A rare earth electrolytic cell state multi-parameter monitoring method based on fused salt image features

Abstract

The application discloses a rare earth electrolytic cell state multi-parameter monitoring method based on fused salt image features, according to the monitoring requirement of the rare earth electrolytic cell operation state, the rare earth electrolytic cell state monitoring is divided into rare earth fused salt temperature monitoring and rare earth fused salt reaction state monitoring; a nonlinear mapping model of the rare earth fused salt temperature is constructed through a GA-BP model to realize the rare earth fused salt temperature monitoring; the intensity of the rare earth fused salt reaction is quantified through a dense optical flow method Farneback, a scale transformation and mean filtering, and a nonlinear mapping model between the rare earth fused salt motion field characteristics and the rare earth fused salt reaction state is established through a gated recurrent unit neural network GRU to realize the rare earth fused salt reaction state monitoring. The above method combines image processing, artificial intelligence and data visualization technology to realize the intelligent monitoring of the rare earth electrolytic cell operation state based on machine vision.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of rare earth electrolytic cell, and particularly relates to a rare earth electrolytic cell state multi-parameter monitoring method based on molten salt image features. BACKGROUND

[0002] In current industrial production, rare earth products are mainly obtained by molten salt electrolysis method. This method mainly converts electrical energy into chemical energy to electrolyze rare earth oxides, so as to separate and purify rare earth metals. The rare earth electrolytic molten salt system mainly includes a chlorinated rare earth molten salt system (a two-component system) and a fluorinated rare earth molten salt system (a three-component system). With the continuous development of the rare earth industry, the chlorinated rare earth molten salt system has been gradually eliminated, and the fluorinated rare earth molten salt system has become the mainstream process of rare earth electrolysis. The rare earth electrolytic cell completes a complex physical, chemical and dynamic production process under harsh conditions such as high temperature, strong light and strong corrosion, and the electrolysis reaction process is still a "black box" model. With the continuous development of sensing technology, people continue to use existing detection technologies to judge and predict various parameters and reaction states in the rare earth electrolytic cell according to actual production experience. In the electrolysis environment, many detection devices are easily corroded, and front-line technical personnel are also faced with serious safety hazards.

[0003] The monitoring targets of the rare earth electrolytic cell state in the prior art mainly include current, voltage, molten salt temperature and molten salt reaction state (whether the molten salt chemical reaction is sufficient). The current and voltage data can be directly obtained from the power input end, while the molten salt temperature and the molten salt reaction state need to be monitored by technical personnel on site. Among them, technical personnel often use infrared thermometers to obtain the rare earth molten salt temperature and use expert experience to judge the rare earth molten salt reaction state. However, these two types of monitoring methods have defects such as high cost, great safety hazards, non-intuitive monitoring information, lack of data accumulation and data traceability, etc., which make it difficult for enterprises to obtain closed-loop regulation feedback, difficult to ensure that the electrolytic cell runs in a high-yield and low-consumption state for a long time, and difficult to improve the production safety factor. Therefore, seeking an efficient and low-cost rare earth electrolytic cell state monitoring method has important guiding significance for optimizing the rare earth electrolysis process. SUMMARY

[0004] The purpose of the present application is to provide a rare earth electrolytic cell state multi-parameter monitoring method based on molten salt image features. This method combines image processing, artificial intelligence and data visualization technology to intelligently monitor the running state of the rare earth electrolytic cell based on machine vision, provides important closed-loop feedback parameters for online control of the rare earth electrolytic cell, and provides an important reference for improving the production safety factor and realizing unmanned factories.

[0005] The purpose of the present application is achieved by the following technical solutions:

[0006] A rare earth electrolytic cell state multi-parameter monitoring method based on molten salt image features, the method comprises:

[0007] Step 1, according to the monitoring requirement of the rare earth electrolytic cell running state, the rare earth electrolytic cell state monitoring is divided into rare earth molten salt temperature monitoring and rare earth molten salt reaction state monitoring;

[0008] Step 2, a nonlinear mapping model of rare earth molten salt temperature is constructed by GA-BP model to realize rare earth molten salt temperature monitoring;

[0009] Step 3, the intensity of rare earth molten salt reaction is quantified by dense optical flow method Farneback, scale transformation and mean filtering, and a nonlinear mapping model between rare earth molten salt motion field features and rare earth molten salt reaction state is established by gated recurrent unit neural network GRU to realize rare earth molten salt reaction state monitoring.

