Water channel parameter automatic generation system based on thermogram analysis and deep learning

Abstract

The invention discloses a water channel parameter automatic generation system based on thermogram analysis and deep learning, and relates to the field of thermal management in industrial manufacturing. According to the water channel parameter automatic generation system based on thermogram analysis and deep learning, thermogram parameters and water channel parameters are obtained; the method comprises the following steps: associating with a network target frame, acquiring a target frame and a related thermogram labeling data set, preprocessing the acquired data set to construct a backbone network structure, analyzing output elements, designing a mixed loss function, calculating a network parameter gradient of the mixed loss function through a back propagation algorithm, and acquiring accurate temperature field distribution data. The collected thermogram is preprocessed, the prediction frame is screened, unique indexing, output and parameter calculation are carried out, finally, the accurate tapping position on the mold is determined, the number of industrial mold testing times is reduced by optimizing the mold cooling water channel position design and the water channel process parameter setting, and the production efficiency is improved. The method is especially suitable for manufacturing scenes such as precision die casting and injection molding which have strict requirements on mold temperature field control.

Description

Technical Field

[0001] The present invention relates to the technical field of thermal management in industrial manufacturing, and particularly to an automatic water channel parameter generation system based on thermal image analysis and deep learning. Background Technique

[0002] In the die-casting and injection molding processes, precise control of the mold temperature field is crucial for ensuring product quality. In the traditional production process, the design of the mold cooling water channel position and the setting of water channel process parameters mainly rely on manual experience. The operator adjusts the water channel parameters by setting fixed rules according to experience, lacking data basis, resulting in large temperature deviations and affecting product quality. Although many systems currently use thermal imagers + manual experience to adjust the water channel design and parameter settings, this method requires a large number of tests, consumes a lot of resources, and may not necessarily find the optimal solution. There is an urgent need for a more advanced, precise and efficient water channel parameter regulation system to solve the above problems. Summary of the Invention

[0003] The purpose of the present invention is to provide an automatic water channel parameter generation system based on thermal image analysis and deep learning, so as to solve the problems of strong empirical dependence, insufficient multi-parameter coupling optimization and a large number of industrial trial moldings existing in the existing traditional mold temperature field control technology mentioned in the above background technique.

[0004] To achieve the above purpose, the present invention provides the following technical solution: An automatic water channel parameter generation system based on thermal image analysis and deep learning, the system includes the following steps: Step 1, obtain industrial data and perform preprocessing; Step 2, train the network; Step 3, online prediction of the model.

[0005] Step 1 in the system further includes the following steps:

[0006] Step 1.1: Obtain thermal image parameters and water channel parameters, and then associate them with the network target box;

[0007] Step 1.2: Obtain the target box and the relevant thermal image annotation data set;

[0008] Step 1.3: Perform preprocessing on the obtained thermal image annotation data set:

[0009] Step 2 in the system further includes the following steps:

[0010] Step 2.1: Construct the backbone network structure:

[0011] Step 2.2: Analyze the elements in the output:

[0012] Step 2.3: Design the mixed loss function;

[0013] Step 2.4: Calculate the gradient of the hybrid loss function with respect to the network parameters through the backpropagation algorithm;

[0014] The following steps are also included in Step 3 of the system:

[0015] Step 3.1: Obtain accurate temperature field distribution data;

[0016] Step 3.2: Preprocess the collected thermal images according to Step 1.2;

[0017] Step 3.3: Screening and unique indexing of prediction boxes;

[0018] Step 3.4: Calculate the parameters according to the output of Step 3.3;

[0019] Step 3.5: Use the calculated position coordinates to determine the precise opening position on the mold.

[0020] Furthermore, in Step 1.1, obtaining the thermal image parameters and water channel parameters, and subsequent association with the network target box mainly involves: using an infrared thermal imager, fixedly installed directly above the die-casting or injection mold to be measured, ensuring that the field of view covers the entire target area, collecting the thermal image of the mold surface during operation, selecting different mold types, different water valve position settings, different cooling parameters, different production batches, etc., and randomly setting parameters such as flow rate and water passing time at the same time to construct N (N>20000) groups of classical working conditions. Obtain the thermal image data at the time of process stability for each working condition once, denoted as p a (a = 1, 2, …, N), and finally obtain P = {p1, p2, …, p N} pictures, and record the water channel parameters (flow rate and water passing time) at that time through the PLC, and then associate them with the grid target box.

