Boiling phenomenon judgment device and method based on deep learning and optical reflection structure
Through deep learning and the boiling phenomenon judgment device of optical reflection structure, we can directly observe the internal changes of the liquid, and use the convolutional neural network to determine whether the liquid is boiling, solving the problem of inaccurate judgment of indirect parameters in the existing technology, achieving efficient and accurate boiling judge.
Patent Information
- Application Number
- CN202010985461.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-09-18
- Publication Date
- 2024-11-29
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing boiling judgment technology is based on indirect parameters, making it difficult to ensure accuracy when the chemical properties of the solution are unstable or the external environment changes, especially when the liquid temperature and pressure change, resulting in inaccurate judgment results.
Using a boiling phenomenon judgment device based on deep learning and optical reflection structure, a single-channel convolutional neural network and twin convolutional neural network are built through real-time image acquisition device, data cache device and machine learning device to directly observe the changes in the liquid and judge the liquid. Whether it is boiling, and calculate the boiling level based on the continuous frame images.
It achieves high-time and accurate boiling judgment, reduces the influence of changes in external environment and liquid traits, improves judgment efficiency and accuracy, avoids misjudgment, and adapts to multi-angle observations.
Smart Images

Figure CN112101469B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image information processing, and in particular to a boiling phenomenon judgment device and method based on deep learning and optical reflection structure. Background Art
[0002] Liquid boiling is a basic physical phenomenon. In the fields of metallurgy, pharmaceuticals, and chemicals, there are scenarios where it is necessary to accurately determine whether boiling occurs.
[0003] Existing boiling judgment technical solutions mainly indirectly judge the boiling phenomenon of the object to be tested by monitoring the temperature upper limit and changes of the solution. When the chemical properties of the solution are unstable or the external environment (pressure, temperature, etc.) changes frequently, the accuracy of such methods based on indirect parameters to judge the boiling phenomenon is difficult to ensure. For example, a temperature sensor is used to judge whether the liquid in the pot is boiling. At this time, the liquid temperature reaches the boiling point and no longer rises, then it is considered that the liquid is boiling. However, due to the increase in steam causing the air pressure in the pot, after boiling for a period of time, the liquid no longer boils because the boiling point increases due to the increase in pressure; afterwards, due to the increase in temperature, the properties of the liquid change, which will cause the boiling point to change again. Therefore, the accuracy of such methods based on indirect parameters to judge the boiling phenomenon is difficult to ensure, and the data information obtained cannot meet the needs. Summary of the invention
[0004] In order to achieve the above object, the technical solution of the present invention is as follows:
[0005] The boiling phenomenon judgment device based on deep learning and optical reflection structure includes: a real-time image acquisition device, a data buffer device and a machine learning device, wherein:
[0006] The real-time image acquisition device includes: a lens, a pipeline, a mirror surface arranged at the turning point of the pipeline, and a photographing device. A lens is arranged at the opening of one end of the pipeline, and the photographing device faces the opening of the other end of the pipeline. The video stream inside the liquid is collected in real time through the lens by optical reflection of the mirror surface, and sent to the data cache device.
[0007] The data cache device is used to store the video stream data inside the liquid collected by the real-time image acquisition device;
[0008] The machine learning device is used to extract the video stream data in the data cache device, analyze and process the video stream data, and determine the boiling level of the liquid based on the results of the analysis and processing; and is used to perform self-optimization based on the updated data of the data cache device.
[0009] Preferably, the pipeline includes a first pipeline, a second pipeline and a third pipeline, the first pipeline, the second pipeline and the third pipeline are vertically connected to each other, and the first pipeline is "L"-shaped.
[0010] Preferably, the second pipeline is provided with a rotating part, and the third pipeline is driven by the rotating part to rotate 360 degrees.
[0011] Preferably, it also includes a visualization device for displaying the judgment result of the machine learning device on the boiling of the liquid to be tested.
[0012] Preferably, it also includes an auxiliary light source supplement device, which is arranged above the real-time image acquisition device and away from the side of the pipeline.
