Particle concentration prediction method based on SSA-BP micro-nano particle light scattering

Abstract

The invention discloses a particle concentration prediction method based on SSA-BP micro-nano particle light scattering, and the prediction method comprises the steps: building an estimation model of the concentration and particle size of micro-nano suspended particles in an industrial environment based on a BP neural network model; predicting the particle size and the concentration of the micro-nano particles by adopting an SSA algorithm; a random forest algorithm is used for calculating feature importance indexes of the influence factors of the extinction coefficient, the scattered light intensity, the particle size and the refractive index. The prediction precision of the particle size and the concentration of the micro-nano particles in optical detection is improved.

Description

Technical Field

[0001] The present invention relates to the field of particle concentration prediction, and in particular to a particle concentration prediction method based on SSA-BP micro-nano particle light scattering. Background Art

[0002] At present, in order to accurately predict the concentration and size of micro-nano particles, mechanistic models such as the Fredholm equation have become the mainstream solution method in neural networks. However, in order to establish an accurate model suitable for industrial environments, external environmental factors such as physical properties, chemistry and biology must be taken into account, and this must be achieved through the simulation of system structure data. Therefore, non-mechanical methods are usually used to estimate the concentration and size of micro-nano particles.

[0003] In the prior art, non-mechanical methods for estimating the concentration and size of micro-nanoparticles mainly include multi-function least squares fitting, which is used to obtain the relationship between the concentration, size and extinction coefficient of micro-nanoparticles

[134] . However, the relationship between the concentration, size and extinction coefficient of micro-nanoparticles is usually nonlinear, and the multi-function least squares fitting can only fit the linear relationship between the concentration, size and extinction coefficient of micro-nanoparticles, resulting in inaccurate concentration and size of micro-nanoparticles. Summary of the invention

[0004] In view of the above problems, the present invention is proposed to provide a particle concentration prediction method based on SSA-BP micro-nanoparticle light scattering that overcomes the above problems or at least partially solves the above problems.

[0005] According to one aspect of the present invention, a particle concentration prediction method based on SSA-BP micro-nano particle light scattering is provided, and the prediction method comprises:

[0006] Based on the BP neural network model, an estimation model for the concentration and size of micro-nano suspended particles in industrial environments is established;

[0007] The SSA algorithm is used to predict the particle size and concentration of micro-nano particles;

[0008] The feature importance index of the influencing factors of extinction coefficient, scattered light intensity, particle size and refractive index was calculated using the random forest algorithm.

[0009] Optionally, the establishment of an estimation model for the concentration and particle size of micro-nano suspended particles in an industrial environment based on a BP neural network model specifically includes:

[0010] Determining initial input information of the neural network model according to the temperature of the sample particles and the extinction coefficient at multiple wavelengths;

[0011] The initial input information is fed into the neural network model, and the target nanoparticle size is output;

[0012] The extinction coefficient, scattered light intensity, particle size and refractive index are used as input parameters to predict the concentration of nanoparticles.

[0013] Optionally, the method of using the SSA algorithm to predict the particle size and concentration of micro-nano particles specifically includes:

[0014] SSA is used to optimize the initial weight value and threshold of the BP neural network. In the prediction of the particle size of micro-nano particles, the SSA-BP model has higher accuracy in predicting the particle size of micro-nano particles than the BP model.

[0015] Optionally, the method of establishing an estimation model for the concentration and particle size of micro-nano suspended particles in an industrial environment is to characterize the light scattering parameters of micro-nano particle dust by irradiating with a weak light source in a flammable and explosive environment.

[0016] Optionally, the BP neural network model specifically includes:

[0017] Each weight and threshold The initial value of is set to a small random number;

[0018] Input training sample (I q ,d q ), 3 to 5 steps for each sample;

[0019] Calculate by inputting variable parameters to find the actual output of each network layer:

[0020] y (l) =f(s (l) )=f(w (l) x (l-1) +θ (l) ) (5-1)

[0021] Calculate the training error:

[0022] The first step starts with the output layer:

[0023]

[0024] The second step is to trace back to the hidden layer and finally reach the input layer:

[0025]

[0026] Modify weights and thresholds:

[0027]

[0028] When all samples in the sample set have gone through steps (3) to (5), a training cycle is completed and the performance index is calculated. in

[0029] If the performance index meets the accuracy requirement of E≤ε, the training ends and the next training cycle continues; ε takes an appropriate small integer and is adjusted according to the actual situation.

[0030] Optionally, the method of using the SSA algorithm to predict the particle size and concentration of micro-nano particles specifically includes:

[0031] (1) Initialize the population;

[0032] (2) Determine the fitness function and calculate the initial fitness value;

[0033] (3) Select food, sort the slugs according to their fitness values, and set the position of the slug with the highest fitness value at the top as the current food position;

[0034] (4) Determine the leader and followers. After selecting the food location, the remaining N-1 flounders in the group are sorted according to the fitness value of the flounder group. The flounders in the first half are regarded as leaders, and the flounders in the second half are regarded as followers.

[0035] (5) Update parameter c 1 ;

[0036] (6) Update of the location of the sea urchin;

[0037] (7) Calculate the fitness of the updated population;

[0038] (8) Compare the fitness value of each salp individual after the update with the fitness value of the current food. If the fitness value of the salp after the update is better than that of the food, the position of the salp with the better fitness value is used as the new food position;

[0039] (9) Repeat steps (4)-(8) until a certain number of iterations is reached and the termination condition is met, then output the current food position as the estimated position of the target.

