Adipic acid reactor chamber crystallization abnormality monitoring system and method

Abstract

The application discloses an adipic acid reaction chamber crystallization abnormality monitoring system and method, which comprises collecting a video picture; single-frame image sampling; processing the single-frame image by using a reaction chamber crystallization monitoring method to determine the crystallization condition in the reaction chamber; S302: performing frequency domain image conversion on the single-frame gray image by using fast Fourier transform and performing filter processing on the frequency domain image; S303: inputting the frequency domain image after the filter processing into a long short-term memory network for prediction, and obtaining a predicted image in the time domain by using fast Fourier inverse transform processing; S304: inputting the frequency domain image into a denoising auto-encoding network, and obtaining a denoised image by using fast Fourier inverse transform processing; S305: calculating a displacement rate matrix of the crystallization boundary pixel points in the predicted image and the denoised image; S306: inputting the displacement rate matrix and the denoised image after fusion into a binary classifier, and determining whether it is a crystallization abnormality by using the binary classifier; and if it is an abnormality, issuing a warning. The application realizes intelligent monitoring of the crystallization abnormality.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of adipic acid production process, in particular to an adipic acid reaction chamber crystallization abnormality monitoring system and method thereof. BACKGROUND

[0002] Adipic acid, also known as fatty acid, is an important organic dibasic acid with the structural formula HOOC(CH2)4COOH. It plays an important role in chemical production, organic synthesis industry, medicine, and lubricant manufacturing. The production process of adipic acid is as follows: cyclohexanol is oxidized by nitric acid in the presence of V(V) and Cu(II) catalysts, and a series of processes such as crystallization, concentration, centrifugation, activated carbon decolorization, recrystallization, concentration, centrifugation, and drying are performed to obtain the finished product adipic acid. In large-scale industrial production of adipic acid, purification treatment of adipic acid is required. The existing purification equipment for adipic acid has continuous multiple crystallization chambers on one side, each including one crystallization chamber (reaction chamber) observation mirror and one siphon observation mirror. In the production process of adipic acid, due to the many and frequent production processes, the existing technology relies on the workers' sensory and experience to judge the crystallization abnormality, which often leads to inaccurate and time-consuming situations. With the development of image technology, it is gradually possible to assist workers in identifying and analyzing crystallization abnormalities through image recognition, reducing the labor intensity of frontline workers and improving the working environment.

[0003] In actual production of adipic acid, the shaking of the stirring rod during stirring and the boiling of adipic acid during heating not only cause the shaking of the camera image, but also cause the shaking of the crystalline body with the liquid, resulting in a serious decrease in recognition accuracy. The existing image recognition technology for crystalline bodies is still in the stage of simple image segmentation and cannot accurately describe the dynamic changes of the crystalline body, which leads to untimely and inaccurate abnormality recognition. SUMMARY

[0004] The purpose of the present application is to provide an adipic acid reaction chamber crystallization abnormality monitoring system and method thereof, which utilizes deep learning technology for image processing, filters shaking noise and adds a prediction algorithm to realize online intelligent monitoring of the crystallization abnormality of the reaction chamber of the adipic acid purification equipment, improve the monitoring accuracy, reduce the labor intensity of frontline workers, and improve the working environment.

[0005] In order to achieve the above purpose, the present application adopts the following technical solutions:

[0006] An adipic acid reaction chamber crystallization abnormality monitoring system, comprising: an infrared camera, a network module, and a central control system,

[0007] The infrared camera is installed at a reaction chamber observation port, is used for axial position adjustment along the reaction chamber observation port, monitors a liquid level warning line, a stirring rod lower half and a contact area thereof, and a complete reaction chamber inner wall picture;

[0008] The infrared camera is connected with the central control system through a network module, and the central control system is used for analyzing the adipic acid reaction chamber crystallization by using image processing technology according to image information collected by the infrared camera, judging whether the reaction chamber crystallization is abnormal, and realizing online monitoring of the reaction chamber crystallization abnormality.

[0009] Further, the central control system further comprises pushing reaction chamber crystallization abnormal data to a user end and issuing an abnormal alarm.