[0013] The boiling phenomenon judgment method based on deep learning and optical reflection structure includes the following steps:
[0014] Construct training sets and test sets of single-channel convolutional neural network samples, and perform grayscale processing and normalization processing on the images in the training sets and test sets of single-channel convolutional neural network samples respectively;
[0015] A single-channel convolutional neural network is constructed, and a training set processed in the single-channel convolutional neural network sample is used to train the constructed single-channel convolutional neural network to obtain a boiling judgment model; a test set processed in the single-channel convolutional neural network sample is used to test the obtained boiling judgment model, to verify the performance of the single-channel convolutional neural network, and to obtain the optimal parameters of the boiling judgment model;
[0016] Construct training sets and test sets of twin convolutional neural network samples, perform grayscale processing and normalization on the images in the training sets and test sets of twin convolutional neural network samples, and form data pairs of continuous frame images;
[0017] Construct a twin convolutional neural network, and use the processed training set in the twin convolutional neural network sample to train the constructed twin convolutional neural network to obtain a boiling level judgment model; use the processed test set in the twin convolutional neural network sample to test the obtained boiling level judgment model, verify the performance of the twin convolutional neural network, and obtain the optimal parameters of the boiling level judgment model;
[0018] The internal image data of the liquid to be tested is obtained for grayscale processing and normalization. The generated image data is used as a boiling test set, and the boiling test set is input into the tested single-channel convolutional neural network to identify whether the liquid to be tested is boiling. When it is judged as no, the boiling level 0 is directly returned; if it is judged as yes, the continuous frame images with continuous judgment results of yes are combined into a boiling level test set, and the boiling level test set is input into the tested twin convolutional neural network to calculate and return the boiling level of the liquid to be tested.
[0019] Preferably, the single-channel convolutional neural network includes 7 layers in total, 6 of which include 2 convolutional layers, 2 maximum pooling layers, 2 fully connected layers and 1 output layer; wherein the 1st and 3rd layers of the single-channel convolutional neural network are convolutional layers, the number of convolutional kernels corresponding to the convolutional layers is 64, the size of the convolutional kernel is 3×3, the step size is 1, and the activation function is the ReLU function;
[0020] The second and fourth layers of the single-channel convolutional neural network are maximum pooling layers, and the convolution kernel size of the maximum pooling layer is 2×2 and the stride is 2;
[0021] The 5th and 6th layers of the single-channel convolutional neural network are fully connected layers. The number of neurons in each fully connected layer is 32, and the activation function is the ReLU function;
[0022] The 7th layer is the output layer. The output layer has 1 neuron and the activation function is the sigmoid function.
[0023] Preferably, the twin convolutional neural network includes a total of 22 layers, including 13 convolutional layers, 5 maximum pooling layers, 3 fully connected layers and 1 output layer; wherein the 1st, 2nd, 4th, 5th, 7th, 8th, 9th, 11th, 12th, 13th, 15th, 16th, and 17th layers of the twin convolutional neural network are convolutional layers, and the number of convolution kernels corresponding to the convolutional layers is 32, 32, 64, 64, 128, 128, 128, 256, 256, 256, 256, 256, 256, respectively, the size of the convolution kernel is 3×3, the step size is 1, and the activation function is the ReLU function;
[0024] The 3rd, 6th, 10th, and 14th layers of the twin convolutional neural network are maximum pooling layers, and the convolution kernel size of the maximum pooling layer is 2×2 and the stride is 2;
[0025] The 18th, 19th, and 20th layers of the twin convolutional neural network are fully connected layers. The number of neurons in each fully connected layer is 1024, 1024, and 512, and the activation function is the ReLU function.
[0026] The 7th layer is the output layer, which has 4 neurons and the activation function is the sigmoid function.
[0027] Preferably, a weighted loss function Loss is used to train the single-channel convolutional neural network, and the weighted loss function Loss used is:
[0028]
[0029] Where output size is the number of training set samples. When the i-th training sample belongs to boiling liquid, y i =1, otherwise y i =0, represents the probability that the i-th training sample is classified as boiling liquid.
[0030] Preferably, a weighted loss function Loss′ is used to train the twin convolutional neural network, and the weighted loss function Loss′ used is:
[0031]
[0032] Where n is the number of training set samples, M is the number of boiling level categories, and when the observed image k belongs to category c, y kc =1, otherwise y kc =0,P kc is the predicted probability that the observed image k belongs to category c.
