Flooding identification method for PFA (Polyfluoroalkoxy) pipeline for electronic grade sulfuric acid

Abstract

The invention relates to the technical field of electronic-grade sulfuric acid production, and particularly discloses a full PFA pipeline identification method for electronic-grade sulfuric acid, which comprises the following steps: collecting PFA pipeline image data, manually marking an image name and a label, and importing the image name and the label into a CSV file to facilitate model extraction; randomly dividing the data into a training set, a verification set and a test set in proportion; constructing a classification prediction model based on a convolutional neural network, and taking the image data as input and the image label as output for training; carrying out optimization iteration on model parameters, so that the parameters of the prediction model are continuously learned and updated; and finally, obtaining an optimal prediction model of the electronic-grade sulfuric acid full pipe in the PFA pipeline. When a follow-up device runs, the flow state of the electronic-grade sulfuric acid fluid in the PFA pipeline can be judged through online monitoring, so that the pipeline flow, the valve opening and the like are operated in advance, and a certain guiding effect on the current production process is achieved.

Description

Technical Field

[0001] The invention belongs to the technical field of electronic-grade sulfuric acid production, and in particular relates to a method for identifying a full liquid of a PFA pipeline for electronic-grade sulfuric acid. Background Art

[0002] During the production of electronic-grade sulfuric acid, the requirements for pipeline materials and transportation systems are extremely strict to ensure product quality and stability. PFA pipes are made of polytetrafluoroethylene materials and have excellent chemical resistance, temperature resistance, non-stickiness, transparency, low friction coefficient and other characteristics. As a high-performance material, PFA pipes are widely used in fluid transportation during the production, filling and testing of electronic-grade sulfuric acid. In addition to the influence of electronic-grade sulfuric acid production process parameters on the quality of electronic-grade sulfuric acid products, the flow state of electronic-grade sulfuric acid in PFA pipes will also affect product quality and testing, especially during canning testing. Electronic-grade sulfuric acid is transported to the tank truck through PFA pipes during filling. After filling, the electronic-grade sulfuric acid in the tank truck is tested by a particle counter. During the transportation process, if the PFA pipe liquid is not full, gas will exist in the upper part of the pipe. During filling, a small amount of bubbles will be brought into the tank truck with the fluid, affecting the level of particulate matter in the electronic-grade sulfuric acid in the tank truck (increase) and reducing product quality. If there is a lot of gas in the PFA pipeline, it will impact the electronic-grade sulfuric acid during filling, causing turbulence and bubbling of the fluid in the tank truck, which will have a greater impact on the particulate matter in the electronic-grade sulfuric acid in the tank truck.

[0003] In the existing fluid transportation and storage industry, the methods for identifying whether a liquid is full include float switch method, capacitive liquid level sensor, ultrasonic liquid level sensor, radar liquid level sensor, pressure sensor, etc. However, when applied to the field of electronic-grade sulfuric acid production, the float switch method may cause mechanical wear, liquid adhesion, and pollution introduction; capacitive, ultrasonic and radar liquid level sensors have high accuracy, but may be affected by liquid properties (such as dielectric constant, temperature, etc.), and are also expensive. The image recognition method for whether the PFA pipeline for electronic-grade sulfuric acid is full of liquid proposed in the present invention directly recognizes the real-time liquid image in the pipeline, uses a convolutional neural network algorithm to process, identify and judge the real-time image, and outputs the signal, thereby improving the accuracy and real-time performance of the recognition, and does not introduce any pollution factors. Summary of the invention

[0004] A method for identifying a full liquid in a PFA pipeline for electronic grade sulfuric acid comprises the following steps: S1. Collect image data of PFA pipeline for electronic grade sulfuric acid in different scenes on site as input data. After data collection, manually mark the image data as output data.

[0005] The collected PFA pipeline images require that the electronic-grade sulfuric acid in the pipeline is clearly visible, and image data in different pipeline scenarios are collected: when filling the electronic-grade sulfuric acid, before filling the electronic-grade sulfuric acid, when the electronic-grade sulfuric acid tank truck is under back pressure, when the electronic-grade sulfuric acid particle size is detected, etc. The difference between each scenario is that the machine has not been started before the electronic-grade sulfuric acid is filled, and the electronic-grade sulfuric acid in the pipeline is empty; when filling the electronic-grade sulfuric acid, the amount of electronic-grade sulfuric acid in each section of the pipeline is different, there is liquid flowing in the large-diameter horizontal pipe but the pipe is not full, the small-diameter horizontal pipe is full of liquid, and the vertical pipe is full of liquid, etc.