[0040] Optionally, the random forest algorithm specifically includes:

[0041] (1) Construct N decision trees;

[0042] (2) Current decision tree k tree =1, calculate the corresponding out-of-bag data OOB k ;

[0043] (3) Calculate the current decision tree for OOB k The prediction error Error_OOB;

[0044] (4) OOB k The random perturbation of the i-th statistical feature in is marked as Calculate the current decision tree pair The prediction error

[0045] (5) For each decision tree, k tree =2,3,...,M, repeat steps (2) to (4);

[0046] (6) Finally, the characteristic importance index of the influencing factors is calculated according to the following formula;

[0047]

[0048] The present invention provides a particle concentration prediction method based on SSA-BP micro-nano particle light scattering, which includes: establishing an estimation model for the concentration and particle size of micro-nano suspended particles in an industrial environment based on a BP neural network model; using an SSA algorithm to predict the particle size and concentration of micro-nano particles; and using a random forest algorithm to calculate the characteristic importance index of factors affecting the extinction coefficient, scattered light intensity, particle size, and refractive index. This improves the prediction accuracy of optical detection of micro-nano particle size and concentration.

[0049] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0051] Figure 1 A schematic diagram of a feedforward network structure provided by an embodiment of the present invention;

[0052] Figure 2 The Hopfield neural network structure provided by the embodiment of the present invention;

[0053] Figure 3 A schematic diagram of the BP neural network structure provided by an embodiment of the present invention;

[0054] Figure 4 A schematic diagram of a neuron model provided by an embodiment of the present invention;

[0055] Figure 5 BP neural network calculation flow chart provided for the embodiment of the present invention;

[0056] Figure 6 BP neural network characteristic parameters provided by the embodiment of the present invention;

[0057] Figure 7 A flowchart of a calculation algorithm provided by an embodiment of the present invention;

[0058] Figure 8 Comparison of the prediction errors between the neural network prediction and the cubic function prediction provided by the embodiment of the present invention;

[0059] Fig. 9 A BP neural network calculation flow chart for calculating the concentration and particle size of micro-nano particles provided in an embodiment of the present invention;

[0060] Fig.10 A comparison chart of the true value and the fitted value of the training set provided by the embodiment of the present invention;

[0061] Fig.11 A comparison chart of the true value and the fitted value of the test set provided by the embodiment of the present invention;

[0062] Fig.12 A flow chart of the SSA algorithm provided by an embodiment of the present invention;

[0063] Fig.13 Comparison of the prediction errors between the SSA-BP model and the neural network prediction provided by the embodiment of the present invention;

[0064] Fig.14 An evolution diagram of the SSA algorithm provided by an embodiment of the present invention;

[0065] Fig.15 A comparison chart of the SSA-BP model training set provided by an embodiment of the present invention;

[0066] Fig.16 A comparison chart of the SSA-BP model test set provided by an embodiment of the present invention;

[0067] Fig.17 A comparison chart of the SSA-BP model test set provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0068] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0069] The terms "comprises" and "having" and any variations thereof in the description embodiments, claims and drawings of the present invention are intended to cover non-exclusive inclusions, for example, including a series of steps or units.

[0070] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0071] The present invention proposes a multi-layer hidden BP artificial neural network structure method, and constructs a neural network method for predicting the concentration and particle size of micro-nano particles with high accuracy through the selection of characteristic parameters of micro-nano particles and technical details such as multi-weight adjustment.

[0072] Artificial neural network theoretical models, including:

[0073] The artificial neural network model is a computational method based on the theory of bionics. It is a hierarchical structure established under the biological neural network. It simulates the human brain neurons as the basic unit, including the input layer, hidden layer and output layer. It is composed of a series of the most basic neuron units, and adjusts its own structure according to the changes in the outside world. Each unit is connected through the weight value. In the artificial neural network, the nodes and weight values ​​of neurons are very important concepts. Under normal conditions, the input layer and the output layer are the first layer of each layer. The input layer is the input feature vector, in which each neural unit is represented as a feature parameter, and the output layer is the output result of the entire network. If complex nonlinear mathematics is to be clearly expressed, at least one hidden layer is required. Due to the existence of the hidden layer, the neural network can achieve more complex operations. In addition, artificial neural networks have strong learning ability, the ability to obtain information from experience, generalization ability, data organization ability, fault tolerance, distributed storage and the promotion of prototyping.

[0074] Artificial neural networks have several important basic characteristics as key performance indicators, including self-learning and adaptability, nonlinearity, robustness, fault tolerance, parallel computing and distributed storage, and distributed storage. The performance of artificial neural networks is evaluated through the above performance indicators to analyze whether they are suitable for specific workplaces.

[0075] The neural network model is not only based on the back propagation (BP) model, but also other models can be constructed based on the recurrent neural network to provide more accurate decisions and more efficient results.

[0076] The basic unit of artificial neural network

[0077] Neural network is a complex information transmission system. Neurons are the basic structure of neural network. They are responsible for processing and transmitting information and can operate independently to realize complex neural network functions. The interconnection between neurons will have an important impact on the function and nature of the network. In artificial neural network system, typical network designs include feedforward network and feedback network.