[0033] Based on the above technical scheme, the beneficial effects of the present invention are: through the characteristics of boiling, that is, the liquid vaporizes violently, a large number of bubbles rise and expand, and then burst and disappear on the surface of the liquid, the changes inside the liquid can be directly observed to directly determine whether the solution is boiling. It has high timeliness and accuracy, and can adapt to situations where the properties of the solution and the external environment change.
[0034] 1. The real-time image acquisition device is combined with AI technology to realize a method for judging the boiling phenomenon of liquid based on artificial intelligence training. Compared with the traditional judgment method, it directly judges whether it is boiling by reading the image, which is not affected by pressure, liquid properties and external environmental factors, and improves the judgment efficiency and accuracy;
[0035] 2. Based on the judgment of continuous frames by the twin neural network, the probability of misjudgment caused by the contents in the liquid resembling boiling bubbles is effectively reduced;
[0036] 3. Through the periscope structure, the internal image of the liquid is directly obtained, which avoids the obstruction of water vapor when directly photographing the top liquid surface of the liquid to be tested, and effectively avoids the influence of boiling liquid vapor on image recognition;
[0037] 4. Set up a rotating part to achieve multi-angle observation. By adjusting the rotation angle, the best viewing angle for collecting images can be achieved, which is conducive to image processing and analysis. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] The specific implementation modes of the present invention are further described in detail below in conjunction with the accompanying drawings.
[0039] Figure 1 : A schematic diagram of the structure of a boiling phenomenon determination device based on deep learning and optical reflection structure of the present invention;
[0040] Figure 2 :Flowchart of boiling phenomenon judgment method based on deep learning and optical reflection structure;
[0041] Among them: 1. Lens; 2. First pipeline; 3. Second pipeline; 4. Third pipeline; 5. First mirror; 6. Second mirror; 7. Third mirror; 8. Rotating part; 9. Photographic device; 10. Auxiliary light source supplement device; 11. Data caching device; 12. Machine learning device; 13. Visualization device. DETAILED DESCRIPTION
[0042] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.
[0043] Embodiment 1
[0044] like Figure 1 As shown, the boiling phenomenon judgment device based on deep learning and optical reflection structure includes: a real-time image acquisition device, a data buffer device 11 and a machine learning device 12, wherein:
[0045] The real-time image acquisition device comprises: a lens 1, a pipe and a mirror disposed at a turning point of the pipe, and a photographing device 9. The lens 1 is disposed at an opening at one end of the pipe, and the photographing device 9 faces the opening at the other end of the pipe. The video stream inside the liquid is collected in real time through the lens by optical reflection of the mirror, and sent to a data cache device 11.
[0046] The auxiliary light source supplement device 10 is arranged above the real-time image acquisition device and away from the side of the pipeline. When the light is insufficient, the auxiliary light source is used to provide continuous lighting to fill in the light so that continuous shooting can be performed;
[0047] A data cache device 11, used to store video stream data inside the liquid collected by the real-time image acquisition device;
[0048] A machine learning device 12, used for extracting the video stream data in the data cache device 11, analyzing and processing the video stream data, and determining the boiling level of the liquid according to the results of the analysis and processing; and for performing self-optimization based on the updated data of the data cache device 11;
[0049] The visualization device 13 is used to display the judgment result of the machine learning device 12 on the boiling of the liquid to be tested.
[0050] Furthermore, the pipeline includes a first pipeline 2, a second pipeline 3 and a third pipeline 4. The first pipeline 2 and the second pipeline 3 are vertically connected and a second mirror 6 is arranged at the connection. The second pipeline 3 and the third pipeline 4 are vertically connected and a third mirror 7 is arranged at the connection. The first pipeline 2 is "L"-shaped and a first mirror 5 is arranged at the internal turning point. The second pipeline 3 is provided with a rotating part 8. The third pipeline 4 is driven by the rotating part 8 to rotate 360 degrees, so that the observation angle can be adjusted and the optimal viewing angle for collecting images can be adjusted.