[0010] In order to achieve the above purpose, another technical scheme of the present application is:

[0011] An adipic acid reaction chamber crystallization abnormality monitoring method based on the above-mentioned adipic acid reaction chamber crystallization abnormality monitoring system comprises the following steps:

[0012] Step S1: collecting a real-time video picture in an adipic acid reaction chamber;

[0013] Step S2: performing single-frame image sampling on the video picture;

[0014] Step S3: processing the single-frame image by using a reaction chamber crystallization monitoring method to judge the crystallization condition in the reaction chamber; specifically comprising:

[0015] S301: performing gray scale processing on the single-frame image to obtain a single-frame gray scale image;

[0016] S302: performing frequency domain image conversion on the single-frame gray scale image by using fast Fourier transform, and performing filter processing on the frequency domain image by using a high-pass filter;

[0017] S303: inputting the filtered frequency domain image in S302 into a long short-term memory network for prediction, and obtaining a predicted image in the time domain by using fast inverse Fourier transform;

[0018] S304: inputting the filtered frequency domain image in S302 into a denoising auto-encoding network for processing, and obtaining a denoised image by using fast inverse Fourier transform;

[0019] S305: calculating a displacement rate matrix of crystallization boundary pixel points in the predicted image obtained in S303 and the denoised image obtained in S304;

[0020] S306: input the displacement rate matrix in S305 and the denoising image in S304 into a binary classifier after fusion, and determine whether it is a crystallization abnormality by using the binary classifier;

[0021] Step S4: if the reaction chamber crystallization is abnormal, issue a warning; otherwise, repeat steps S1-S3 for the next monitoring.

[0022] Further, the S302 further includes normalization processing after the single-frame gray-scale image is converted into a frequency domain image by using fast Fourier transform, specifically: c i = x i / x max ; wherein c i is a normalized index parameter, x i is the image gray value before processing, and x max is the maximum gray value of x i .

[0023] Further, before the step S3, the method further includes collecting historical data, and training the high-pass filter, the long short-term memory network, the denoising auto-encoder network, and the binary classifier by using the historical data.

[0024] Further, the collecting historical data further includes constructing a training set and a test set.

[0025] Further, the S305: calculating the displacement rate matrix of the crystallization boundary pixel points in the predicted image obtained in S303 and the denoising image obtained in S304, includes:

[0026] S3051: using image segmentation technology to respectively segment the crystallization boundary in the predicted image obtained in S303 and the denoising image obtained in S304;

[0027] S3052: calculating the relative displacement speed of each crystallization pixel point in the denoising image and the predicted image by using the optical flow method, and recording it in the relative displacement rate matrix; wherein the relative displacement rate matrix is equal to the image size, and the element value of the non-crystallization region is directly set to 0.

[0028] Further, the optical flow method in S3052 is replaced by a background difference method, which calculates the relative change of the crystallization pixel point in the denoising image and the predicted image by using the background difference method, and generates a corresponding matrix, which is the relative displacement rate matrix.

[0029] Further, the fusion in S306 is a convolutional neural network-based fusion, which overlaps two equal matrices into one matrix; or uses concat connection.

[0030] Further, the step S2 adopts an adaptive sampling principle or a fuzzy sampling principle for the single-frame image sampling of the video picture,

[0031] The adaptive sampling principle is that the sampling frequency is positively correlated with the crystallization area size, the crystallization area change rate, the stirring rod offset and the adipic acid boiling confidence, that is, the larger the crystallization area size, the higher the sampling frequency, the larger the crystallization area change rate, the higher the sampling frequency, the larger the stirring rod offset, the higher the sampling frequency, and the higher the adipic acid boiling confidence, the higher the sampling frequency.

[0032] The fuzzy sampling principle divides the sampling frequency into a plurality of fixed categories, each category corresponds to a corresponding crystallization area size range and crystallization area change rate range, and also satisfies that the sampling frequency is positively correlated with the crystallization area size, the crystallization area change degree, the stirring rod offset and the adipic acid boiling confidence; the category division can be artificially experienced or historically planned.