[0010] It can be seen from the above technical solution provided by the present application that the above method combines image processing, artificial intelligence and data visualization technology to realize intelligent monitoring of the running state of the rare earth electrolytic cell based on machine vision, provides important closed-loop feedback parameters for online control of the rare earth electrolytic cell, and provides important reference for improving the production safety factor and realizing unmanned factory. BRIEF DESCRIPTION OF DRAWINGS

[0011] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0012] Figure 1 A rare earth electrolytic cell state multi-parameter monitoring method based on molten salt image features is provided for the embodiments of the present application;

[0013] Figure 2 The displacement statistical histograms of the two types of motion fields of sufficient reaction and insufficient reaction are compared for the embodiments of the present application;

[0014] Figure 3 The result diagram of rare earth molten salt cathode rod profile fitting is described for the embodiments of the present application;

[0015] Figure 4 The comparison diagram of normalized velocity features before and after median filtering is described for the embodiments of the present application;

[0016] Figure 5 The diagram of model prediction result is described for the embodiments of the present application. DETAILED DESCRIPTION

[0017] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments of the present application, which do not constitute a limitation of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative efforts fall within the protection scope of the present application.

[0018] As Figure 1 Fig. 1 is a flowchart of a method for monitoring a rare earth electrolytic cell state based on molten salt image features according to an embodiment of the present application. The method comprises the following steps:

[0019] Step 1: According to the monitoring requirements of the rare earth electrolytic cell state, the rare earth electrolytic cell state monitoring is divided into rare earth molten salt temperature monitoring and rare earth molten salt reaction state monitoring.

[0020] Step 2: A nonlinear mapping model of the rare earth molten salt temperature is constructed by using a GA-BP model to realize the rare earth molten salt temperature monitoring.

[0021] In this step, due to the defect that the traditional colorimetric formula is easily affected by the ambient light, the present application is based on the colorimetric features, and auxiliary features are constructed to improve the mapping ability of the model. First, the colorimetric features are extracted as the basic variables by using the colorimetric formula (1);

[0022]

[0023] Then, several types of auxiliary features are preliminarily established by consulting a large number of literatures: HSI (Hue, Saturation and Intensity), RR (Relative Red), CVA (Color Vector Angle) and YUV color features. HSI compensates for the defect of less image information in the RGB color space by describing the hue, brightness and saturation information of the object, as shown in formula 2; RR can effectively eliminate the influence of light and improve the robustness of the prediction model, as shown in formula 5; CVA is used to represent the visual difference between different colors, as shown in formula 4; and the YUV color model conforms to the feature that the human vision is more sensitive to brightness information, as shown in formula 3:

[0024]

[0025]

[0026]

[0027]

[0028] Wherein, Y represents the brightness of the pixel; U represents the difference between the red component and the brightness; V represents the difference between the blue component and the brightness; K is a correction coefficient, which is related to the camera parameters, exposure time and the emissivity of the object; when the measured object and the hardware and software parameters are determined, the mapping relationship of K, R / G and T can be theoretically established, and the correction coefficient K can be obtained by experiment calibration; H, S and I represent the hue, brightness and saturation, respectively; R, G and B represent the gray values of the red, green and blue channels, respectively; Grey represents the gray image; RR represents the relative red feature; CVA represents the color vector angle feature;

[0029] Based on the limitation of the colorimetric formula, a nonlinear mapping model of rare earth molten salt temperature is constructed by a genetic algorithm GA-BP neural network model, in the field of nonlinear mapping model, compared with polynomial regression, logistic regression and support vector machine model, BP neural network has better nonlinear mapping ability, the specific process is as follows:

[0030] The BP neural network model is composed of input layer, hidden layer and output layer, the input layer is determined by the dimension of independent variable (feature vector), and the output layer is determined by the dimension of dependent variable, the number of input layer nodes of the model is 8, the number of output layer nodes is 6, and the number of hidden layer nodes is determined according to the empirical formula 6:

[0031]

[0032] I and O represent the number of input and output nodes respectively; according to formula 6, the value range of hidden layer node N is (1, 13); the model is composed of input layer, hidden layer and output layer nodes connected together, the calculation process of the model mainly consists of weight forward propagation and error back propagation, and the calculation formula is as follows:

[0033]

[0034]