[0051] like Figure 2 As shown in the figure, the boiling phenomenon judgment method based on deep learning and optical reflection structure has the following basic principles:
[0052] Judgment principle 1: Use a neural network to judge whether a single image is a boiling liquid frame by frame. This process can use a single-channel convolutional neural network to obtain a model after training. This process can be used as the initial judgment logic for liquid boiling. Its defect is that when the observed liquid contains contents similar to bubbles, such as translucent colloids, or the bottom of the container containing the observed liquid is uneven, this will cause the neural network to make a misjudgment and mistakenly regard the non-boiling observed liquid as boiling;
[0053] Judgment principle 2: Since boiling liquid will produce a large number of bubbles and float to the surface of the liquid, the video effect of boiling liquid is more dynamic (compared to non-boiling liquid). By comparing frame by frame, it can be found that the changes inside the boiling liquid are relatively more drastic. At the same time, the more drastic the boiling, the more significant the changes. Based on the above characteristics, this solution uses a twin convolutional neural network, inputs multiple continuous frame images of the liquid to be judged for comparison, and judges the severity of boiling by comparing the similarity of multiple images, and finally outputs the boiling level.
[0054] Based on the above principle, the present invention specifically includes the following steps:
[0055] Step 1: Construct a training set and a test set of single-channel convolutional neural network samples, wherein the training set of the single-channel convolutional neural network samples includes two parts, namely, positive samples consisting of pictures with boiling bubbles and negative samples consisting of pictures without boiling bubbles; the test set of the single-channel convolutional neural network samples includes two parts, namely, positive samples consisting of pictures with boiling bubbles and negative samples consisting of pictures without boiling bubbles.
[0056] Step 2: Perform grayscale processing on the images in the training set and test set of the single-channel convolutional neural network samples in step 1, so that the images change from three channels to single channels, and the image data changes from three channels to single channels, reducing the amount of calculation for subsequent training. The specific method is: the images in the training set and the images in the test set of the single-channel convolutional neural network samples are both RGB images. For the RGB input image J, X is a pixel point of the image, and the pixel J of point X in the corresponding grayscale image gray (X) is:
[0057] J gray (X) = (X r +X g +X b ) / 3
[0058] Where, X r ,X g ,X b are the three primary color values of the X pixel.
[0059] Step 3: Normalize the images in the training set and test set of the single-channel convolutional neural network samples in step 2, respectively, where:
[0060] Point X is at the corresponding normalized pixel J norm (X) is:
[0061]
[0062] Where, X i is the gray value of the i-th pixel, X min is the minimum color value in the grayscale image, X max is the maximum color value in the grayscale image,
[0063] The process of generating corresponding grayscale images from the images in the training set and the images in the test set of the single-channel convolutional neural network samples is as follows:
[0064] The RGB image is represented as a three-dimensional matrix, and then the mean of the RGB components of each pixel is stored in a grayscale image of the same size as the original RGB image to obtain a grayscale image of the RGB image.
[0065] Step 4: Construct a single-channel convolutional neural network: The constructed single-channel convolutional neural network includes 7 layers in total, 6 of which include 2 convolutional layers, 2 maximum pooling layers, 2 fully connected layers and 1 output layer; the 1st and 3rd layers of the single-channel convolutional neural network are convolutional layers, the number of convolutional kernels corresponding to the convolutional layers is 64, the size of the convolutional kernel is 3×3, the step size is 1, and the activation function is the ReLU function;
[0066] The second and fourth layers of the single-channel convolutional neural network are maximum pooling layers, and the convolution kernel size of the maximum pooling layer is 2×2 and the stride is 2;
[0067] The 5th and 6th layers of the single-channel convolutional neural network are fully connected layers. The number of neurons in each fully connected layer is 32, and the activation function is the ReLU function;
[0068] The 7th layer is the output layer. The output layer has 1 neuron and the activation function is the sigmoid function.