[0033] The beneficial effects of the present application are:

[0034] The adipic acid reaction chamber crystallization abnormality monitoring system and method thereof of the present application adds a camera and a central control system, replaces the ocular lens with the camera, processes images by using deep learning technology, and combines the denoising recognition of the crystallization area and the crystallization area change rate to realize the rapid and accurate intelligent monitoring of the crystallization abnormality of the adipic acid purification equipment reaction chamber, reduce the labor intensity of the front-line workers, and improve the working environment. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 It is a principle block diagram of the adipic acid reaction chamber crystallization abnormality monitoring system in embodiment 1 of the present application;

[0036] Figure 2 It is a flowchart of the adipic acid reaction chamber crystallization abnormality online monitoring method in embodiment 2 of the present application;

[0037] Figure 3 It is a crystallization image recognition flowchart of the adipic acid in embodiment 2 of the present application.

[0038] Note: 1 is an infrared camera, 2 is a network module, 3 is a central control system, 4 is an external power supply, and 5 is a reaction chamber observation port. DETAILED DESCRIPTION

[0039] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments.

[0040] Embodiment 1:

[0041] Referring to Figure 1 As shown in the drawings, the present application provides an adipic acid reaction chamber crystallization abnormality monitoring system, comprising an infrared camera 1, a network module 2, a central control system 3, an external power supply 4,

[0042] The sight glass of the reaction chamber observation port 5 is replaced by an infrared camera 1, which is connected to the central control system 3 through the network module 2, and the infrared camera 1 is powered by the external power supply 4. The network module 2 can be a wireless or wired network module, and the central control system 3 is a computer device.

[0043] The infrared camera 1 is installed on the reaction chamber observation port 5 and can be adjusted in axial position along the reaction chamber observation port, and needs to meet the requirements of including the monitoring liquid level warning line, the lower half of the stirring rod and its contact area, and the complete reaction chamber inner wall picture in the shooting picture.

[0044] The central control system 3 is used to analyze the adipic acid reaction chamber crystallization according to the image information collected by the infrared camera, judge whether the reaction chamber crystallization is abnormal, and realize online monitoring of the reaction chamber crystallization abnormality.

[0045] The central control system 3 also includes pushing the reaction chamber crystallization abnormality data to the user end and issuing an abnormal alarm. In this embodiment, the central control system 3 can push the abnormal data to the user end in the FTP / MQ / Http mode, and the video data provides a standard video stream interface (RTSP). The central control system 3 is connected with an alarm device, which uses sound and light alarm device to issue an alarm.

[0046] Embodiment 2:

[0047] As shown in Figure 2 and Figure 3 A kind of adipic acid reaction chamber crystallization abnormality monitoring method is provided, based on the adipic acid reaction chamber crystallization abnormality monitoring system of this embodiment 1, comprising the following steps:

[0048] Step S1: the infrared camera 1 collects real-time video picture in adipic acid reaction chamber, and sends real-time video picture in reaction chamber to central control system 3.

[0049] Step S2: the central control system 3 carries out single frame image sampling to the video picture.

[0050] In order to reduce unnecessary image processing, the sampling amount of image can be reduced, while ensuring the flexibility of sampling, therefore the single frame image sampling principle of the video picture adopts adaptive sampling principle or fuzzy sampling principle.

[0051] The adaptive sampling principle is that the sampling frequency is positively correlated with the crystallization area size, the crystallization area change rate, the stirring rod offset and the boiling confidence of adipic acid, that is, the larger the crystallization area size, the higher the sampling frequency, the larger the crystallization area change rate, the higher the sampling frequency, the larger the stirring rod offset, the higher the sampling frequency, and the higher the boiling confidence of adipic acid, the higher the sampling frequency, and vice versa. The stirring rod offset is the offset angle of the stirring rod straight line after image segmentation by using the image segmentation technology. The boiling confidence of adipic acid is to judge whether it is boiling by using the convolution network image segmentation to identify the change of liquid surface ripples, and the output result is the boiling confidence.

[0052] The fuzzy sampling principle divides the sampling frequency into multiple fixed categories, each category corresponds to a corresponding crystallization area size range and crystallization area change rate range, and also satisfies that the sampling frequency is positively correlated with the crystallization area size, the crystallization area change rate, the stirring rod offset and the boiling confidence of adipic acid. The category division can be artificially experienced or historically planned.