[0006] The method for labeling various types of image data is as follows: the image of the horizontal pipeline filled with electronic grade sulfuric acid is marked as 1, indicating that the pipeline is full of liquid; the image of the horizontal pipeline not filled with electronic grade sulfuric acid is marked as 0, indicating that the pipeline is not full of liquid.

[0007] S2. Perform preliminary processing on the image data and annotation values ​​obtained in S1. Manually summarize the image names and corresponding labels into a CSV file for model input. The image data is saved in a separate folder for subsequent use. The data in the CSV file is saved in the format of the first column being the name of the image data file, and the second column being the manually made label value (0 or 1) for the image data. Each row in the CSV file represents an image input. The label value is automatically extracted when the model is run; the image data is summarized and used as model input.

[0008] S3, randomly dividing the processed image data obtained in S3 into a training set, a validation set, and a test set according to a ratio of 7:2:1; S4. Construct a classification prediction model based on a convolutional neural network. The model includes an input layer, a convolution layer, an activation layer, a pooling layer, a normalization layer, a fully connected layer, and an output layer, which are stacked together in a specific order.

[0009] The input layer receives raw image data, which is usually three-dimensional (height × width × number of channels), where the number of channels is 3 (corresponding to the red channel, blue channel, and green channel respectively).

[0010] The convolution layer receives the input layer 3D data and uses a set of learnable convolution kernels to perform convolution operations on the input image data to extract features. In convolutional neural networks, the convolution operation can be viewed as sliding the convolution kernel on the image and calculating the dot product between the convolution kernel and the local area of ​​the image at each position. Each convolution kernel extracts a specific feature in the image, such as edges, corners, or textures. The output is a feature map, which is the spatial mapping of the filter.

[0011] The steps of the convolution operation are as follows: (1) Initialize the output feature map: The size of the output feature map is determined by the size of the convolution kernel, the step size, and the padding. Assume that the size of the output feature map is ,in is the number of output channels (3 for color images and 2 for grayscale images).

[0012] (2) Convolution for each channel: For each channel i (i=1, 2, 3, corresponding to the red channel, green channel, and blue channel respectively) of the input image, a convolution operation is performed on the channel using the convolution kernel.

[0013] ; Among them, a and b are half of the size of the convolution kernel respectively (set the size of the convolution kernel yourself, the convolution kernel is an odd size, such as 3x3, 5x5. Assuming the convolution kernel size is 3x3, a and b are equal to 1). Represents the pixel value of the i-th color channel of the input image, where i is 1 (red channel), 2 (green channel), or 3 (blue channel). represents the weight of the convolution kernel at position (i, j) for the i-th color channel. represents the output of the convolution operation at position (x, y) on the i-th color channel of the input image. The specific implementation operation is as follows: 1) For each color channel i, the convolution kernel The center is placed on the input image 2) For each element in the convolution kernel , and compare it with the pixel value at the corresponding position in the input image Multiply; 3) Sum all the products to get the value of the convolution output at position (x, y).

[0014] (3) Accumulate the results of all channels: For each pixel value at a position (x, y), add the convolution results of all channels to obtain the output feature value of that position.

[0015] ; It is the input after convolution operation between the input image and the convolution kernel on channel i. is the summed characteristic value of the input results on 1 (red channel), 2 (green channel), or 3 (blue channel).

[0016] (4) Add bias term: If the convolutional layer contains a bias term b, the bias term is added to the above result.

[0017] ; in, is the output feature value after activation, is the activation function, and b is the bias constant.

[0018] The activation layer is after the convolutional layer, and usually a nonlinear activation function is applied, such as ReLU, Tanh, and Sigmoid functions. The activation function helps introduce nonlinearity, allowing the network to learn more complex functions.

[0019] ; ; ; The pooling layer downsamples the feature map to reduce the dimension of the data while retaining important feature information. Common pooling methods include maximum pooling and average pooling.

[0020] The average pooling operation slides a fixed-size window over the input data and calculates the average of all elements within each window. The step size of the pooling window sliding over the input data determines the size of the output feature map. If the step size is 1, the window moves one element at a time; if the step size is greater than 1, the window moves multiple elements at a time. After the average pooling operation, the size of the output feature map will be determined by the size and step size of the pooling window. If the size of the input feature map is H×W, the size of the pooling window is K×K, and the step size is S, the size of the output feature map is It can be calculated by the following formula: ; ; in Represents the floor function.