[0078] (1) Feedforward network

[0079] Feedforward Neural Network (FNN), referred to as feedforward network, is a complex artificial neural network. It is composed of multiple neuron cells and has the characteristics of unidirectional multi-level structure. The direction of network information processing is layer-by-layer calculation from the input layer to the output layer. Each layer has its own unique function and can effectively process complex information. The "information processing" node is responsible for receiving external information, while the input node is responsible for passing external information to the first hidden layer, thereby realizing the effective processing of complex information. According to its information processing method, network nodes can be divided into two categories: one is responsible for obtaining information from the outside world and inputting it to the first layer, called input nodes; the second category includes each hidden layer and output layer node, which is a node with data processing capabilities. The feedforward network can be divided into multiple levels, each layer is arranged in the order of the signal, and each layer only receives input information from the neuron cells in the previous layer. There is no information feedback between the layers, thereby realizing unidirectional multi-level transmission. There is no information feedback between adjacent layers. It is a directed acyclic path, such as Figure 1 As shown, this is a schematic diagram of the feedforward network structure.

[0080] from Figure 1 As can be seen from the figure, the visible layers of the feedforward neural network include the node layers of the input layer and the output layer, the other layers are called hidden layers, and the node units connected to form the hidden layers are called hidden nodes.

[0081] Feedforward network technology has a wide range of applications. As a static nonlinear mapping, it can process complex nonlinear data through composite mapping. Its calculation accuracy can reach the level of any continuous function and square integrable variable, thus greatly improving the calculation efficiency. Feedforward network has stronger learning ability, pattern recognition ability and classification ability than feedback network.

[0082] (2) Feedback Network

[0083] Feedback neural network is a complex dynamic system. Each node in the network has the ability to process information. It can adjust the behavior of the model through the feedback relationship between input and output to achieve stable calculation results. It can achieve efficient operation in a short time, thereby improving the accuracy and reliability of the model.

[0084] Elman and Hopfield are two classic feedback neural networks that can provide accurate calculation results. The Elman network is a two-layer reverse propagation network. The model consists of an entry layer, a hidden layer, a carrier layer, and an input-output layer. The neurons are connected by the hidden layer and the input vector, and are fed back to the input part of the hidden layer. Since the input signal represents spatial information, the carrier layer can act as a one-step delay operator, giving the system a stronger memory capacity, thereby improving the efficiency and reliability of the network. Therefore, the Elman network can perform pattern recognition in both the spatial and temporal domains.

[0085] Hopfield neural network is a complex neural network model with strong associative memory power. It stores one or more stable target vectors. It can imitate the memory mechanism of biological neural network. When the stable vector is input, the target will "awaken" the network memory mode. This model has been widely used in practical engineering. This is a single-layer feedback network system. The input and output of each node will be passed to another node instead of self-feedback. Figure 2 As shown, the neural network structure has good stability and reliability, and effectively improves the efficiency and accuracy of the network system.

[0086] Basic model of artificial neural network

[0087] Based on the physical characteristics of dust particles in industrial sites, this study uses a feedforward neural network model, which can achieve that each neuron is connected once to the neurons in the previous layer. This fully connected layer divides the signal transmission into two main steps: the first step is to start from the input layer, pass through a series of hidden layers, and finally reach the output layer after multiple layers of propagation; the second step is a reverse transmission process, that is, backpropagation from the output layer, all the way back to the hidden layer, and finally enter the input layer. In this calculation process, the hidden layer will adjust the parameters of other layers according to the parameter weights and biases of the input layer. Figure 3 As shown, a typical double hidden BP network structure diagram.

[0088] In a neural network, except for the input layer, all neurons in the previous layer will make an output action to the neurons in the next layer, such as Figure 4 Shown is a schematic diagram of the neuron model.

[0089] exist Figure 4In the neuron model shown, the input of the second layer of neurons is equal to the previous layer of neurons and the weight value The sum of products, the threshold After processing, the Relu activation function f(x)=max(0,x) is used to obtain the actual output y (l) At this time, y, which is the output value of the previous layer, (l) It is also part of the input value of the next layer of neurons. According to the neural network theory, the model obtained by BP neural network training is all the thresholds and weights.

[0090] Taking BP neural network as the benchmark, it is used to solve the defect that the multi-function least squares method used in the existing technology can only fit the linear relationship between dust particle concentration, particle size and extinction coefficient, resulting in low calculation accuracy, and to achieve accurate estimation of nanoparticle concentration and particle size.

[0091] The BP neural grid design calculation process is as follows:

[0092] (1) Set each weight and threshold The initial value of is set to a small random number.

[0093] (2) Input training samples (I q ,d q ), and perform 3 to 5 steps for each sample.

[0094] (3) Calculate by inputting variable parameters to find the actual output of each network layer:

[0095] y (l) =f(s (l) )=f(w (l) x (l-1) +θ (l) ) (5-1)

[0096] (4) Calculate the training error:

[0097] The first step starts with the output layer:

[0098]

[0099] The second step is to trace back to the hidden layer and finally reach the input layer:

[0100]

[0101] (5) Modify weights and thresholds:

[0102]

[0103] (6) When all samples in the sample set have gone through steps (3) to (5), a training cycle (Epochs) is completed and the performance index is calculated. in

[0104] (7) If the performance index can meet the accuracy requirement of E ≤ ε, the training ends and the next training cycle continues. ε is an appropriate small integer and is adjusted according to the actual situation.