[0053] The above two sampling strategies are selected according to the specific application scene requirements. Because the fuzzy sampling has low resolution and high speed, the adaptive sampling can selectively improve the resolution, but the speed is also correspondingly reduced. It is necessary to see whether the real-time requirement or the precision requirement of the application scene is high, so as to select the adaptive sampling principle or the fuzzy sampling principle.

[0054] Step S3: The central control system 3 processes the single frame image by using the reaction chamber crystallization monitoring method to judge the crystallization condition in the reaction chamber.

[0055] In combination with Figure 3 , the step S3 specifically comprises the following steps:

[0056] S301: The single frame image collected in step S2 is subjected to gray scale processing to obtain a single frame gray scale image;

[0057] S302: The single frame gray scale image is subjected to frequency domain image conversion by using fast Fourier transform FFT, and the frequency domain image is subjected to filter processing by using a high-pass filter;

[0058] After the single frame gray scale image is subjected to frequency domain image conversion by using fast Fourier transform FFT in the S302, it further comprises normalization processing, specifically: c i = x i / x max ; wherein, c i is a normalization index parameter, x i is the image gray scale value before processing, and x max is the maximum gray scale value of x i .

[0059] S303: input the frequency domain image in S302 into a long short-term memory network (LSTM) for prediction, and obtain a predicted image in the time domain through inverse fast Fourier transform (IFFT) processing after the prediction;

[0060] S304: input the frequency domain image in S302 into a denoising auto-encoding network for processing, and obtain a denoised image through inverse fast Fourier transform (IFFT) processing after the processing;

[0061] S305: calculate the displacement rate matrix of the crystal boundary pixel points in the predicted image obtained in S303 and the denoised image obtained in S304; specifically including:

[0062] S3051: use a YOLO network in image segmentation technology to respectively segment the crystal boundaries in the predicted image obtained in S303 and the denoised image obtained in S304;

[0063] S3052: use a deep learning convolutional neural network (CNN) to build an optical flow method solving model, calculate the relative displacement velocity of each crystal pixel point in the denoised image and the predicted image, and record it in the relative displacement rate matrix; wherein the relative displacement rate matrix is equal to the image size, and the element values of the non-crystal region are directly set to 0.

[0064] S306: input the displacement rate matrix in S305 and the denoised image in S304 into a binary classifier SVM after fusion, the fusion is concat connection, use the binary classifier SVM to judge whether it is a crystal abnormal situation, the crystal abnormal situation is: judge whether the crystal area is greater than a preset threshold, or whether the crystal rate is greater than a preset threshold, any one of the two conditions greater than the preset threshold is a crystal abnormality. The preset threshold is obtained by training the training set, the training set is marked by artificial or machine, and the normal and abnormal situations of the crystal are marked in it, and the purpose of the threshold is to ensure the accuracy of identification.

[0065] Before the step S3: the central control system 3 processes a single frame image using the reaction chamber crystal monitoring method to judge the crystal situation in the reaction chamber, it further includes collecting historical data, constructing a training set and a test set, training a high-pass filter, training a long short-term memory network, training a denoising auto-encoding network, and a binary classifier, specifically including:

[0066] First, collect historical data to construct a training set and a test set, the historical data is an infrared camera historical video image marked by artificial method. When the historical data is not enough, random jitter signals can also be added artificially to enrich the training set and the test set.

[0067] Then, the high-pass filter is trained, the long short-term memory network is trained, the denoising auto-encoding network is trained, and the binary classifier is trained respectively by using the training set constructed by the historical data.

[0068] The long short-term memory network is trained by using the time sequence sampling image, and the loss function needs to satisfy that the prediction accuracy of the image at the next moment is as accurate as possible.

[0069] The denoising auto-encoding network needs to restore and smooth the frequency domain image as much as possible.

[0070] The binary classifier adopts the convolutional neural network, and the loss function needs to satisfy the accuracy of the artificial marking.

[0071] Finally, the high-pass filter, the long short-term memory network, the denoising auto-encoding network and the binary classifier are tested respectively by using the test set.