[0021] The maximum pooling operation also slides a fixed-size window on the input data and selects the largest element in each window as the output of the window. The calculation formula is: ; in, are the coordinates of the pooling window, and max is the maximization function.

[0022] The normalization layer is used after the convolution layer or the pooling layer to standardize the activation values ​​of the feature map to speed up the training process, improve the stability of the model, and reduce the problem of gradient disappearance or explosion.

[0023] The steps of batch normalization are: (1) Calculate the mean and standard deviation of each mini-batch data; (2) The mini-batch normalization calculation formula is: ; Among them, x is the input batch data, is the normalized data, is the mean of each batch of data, is the standard deviation of each batch of data.

[0024] The fully connected layer is at the end of the model, accepting the normalized input value and outputting the probability value of the corresponding model prediction belonging to 0 or 1 respectively, and the sum of the two probability values ​​is 1. The model takes the prediction with the higher probability value as the model output, and then gives the recognition of the input image input 0 or 1, indicating that the pipe in the image is not full or full of liquid. The fully connected layer connects all the activation values ​​of the previous layer to each neuron, and the number of neurons in the last fully connected layer corresponds to the number of categories, which is 2.

[0025] In the fully connected layer, each neuron in each layer is a linear algebraic function, which is activated by a nonlinear activation function and input to the next layer. The last layer outputs the probability value of the classification result. It is specifically expressed as: ; in, is the weight of the j-th input to the ith neuron, is the jth input, is the bias of the ith neuron, f is the activation function, and i and j represent the indices of the output neuron and input neuron, respectively.

[0026] The activation function f is used to introduce nonlinear factors so that the neural network can learn and simulate complex processes. Common activation functions include: Sigmoid, Tanh, ReLU, etc.

[0027] Finally, the model parameters are updated in reverse by defining the minimization of the loss function. Common loss functions include SoftmaxLoss and Sigmoid Loss.

[0028] ; Where y is the true label (0 or 1), is the predicted probability after the Sigmoid function. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 The figure is a flow chart of the prediction method of the present invention.

[0030] Figure 2 Schematic diagram of the convolution operation for a 3-channel image.

[0031] Figure 3 Image of a PFA tube with electronic grade sulfuric acid liquid not filling the tube (marked as 0).

[0032] Figure 4 Image of a full tube of electronic grade sulfuric acid liquid in a PFA tube (marked as 1).

[0033] Figure 5A live view of the input image (PFA_1.jpg) received by the prediction model.

[0034] Figure 6 A live view of the input image (PFA_pipeline_2.jpg) received by the prediction model.

[0035] Figure 7 A live view of the input image (PFA_pipeline_3.jpg) received by the prediction model.

[0036] Figure 8 A live view of the input image (PFA_pipeline_4.jpg) received by the prediction model.

[0037] Fig. 9 A live view of the input image (PFA_5.jpg) received by the prediction model.

[0038] Fig.10 A live view of the input image (PFA_6.jpg) received by the prediction model. DETAILED DESCRIPTION

[0039] The present invention is described in detail below in conjunction with specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and a specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0040] The present invention provides a method for identifying a full liquid in a PFA pipeline for electronic-grade sulfuric acid, comprising the following steps: S1. Collect image data of PFA pipeline for electronic grade sulfuric acid in different scenes on site as input data. After data collection, manually mark the image data as output data.

[0041] The collected PFA pipeline images require that the electronic-grade sulfuric acid in the pipeline is clearly visible, and image data in different pipeline scenarios are collected: when filling the electronic-grade sulfuric acid, before filling the electronic-grade sulfuric acid, when the electronic-grade sulfuric acid tank truck is under back pressure, when the electronic-grade sulfuric acid particle size is detected, etc. The difference between each scenario is that the machine has not been started before the electronic-grade sulfuric acid is filled, and the electronic-grade sulfuric acid in the pipeline is empty; when filling the electronic-grade sulfuric acid, the amount of electronic-grade sulfuric acid in each section of the pipeline is different, there is liquid flowing in the large-diameter horizontal pipe but the pipe is not full, the small-diameter horizontal pipe is full of liquid, and the vertical pipe is full of liquid, etc.