[0069] Step 5: Use the final training set obtained in step 3 to train the single-channel convolutional neural network constructed in step 4 to obtain a boiling judgment model; and use the final test set obtained in step 3 to test the boiling judgment model obtained in step 4 to verify the performance of the single-channel convolutional neural network and obtain the optimal parameters of the boiling judgment model;
[0070] In step 5, the weighted loss function Loss is used to train the single-channel convolutional neural network. The loss function Loss used is:
[0071]
[0072] Where output size is the number of training set samples. When the i-th training sample belongs to boiling liquid, u i =1, otherwise y i =0, represents the probability that the i-th training sample is classified as boiling liquid;
[0073] In step 5, the training process of the single-channel convolutional neural network includes the following steps:
[0074] 501), initializing a single channel convolutional neural network;
[0075] 502), using the final training set obtained in step 3 as the input of the single-channel convolutional neural network, and calculating the output;
[0076] 503) Use the chain rule to calculate the gradient of the actual result of the training sample in step 3 relative to the loss function; adjust the parameters with an initial learning rate of 0.001 and a continuously decreasing learning rate;
[0077] 504), continuously repeating step 502) to step 503), iterating in a loop until the loss function converges, and finally obtaining a liquid boiling judgment model,
[0078] When testing the liquid boiling judgment model, the accuracy of the single-channel convolutional neural network classification is A c , Sensitivity S n Specificity S pAnd speed are tested, among which,
[0079] Sensitivity n The expression is as follows:
[0080]
[0081] Specificity S p The expression is as follows:
[0082]
[0083] Where TP, FP, FN and TN represent the number of true positive samples, false positive samples, false negative samples and true negative samples in the final test set, respectively.
[0084] Step 6: Construct the training set and test set of the twin convolutional neural network samples, where the training set and the test set both include four parts, namely, a group of continuous frame images of violently boiling liquid (the boiling level is marked as 2), a group of continuous frame images of non-boiling liquid (the boiling level is marked as 0), a group of continuous frame images of locally slightly boiling liquid (the boiling level is marked as 1), and other pseudo-boiling image groups (the boiling level is marked as 0).
[0085] Step 7: Perform grayscale processing and normalization on the images in the training set and test set of the twin convolutional neural network samples in step 6. The processing process can refer to steps 2 and 3, and the continuous frame image data are combined into data pairs to form a continuous frame image group of violently boiling liquid, a continuous frame image group of non-boiling liquid, a continuous frame image group of locally slightly boiling liquid, and other pseudo-boiling image groups.
[0086] Step 8: Construct a twin convolutional neural network for judging the boiling degree of the liquid. The constructed twin convolutional neural network includes 22 layers, including 13 convolutional layers, 5 maximum pooling layers, 3 fully connected layers and 1 output layer. The 1st, 2nd, 4th, 5th, 7th, 8th, 9th, 11th, 12th, 13th, 15th, 16th and 17th layers of the twin convolutional neural network are convolutional layers. The number of convolutional kernels corresponding to the convolutional layers are 32, 32, 64, 64, 128, 128, 128, 256, 256, 256, 256, 256, 256, respectively. The size of the convolutional kernel is 3×3, the step size is 1, and the activation function is the ReLU function.
[0087] The 3rd, 6th, 10th, and 14th layers of the twin convolutional neural network are maximum pooling layers, and the convolution kernel size of the maximum pooling layer is 2×2 and the stride is 2;
[0088] The 18th, 19th, and 20th layers of the twin convolutional neural network are fully connected layers. The number of neurons in each fully connected layer is 1024, 1024, and 512, and the activation function is the ReLU function.
[0089] The 7th layer is the output layer, which has 4 neurons and the activation function is sigmoid.
[0090] Step 9: Use the final training set obtained in step 7 to train the twin convolutional neural network constructed in step 8 to obtain a boiling level judgment model; use the final test set obtained in step 7 to test the boiling level judgment model obtained in step 8, verify the performance of the twin convolutional neural network, and obtain the optimal parameters of the boiling level judgment model.
[0091] In step 9, the weighted loss function Loss′ is used to train the twin convolutional neural network. The weighted loss function Loss′ used is:
[0092]
[0093] Where n is the number of training set samples, M is the number of boiling level categories, and when the observed image k belongs to category c, y kc =1, otherwise y kc =0,P kc is the predicted probability that the observed image k belongs to category c.
