Powder fuel mass flow rate prediction model based on physical information and data driving

Abstract

The invention discloses a powder fuel mass flow rate prediction model based on physical information and data driving, and relates to the technical field of laser spectrum application. The laser system generates laser beams, the light modulation and shaping system modulates the laser beams, the signal detection system suppresses background light noise, the camera system collects scattering images, the image preprocessing module carries out cutting preprocessing, the feature extraction module learns image features through a deep belief network, and the XGBoost regression prediction module considers physical constraint parameters. And the learned image features and real-time quality flow rate data are input into an XGBoost model for training calculation, and a nonlinear quality flow rate prediction model is constructed. A scattering image and a real-time quality flow rate serve as original data, physical constraint parameters are considered at the same time, image features are extracted through a deep belief network, training is carried out based on an XGBoost model, a nonlinear quality flow rate prediction model based on physical information and data driving is constructed, and the quality flow rate prediction precision is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of laser spectroscopy applications, specifically a prediction model for the mass flow rate of powder fuel based on physical information and data-driven. Background Art

[0002] Powder fuel engines have the advantages of adjustable flow rate, high combustion efficiency, and high specific impulse. They are suitable for combat missions such as large airspace, wide speed range, and multi-ballistic flight, and have broad application scenarios in new-generation long-range air-to-air missiles, air-to-ground missiles, and air defense missiles. This type of engine uses high-energy metals or powders such as boron as fuel. Under the action of the fluidizing gas, the powder fuel enters the combustion chamber in a gas-solid two-phase form to organize combustion, thereby generating thrust. As the core component of the powder fuel engine, the performance of the fuel supply system directly determines the thrust adjustment ability and multi-pulse start characteristics of the engine. To meet the complex requirements in multi-task scenarios, the fuel supply system not only needs to have the flexible adjustment ability of the mass flow rate, but also must maintain a high degree of flow stability under constant operating conditions. Therefore, achieving real-time high-precision prediction of the mass flow rate of powder fuel is of great significance for developing high-reliability fuel supply technology and realizing precise control of engine thrust. By accurately predicting the mass flow rate of powder fuel, key data support can be provided for the control strategy of the fuel supply system, thereby optimizing the overall performance of the engine and ensuring its stable and efficient operation under diverse working conditions.

[0003] At present, the research on the measurement of the mass flow rate of powder fuel mainly focuses on the pneumatic conveying pipeline environment. There is less measurement research carried out under the fuel supply mode around the characteristics of the powder fuel engine, and it is limited to steady-state measurement methods such as Coriolis flow meters and weighing methods. However, the mass flow rate of the dense gas-solid two-phase flow field is affected by various factors, such as fluidizing gas pressure, piston speed, powder fuel characteristics, and fluidization parameters, resulting in large differences in the characteristics of the dense gas-solid two-phase flow field and unclear importance of characteristics. Therefore, the existing models developed based on steady-state measurement results cannot achieve high-precision prediction of the mass flow rate of powder fuel.

[0004] In recent years, machine learning methods have shown great potential as a new idea for predictive modeling. Especially in the case of multi-factor coupling and strong interaction effects between variables, machine models have shown more advantages. Machine learning models pay more attention to the associations and correlations between data samples. By mining the potential laws and associations between input and output variables in a high-dimensional space, complex mapping relationships are established, and great progress has been made in the characterization of powder fuel mass flow rate. For example, using support vector machines, convolutional neural networks, etc., by mining characteristic parameters, the powder fuel mass flow rate under the influence of various factors is obtained, and finally it is applied in a pneumatic conveying pipeline with good results. However, up to now, no relevant research reports have been seen in the field of powder fuel engines. Although machine learning models provide a new idea for the construction of mass flow rate prediction models with their powerful data mining capabilities, as black box systems lacking physical mechanism constraints, people have always had doubts about the reliability of prediction results.

[0005] In view of the above analysis, in order to achieve high-precision prediction of powder fuel mass flow rate, it is urgent to establish a physical information and data-driven powder fuel mass flow rate prediction model. By combining the physical information to describe the mechanism of the powder fuel conveying process and the data-driven fitting ability for complex nonlinear relationships, reliable technical support can be provided for the optimal control of the fuel supply system. Summary of the Invention

[0006] To solve the deficiencies in the background technology, the present invention provides a physical information and data-driven powder fuel mass flow rate prediction model, which takes scattering images and real-time mass flow rate as original data, and at the same time considers physical constraint parameters. Image features are extracted through a deep belief network and trained based on the XGBoost model to construct a physical information and data-driven nonlinear mass flow rate prediction model, improving the accuracy of mass flow rate prediction.

