obtain the predicted values ​​of key detonation initiation parameters of the target detonation process is as follows: Obtain the historical operational dataset of the target detonation process, including: time series of several sets of input feature variables, and measured values ​​of key detonation initiation parameters corresponding to each set of time series; The key parameters for detonation initiation include: time domain parameters and intensity domain parameters; For any set of input feature variables in the historical operation dataset, the time series is preprocessed to obtain the preprocessed time series data block, and together with the measured values ​​of the detonation initiation key parameters corresponding to the time series, a set of samples is obtained, thus obtaining several sets of samples. All samples are divided into training and test sets according to a preset ratio; Based on the adaptive hybrid loss function, the training set is input into the two-stream neural network model for iterative training, and the following operations are performed in each iteration: For the samples in the training set, take a sample that contains... The training batches of samples are input into a two-stream neural network model to obtain the predicted values ​​corresponding to the global feature extraction branches of each sample. Predicted values ​​corresponding to the time feature extraction branch ; Predicted values ​​based on the global feature extraction branch for each sample Predicted values ​​corresponding to the time feature extraction branch The adaptive mixture loss is calculated using the adaptive mixture loss function. ; An optimization algorithm is used for backpropagation to minimize the adaptive mixing loss. To achieve this, update all learnable parameters in the two-stream neural network model; When the model converges or reaches the preset number of training rounds, training is stopped, and a trained two-stream neural network model is obtained. The trained two-stream neural network model is evaluated using a test set to obtain the final two-stream neural network model. The preprocessed time-series data blocks are input into the final two-stream neural network model for prediction, to obtain the predicted values ​​of the key parameters for detonation initiation in the target detonation process.

9. The detonation initiation prediction method based on a dual-stream neural network model according to claim 8, characterized in that, The adaptive hybrid loss function is: The Huber loss function is used to extract the predicted value corresponding to each branch based on the global features of each sample. Calculate the global characteristic flow loss using measured values ​​of key detonation initiation parameters. ; The Log-Cosh loss function is used to extract the predicted value corresponding to each branch based on the temporal characteristics of each sample. Calculate the time characteristic flow loss based on the measured values ​​of key detonation initiation parameters. ; The cosine similarity loss function is used to extract the predicted value corresponding to each branch based on the global features of each sample. Predicted values ​​corresponding to the time feature extraction branch Calculate consistency loss ; Global Feature Flow Loss Temporal characteristic flow loss and consistency loss We perform weighted summation to obtain the adaptive hybrid loss. .

10. A detonation initiation prediction system based on a dual-stream neural network model, used to implement the detonation initiation prediction method based on a dual-stream neural network model as described in any one of claims 1-9, characterized in that, The system includes: The data preprocessing module is used to acquire the input characteristic variables of the target detonation process in real time, form the time series of the input characteristic variables, and perform preprocessing to obtain the preprocessed time series data block; The data preprocessing module includes: The data acquisition unit is used to acquire several input characteristic variables of the target detonation process in real time and form a time series of each input characteristic variable; The filtering unit is used to filter the time series of each input feature variable using the Hampel filtering method based on the median and absolute median deviation to obtain the filtered time series. The batch normalization unit is used to perform batch normalization on the filtered time series to obtain the batch normalized time series of all input feature variables; Sliding cut-off unit, used to extract segments of length... A window is used to slide and truncate the time series of all input feature variables after batch normalization, with each window trunculating a fixed time step. All data points of the input feature variables within a continuous time step are used to form a feature variable matrix; all feature variable matrices are used as preprocessed time series data blocks. The model building module is used to build and train a two-stream neural network model to obtain a trained two-stream neural network model. The parameter prediction module is used to extract features and perform regression prediction on the preprocessed time series data blocks using a dual-stream neural network model to obtain the predicted values ​​of the key parameters for detonation initiation in the target detonation process.

Images