[0094] In step 9, the training process of the twin convolutional neural network includes the following steps:
[0095] 901), Initialize the twin convolutional neural network;
[0096] 902), using the final training set obtained in step 7 as the input of the twin convolutional neural network, and calculating the output;
[0097] 903), using the chain rule to calculate the gradient of the boiling level of the training samples in the training set in step 7 relative to the loss function; adjusting the parameters with an initial learning rate of 0.001 and a continuously decreasing learning rate;
[0098] 904), continuously repeating step 902) to step 903), iterating in a loop until the loss function converges, and finally obtaining a liquid boiling level judgment model,
[0099] When testing the liquid boiling judgment model, the accuracy of the twin convolutional neural network classification is A c Conduct testing.
[0100] Step 10, when the boiling of the liquid is identified, a camera is used to capture a video of the inside of the liquid to be tested, the effective area of the image is extracted from the internal video, and the effective area is gray-scaled and normalized, and the generated data is used as a test set. The test set is input into the convolutional neural network tested in step 5 to identify whether the liquid is boiling. When the judgment result is no, the boiling level 0 is directly returned and the liquid to be tested is not boiling; when the judgment result is yes, continue to step 11.
[0101] Step 11: The continuous frame images with a continuous judgment result of yes in step 10 are combined into a boiling level test set, and the boiling level test set is input into the twin convolutional neural network tested in step 9 to calculate and return the boiling level of the liquid to be tested.
[0102] The above description is only a preferred implementation of the boiling phenomenon judgment device and method based on deep learning and optical reflection structure disclosed in the present invention, and is not intended to limit the protection scope of the embodiments of this specification. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of this specification shall be included in the protection scope of the embodiments of this specification.
[0103] The systems, devices, modules or units described in the above embodiments may be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, the computer may be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.
[0104] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.
[0105] Each embodiment in the embodiments of this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.
Claims
1. A boiling phenomenon judgment device based on deep learning and optical reflection structure, characterized in that: include: A real-time image acquisition device, a data caching device, and a machine learning device, wherein: The real-time image acquisition device includes: a lens, a pipeline, a mirror surface arranged at the turning point of the pipeline, and a photographing device. A lens is arranged at the opening of one end of the pipeline, and the photographing device faces the opening of the other end of the pipeline. The video stream inside the liquid is collected in real time through the lens by optical reflection of the mirror surface, and sent to the data cache device. The data cache device is used to store the video stream data inside the liquid collected by the real-time image acquisition device; The machine learning device is used to extract the video stream data in the data cache device, analyze and process the video stream data, and determine the boiling level of the liquid according to the results of the analysis and processing; and to perform self-optimization based on the updated data of the data cache device. The judgment method of the boiling phenomenon judgment device based on deep learning and optical reflection structure comprises the following steps: Construct training sets and test sets of single-channel convolutional neural network samples, and perform grayscale processing and normalization processing on the images in the training sets and test sets of single-channel convolutional neural network samples respectively; A single-channel convolutional neural network is constructed, and the single-channel convolutional neural network is trained using the training set processed in the single-channel convolutional neural network sample to obtain a boiling judgment model; the boiling judgment model is tested using the test set processed in the single-channel convolutional neural network sample to verify the performance of the single-channel convolutional neural network and obtain the optimal parameters of the boiling judgment model, wherein the single-channel convolutional neural network includes 7 layers in total, 6 of which include 2 convolutional layers, 2 maximum pooling layers, 2 fully connected layers and 1 output layer; wherein the 1st and 3rd layers of the single-channel convolutional neural network are convolutional layers, and the number of convolution kernels corresponding to the convolutional layers is The number of convolution kernels is 3×3, the step size is 1, and the activation function