[0035] Wherein, h i represents the i-th neuron of the hidden layer; w ti represents the connection weight between the t-th temperature input feature and the i-th hidden layer neuron; b ti represents the bias value of the t-th temperature input feature and the i-th hidden layer neuron; w it represents the connection weight between the i-th hidden layer neuron and the t-th output; b it represents the bias value of the i-th hidden layer neuron and the t-th output; N represents the number of hidden layer nodes; K1 represents the number of input nodes; K2 represents the number of output nodes; I t represents the t-th temperature input feature; O tdenotes the tth output node; ReLu and σ are both nonlinear activation functions, whose expressions are shown in Equations 9 and 10, respectively:

[0036]

[0037]

[0038] In order to improve the prediction accuracy of the nonlinear mapping model, the genetic algorithm GA is introduced to optimize the hidden layer node number, the initial connection weight parameter and the initial bias parameter in the BP neural network model. The GA optimizes and solves according to the optimization parameters provided by the BP neural network model, the parameter bounds and the optimization function. The optimization process includes the processes of initializing the population, calculating the fitness function, selection, crossover and mutation, etc. Assuming that the total number of optimization parameters is N, the calculation method is as follows:

[0039] N = wnum1 + bnum1 + wnum2 + bnum2 + hiddenlayersize (11)

[0040] wnum1 = inputnum * hiddenlayersize, wnum2 = hiddenlayersize * outputnum (12)

[0041] bnum1 = hiddenlayersize, bnum2 = outputnum (13)

[0042] In the above formula, wnum1 represents the number of connection weights between the input layer and the hidden layer; bnum1 represents the number of bias weight between the input layer and the hidden layer; wnum2 represents the number of connection weights between the hidden layer and the output layer; bnum2 represents the bias weight parameter between the hidden layer and the output layer; hiddenlayersize represents the number of hidden layer nodes; inputnum represents the number of input layer nodes; outputnum represents the number of output layer nodes; the theoretical bounds of each weight parameter are (-1, 1); in order to expand the search range, the bounds of all weight parameters are set to (-3, 3) in the experiment.

[0043] It is worth noting that, due to the upper and lower limits of the number of hidden layer nodes (positive integers determined by the empirical formula) being different from the upper and lower limits of the weight parameters, the experiment refers to the greedy strategy, and after optimizing the number of hidden layer nodes, the other weight parameters are optimized. Among them, the optimization objective function of the genetic algorithm adopts the mean square error function, as shown in Equation 14. In addition, the recognition rate of the model is set to 0.001, and the maximum number of iterations is 3000. The population size of the optimizer is set to 30, and the maximum number of iterations is set to 100.

[0044]

[0045] where MES represents mean square error; n represents the length of sample data; Y i represents the true value of the i-th data label; Y i represents the predicted value of the i-th data label.

[0046] The image brightness range of the electrolytic cell is the entire furnace table. To ensure the consistency of the data as much as possible, the experiment needs to take the maximum value after intercepting the same size pixel block in the center of the molten salt. The steps of taking the same pixel block include: image binarization, extracting the molten salt contour, calculating the contour center point, intercepting two pixel blocks of the same size, and merging the pixel blocks. In order to compare the prediction results with the classical image recognition algorithm, the size of the pixel block refers to the minimum size (16x16) commonly used in image feature extraction network. In order to maintain the model input dimension unchanged while taking into account the image information of the left and right half contours of the electrolytic cell port, the experiment merges the two pixel blocks by the mean algorithm.

[0047] In the specific implementation, in order to solve the problem of large fluctuation of image sampling data, the mean filter is set to filter the experimental data. Specifically, the maximum value of each channel of the pixel block intercepted in the previous section is taken to approximate the peak value of the camera response curve, and then the maximum value sampling data (hereinafter referred to as sampling data) of the same temperature segment is filtered. The filter width interval is 10 frames, and the maximum width value is the maximum frame rate value of the color industrial camera. With the increase of the filter width, the variance of the sampling data will decrease, which shows that the mean filter can effectively reduce the fluctuation value of the sampling data in the same temperature segment. In order to reduce the fluctuation of the data as much as possible and reduce the sampling time, the maximum frame rate of the color industrial camera is taken as the filter width in this embodiment.