[0105] Based on the above calculation process, the following Figure 5 The calculation flow chart is shown.

[0106] Artificial neural network model for micro-nano particles

[0107] 5.3.1 Selection of characteristic parameters of micro-nano particles

[0108] According to the analysis of the above research, at the beginning of building a neural network model, the relationship between characteristic parameters and output is analyzed, and the key is to determine the input characteristic parameters of the model. In the field of dust explosion, the focus is on the five elements of its explosion, namely oxygen, confined space, combustible dust, dust cloud and fire source, and how to judge when the combustible dust and dust cloud reach the critical point of explosion is a more important key. As studied in this article, the light scattering method is used to study the physical properties of dynamic suspended particles, so as to accurately judge the dynamic changes of particle concentration. Based on the study of the optical properties of micro-nano suspended particles, a number of key parameters have been determined. In order to simplify the model, this article will select the parameters and summarize the factors affecting the optical detection of the physical properties of micro-nano suspended particles as follows.

[0109] According to the measurement principle of Lambert-Beer scattering method, the relationship between incident light intensity and outgoing light intensity is:

[90] :

[0110] I=I 0 exp(-K exp L) (5-6)

[0111] Where I is the intensity of scattered light, in W / cm 2 ;I 0 is the intensity of the incident light, in W / cm 2 ; K ext is the extinction coefficient; L is the thickness of the granular medium, in cm.

[0112] Assume that there are N particles of diameter D in a unit volume of the scattering medium. ball If there is a spherical micro-nano particle and a parallel monochromatic light beam is irradiated to the particle, the extinction coefficient is:

[0113]

[0114] Where σ is the light-facing area of ​​monodisperse particles, in cm 2 ;D ball is the particle size of the measured particle, in cm; N is the number of particles per unit volume, called volume concentration, in cm -3 .

[0115] The scattering and absorption of light causes the intensity of the transmitted light to attenuate, that is, extinction, which is defined as:

[0116]

[0117] in, is the transmittance; λ is the wavelength; T is the ambient temperature.

[0118] When two beams of light with different wavelengths are incident on the same particle system, we can obtain:

[0119]

[0120] The ratio of the extinction of two wavelengths is calculated as:

[0121]

[0122] The calculation equation for the particle concentration N is:

[0123]

[0124] According to the measurement principle of Lambert-Beer scattering method, it can be known that, ideally, there is a correlation between the extinction coefficient and the particle size; there is a correlation between the extinction coefficient and the particle size and concentration. However, in the actual measurement process, the ambient temperature will directly affect the hazard factor of combustible dust. For specific particles, ensuring the appropriate ambient temperature and humidity will affect the safety of the entire industrial environment. For optical measurement, the particle size and concentration are also closely related to the ambient temperature.

[0125] When the measured particles have a certain degree of transmittance, the transmission method can achieve accurate measurement of the concentration of suspended particles, thereby better controlling production safety. For example, the semiconductor particles studied in Chapter 4, or some metal particles, are suspended in a closed space. After analysis, the important calculation parameters are the refractive index n and its related characteristic parameters (such as Figure 6 shown).

[0126] The overall algorithm is divided into two BP neural network models (such as Figure 7As shown in Figure 2, the input parameters of the first BP neural network model are: ambient temperature (T), wavelength (λ); the output parameter is: particle size (D). The input parameters of the second BP neural network model are: extinction coefficient (K), scattered light intensity (I), particle size (D), refractive index (n); the output parameter is particle concentration (N).

[0127] Construction of light scattering parameter prediction model, including:

[0128] According to the demand for real-time estimation of the critical explosion concentration of micro-nano particles in industrial dust explosion hazardous environments, after repeatedly measuring the ambient temperature of micro-nano suspended particles and collecting the light intensity generated by the particles after irradiation with each incident light wavelength, the extinction coefficient of micro-nano suspended particles under various incident light wavelengths is obtained based on the BP neural network calculation.

[0129] (1) Place flammable and explosive micro-nano dust particles in a closed pipe cabin and measure the ambient temperature in the cabin as one of the input information. At the same time, determine the extinction coefficient of the micro-nano particles under multiple incident wavelengths as another initial input information of the neural network model. Alternatively, the ambient temperature in the closed annular pipe layer where the micro-nano particles are located and their extinction coefficients at multiple wavelengths can be calculated and processed, and the processed calculation results can be used as unified initial input information.

[0130] (2) In order to reduce the impact of the dimension of the input data, the input data and output data are uniformly normalized, with the goal of normalizing to [0,1]. The initial calculation information after the above normalization is input into the BP neural network model, and the normalized concentration and particle size of the explosion-related micro-nano particles are output, and then accurately denormalized to the original space; wherein, the BP neural network model is trained by obtaining the actual concentration and actual particle size of the micro-nano particles through the second input information determined according to the temperature of the micro-nano particles and the extinction coefficient at multiple wavelengths. The BP neural network model can be composed of one or more neurons, which can be used to solve complex problems.

[0131] (3) Before the initial input information is input into the BP neural network model, the model must be trained multiple times without distinction using sample data to ensure that the model can accurately express the expected results, thereby improving the performance and accuracy of the model. Therefore, we conducted a standardized production investigation on dust-related enterprises in Beijing, and selected high-tech production industrial bases related to semiconductor products to determine the second input information based on the temperature of the sample suspended micro-nano particles in the industrial environment and the extinction coefficient of the particles at multiple wavelengths. The method for determining the second input information is the same as the method for determining the initial input information.