[0042] The method for labeling various types of image data is as follows: the image of the horizontal pipeline filled with electronic grade sulfuric acid is marked as 1, indicating that the pipeline is full of liquid; the image of the horizontal pipeline not filled with electronic grade sulfuric acid is marked as 0, indicating that the pipeline is not full of liquid.

[0043] S2. Perform preliminary processing on the image data and annotation values ​​obtained in S1. Manually summarize the image names and corresponding labels into a CSV file for use as model input. The image data is saved in a separate folder for subsequent use. The data in the CSV file is saved in the format of the first column being the name of the image data file, and the second column being the manually made label value (0 or 1) for the image data. Each row in the CSV file represents an image input. The label value is automatically extracted when the model is run; the image data is summarized and used as model input.

[0044] S3, randomly divide the processed image data obtained in S2 into a training set, a validation set, and a test set according to a ratio of 7:2:1; S4. Construct a classification prediction model based on a convolutional neural network. The model includes an input layer, a convolution layer, an activation layer, a pooling layer, a normalization layer, a fully connected layer, and an output layer, which are stacked together in a specific order.

[0045] The input layer receives the original image data, which is three-dimensional (height × width × number of channels), where the number of channels is 3 (RGB color image, corresponding to the red channel, blue channel, and green channel respectively). In this example, the input image data size is 1707×1280×3, and the convolution kernel size is 3×3×3.

[0046] The convolution layer receives the input layer 3D data and uses a set of learnable convolution kernels to perform convolution operations on the input image data to extract features. In convolutional neural networks, the convolution operation can be viewed as sliding the convolution kernel on the image and calculating the dot product between the convolution kernel and the local area of ​​the image at each position. Each convolution kernel extracts a specific feature in the image, such as edges, corners, or textures. The output is a feature map, which is the spatial mapping of the filter.

[0047] The steps of the convolution operation are as follows: (1) Randomly initialize the output feature map: The size of the output feature map is determined by the size of the convolution kernel, the step size, and the padding. Assume that the size of the output feature map is ,in is the number of output channels (3 for color images and 2 for grayscale images). (2) Perform convolution on each channel: For each channel i (i = 1, 2, 3, corresponding to the red channel, green channel, and blue channel in RGB, respectively) of the input color image, use the convolution kernel to perform a convolution operation on the channel.

[0048] ; Among them, a and b are half of the size of the convolution kernel respectively (the size of the convolution kernel is set by yourself, and the convolution kernel is an odd size, such as 3*3, 5*5. In this case, the convolution kernel size is 3*3, then a and b are equal to 1). Represents the pixel value of the i-th color channel of the input image, where i is 1 (red channel), 2 (green channel), or 3 (blue channel). represents the weight of the convolution kernel at position (i, j) for the i-th color channel. represents the output of the convolution operation at position (x, y) on the i-th color channel of the input image. The specific implementation operation is as follows: 1) For each color channel i, the convolution kernel The center is placed on the input image 2) For each element in the convolution kernel , and compare it with the pixel value at the corresponding position in the input image Multiply; 3) Sum all the products to get the value of the convolution output at position (x, y).

[0049] (3) Accumulate the results of all channels: For each position (x, y), add the convolution results of all channels to obtain the output feature value of that position.

[0050] ; It is the input after convolution operation between the input image and the convolution kernel on channel i. is the summed characteristic value of the input results on 1 (red channel), 2 (green channel), or 3 (blue channel).

[0051] (4) Add bias term: If the convolutional layer contains a bias term b, the bias term is added to the above result.

[0052] ; in, is the output feature value after activation, is the activation function and b is the bias constant.

[0053] The activation layer is usually followed by a convolutional layer and a nonlinear activation function is applied, such as ReLU, Tanh, or Sigmoid. The activation function helps introduce nonlinearity, allowing the network to learn more complex functions. The activation function used in this example is the Sigmoid function.

[0054] ; ; ; The pooling layer downsamples the feature map to reduce the dimension of the data while retaining important feature information. Common pooling methods include maximum pooling and average pooling. The pooling method used in this example is maximum pooling.