[0048] Among them, since the rare earth electrolytic cell has a certain temperature inertia, that is, the fluctuation of the molten salt temperature in a short time will be maintained in a relatively stable range. When the temperature label is obtained by the thermocouple, it is found that the temperature fluctuation of the electrolytic cell in a short time is about 10℃. Since the normal working temperature of the electrolytic cell is about 1060-1100℃, the temperature soft measurement target is divided into the following temperature segments:

[0049] <1060℃, 1060-1070℃, 1070-1080℃, 1080-1090℃, 1090-1100℃, and >1100℃.

[0050] In a specific implementation, 23490 samples of data are collected for the above six temperature ranges. It can be seen from the sampling data that the mean filter can effectively filter out the influence of sampling data fluctuations. Overall, the filtered sampling data has a relatively obvious distinction. Among them, the red channel meets the theoretical requirement that the higher the temperature, the greater the gray response value; while the blue-green channel cannot meet this rule. The reason may be that the main color of the temperature image of the electrolytic cell is red, so the anti-interference ability of the red channel is relatively strong, while the other two channels will be disturbed by the environment light and the complex working environment to some extent. This is also an important reason for the larger error of colorimetric method in practical application.

[0051] In addition, through a large number of experiments, it is found that the nonlinear mapping model constructed by the GA-BP algorithm of the present application has good prediction performance. For example, the data set is divided into a training set and a validation set in a ratio of 8:2, the input dimension of the data is 8 (basic features + auxiliary features), and the output dimension is 6. After training the model with the training set, the converged model is used to predict and verify the validation set, and the results of K nearest neighbor method (KNN), support vector machine (SVM) and their optimized models are compared, wherein:

[0052] The distance of K nearest neighbor method is Euclidean distance, and the optimization parameter is K, and the optimization range is [1, 200]; the support vector machine kernel function adopts Gaussian kernel, and the optimization parameters are BoxConstraint and KernelScale, and the optimization range is [0.001, 1000]. BoxConstraint is related to the number of support vector machines. The more the number of support vector machines, the better the nonlinear fitting ability of the model, but the time cost of calculating the model will also increase. KernelScale is the geometric progression of the sigma parameter in the Gaussian kernel function, which is used to adjust the scaling ratio of the kernel function scale. The optimization model adopts the classical global optimization algorithm-genetic algorithm, and the population size of the genetic algorithm is set to 30, and the maximum iteration number is 100. The solution results of each model are shown in Table 1:

[0053] Table 1 Comparison of model prediction accuracy

[0054]

[0055]

[0056] From the results in Table 1 above, it can be seen that before optimization, the performance of the BPNN model is the best, and the training set accuracy and the prediction set accuracy both reach the enterprise prediction accuracy requirement (greater than or equal to 90%). After optimization by genetic algorithm, the performance of each model has been significantly improved, among which the GA-BP model described in the present application has the best performance, and the training set prediction accuracy and the test set prediction accuracy are 98.75% and 98.41% respectively.

[0057] Step 3, quantifying the intensity of the rare earth molten salt reaction by the dense optical flow method Farneback, scale transformation and mean filter, and establishing a nonlinear mapping model between the rare earth molten salt motion field characteristics and the rare earth molten salt reaction state by the Gated recurrent unit (GRU), to realize the monitoring of the rare earth molten salt reaction state.

[0058] In this step, the optical flow method is mainly divided into dense optical flow method and sparse optical flow method, the dense optical flow method needs to solve the optical flow of all pixels of the image, while the sparse optical flow method only needs to solve the optical flow of part of the image, so the calculation result of the dense optical flow method is more accurate. The classical dense optical flow method has Horn-Schunck and Farneback. The Horn-Schunck optical flow method derives the optical flow iteration formula by introducing global smoothing constraint on the basis of the basic equation of optical flow. The Farneback optical flow method derives a new velocity field 1 estimation equation based on polynomial expansion. Therefore, first, the rare earth molten salt velocity field characteristics are extracted by the Farneback optical flow method, and the image field information expression is:

[0059] f(x)=x T Ax+b T x+c (15)

[0060] If the field information expression of the previous frame image is:

[0061] f1(x)=x T A1x+b1 T x+c1 (16)

[0062] Assume that the next frame image has a displacement d, based on the basic assumption of optical flow (image gradient constant and local optical flow constant), its image field information is expressed as:

[0063]

[0064] According to the assumption, the above polynomial coefficients are equal:

[0065]

[0066] From the above formula, the image displacement calculation formula is:

[0067]

[0068] Although the above formula has theoretically derived the formula for calculating the displacement between two frames of images, in practical solutions, using a single polynomial model to approximate complex image information often leads to significant calculation errors. To solve this problem, a local polynomial is introduced to replace the global polynomial in formula (17). The two images are expanded using local polynomials to obtain the expansion coefficients A1(x), b1(x), c1(x) and A2(x), b2(x), c2(x) for the two images respectively. Ideally, based on the optical flow assumption, A2 = A1 should be satisfied, but in practice, the following approximation is required to reduce the error:

[0069]

[0070] In addition, Δb(x) is introduced:

[0071]

[0072] At this point, the image displacement calculation formula is updated from (21) to:

[0073] A(x)d(x)=Δb(x) (22)

[0074] Since calculating each pixel using formula (24) would result in a huge computational load, in order to minimize the neighborhood range of image computation as much as possible, assuming that the image displacement changes slowly enough, then d(x) needs to satisfy the following function minimization objective:

[0075]

[0076] In the formula, w(Δx) is the weight parameter corresponding to the image pixel. The displacement calculation formula obtained by the least squares method is:

[0077] d(x)=(∑wA T A) -1 ∑wA T Δb (24)

[0078] To improve the robustness of the model, the 8-parameter 2D motion parameterization model is represented as follows:

[0079]

[0080] The above formula can be expressed in matrix form as follows:

[0081] d = Sp (26)

[0082]

[0083] p = (a1, a2, a3, a4, a5, a6, a7, a8) T (28)

[0084] Substituting the 8-parameter 2D motion parameterization model into formula (25), we obtain the new weighted least squares expression as follows:

[0085]

[0086] Solving using the least squares method, the parameterized model is expressed as:

[0087]

[0088] In solving the parametric model, first calculate separately A i S i and Δb i Then, the displacement d is calculated using a weighted average. At this point, the model can calculate the displacement change between two frames relatively well. However, when the displacement changes significantly, the model still introduces a large error. Generally, the displacement change between two frames in industrial cameras with high frame rates is relatively small. To overcome the error caused by large displacements, prior displacement can be introduced to update the model, and image pyramid techniques can be used to construct multi-resolution input images.

[0089] To more intuitively reflect the intensity of the electrochemical reaction, this paper statistically analyzes the displacement of the motion field, such as... Figure 2 The figure shown is a displacement statistical histogram comparing two types of motion fields with sufficient and insufficient responses, as described in an embodiment of the present invention. Figure 2 As can be seen, the maximum displacement of a fully reacted molten salt is greater than that of an incompletely reacted salt. This verifies that a fully reacted molten salt exhibits more vigorous physical characteristics, which is reflected in the larger image displacement. Figure 2 Therefore, the experiment extracts the maximum displacement of each motion field as a feature. Assuming the industrial camera has a frame rate of N, it can capture N images per second, while the experiment can acquire N-1 motion fields. Considering both time and computational costs, it was found that setting the sampling time to 1 second effectively reduces computational cost and experimental time while still obtaining sufficient information to distinguish between the two types of motion fields. Therefore, this experiment can acquire N-1 feature values ​​in a single sampling.

[0090] Since the optical flow calculation is performed in pixels, the experiment needs to find a conversion factor to convert the image scale to the actual physical field scale, thereby normalizing the molten salt motion field obtained from different electrolytic cells to the same scale range. Because the cathode rod in rare earth electrolysis has a fixed diameter and is installed vertically, the ratio of the cathode rod diameter in the image to the actual cathode rod diameter is used as the normalization conversion factor. First, the molten salt region is separated by threshold segmentation, then the contour of the molten salt region is extracted, and finally, two parallel lines with the most points distributed in the vertical direction are fitted using the contours; for example...Figure 3 The diagram shown is a schematic diagram of the result of fitting the rare earth molten salt cathode rod profile according to an embodiment of the present invention, wherein the two approximate semicircles are the profile of the molten salt region, and the two parallel vertical lines are the profile of the cathode rod fitting.