[0132] (4) The experimental sample particles collected in the industrial environment are used as the second input information of the model, and are input into the neural network model together with important parameters such as the concentration and particle size of the dust particles in the working environment, so as to compare the difference between the concentration output by the neural network model and the actual concentration, and the difference between the particle size output by the neural network model and the actual particle size. Based on the comparison of the difference between these parameters, the BP neural network model is adjusted in key areas to achieve the goal of deeply optimizing the parameters of the neural network model.

[0133] (5) After the above work, the initial input information is input into the optimized neural network model. After multiple layers of calculation, the calculation results are output to the concentration and average particle size of the micro-nano particles in the industrial environment. The standards of the five elements of explosion are combined to make an accurate quantitative judgment on the lower limit of the dust combustion explosion hazard in the explosion-hazardous place.

[0134] The extinction coefficient and ambient temperature of micro-nano particles are used as the input of the neural network model, while the particle size and concentration of micro-nano particles are used as the output of the neural network model. By learning the nonlinear relationship between the extinction coefficient, particle size, ambient temperature and concentration of micro-nano particles, the particle size and concentration of suspended micro-nano particles can be accurately estimated.

[0135] A method for inputting initial input information into a neural network model is established, and the neural network model is improved by learning the first input information and estimates the particle size based on the initial input information. Then, the particle size and extinction coefficient, scattered light intensity and refractive index generated by the first neural network model are input into the second neural network model for learning to obtain a more accurate nonlinear relationship between the particle size, extinction coefficient, scattered light intensity and refractive index and concentration, thereby estimating the concentration of the target nanoparticles.

[0136] In order to verify the accuracy of the estimation of the neural network model, the particle size obtained by the micro-nano particle concentration and particle size estimation method is compared with the particle size obtained by the cubic function fitting model algorithm. The comparison results of the test set are shown in Table 5-1.

[0137] Table 5-1 Error analysis of particle size estimation of micro-nano particles

[0138]

[0139] Plot the relative error comparison of the two methods, such as Figure 8As shown in Table 5-1, it can be seen that the relative errors of the neural network prediction method adopted by the present invention are less than or equal to the cubic function prediction method. When the particle size is small (less than 5μm), the neural network estimation error is significantly smaller than the estimation error of the cubic function fitting model. In the case of a larger particle size, the error results of the two methods are similar, but the estimation error in this study is still smaller than the estimation error of the cubic function fitting model. In general, the concentration and particle size estimation method of micro-nano particles based on the BP neural network is more accurate than the cubic function prediction method, but the errors when the particle size is 1.16, 1.30 and 2.47 are all more than 5%, which is a large error. It will cause a large error in the subsequent prediction of particle concentration, and further optimization is required.

[0140] The present invention has established a BP neural network model that can perform preliminary self-learning of the nonlinear relationship between particle size, extinction coefficient, scattered light intensity, refractive index and concentration. At the same time, multiple dust particles generated by processing are collected in the factory, and the optical parameters of the particles are collected by light scattering method. After inputting into the model, the particle size and concentration of the particles are estimated. In the calculation process, the influence of ambient temperature on the concentration estimation result is fully considered, which not only ensures that the estimated micro-nano particle size and dust concentration have good accuracy, but also the trained model can be reused. When it is necessary to estimate the concentration and average particle size of suspended particles in the industrial environment, it is only necessary to input the corresponding input information into the trained neural network model to output the estimated concentration and average particle size, which can improve a part of the execution efficiency under specific conditions. Compared with the previous forecasting methods, the explosion hazard of some dust in the industrial production process is reduced.

[0141] On the basis of the present invention, in addition to inputting the initial input information into the neural network model and outputting the concentration and particle size of the micro-nano suspended particles, the second input information is also input into the initial neural network model, and the particle size, temperature and extinction coefficient output by the first neural network model are used as the input of the second neural network model to output the particle concentration; in addition, the loss function of the first neural network model is obtained according to the difference between the particle size output by the first neural network model and the actual result, and the loss function of the second neural network model is obtained according to the difference between the particle concentration output by the second neural network model and the actual concentration. Finally, the parameters of the complete BP neural network model are optimized based on the two. The specific process is as follows: Fig. 9 shown.

[0142] The present invention selects 103 experimental samples of multiple workstations from semiconductor production sites for neural network calculation, aiming to better simulate the changing state of dust concentration in industrial production sites. In order to improve the robustness of the BP neural network, the samples are randomly divided into a training set and a test set in a ratio of 9:1.

[0143] There are 93 samples in the training set. This summary selects 27 samples for display, including the particle size distribution, refractive index, extinction coefficient, scattered light intensity, concentration, etc. of the particles, as shown in Table 5-2.

[0144] Table 5-2 Distribution of data in the training set

[0145]

[0146]

[0147] The test set samples total 10 data, and the selection parameters are the same as those of the training set, as shown in Table 5-3.

[0148] Table 5-3 Distribution of data in the test set

[0149]

[0150]

[0151] In order to verify the rationality of sample division, the maximum, minimum, mean and variance of the training set and test set samples were calculated. From the results in Table 5-4, it can be seen that the value ranges of the five parameters in the test set are all subsets of the value ranges of the five corresponding parameters in the training set, and the mean values ​​of the five parameters in the training set and the test set are slightly different, so the sample division of the training set and the test set is reasonable.