[0007] To achieve the above object, the present invention adopts the following technical solutions: A physical information and data-driven powder fuel mass flow rate prediction model includes a data acquisition system and a computer system. The data acquisition system includes a laser system, an optical modulation and shaping system, a signal detection system, and a camera system. The computer system includes an image preprocessing module, a feature extraction module, and an XGBoost regression prediction module;

[0008] The laser system generates a laser beam according to the type of powder fuel in the combustion chamber;

[0009] The optical modulation and shaping system is arranged between the laser system and the engine to modulate the laser beam;

[0010] The signal detection system is arranged behind the area to be measured to suppress background light noise;

[0011] The camera system collects the scattering images of the powder fuel particles in the combustion chamber, and combines the particle density and the particle flow rate to obtain the real-time mass flow rate of the scattering images;

[0012] The image preprocessing module performs cropping preprocessing on the collected scattering images;

[0013] The feature extraction module learns the image features of the scattering images through a deep belief network;

[0014] The XGBoost regression prediction module considers the physical constraint parameters, and inputs the learned image features and the real-time mass flow rate data into the XGBoost model for training and calculation to construct a non-linear mass flow rate prediction model;

[0015] The specific implementation steps are as follows:

[0016] Step 1: Acquisition and processing of original data

[0017] Use the camera system to collect the scattering images of multiple groups of powder fuels, measure the particle density of the powder fuel according to the extinction principle, measure the particle flow rate through the PIV and optical flow algorithms, combine the particle density and the particle flow rate to obtain the real-time mass flow rate of the scattering images, and consider the physical constraint parameters to label the scattering images as the original data;

[0018] Step 2: Extract deep-level image features of the data

[0019] Crop the scattering images into an image format suitable for input to deep learning, and extract the image features through a deep belief network without considering the labels. It is composed of multiple RBMs stacked together. Through layer-by-layer training, the features learned by each RBM are used as the input for the next layer;

[0020] Step 3: Construct a non-linear mass flow rate prediction model

[0021] Form a data set with the extracted image features, the real-time mass flow rate considering the labels, and the physical constraint parameters as input variables, and use the predicted mass flow rate as the output variable. Train through the XGBoost model, and its objective function is where n is the total number of training data, K is the total number of decision trees trained, is the loss function, y i is the i-th real-time mass flow rate data, is the i-th predicted mass flow rate data, Ω(f k ) is the regularization term, f k is the decision tree trained in the k-th round, T k is the number of leaf nodes of the decision tree trained in the k-th round, w jis the weight of the j-th leaf node, and γ and λ are hyperparameters;

[0022] The XGBoost model optimizes through forward stagewise, and trains a new regression tree f k (x) in each round to optimize the residual of the previous round. Then the predicted value in the k-th round is is the mass flow rate obtained from the predictions in the previous k - 1 rounds, and f k (x i ) is the predicted value of the decision tree newly trained in the k-th round for the input x i , and x i are the image features and physical constraint parameters in the i-th group of data;

[0023] The newly trained regression tree f k (x) optimizes the prediction error by fitting the negative gradient of the current loss. The negative gradient of the i-th group of data in the k-th round Through continuous iteration, the final predicted mass flow rate data is represented by the weighted sum of all decision trees as to obtain a trained non-linear mass flow rate prediction model;

[0024] Step 4: Model accuracy evaluation

[0025] Input the scattered image of the preprocessed powder fuel into the trained non-linear mass flow rate prediction model, output the predicted mass flow rate of the powder fuel, and compare the data with the real-time mass flow rate considering the label to evaluate the prediction accuracy. If the requirements are not met, adjust the hyperparameters γ and λ in the objective function.

[0026] Furthermore, the physical constraint parameters include the piston movement speed of the engine, the caliber size, and the particle size scale of the powder fuel particles.