is ReLU function; the second and fourth layers of the single-channel convolutional neural network are maximum pooling layers, and the convolution kernel size of the maximum pooling layer is 2×2, and the step size is 2; the fifth and sixth layers of the single-channel convolutional neural network are fully connected layers, and the number of neurons in each fully connected layer is 32, and the activation function is ReLU function; the seventh layer is the output layer, and the output layer has 1 neuron, and the activation function is sigmoid function. The weighted loss function Loss is used to train the single-channel convolutional neural network, and the weighted loss function Loss used is: Where output size is the number of training set samples. When the i-th training sample belongs to boiling liquid, yi = 1, otherwise y i =0, represents the probability that the i-th training sample is classified as boiling liquid; A training set and a test set of twin convolutional neural network samples were constructed, and the images in the training set and the test set of the twin convolutional neural network samples were gray-scale processed and normalized respectively, and the continuous frame picture data were combined into data pairs. The twin convolutional neural network includes 22 layers in total, including 13 convolutional layers, 5 maximum pooling layers, 3 fully connected layers and 1 output layer. Among them, the 1st, 2nd, 4th, 5th, 7th, 8th, 9th, 11th, 12th, 13th, 15th, 16th and 17th layers of the twin convolutional neural network are convolutional layers, and the number of convolution kernels corresponding to the convolutional layers are 32, 32, 64, 64, 128, 128, 128, 256 ... 6, 256, 256, the size of the convolution kernel is 3×3, the step size is 1, and the activation function is the ReLU function; the 3rd, 6th, 10th, and 14th layers of the twin convolutional neural network are maximum pooling layers, and the convolution kernel size of the maximum pooling layer is 2×2, and the step size is 2; the 18th, 19th, and 20th layers of the twin convolutional neural network are fully connected layers, and the number of neurons in each fully connected layer is 1024, 1024, and 512, and the activation function is the ReLU function. The 7th layer is the output layer, and the output layer has 4 neurons. The activation function is the sigmoid function. The weighted loss function Loss′ is used to train the twin convolutional neural network. The weighted loss function Loss′ used is: Where n is the number of training set samples, M is the number of boiling level categories, and when the observed image k belongs to category c, y kc =1, otherwise y kc =0,P kc is the predicted probability that the observed image k belongs to category c; Construct a twin convolutional neural network, and use the processed training set in the twin convolutional neural network sample to train the constructed twin convolutional neural network to obtain a boiling level judgment model; use the processed test set in the twin convolutional neural network sample to test the obtained boiling level judgment model, verify the performance of the twin convolutional neural network, and obtain the optimal parameters of the boiling level judgment model; The internal image data of the liquid to be tested is obtained for grayscale processing and normalization. The generated image data is used as a boiling test set, and the boiling test set is input into the tested single-channel convolutional neural network to identify whether the liquid to be tested is boiling. When it is judged as no, the boiling level 0 is directly returned; if it is judged as yes, the continuous frame images with continuous judgment results of yes are combined into a boiling level test set, and the boiling level test set is input into the tested twin convolutional neural network to calculate and return the boiling level of the liquid to be tested.
2. The boiling phenomenon judgment device based on deep learning and optical reflection structure according to claim 1 is characterized in that: The pipeline includes a first pipeline, a second pipeline and a third pipeline. The first pipeline, the second pipeline and the third pipeline are vertically connected to each other. The first pipeline is "L" shaped, and a lens is arranged at the opening of the third pipeline.
3. The boiling phenomenon determination device based on deep learning and optical reflection structure according to claim 2 is characterized in that: The second pipeline is provided with a rotating part, and the third pipeline is driven by the rotating part to rotate 360 degrees.
4. The boiling phenomenon judgment device based on deep learning and optical reflection structure according to claim 1 is characterized in that: It also includes a visualization device for displaying the judgment result of the machine learning device on the boiling of the liquid.
5. The boiling phenomenon determination device based on deep learning and optical reflection structure according to claim 1 is characterized in that: It also includes an auxiliary light source supplement device, which is arranged above the real-time image acquisition device and away from the side of the pipeline.