[0091] Let displacement d represent the image displacement obtained by the Farneback optical flow method. n D represents the distance between the fitted profiles of the cathode rods. t Let d represent the actual diameter of the cathode rod (85 mm), then the normalized image displacement d t The expression is:

[0092]

[0093] Assuming the frame rate of the image is N fps, then the time interval between two frames is 1 / N s. Therefore, the expression for the normalized speed v is:

[0094]

[0095] At this point, the normalized velocity features in the rare earth molten salt video stream can be extracted;

[0096] Because the camera is affected by pulse noise from the current and random environmental noise during imaging, median filtering is required for the data. Since the sampling time in the experiment is 1 second, calculations show that N-1 (N = 90 fps in this paper) normalized velocity values ​​can be obtained within 1 second. Because the sampling time is sufficiently small, the order of the median filter is set to N-1; for example... Figure 4 The image shown is a comparison of the normalized velocity features described in this embodiment of the invention before and after median filtering. Figure 4 As can be seen, the median filter can effectively distinguish between the two types of signals, laying the foundation for the neural network to identify signals in the following text.

[0097] The gated recurrent unit (GRU) neural network consists of reset gates and update gates, and its computation process is as follows:

[0098] The input value X of the current cell t and the state h output by the previous unit t-1 The inputs to the reset and update gates are composed of nonlinear combinations:

[0099] r t =σ(W t ·[h t-1 ,x t ]+b r (33)

[0100] z t =σ(W z ·[h t-1 ,xt ]+b z (34)

[0101] In the formula, r t Indicates the output of the reset gate; z t Indicates updating the gate's output; W t and W z These represent the weight parameters for the reset gate and the update gate, respectively; b r and b z represents the bias matrices of the reset gate and update gate, respectively; · represents matrix dot product; [] represents matrix concatenation; σ represents the nonlinear activation function sigmoid, expressed as:

[0102]

[0103] Then update the candidate state as follows:

[0104]

[0105] In the formula, H t This represents the candidate state of this unit; W represents the weight parameter corresponding to the candidate state. b represents the Hadamard product of matrices; n The weight parameters corresponding to the candidate states are represented by tanh; tanh represents the bitangent function, expressed as:

[0106]

[0107] Finally, the output of the hidden layer (unit) is:

[0108]

[0109] Gated recurrent unit neural network (GRU) controls the filtering of output information through reset gates and update gates, enabling it to maintain a relatively stable amount of effective information in the increasing time series information.

[0110] The 106 extracted features were randomly shuffled, and then the 106 features were divided into a training set and a test set in an 8:2 ratio. The training set was used to train a nonlinear mapping model between the rare earth molten salt motion field features and the rare earth molten salt reaction state. Then, the trained nonlinear mapping model was used to predict and verify the test set data.

[0111] like Figure 5 The diagram shown is a schematic representation of the prediction results of the model described in an embodiment of the present invention. Figure 5 As can be seen, the model's predictions hit the true values ​​very well, with an accuracy rate of up to 95.23%, which meets the requirements of the production enterprise's testing standards (90%).

[0112] This embodiment also compares the solution results of several classic pattern recognition algorithms in the field of machine learning. The support vector machine (SVM) classification kernel function uses a Gaussian kernel with good nonlinear performance; the backpropagation neural network (BPNN) adopts a classic 3-layer structure with 30 hidden layer neurons; the long short-term memory neural network (LSTM) is similar in structure to the gated recurrent unit neural network (GRU), with the former replacing the gated unit of the latter with a long short-term memory unit.

[0113] After each model has been trained, the trained model is used to predict the test set data. Table 2 below shows a comparison of the model prediction results:

[0114] Table 2 Comparison of Model Prediction Results

[0115]

[0116] As can be seen from the comparison results in Table 2, recurrent neural networks have better learning and prediction capabilities for time-series information. The performance of the gated recurrent unit neural network in this application is better than that of the long short-term memory neural network, and its prediction accuracy can reach up to 95.23%.

[0117] It is worth noting that the contents not described in detail in the embodiments of the present invention belong to the prior art known to those skilled in the art.

[0118] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims. The information disclosed in the background section is intended only to enhance the understanding of the overall background technology of the present invention and should not be construed as an admission or implication in any way that such information constitutes prior art known to those skilled in the art.