[0027] Furthermore, the RBM consists of a visible layer and a hidden layer. The weight value between any two connected neurons in the visible layer and the hidden layer is w ij , the neuron states of the visible layer and the hidden layer are v i and hj respectively. The energy function of the RBM is defined as: E(v, h) = -∑a i v i -∑b j h j -∑v i w ij h j , where ai and b j are bias terms.

[0028] Furthermore, the evaluation of the accuracy is based on the average error. If the average error ≤ 5%, it is considered that the prediction result of the non-linear mass flow rate prediction model meets the requirements.

[0029] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention uses a camera system to collect the scattering images of powder fuel, and uses a deep belief network to perform feature representation learning on the original image data. The extracted features are input into a regression model of supervised learning to establish a mapping relationship between the scattering image and the mass flow rate. The calibration of the mass flow rate is obtained by using the real-time flow rate data acquired by the extinction method and used as the target value (label) of the regression model. At the same time, physical parameters closely related to the mass flow rate, such as the piston movement speed of the engine, the caliber size, and the particle size scale of the powder fuel particles, are considered as auxiliary physical constraint conditions to construct a powder fuel mass flow rate prediction model based on physical information and data-driven, which helps to improve the accuracy of mass flow rate prediction and solve the problem of unstable operation of the powder fuel engine caused by the current mass flow rate deviation. Description of the Drawings

[0030] Figure 1 is a schematic structural diagram of the prediction model of the present invention;

[0031] Figure 2 is a schematic diagram of the relationship between RBM and DBN in the present invention. Detailed Embodiments

[0032] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0033] As Figures 1 to 2 shown, the powder fuel mass flow rate prediction model based on physical information and data-driven includes a data acquisition system and a computer system, wherein: the data acquisition system includes a laser system, an optical modulation and shaping system, a signal detection system, and a camera system for real-time acquisition of the scattering images of powder fuel; the computer system includes an image preprocessing module, a feature extraction module, and an XGBoost regression prediction module for image preprocessing and model training.

[0034] The laser system generates a laser beam according to the type of powder fuel in the combustion chamber to determine the wavelength;

[0035] The optical modulation and shaping system is arranged between the laser system and the engine for modulating the laser beam;

[0036] The signal detection system is arranged at the same horizontal height as the laser system behind the area to be measured for suppressing the background light noise without modulation and improving the signal detection signal-to-noise ratio;

[0037] The camera system collects the scattering images of powder fuel particles in the combustion chamber, and combines the particle density and particle flow rate to obtain the real-time mass flow rate of the scattering images;

[0038] The image preprocessing module performs cropping preprocessing on the collected scattering images;

[0039] The feature extraction module learns the image features of the scattering images through a deep belief network;

[0040] The XGBoost regression prediction module considers physical constraint parameters such as the piston movement speed, bore size of the engine, and particle size scale of the powder fuel particles, and inputs the learned image features and real-time mass flow rate data into the XGBoost model for training and calculation to construct a non-linear mass flow rate prediction model to achieve the prediction of the mass flow rate.

[0041] When the prediction model of the present invention is applied, first, multiple sets of scattering images of powder fuel and the associated real-time mass flow rate are obtained as original data. At the same time, physical constraint parameters such as the piston movement speed, bore size of the engine, and particle size scale of the powder fuel particles are considered, and these scattering images are labeled. Then, the obtained scattering images are input into the deep belief network for unsupervised learning without considering the labels to extract the image features of the original data and obtain the feature law closer to the original nature of the data. Finally, through the XGBoost model based on supervised learning, the model is trained using a data set including image features, real-time mass flow rate considering labels, and physical constraint parameters to construct a non-linear mass flow rate prediction model. The specific implementation steps are as follows:

[0042] Step 1: Acquisition and processing of original data

[0043] The laser beam generated by the laser system is incident on the area to be measured in the combustion chamber as a light source after passing through the optical modulation and shaping system, exciting the powder fuel particles in the same plane. The camera system collects multiple sets of scattering images of the powder fuel. The signal detection system measures the particle density of the powder fuel by comparing the relationship between the transmitted light intensity and the incident light intensity according to the extinction principle. For small particles and low-concentration regions, the particle flow rate is measured by PIV. For a dense gas-solid two-phase flow field, the optical flow algorithm is used to calculate the movement of pixel points or feature points in the continuous image frames to obtain the velocity vector field. Combining the particle density and particle flow rate, the real-time mass flow rate of the scattering images is obtained. Considering physical constraint parameters such as the piston movement speed, bore size of the engine, and particle size scale of the powder fuel particles, these scattering images are labeled as the original data for data preparation for subsequent XGBoost model training;