[0021] Furthermore, in Step 1.2: Obtaining the target box and the related thermal image annotation dataset mainly involves: uniformly scaling the thermal image collection P to X×X pixels (recommended to be 512×512), converting the original radiation value to a temperature value (unit: °C), and linearly normalizing the temperature to [0, 1] based on global statistics; manually annotate the position where the water channel projects onto the thermal image on each image p a ; specifically, use a rectangular annotation box to mark the position of the water channel, and call this rectangular box the target box. Let the number of target boxes marked on the thermal image be represented by b (b = 0, 1, 2), and each grid can mark at most B = 3 target boxes (when annotating, it is necessary to ensure that the annotation box is the smallest matrix that can enclose the target). For the image p aDivide it into S×S (recommended to be 8×8) grids, and assign a unique coordinate (i, j) to each grid, where i represents the row index (ranging from 0 to S - 1), and j represents the column index for each target box (also ranging from 0 to S - 1).

[0022] Furthermore, for the target box, according to the situation of the target box, set the following parameters in each grid of the ath thermal image:

[0023] c a,i,j,b : If there is a bth target box in the grid (i, j), then c a,i,j,b = 1, otherwise c a,i,j,b = 0

[0024] Bounding box parameters x a,i,j,b , y a,i,j,b , w a,i,j,b , h a,i,j,b : Where x a,i,j,b is the horizontal offset of the center point of the target box relative to the upper left corner of the grid (i, j); y a,i,j,b is the vertical offset of the center point of the target box relative to the upper left corner of the grid (i, j); w a,i,j,b is the width of the target box; h a,i,j,b is the height of the target box;

[0025] Water channel parameters f a,i,j,b , t a,i,j,b : Where f a,i,j,b represents the water channel flow rate corresponding to this target box (water channel opening); t a,i,j,b represents the water passing time of the water channel corresponding to this target box (water channel opening);

[0026] If there is no target in the grid, then for this target box c a,i,j,b = 0, and other parameters are also set to 0;

[0027] Repeat the above labeling process to finally obtain the entire dataset:

[0028] Where N is the number of thermal images, and B = 3 is the number of predictions for each grid.

[0029] Furthermore, for the dataset, preprocess the thermal image annotation dataset Q; linearly map the bounding box parameters x a,i,j,b , y a,i,j,b , w a,i,j,b , h a,i,j,b to the range of (0, 1) based on the thermal image size; scale the water channel parameters f a,i,j,b , t a,i,j,b from the maximum and minimum values of the actual process parameters to (0, 1); repeat the above steps to finally obtain the preprocessed dataset Among them represents the q-th picture of the a-th picture a The corresponding labeled information after preprocessing

[0030] Furthermore, step 2.1: Construct the backbone network structure as follows:

[0031] The output shape of the input layer is X×X×1 (X is recommended to be 512), which is equal to the size of the thermal image;

[0032] The first layer is a Stem convolution with a kernel size of 3×3, a stride of 2, and 32 output channels, using the SiLU activation function;

[0033] The second layer is a CSP module with 1 residual unit (including 1×1 and 3×3 convolutions), and the channels are compressed to 64;

[0034] The third layer is a Transition layer with a kernel size of 3×3, a stride of 2, and 64 output channels;

[0035] The fourth layer is a CSP module with 1 residual unit (including 1×1 and 3×3 convolutions with a dilation rate of 2), and the channels are compressed to 128;

[0036] The fifth layer is a Transition layer with a kernel size of 3×3, a stride of 2, and 128 output channels;

[0037] The sixth layer is an SPP fast pooling with 5×5, 9×9, and 13×9 parallel pooling;

[0038] The seventh layer is an upsampling layer. After 2x upsampling, it is followed by a 1×1 convolution with 128 output channels;

[0039] The eighth layer is a cross-layer feature fusion, concatenating the output of the upsampling layer and the fourth layer;

[0040] The ninth layer is a downsampling layer with a kernel size of 4×4 and a stride of 4;

[0041] The ninth layer is a prediction head with a 1×1 convolution, and the output is S×S×(B×7) (recommended S = 8, B = 3)