Claims

1. A method for multi-parameter monitoring of the state of a rare earth electrolyzer based on molten salt image features, characterized in that, The method includes: Step 1: Based on the monitoring requirements of the rare earth electrolytic cell's operating status, the rare earth electrolytic cell status monitoring is divided into rare earth molten salt temperature monitoring and rare earth molten salt reaction status monitoring. Step 2: Construct a nonlinear mapping model for rare earth molten salt temperature using the GA-BP model to achieve rare earth molten salt temperature monitoring; the specific process is as follows: First, colorimetric features are extracted as basic variables using colorimetric formula 1: （1） Then, several auxiliary features were established as: HSI, RR, CVA, and YUV color features; HSI compensates for the lack of image information in the RGB color space by describing the hue, brightness, and saturation information of an object, as shown in Equation 2; RR can effectively eliminate the influence of illumination and improve the robustness of the prediction model, as shown in Equation 5; CVA is used to represent the visual differences between different colors, as shown in Equation 4; the YUV color model conforms to the characteristic that human vision is more sensitive to brightness information, as shown in Equation 3. （2） （3） （4） （5） Where Y represents the pixel brightness; U represents the difference between the red component and the brightness; V represents the difference between the blue component and the brightness; K is a correction coefficient, which is related to camera parameters, exposure time, and the emissivity of the object. The correction coefficient K is obtained through experimental calibration; H, S, and I represent hue, brightness, and saturation, respectively; R, G, and B represent the grayscale values ​​of the red, green, and blue channels, respectively; and Grey represents the grayscale image. , , All are intermediate variables; RR represents the relative red feature; CVA represents the color vector angle feature; Due to the limitations of colorimetric formulas, a nonlinear mapping model for rare earth molten salt temperature is constructed using a genetic algorithm-background algorithm (GA-BP) neural network model. The specific process is as follows: A BP neural network model consists of three parts: an input layer, hidden layers, and an output layer. The input layer is determined by the dimension of the independent variable, and the output layer is determined by the dimension of the dependent variable. The number of nodes in the hidden layer is shown in Formula 6. （6） I and O represent the number of input and output nodes, respectively; according to Formula 8, the value range of the hidden layer node N is (1, 13); the model is composed of fully connected nodes in the input layer, hidden layer, and output layer; the calculation process of the model consists of two parts: forward propagation of weights and backward propagation of errors, and its calculation formula is as follows: （7） （8） in, This represents the i-th neuron in the hidden layer; This represents the connection weight between the t-th temperature input feature and the i-th hidden layer neuron; This represents the bias value between the t-th temperature input feature and the i-th hidden layer neuron; This represents the connection weight between the i-th hidden layer neuron and the t-th output; K1 represents the bias value between the i-th hidden layer neuron and the t-th output; K2 represents the number of hidden layer nodes; K1 represents the number of input nodes; K2 represents the number of output nodes. This represents the t-th temperature input feature; Represents the t-th output node; ReLU and Both are nonlinear activation functions, and their expressions are shown in Equations 11 and 12, respectively: （9） （10） The number of hidden layer nodes, initial connection weight parameters, and initial bias parameters in the BP neural network model are optimized by introducing a genetic algorithm (GA). The optimization process includes initializing the population, calculating the fitness function, selection, crossover, and mutation. Let the total number of optimization parameters be N, and the calculation method is as follows: （11） （12） （13） In the above formula, This represents the number of connection weights between the input layer and the hidden layer; This represents the number of bias weights between the input layer and the hidden layer; This indicates the number of connection weights between the hidden layer and the output layer; This represents the bias weight parameters between the hidden layer and the output layer; Indicates the number of hidden layer nodes; Indicates the number of nodes in the input layer; This represents the number of nodes in the output layer; the theoretical limits for each weight parameter are (-1, 1). The objective function of the genetic algorithm is expressed using the mean square error function as follows: （14） In the formula, MES represents the mean squared error; n represents the length of the sample data; This represents the true value of the i-th data label; This represents the predicted value of the i-th data label; The image brightness range of the rare earth electrolytic cell is the entire furnace surface. In order to ensure the consistency of the data as much as possible, it is necessary to extract a pixel block of the same size from the center of the molten salt and then take the maximum value. Step 3: The intensity of the rare earth molten salt reaction is quantified by the dense optical flow method Farneback, scaling transformation and mean filtering, and a nonlinear mapping model between the rare earth molten salt motion field characteristics and the rare earth molten salt reaction state is established by the gated recurrent unit neural network GRU to realize the monitoring of the rare earth molten salt reaction