[0044] Step 2: Extract deep-level image features of the data

[0045] The scattered image is cropped by the image preprocessing module, reshaped into an image format suitable for input to deep learning, and then input into the feature extraction module without considering the labels. The feature extraction module uses a deep belief network (DBN) to extract image features. As shown in Figure 2 , it is composed of multiple restricted Boltzmann machines (RBMs) stacked together. An RBM consists of a visible layer and a hidden layer. The number of neurons in the first visible layer is determined by the dimensionality of the input data. The weight value between any two connected neurons in the visible layer and the hidden layer is w ij , the neuron states of the visible layer and the hidden layer are v i and h j respectively. Then, the energy function of the RBM is defined as: E(v, h) = -∑a i v i - ∑b j h j - ∑v i w ij h j , where ai and bj are bias terms. Through layer-by-layer training, the features learned by each RBM are used as the input for the next layer, and a DBN is formed by cascading multiple RBMs. The DBN uses a non-linear structure to extract image features and map the data from a high-dimensional space to a low-dimensional space, maximizing the finding of the best feature representation that can reconstruct the original data;

[0046] Step 3: Construct a non-linear mass flow rate prediction model

[0047] The extracted image features, the real-time mass flow rate considering the labels, and the physical constraint parameters are formed into a data set and used as input variables together, and the predicted mass flow rate is used as the output variable. The prediction of the powder fuel mass flow rate is realized through the XGBoost regression prediction module. The core of the XGBoost model is to iteratively train multiple decision tree models and combine them to form a powerful integrated model. Its objective function consists of a loss function and a regularization term, and can be expressed as where n is the total number of training data, K is the total number of decision trees trained, is the loss function, and the mean squared error (MSE) is adopted, y i is the i-th real-time mass flow rate data, is the i-th predicted mass flow rate data, Ω(f k ) is the regularization term used to prevent overfitting, and is defined as: f k is the decision tree trained in the k-th round, T k is the number of leaf nodes of the decision tree trained in the k-th round, w jis the weight of the j-th leaf node, and γ and λ are hyperparameters used to control the complexity of the model. Here, f k represents the entire decision tree, which is used to calculate the structural complexity of the tree, but the specific input x is not involved at this time.

[0048] The XGBoost model optimizes step by step forward. In each round, a new regression tree f k (x) is trained to optimize the residuals of the previous round. Then, the predicted value in the k-th round is is the mass flow rate obtained from the predictions in the previous k - 1 rounds, and f k (x i ) is the predicted value of the decision tree newly trained in the k-th round for the input x i , and x i is the image feature and physical constraint parameter in the i-th group of data. Here, f k (x) represents the predicted value after the decision tree f k acts on the input x, which is different from the meaning of f k in the regularization term.

[0049] The newly trained regression tree f k (x) optimizes the prediction error by fitting the negative gradient of the current loss. The definition of the negative gradient of the i-th group of data in the k-th round is Through continuous iteration, the final predicted mass flow rate data is represented by the weighted sum of all decision trees as and a trained non-linear mass flow rate prediction model is obtained;

[0050] Step Four: Model Accuracy Evaluation

[0051] Input the preprocessed scattering images of the powdered fuel into the trained non-linear mass flow rate prediction model, output the predicted mass flow rate of the powdered fuel, and compare the data with the real-time mass flow rate considering the label to evaluate the prediction accuracy. Use the average error as the basis for determining accuracy. If the average error ≤ 5%, it is determined that the prediction result of the non-linear mass flow rate prediction model is accurate; otherwise, adjust the two hyperparameters γ and λ in the objective function until the requirements are met.

[0052] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, in any aspect, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent conditions of the claims are intended to be included in the present invention. Any reference signs in the claims should not be regarded as limiting the claims involved.