[0072] Step S4: when it is judged that the reaction chamber crystallization is abnormal, a warning is given, and the network module 2 sends the working staff of the user end to the central control system 3; otherwise, steps S1-S3 are repeated to monitor the next time.

[0073] The present application fully considers the combination of frequency domain and time domain, considers from multiple dimensions, simultaneously uses the static noise reduction and the dynamic combination of displacement change rate, thereby improving the recognition accuracy and confidence, and realizes the intelligent monitoring of the adipic acid reaction chamber crystallization.

[0074] Embodiment 3:

[0075] An adipic acid reaction chamber crystallization abnormality monitoring method,

[0076] The feature of the embodiment is that: the fusion in S306 is convolutional neural network (CNN) based fusion, which overlaps two equal matrices into one matrix. That is, by assigning different weights to each input matrix, the influence degree of the crystallization area and the crystallization rate on the crystallization abnormality is distinguished, so that the abnormality recognition accuracy is more accurate. In the case of ensuring a certain accuracy, the difference format is simple, which can reduce the complexity of calculation, thereby reducing the calculation amount, and the optical flow method in S3052 is replaced by the background difference method, which calculates the relative change amount of the crystallization pixel points in the denoising image and the prediction image by using the background difference method, and generates a corresponding matrix, which is a relative displacement rate matrix.

[0077] The embodiment has the same other features as embodiment 2, and thus is not repeated here.

[0078] The technical solutions provided by the present application are described in detail above. The principles and implementation modes of the present application are described by applying specific examples, and the above examples are only used to help understand the method of the present application and its core idea. It should be pointed out that, for ordinary skilled persons in the technical field, some improvements and modifications can be made to the present application without departing from the principles of the present application, and these improvements and modifications also fall within the protection scope of the claims of the present application.

Claims

1. An adipic acid reactor chamber crystallization abnormality monitoring system characterized by, The system comprises: an infrared camera, a network module and a central control system, the infrared camera is installed on the observation port of the reaction chamber, is used for axial position adjustment along the observation port of the reaction chamber, monitors the liquid level warning line, the lower half of the stirring rod and the contact area thereof, and the complete reaction chamber inner wall picture; the infrared camera is connected with the central control system through the network module, and the central control system is used for analyzing the adipic acid reaction chamber crystallization by using image processing technology according to the image information collected by the infrared camera, judging whether the reaction chamber crystallization is abnormal, and realizing online monitoring of the reaction chamber crystallization abnormality. The system comprises the following steps: Step S1: collecting real-time video pictures in the adipic acid reaction chamber; Step S2: single-frame image sampling is performed on the video pictures; Step S3: the single-frame image is processed by using a reaction chamber crystallization monitoring method to judge the crystallization condition in the reaction chamber; specifically comprising: S301: performing gray processing on the single-frame image to obtain a single-frame gray image; S302: performing frequency domain image conversion on the single-frame gray image by using fast Fourier transform, and performing filter processing on the frequency domain image by using a high-pass filter; S303: inputting the filtered frequency domain image in S302 into a long short-term memory network for prediction, and obtaining a predicted image in the time domain by using fast Fourier inverse transform; S304: inputting the filtered frequency domain image in S302 into a denoising auto-encoding network for processing, and obtaining a denoised image by using fast Fourier inverse transform; S305: calculating the displacement rate matrix of the crystallization boundary pixel points in the predicted image obtained in S303 and the denoised image obtained in S304; S306: fusing the displacement rate matrix in S305 and the denoised image in S304, and inputting the fused image into a binary classifier to judge whether it is a crystallization abnormal condition; Step S4: if the reaction chamber crystallization is abnormal, a warning is issued; otherwise, the next monitoring is repeated by repeating steps S1 to S3.

2. The adipic acid reaction chamber crystallization anomaly monitoring system of claim 1, wherein: The central control system further comprises pushing the reaction chamber crystallization abnormal data to a user terminal and issuing an abnormal alarm.