state. Specifically, step 3 involves the following process: First, the velocity field features of rare earth molten salt are extracted using the Farneback optical flow method. The image neighborhood information expression is then defined as follows: （15） If the neighborhood information expression of the previous frame image is: （16） Assuming the subsequent image frame has a displacement d, based on the fundamental assumptions of optical flow, its image neighborhood information is represented as follows: （17） Based on the assumption that the coefficients of the above polynomials are equal: （18） From the above formula, we obtain the formula for calculating image displacement: （19） By introducing a local polynomial to replace the global polynomial in formula (17), the two graphs are expanded using local polynomials, and the expansion coefficients of the two graphs are obtained respectively. and Ideally, based on the optical flow assumption, it should satisfy... However, in practice, the following approximations are needed to reduce the error: （20） In addition, the introduction : （21） At this point, the image displacement calculation formula is updated from (21) to: （22） Assuming the image displacement changes slowly enough, then The following function must be minimized: （23） In the formula, The weight parameters corresponding to the image pixels are used, and the displacement calculation formula obtained by the least squares method is: （24） To improve the robustness of the model, the 8-parameter 2D motion parameterization model is represented as follows: （25） The above formula can be expressed in matrix form as follows: （26） （27） （28） Substituting the 8-parameter 2D motion parameterization model into formula (25), we obtain the new weighted least squares expression as follows: （29） Solving using the least squares method, the parameterized model is expressed as: （30） In solving the parametric model, first calculate separately and Then, the displacement d is solved by weighted average. Since the diameter of the cathode rod in rare earth electrolysis is fixed and it is installed vertically, the ratio of the cathode rod diameter in the image to the actual cathode rod diameter is used as a normalization conversion factor. First, the molten salt region is separated by threshold segmentation, then the contour of the molten salt region is extracted, and finally, two parallel lines with the most points in the vertical direction are fitted by the contour. Let displacement d represent the image displacement obtained by the Farneback optical flow method. This represents the distance between the fitted profiles of the cathode rods. The actual diameter of the cathode rod (85 mm) is then represented by the normalized image displacement. The expression is: （31） Assuming the frame rate of the image is N fps, then the time interval between two frames is 1 / N s. Therefore, the expression for the normalized speed v is: （32） At this point, the normalized velocity features in the rare earth molten salt video stream have been extracted. Since the camera is affected by pulse noise from the current and random noise from the environment during imaging, median filtering is also required for the data. Since the sampling time is small enough, the order of the median filter is set to N-1. The gated recurrent unit (GRU) neural network consists of reset gates and update gates, and its computation process is as follows: The input value X of the current cell t and the state h output by the previous unit t-1 The inputs to the reset and update gates are composed of nonlinear combinations: （33） （34） In the formula, Indicates that the output of the reset gate is being reset; This indicates that the output of the updated gate is being updated; and These represent the weight parameters for the reset gate and the update gate, respectively. and These represent the bias matrices for the reset gate and the update gate, respectively; [] represents matrix dot product; [] represents matrix concatenation; The non-linear activation function sigmoid is represented as: （35） Then update the candidate state as follows: （36） In the formula, Indicates the candidate state of this unit; This represents the weight parameters corresponding to the candidate states; Represents the Hadamard product of matrices; The weight parameters corresponding to the candidate states are represented by tanh; tanh represents the bitangent function, expressed as: （37） Finally, the output of the hidden layer (unit) is: （38） Gated recurrent unit neural network (GRU) controls the filtering of output information through reset gates and update gates, enabling it to maintain a relatively stable amount of effective information in the increasing time series information. The extracted features are randomly shuffled, and then the features are divided into training and test sets in an 8:2 ratio. The training set is used to train a nonlinear mapping model between the rare earth molten salt motion field features and the rare earth molten salt reaction state. Then, the trained nonlinear mapping model is used to predict and verify the test set data.

2. The method for multi-parameter monitoring of rare earth electrolytic cell status based on molten salt image features according to claim 1, characterized in that, In step 2, a mean filter is further set to filter the experimental data. Specifically, the maximum value of each channel of the cropped pixel block is taken to approximate the peak value of the camera response curve. Then, the maximum value sampled data of the same temperature range is filtered. The filter width interval is 10 frames, and its maximum width value is the maximum frame rate value of the color industrial camera. Due to the inherent temperature inertia of rare earth electrolytic cells, the temperature soft measurement target is divided into the following temperature ranges: <1060℃, 1060-1070℃, 1070-1080℃, 1080-1090℃, 1090-1100℃ and >1100℃.

Images