[0053] In addition, it should be understood that although this specification is described in terms of embodiments, not every embodiment contains only an independent technical solution. This narrative manner of the specification is merely for clarity. Those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A physical information-based and data-driven powder fuel mass flow rate prediction model, characterized by: It includes a data acquisition system and a computer system. The data acquisition system includes a laser system, a light modulation and shaping system, a signal detection system and a camera system. The computer system includes an image preprocessing module, a feature extraction module and an XGBoost regression prediction module. The laser system generates a laser beam at a wavelength determined according to the type of powdered fuel in the combustion chamber; The light modulation and shaping system is arranged between the laser system and the engine to modulate the laser beam; The signal detection system is arranged behind the area to be detected to suppress background light noise; The camera system collects scattered images of powdered fuel particles in the combustion chamber and obtains the real-time mass flow rate of the scattered images by combining the particle density and the particle flow rate; The image preprocessing module performs cropping preprocessing on the collected scattering image; The feature extraction module learns the image features of the scattering image through a deep belief network; The XGBoost regression prediction module considers physical constraint parameters, and inputs the learned image features and real-time mass flow rate data into the XGBoost model for training calculation to build a nonlinear mass flow rate prediction model; The specific implementation steps are as follows: Step 1: Collection and processing of raw data The camera system is used to collect multiple groups of scattered images of powdered fuel. The particle density of powdered fuel is measured according to the extinction principle. The particle flow rate is measured by PIV and optical flow algorithm. The real-time mass flow rate of the scattered image is obtained by combining the particle density and particle flow rate. The scattered image is labeled as raw data considering the physical constraint parameters. Step 2: Extract deep image features from data The scatter image is cropped into an image format suitable for deep learning input, and the image features are extracted through a deep belief network without considering the label. The network is composed of multiple RBM stacks. Through layer-by-layer training, the features learned by each RBM are used as the input of the next layer. Step 3: Build a nonlinear mass flow rate prediction model The extracted image features, the real-time mass flow rate considering the labels, and the physical constraint parameters form a data set as input variables, and the predicted mass flow rate is used as the output variable. The XGBoost model is trained, and its objective function is: Among them, n is the total number of training data, K is the total number of decision trees trained, is the loss function, y i is the ith real-time mass flow rate data, is the i-th predicted mass flow rate data, Ω(f k ) is the regularization term, f k is the decision tree trained in the kth round, T k is the number of leaf nodes of the decision tree trained in the kth round, w j is the weight of the jth leaf node, γ and λ are hyperparameters; The XGBoost model uses forward step-by-step optimization to train a new regression tree f in each round. k (x) To optimize the residual of the previous round, the prediction value of the kth round is is the mass flow rate predicted in the first k-1 rounds, f k (x i ) is the newly trained decision tree for the kth round for the input x i The predicted value, x i are the image features and physical constraint parameters in the i-th group of data; The newly trained regression tree f k (x) Optimize the prediction error by fitting the negative gradient of the current loss. The negative gradient of the i-th group of data in the k-th round Through continuous iterations, the final predicted mass flow rate data is expressed by the weighted sum of all decision trees as A trained nonlinear mass flow rate prediction model is obtained; Step 4: Model Accuracy Evaluation The preprocessed scattering image of powdered fuel is input into the trained nonlinear mass flow rate prediction model, and the predicted mass flow rate of powdered fuel is output. The data is compared with the real-time mass flow rate considering the label to evaluate the prediction accuracy. If the requirements are not met, the hyperparameters γ and λ in the objective function are adjusted.

2. The powder fuel mass flow rate prediction model based on physical information and data-driven according to claim 1, characterized in that: The physical constraint parameters include the engine's piston movement speed, bore size, and particle size of the powdered fuel particles.

3. The powder fuel mass flow rate prediction model based on physical information and data-driven according to claim 1, characterized in that: The RBM consists of a visible layer and a hidden layer. The weight value between any two connected neurons in the visible layer and the hidden layer is w ij , the neuron states of the visible layer and the hidden layer are vi and hj respectively, and the energy function of RBM is defined as: E(v,h)=-∑a i v i -∑b j h j -∑v i w ij h j , where a i and b j is the bias term.

4. The powder fuel mass flow rate prediction model based on physical information and data-driven according to claim 1, characterized in that: The accuracy evaluation is based on the average error. If the average error is ≤5%, it is determined that the prediction result of the nonlinear mass flow rate prediction model meets the requirements.

Images