[0152] Table 5-4 Statistical parameters of training set and test set

[0153]

[0154] Particle concentration calculation results of artificial neural network model

[0155] The parameter settings of the BP neural network are shown in Table 5-5:

[0156] Table 5-5 Statistical parameters of training set and test set

[0157]

[0158]

[0159] The present invention calculates the mean square error RMSE and the determination coefficient R of the sample 2 To indicate the accuracy of the fitting, the calculation results show that the RMSE of the training set is 1.4860, R 2 is 0.7131. The comparison results are as follows Fig.10 shown.

[0160] The RMSE of the test set is 1.3551, R 2 is 0.5818. The specific comparison chart is as follows Figure 5-11 shown.

[0161] The results of the training set and the test set show that the prediction results of the BP neural network are low. The results of the training set show that the prediction results when the particle concentration is high are worse than those when the particle concentration is low. The model falls into a local optimum during training and cannot obtain a more accurate mapping relationship when the particle concentration is low, which ultimately leads to low accuracy of the test set.

[0162] Based on the SSA-BP fast prediction of micro-nano particle size, according to the calculation results, the BP neural network falls into a local optimum and the mapping relationship is inaccurate. In order to further optimize the BP model, this section uses the Salp Squid Algorithm (SSA) to optimize the initial weight value and threshold of the BP neural network to improve the accuracy of the model.

[0163] The principle of the Salp Swarm Algorithm (SSA), salps belong to the family of Neuraminididae. Due to the need for navigation and foraging, salp groups form chains, which are called salp chains. This mechanism of salp groups helps them to move quickly and prey collaboratively in the ocean. Inspired by the intelligent behavior of salp groups, the Salp Swarm Algorithm (SSA) was proposed by Mirjalili and other scholars in 2017. Compared with traditional optimization algorithms, SSA has better convergence characteristics and stability, and the algorithm principle is simple, with fewer adjustment parameters and easy implementation. Many scholars have applied it to photovoltaic system optimization, feature number selection, robot motion control and other fields, and achieved good results.

[0164] The salp colony consists of two parts: the leader and the followers. The leader plays a leading role and gradually approaches the food source (i.e. the optimal solution) in the n-dimensional space by actively searching. The formula is as follows:

[0165]

[0166] Where x represents the position of the leader in the j-dimensional space; F j is the location of the food source in the j-th dimension; ub and lb represent the upper and lower bounds of the j-th dimension, respectively; c 2 and c 3 is a random number uniformly distributed in [0,1]; the coefficient c 1 It is the parameter of the SSA algorithm to balance exploration and exploitation, and its formula is as follows:

[0167]

[0168] Where l represents the current number of iterations, and L represents the maximum number of iterations of the algorithm. The follower position update method follows Newton's motion theorem:

[0169]

[0170] When t=1, set the initial speed v 0 =0, then:

[0171]

[0172] The follower's position update can be expressed as:

[0173]

[0174] The algorithm steps of SSA are as follows, and the algorithm flow chart is as follows Fig.12 shown.

[0175] (1) Initialize the population;

[0176] (2) Determine the fitness function and calculate the initial fitness value;

[0177] (3) Food selection. Since the location of the target (i.e., food) is not clear during actual positioning, the slugs are sorted according to their fitness values, and the location of the slug with the highest fitness value is set as the current food location;

[0178] (4) Determine the leader and followers. After the food location is selected, the remaining N-1 thrips in the group are sorted according to the fitness value of the thrips group. The first half of the thrips are regarded as leaders, and the second half of the thrips are regarded as followers.

[0179] (5) Update parameter c 1 ;

[0180] (6) Update the position of the leader Zun Haishao. Use formula (5-13) to update the position of the leader Zun Haishao, and use formula (5-15) to update the position of the follower Zun Haishao;

[0181] (7) Calculate the fitness value. Calculate the fitness of the updated population;

[0182] (8) Update the location of the food. Compare the fitness value of each salp after the update with the fitness value of the current food. If the fitness value of the salp after the update is better than that of the food, the location of the salp with the better fitness value is used as the new location of the food;

[0183] (9) Repeat steps (4)-(8) until a certain number of iterations is reached and the termination condition is met, then output the current food position as the estimated position of the target.

[0184] 5.4.2 SSA-BP algorithm for rapid prediction of micro-nano particle size

[0185] (1) SSA-BP principle

[0186] When BP neural network is used to solve complex problems, the accuracy of the solution is not ideal. Its initial weight value and threshold have a great influence on the accuracy of the model. Improper selection of weight value and threshold value can easily make the network result in a local optimum and prone to overfitting. Therefore, this paper uses the SSA algorithm to optimize the initial weight value and threshold of the BP neural network, and selects the optimal weight value and threshold combination as the initial value of the BP network to achieve the purpose of improving accuracy. In the optimization process, the mean square error of the output value of the training sample is used as the fitness function. The specific expression is as follows:

[0187]

[0188] Where: Y i is the true value of the concentration and particle size of micro-nano particles, is the fitting value of the concentration and particle size of micro-nano particles, and n is the number of samples.

[0189] (2) SSA-BP algorithm quickly predicts the particle size of micro-nano particles

[0190] First, the SSA-BP model was used to predict the particle size and compared with the prediction results of the BP model. The prediction results are shown in Table 5-6.