[0055] The average pooling operation slides a fixed-size window over the input data and calculates the average of all elements within each window. The step size of the pooling window sliding over the input data determines the size of the output feature map. If the step size is 1, the window moves one element at a time; if the step size is greater than 1, the window moves multiple elements at a time. After the average pooling operation, the size of the output feature map will be determined by the size and step size of the pooling window. If the size of the input feature map is H×W, the size of the pooling window is K×K, and the step size is S, the size of the output feature map is It can be calculated by the following formula: ; ; in Represents the floor function.

[0056] The maximum pooling operation also slides a fixed-size window on the input data and selects the largest element in each window as the output of the window. The calculation formula is: ; in, are the coordinates of the pooling window, and max is the maximization function.

[0057] The normalization layer is used after the convolution layer or the pooling layer to standardize the activation values ​​of the feature map to speed up the training process, improve the stability of the model, and reduce the problem of gradient disappearance or explosion.

[0058] The steps of batch normalization are: (1) Calculate the mean and standard deviation of each mini-batch data; (2) The mini-batch normalization calculation formula is: ; in, is the mean of each batch of data, is the standard deviation of each batch of data.

[0059] The fully connected layer is at the end of the model, accepting normalized input values ​​and outputting corresponding probability values ​​of 0 / 1, thereby identifying 0 / 1 for the image. The fully connected layer connects all activation values ​​of the previous layer to each neuron, and the number of neurons in the last fully connected layer corresponds to the number of categories, which is 2. The number of hidden layers in the fully connected layer ranges from 2 to 6, and the number of hidden layer neurons ranges from 8 to 512.

[0060] In the fully connected layer, each neuron in each layer is a linear algebraic function, which is activated by a nonlinear activation function and input to the next layer. The last layer outputs the probability value of the classification result. It is specifically expressed as: ; in, is the weight of the j-th input to the ith neuron, is the jth input, is the bias of the ith neuron, f is the activation function, and i and j represent the indices of the output neuron and input neuron, respectively.

[0061] The activation function f is used to introduce nonlinear factors so that the neural network can learn and simulate complex processes. Common activation functions include: Sigmoid, Tanh, ReLU, etc.

[0062] Finally, the model parameters are updated in reverse by defining the loss function to minimize. Common loss functions include SoftmaxLoss and Sigmoid Loss. The loss function used in this example is Sigmoid Loss.

[0063] ; Where y is the true label (0 or 1), is the predicted probability after the Sigmoid function.

[0064] The above embodiments describe the preferred implementation of the present invention, but the present invention is not limited thereto. Within the technical concept of the present invention, the technical solution of the present invention can be subjected to a variety of simple modifications, including the combination of various technical features in any other manner, and these simple modifications and combinations should also be regarded as the contents disclosed by the present invention and belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

[0065] Table 1 Input example in CSV file

[0066] Table 2 Prediction results of some input images after model training

Claims

1. A method for identifying a full liquid in a PFA pipeline for electronic grade sulfuric acid, characterized in that: The following steps are involved: S1. Collect image data of PFA pipeline for electronic grade sulfuric acid in different scenes on site as input data, and manually mark the image data as output data after data collection; S2, preliminarily process the image data and annotation values ​​obtained in S1, summarize the image names and corresponding tags into a CSV file, and use it as model input; S3, randomly dividing the processed image data obtained in S3 into a training set, a validation set, and a test set according to a ratio of 7:2:1; S4. Construct a classification prediction model based on a convolutional neural network; the model includes an input layer, a convolution layer, an activation layer, a pooling layer, a normalization layer, a fully connected layer, and an output layer, and is stacked together in a specific order to achieve full liquid recognition of PFA pipeline liquid for electronic grade sulfuric acid.

2. The method for identifying full liquid in a PFA pipeline for electronic grade sulfuric acid according to claim 1, characterized in that: In step S1, in the collected PFA pipeline image, the electronic-grade sulfuric acid in the pipeline is clearly visible, and image data in the pipeline is collected in scenes including when the electronic-grade sulfuric acid is filled, before the electronic-grade sulfuric acid is filled, when the electronic-grade sulfuric acid tank truck is back-pressurized, and when the particle size of the electronic-grade sulfuric acid is detected.

3. The method for identifying full liquid in a PFA pipeline for electronic grade sulfuric acid according to claim 1, characterized in that: In step S2, the data in the CSV file is saved in a format in which the first column is the name of the image data file, and the second column is the label value manually made for the image data; each row in the CSV file represents an image input, and the label value is automatically extracted when the model is running, and the image data is summarized and used as the model input.