6. A boiling phenomenon judgment method based on deep learning and optical reflection structure, characterized in that: The steps include: Construct training sets and test sets of single-channel convolutional neural network samples, and perform grayscale processing and normalization processing on the images in the training sets and test sets of single-channel convolutional neural network samples respectively; A single-channel convolutional neural network is constructed, and a training set processed in a single-channel convolutional neural network sample is used to train the constructed single-channel convolutional neural network to obtain a boiling judgment model; a test set processed in a single-channel convolutional neural network sample is used to test the obtained boiling judgment model, to verify the performance of the single-channel convolutional neural network, and to obtain the optimal parameters of the boiling judgment model, wherein the single-channel convolutional neural network includes 7 layers in total, 6 of which include 2 convolutional layers, 2 maximum pooling layers, 2 fully connected layers and 1 output layer; wherein the single-channel convolutional neural network has a total of 7 layers ... The first and third layers of the neural network are convolutional layers. The number of convolution kernels corresponding to the convolutional layers is 64, the size of the convolution kernel is 3×3, the step size is 1, and the activation function is the ReLU function. The second and fourth layers of the single-channel convolutional neural network are maximum pooling layers. The convolution kernel size of the maximum pooling layer is 2×2, and the step size is 2. The fifth and sixth layers of the single-channel convolutional neural network are fully connected layers. The number of neurons in each fully connected layer is 32, and the activation function is the ReLU function. The seventh layer is the output layer, which has 1 neuron. The activation function is the sigmoid function, and the weighted loss function Loss is used to train the single-channel convolutional neural network. The weighted loss function Loss used is: Where output size is the number of training set samples. When the i-th training sample belongs to boiling liquid, y i =1, otherwise y i =0, represents the probability that the i-th training sample is classified as boiling liquid; Construct training sets and test sets of twin convolutional neural network samples, perform grayscale processing and normalization on the images in the training sets and test sets of the twin convolutional neural network samples, and form data pairs with continuous frame picture data. The twin convolutional neural network includes 22 layers, including 13 convolutional layers, 5 maximum pooling layers, 3 fully connected layers and 1 output layer; the 1st, 2nd, 4th, 5th, 7th, 8th, 9th, 11th, 12th, 13th, 15th, 16th and 17th layers of the twin convolutional neural network are Convolution layer, the number of convolution kernels corresponding to the convolution layer is 32, 32, 64, 64, 128, 128, 128, 256, 256, 256, 256, 256, 256, respectively, the size of the convolution kernel is 3×3, the step size is 1, and the activation function is the ReLU function; the 3rd, 6th, 10th, and 14th layers of the twin convolutional neural network are maximum pooling layers, the convolution kernel size of the maximum pooling layer is 2×2, and the step size is 2; the 18th, 19th, and 20th layers of the twin convolutional neural network are fully connected layers, the number of neurons in each fully connected layer is 1024, 1024, and 512, and the activation function is the ReLU function. The 7th layer is the output layer, the output layer has 4 neurons, and the activation function is the sigmoid function. The weighted loss function Loss′ is used to train the twin convolutional neural network, and the weighted loss function Loss′ used is: Where n is the number of training set samples, M is the number of boiling level categories, and when the observed image k belongs to category c, y kc =1, otherwise y kc =0,P kc is the predicted probability that the observed image k belongs to category c; Construct a twin convolutional neural network, and use the processed training set in the twin convolutional neural network sample to train the constructed twin convolutional neural network to obtain a boiling level judgment model; use the processed test set in the twin convolutional neural network sample to test the obtained boiling level judgment model, verify the performance of the twin convolutional neural network, and obtain the optimal parameters of the boiling level judgment model; The internal image data of the liquid to be tested is obtained for grayscale processing and normalization. The generated image data is used as a boiling test set, and the boiling test set is input into the tested single-channel convolutional neural network to identify whether the liquid to be tested is boiling. When it is judged as no, the boiling level 0 is directly returned; if it is judged as yes, the continuous frame images with continuous judgment results of yes are combined into a boiling level test set, and the boiling level test set is input into the tested twin convolutional neural network to calculate and return the boiling level of the liquid to be tested.
Citation Information
Patent Citations
Gas-liquid two-phase flow measurement method based on high-speed photography technology and device thereof
CN105784707A
Device for analyzing heat exchange characteristics and bubble behaviors under non-steady-state flow
CN110161076A
Grate cooler working condition monitoring method based on image recognition
CN110595397A
Boiling phenomenon judgment device based on deep learning and optical reflection structure
CN212460580U
Transient boiling transition monitoring system for boiling water nuclear reactor and monitoring method
JP2005283269A