[0191] Table 5-6 Comparison of error in particle size estimation of nanoparticles of different models

[0192]

[0193] By the results ( Fig.13 ) It can be seen that the SSA-BP model has a higher accuracy in predicting the particle size of nanoparticles than the BP model, both of which are less than 3%, so SSA-BP can accurately predict the particle size of nanoparticles.

[0194] The SSA-BP model is then trained to predict particle concentration using the extinction coefficient, scattered light intensity, particle size, and refractive index as input.

[0195] Set the evolutionary number of the Pin Hai Shao algorithm to 50, the number of Pin Hai Shao population to 20, the range of the initial weight value and threshold to [-3,3], and the optimization results are as follows:

[0196] Depend on Fig.14 It can be seen that the model reached convergence in the 33rd generation, and the fitness value (i.e., training set RMSE) converged to 0.6825.

[0197] The comparison diagrams of the actual and predicted values ​​of the particle concentrations in the training set and the test set are shown in Figure 2. Fig.15 and Fig.16 As shown, the RMSE of the training set is 0.6825 and the determination coefficient R 2The RMSE of the test set is 0.2434, and the coefficient of determination R 2 It is 0.9865, indicating that the BP neural network model optimized by the SSA algorithm has good robustness and can map the complex relationship between particle concentration and extinction coefficient, scattered light intensity, particle size and refractive index.

[0198] Calculation of the importance of light scattering factors based on the RF algorithm In order to further study the relationship between the extinction coefficient, scattered light intensity, particle size and refractive index and the concentration of micro-nano particles, this section explores the correlation between the above four influencing factors and the concentration of micro-nano particles. The characteristic importance index (PIM) of the influencing factors is calculated by the random forest (RF) algorithm in machine learning to judge the sensitivity of the concentration of micro-nano particles to the above four influencing factors.

[0199] The random forest (RF) algorithm is a learning algorithm based on decision trees, which is usually used to solve regression and classification problems. The algorithm is mainly composed of multiple non-intersecting decision trees. During the calculation process, a random algorithm is used to select samples and calculate eigenvalues ​​for the decision trees, and a separate judgment is made on the learning. The final result is obtained by voting through the calculation results of multiple decision trees. We call it ensemble learning. The operation mode of the random forest algorithm can be summarized as multi-channel calculation, and the final answer is obtained after aggregation.

[0200] Explain the calculation process of random forest in a more intuitive way. Multiple decision trees will be generated in the calculation, and each decision tree is a weak classifier with average performance. For an experimental sample, there will be as many classification results as there are decision trees, and each classifier is a calculation result generated by regression. Usually, the Bagging algorithm is used for "random sampling" to obtain relatively independent training subsets, and the training subsets are used to generate base classifiers to obtain results.

[0201] The advantage of the random forest algorithm model is that it does not require the function model to be set in advance, but effectively realizes the requirements of simple calculation, few parameters, and strong anti-overfitting ability. There are many factors affecting the concentration of combustible suspended dust particles in dust explosion hazardous places, especially the influence of dust concentration on parameters such as extinction coefficient, refractive index, particle size and scattered light intensity. This summary will use the random forest model to identify the important indexes of influencing factors and improve the calculation accuracy of the micro-nano particle light scattering parameter prediction model.

[0202] The specific steps are as follows:

[0203] (1) Construct N decision trees;

[0204] (2) Current decision tree k tree =1, calculate the corresponding out-of-bag data OOBk ;

[0205] (3) Calculate the current decision tree for OOB k The prediction error Error_OOB k ;

[0206] (4) OOB k The random perturbation of the i-th statistical feature in is marked as OOB i k , calculate the current decision tree The prediction error

[0207] (5) For each decision tree, k tree =2,3,...,M, repeat steps (2) to (4);

[0208] (6) Finally, the characteristic importance index of the influencing factors is calculated according to the following formula.

[0209]

[0210] Bring 103 samples into the RF model calculation, and the importance of each influencing factor is ranked as follows Fig.17 shown.

[0211] From the results, we can see that the most important characteristic index of refractive index is 0.74, followed by scattered light intensity and particle size at 0.6, and the extinction coefficient is the least important at only 0.2.

[0212] Beneficial effects: The machine learning method is used to conduct non-contact and rapid prediction of the concentration of micro-nanoparticles detected optically in industrial production.

[0213] Firstly, based on the BP neural network model, an estimation model for the concentration and size of micro-nano suspended particles in industrial environments was established, including determining the initial input information of the neural network model according to the temperature of the sample particles and the extinction coefficient at multiple wavelengths; inputting the initial input information into the neural network model to output the target nanoparticle size; and then using the extinction coefficient, scattered light intensity, particle size and refractive index as input parameters to predict the concentration of nanoparticles. In the nanoparticle size prediction model, the concentration and size estimation method of micro-nano particles based on the BP neural network is more accurate than the cubic function prediction method, but the errors when the particle size is 1.16μm, 1.30μm and 2.47μm are all more than 5%, which is a large error, which will cause a large error in the subsequent prediction of particle concentration and also affect the prediction accuracy of particle concentration.