4. The method for identifying full liquid in a PFA pipeline for electronic grade sulfuric acid according to claim 1, characterized in that: The input layer receives the original image data, which is three-dimensional data of height×width×number of channels, where the number of channels is 3, corresponding to the red channel, blue channel, and green channel respectively.

5. The method for identifying full liquid in a PFA pipeline for electronic grade sulfuric acid according to claim 1, characterized in that: The convolution layer receives the input layer three-dimensional data and uses a set of learnable convolution kernels to perform convolution operations on the input image data to extract features. The steps of the convolution operation are as follows: (1) Initialize the output feature map: The size of the output feature map is determined by the size of the convolution kernel, the step size, and the padding. The size of the output feature map is ,in is the number of output channels, 3 for color images and 2 for grayscale images; (2) Convolution for each channel: For each channel i of the input color image, i=1, 2, 3, corresponding to the red channel, green channel, and blue channel respectively, a convolution kernel is used to perform a convolution operation on the channel. ； Among them, a and b are half of the size of the convolution kernel, represents the pixel value of the input image on the i-th color channel, where i is 1, 2, or 3. represents the weight of the convolution kernel at position (i, j) for the i-th color channel, Represents the output of the convolution operation at position (x, y) on the i-th color channel of the input image. (3) Accumulate the results of all channels: For each pixel value (x, y), add the convolution results of all channels to obtain the output feature value of that position: ； is the input after convolution operation between the input image and the convolution kernel on channel i. is the summed eigenvalue of the results of entering them on 1, 2, or 3; (4) Add bias term: If the convolutional layer contains a bias term b, the bias term is added to the above result; ； in, is the output feature value after activation, is the activation function and b is the bias constant.

6. The method for identifying full liquid in a PFA pipeline for electronic grade sulfuric acid according to claim 1, characterized in that: The activation layer is after the convolution layer, and nonlinear activation functions, ReLU, Tanh, and Sigmoid functions are applied: ； ； 。 7. The method for identifying full liquid in a PFA pipeline for electronic grade sulfuric acid according to claim 1, characterized in that: The pooling layer samples the feature map. Its operation slides a fixed-size window on the input data and calculates the average value of all elements in each window. The size of the input feature map is H×W, the size of the pooling window is K×K, and the step size is S. The size of the output feature map is It can be calculated by the following formula: ； ； in represents the floor function; The maximum pooling operation also slides a fixed-size window on the input data and selects the largest element in each window as the output of the window. The calculation formula is: ； in, are the coordinates of the pooling window, and max is the maximization function.

8. The method for identifying full liquid in a PFA pipeline for electronic grade sulfuric acid according to claim 1, characterized in that: The normalization layer is used after the convolution layer or the pooling layer to standardize the activation values ​​of the feature map, so as to accelerate the training process, improve the stability of the model and reduce the problem of gradient disappearance or explosion; The steps of batch normalization are: (1) Calculate the mean and standard deviation of each mini-batch data; (2) The mini-batch normalization calculation formula is: ； Among them, x is the input batch data, is the normalized data, is the mean of each batch of data, is the standard deviation of each batch of data.

9. The method for identifying full liquid in a PFA pipeline for electronic grade sulfuric acid according to claim 1, characterized in that: The fully connected layer is at the very end of the model, accepts the normalized input values, and outputs the probability values ​​of the corresponding model predictions belonging to 0 or 1 respectively. The sum of the two probability values ​​is 1. The model takes the prediction with a higher probability value as the model output, and then gives the recognition of the input image input 0 or 1, indicating that the pipe in the image is not full of liquid or full of liquid. The fully connected layer connects all the activation values ​​of the previous layer to each neuron, and the number of neurons in the last fully connected layer corresponds to the number of categories, which is 2.

10. The method for identifying full liquid in a PFA pipeline for electronic grade sulfuric acid according to claim 1, characterized in that: In the fully connected layer, each neuron in each layer is a linear algebraic function, which is activated by a nonlinear activation function and input to the next layer. The last layer outputs the probability value of the classification result, which is specifically expressed as: ； in, is the weight of the j-th input to the ith neuron, is the jth input, is the bias of the i-th neuron, f is the activation function, i and j represent the indexes of the output neuron and input neuron respectively; Finally, the model parameters are updated in reverse by minimizing the loss function, and the loss function is: ； Where y is the true label (0 or 1), is the predicted probability after the Sigmoid function.

Images