3. A method of monitoring for an adipic acid reactor chamber crystallization anomaly, the method comprising: The adipic acid reaction chamber crystallization abnormality monitoring system according to any one of claims 1 or 2 comprises the following steps: Step S1: collecting real-time video pictures in the adipic acid reaction chamber; Step S2: single-frame image sampling is performed on the video pictures; Step S3: the single-frame image is processed by using a reaction chamber crystallization monitoring method to judge the crystallization condition in the reaction chamber; specifically comprising: S301: performing gray processing on the single-frame image to obtain a single-frame gray image; S302: performing frequency domain image conversion on the single-frame gray image by using fast Fourier transform, and performing filter processing on the frequency domain image by using a high-pass filter; S303: inputting the filtered frequency domain image in S302 into a long short-term memory network for prediction, and obtaining a predicted image in the time domain by using fast Fourier inverse transform; S304: inputting the filtered frequency domain image in S302 into a denoising auto-encoding network for processing, and obtaining a denoised image by using fast Fourier inverse transform; S305: calculate the displacement rate matrix of the crystal boundary pixel points in the prediction image obtained in S303 and the denoising image obtained in S304; S306: input the displacement rate matrix in S305 and the denoising image in S304 into a binary classifier after fusion, and determine whether it is a crystal abnormality by using the binary classifier; Step S4: if the reaction chamber is abnormal, issue a warning; otherwise, repeat steps S1-S3 for the next monitoring.

4. The method of claim 3, wherein: After performing frequency domain image conversion on the single-frame grayscale image using Fast Fourier Transform, step S302 also includes normalization processing, specifically: ;in, For normalized index parameters, The image grayscale values ​​before processing. for The maximum grayscale value.

5. The method of claim 3, wherein: Before the step S3, further comprising collecting historical data, training the high-pass filter, training the long short-term memory network, training the denoising auto-encoding network, and training the binary classifier using the historical data.

6. The method of claim 5, wherein: The collection of historical data further comprises constructing a training set and a test set.

7. The method of claim 3, wherein the method further comprises: determining a ratio of the first and second signals; and determining a ratio of the third and fourth signals. The S305: calculate the displacement rate matrix of the crystal boundary pixel points in the prediction image obtained in S303 and the denoising image obtained in S304, comprising: S3051: use image segmentation technology to segment the crystal boundary in the prediction image obtained in S303 and the denoising image obtained in S304, respectively; S3052: calculate the relative displacement velocity of each crystal pixel point in the denoising image and the prediction image by using the optical flow method, and record it in the relative displacement rate matrix; wherein the relative displacement rate matrix is equal to the image size, and the element value of the non-crystal region is directly set to 0.

8. The method of claim 7, wherein the method further comprises: determining a ratio of the first and second concentrations of the first and second impurities, respectively, in the first and second samples, respectively; and comparing the ratio to a predetermined ratio. In S3052, the optical flow method is replaced by the background difference method, which calculates the relative change of the crystal pixel points in the denoising image and the prediction image by using the background difference method, and generates a corresponding matrix, which is the relative displacement rate matrix.

9. The method of claim 3, wherein the method further comprises: determining a ratio of the first and second signals; and determining a ratio of the third and fourth signals. The fusion in S306 is based on convolutional neural network fusion, which overlaps two equal matrices into one matrix; or uses concat connection.

10. The method of claim 3, wherein: the crystallization chamber is a adipic acid crystallization chamber. The principle of single-frame image sampling of the video picture in step S2 adopts an adaptive sampling principle or a fuzzy sampling principle, The adaptive sampling principle, the sampling frequency is positively correlated with the crystal area size, the crystal area change rate, the stirring rod offset, and the adipic acid boiling confidence, that is, the larger the crystal area, the higher the sampling frequency, the larger the crystal area change rate, the higher the sampling frequency, the larger the stirring rod offset, the higher the sampling frequency, and the higher the adipic acid boiling confidence, the higher the sampling frequency; The fuzzy sampling principle divides the sampling frequency into multiple fixed categories, each category corresponds to a corresponding crystal area size range and crystal area change rate range, and also satisfies that the sampling frequency is positively correlated with the crystal area size, the crystal area change rate, the stirring rod offset, and the adipic acid boiling confidence; the division of categories can be artificial experience division or historical data unified division.

Images