[0214] Secondly, in order to improve the prediction accuracy of the particle size and concentration of micro-nano particles in optical detection, the SSA algorithm was introduced into the prediction of the particle size and concentration of micro-nano particles. SSA was used to optimize the initial weight value and threshold of the BP neural network to improve the accuracy of the model. In the prediction of the particle size of micro-nano particles, the SSA-BP model has a higher accuracy than the BP model in predicting the particle size of micro-nano particles, and the error is less than 3%. In the prediction of the concentration of micro-nano particles, the SSA-BP model has a high accuracy, with the RMSE of the training set being 0.6825 and the determination coefficient R 2 is 0.9624, the RMSE of the test set is 0.2434, and the coefficient of determination R 2 It is 0.9865, indicating that the BP neural network model optimized by the SSA algorithm has good robustness and can well map the complex relationship between micro-nano particle concentration and extinction coefficient, scattered light intensity, particle size and refractive index.

[0215] Finally, in order to further study the relationship between the extinction coefficient, scattered light intensity, particle size and refractive index and the concentration of micro-nano particles, the random forest algorithm was used to calculate the characteristic importance index of these four influencing factors. The results showed that the refractive index had the highest characteristic index of 0.74, followed by scattered light intensity and particle size at 0.6, and the extinction coefficient had the lowest importance at only 0.2. This achieved the rapid characterization of the light scattering parameters of micro-nano particle dust using weak light sources in flammable and explosive environments.

[0216] The above specific implementation methods further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above are only specific implementation methods of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A particle concentration prediction method based on SSA-BP micro-nanoparticle light scattering, characterized in that: The prediction method comprises: Based on the BP neural network model, an estimation model for the concentration and size of micro-nano suspended particles in industrial environments is established; The SSA algorithm is used to predict the particle size and concentration of micro-nano particles; The feature importance index of the influencing factors of extinction coefficient, scattered light intensity, particle size and refractive index was calculated using the random forest algorithm.

2. The particle concentration prediction method based on SSA-BP micro-nano particle light scattering according to claim 1, characterized in that: The estimation model of the concentration and particle size of micro-nano suspended particles in industrial environments based on the BP neural network model specifically includes: Determining initial input information of the neural network model according to the temperature of the sample particles and the extinction coefficient at multiple wavelengths; The initial input information is fed into the neural network model, and the target nanoparticle size is output; The extinction coefficient, scattered light intensity, particle size and refractive index are used as input parameters to predict the concentration of nanoparticles.

3. The particle concentration prediction method based on SSA-BP micro-nano particle light scattering according to claim 1, characterized in that: The use of the SSA algorithm to predict the particle size and concentration of micro-nano particles specifically includes: SSA is used to optimize the initial weight value and threshold of the BP neural network. In the prediction of the particle size of micro-nano particles, the SSA-BP model has higher accuracy in predicting the particle size of micro-nano particles than the BP model.

4. The particle concentration prediction method based on SSA-BP micro-nano particle light scattering according to claim 1, characterized in that: Said The light scattering parameters of micro-nano particle dust are characterized by using weak light source in flammable and explosive environments.

5. The particle concentration prediction method based on SSA-BP micro-nano particle light scattering according to claim 1, characterized in that: The BP neural network model specifically includes: Each weight and threshold The initial value of is set to a small random number; Input training sample (I q ,d q ), 3 to 5 steps for each sample; Calculate by inputting variable parameters to find the actual output of each network layer: y (l) =f(s (l) )=f(w (l) x (l-1) +θ (l) ) (5-1) Calculate the training error: The first step starts with the output layer: The second step is to trace back to the hidden layer and finally reach the input layer: Modify weights and thresholds: When all samples in the sample set have gone through steps (3) to (5), a training cycle is completed and the performance index is calculated. in If the performance index meets the accuracy requirement of E≤ε, the training ends and the next training cycle continues; ε takes an appropriate small integer and is adjusted according to the actual situation.

6. The particle concentration prediction method based on SSA-BP micro-nano particle light scattering according to claim 1, characterized in that: The use of the SSA algorithm to predict the particle size and concentration of micro-nano particles specifically includes: (1) Initialize the population; (2) Determine the fitness function and calculate the initial fitness value; (3) Select food, sort the slugs according to their fitness values, and set the position of the slug with the highest fitness value at the top as the current food position; (4) Determine the leader and followers. After selecting the food location, the remaining N-1 flounders in the group are sorted according to the fitness value of the flounder group. The flounders in the first half are regarded as leaders, and the flounders in the second half are regarded as followers. (5) Update parameter c1; (6) Update of the location of the sea urchin; (7) Calculate the fitness of the updated population; (8) Compare the fitness value of each salp individual after the update with the fitness value of the current food. If the fitness value of the salp after the update is better than that of the food, the position of the salp with the better fitness value is used as the new food position; (9) Repeat steps (4)-(8) until a certain number of iterations is reached and the termination condition is met, then output the current food position as the estimated position of the target.

7. The particle concentration prediction method based on SSA-BP micro-nanoparticle light scattering according to claim 1, characterized in that: The random forest algorithm specifically includes: (1) Construct N decision trees; (2) Current decision tree k tree =1, calculate the corresponding out-of-bag data OOB k ; (3) Calculate the current decision tree for OOB k The prediction error Error_OOB; (4) OOB k The random perturbation of the i-th statistical feature in is marked as Calculate the current decision tree The prediction error (5) For each decision tree, k tree =2,3,...,M, repeat steps (2) to (4); (6) Finally, the characteristic importance index of the influencing factors is calculated according to the following formula;

Images