[0042] Furthermore, in step 2.2: Parse the elements in the output. Specifically, for each grid (i,j) and each of the 7 parameters of each bounding box b, they are respectively set to Among them is the predicted confidence; is the predicted bounding box parameter; is the predicted waterway parameter;

[0043] The 2.3: Design of the hybrid loss function: The hybrid loss function is used to supervise the prediction accuracy of the model for the waterway position, size, and parameters. The specific formula is as follows:

[0044]

[0045] where S is the grid size of 8, B is the number of bounding boxes predicted for each grid cell of 3, N is the number of images, a is the image index, i, j are the grid indices, b is the bounding box index, and c a,i,j,b represents the true confidence (1 means there is an object, 0 means there is no object); x a,i,j,b , y a,i,j,b , w a,i,j,b , h a,i,j,b are the true bounding box parameters; f a,i,j,b , t a,i,j,b are the true waterway parameters (flow rate and time), λ conf , λ loc , λ param are hyperparameters, recommended to be set as 6, 3, 1;

[0046] The step 2.4: Calculate the gradient of the hybrid loss function with respect to the network parameters through the backpropagation algorithm, and use the AdamW optimizer (initial learning rate of 3e-4, weight decay of 0.05) to update the network weights and bias terms. The training is set with a batch size of 32 and a total of 300 iterations until the loss converges.

[0047] Furthermore, the step 3.1: Obtain accurate temperature field distribution data, specifically: In the actual mold manufacturing process, when designing or optimizing the waterway openings, use an infrared thermal imager to capture the thermal image of the mold during operation in real time, ensuring that the installation position and field of view angle of the infrared thermal imager are consistent with those in the training stage and covering the entire target area to obtain accurate temperature field distribution data;

[0048] The step 3.2: Preprocess the captured thermal image according to step 1.2, including scaling to X×X pixels (recommended to be 512×512), temperature normalization, etc., to obtain the preprocessed thermal image p new , and input p new into the trained deep learning model to obtain the predicted output whose shape is S×S×(B×7), and parse out the parameters according to the method provided in step 2.2. For each grid (i, j), B bounding boxes are predicted, and each bounding box contains 7 parameters:

[0049] Furthermore, the step 3.3: Screening and unique indexing of the predicted boxes are specifically as follows:

[0050] Set the confidence threshold τ = 0.5, and only retain the prediction boxes. Sort all the prediction boxes in descending order of confidence, and use the non-maximum suppression algorithm to screen the boxes. The traditional algorithm uses the intersection over union (IoU). The calculation method of IoU in the present invention is improved as follows:

[0051] For two prediction boxes m and n:

[0052]

[0053] where IoU(m,n) measures the spatial overlap degree of the boxes, and the specific formula is where Area(m∩n) represents the area of the overlapping part of the two bounding boxes, and Area(m∪n) represents the total area covered by the two bounding boxes; is the calculation formula for the flow difference, where f max -f min is the difference between the maximum and minimum values of the flow in the training data; Set α = 0.7, β = 0.3, and finally, through the improved non-maximum suppression algorithm, after processing, we get K k ={(xx k ,yy k ,ww k ,hh k ,ff k ,tt k )} where k represents the unique index of the retained prediction box.

[0054] Furthermore, in step 3.4: According to the output of step 3.3, calculate the parameters, specifically:

[0055] According to the output of 3.3, calculate the following parameters according to the corresponding responsible grid (i k ,j k ):

[0056] The upper left coordinate of the grid:

[0057] The coordinate of the center point relative to the image:

[0058] where W and H are the width and height of the actual mold;

[0059] The size: w′ k = ww k ×W, h′ k = hh k ×H

[0060] The flow and time: f′ k = f min + ffk ×(f max -f min ), t′ k =t min + tt k ×(t max - t min )

[0061] where f max , f min , t max , t min are the maximum and minimum flow rates and times in the training data;

[0062] Step 3.5: Use the calculated position coordinates to determine the exact opening positions on the mold, specifically:

[0063] Use the calculated position coordinates to determine the exact opening positions on the mold; use the calculated opening sizes to determine the sizes of the water channel openings, and give this position and size information to the mold design or processing personnel for manufacturing new molds or modifying existing molds;

[0064] Use the calculated flow rate values to set the flow control valves or water pumps for the corresponding water channels, and use the calculated water passing time values to set the opening duration of the corresponding water channel solenoid valves in the control program (such as PLC) of the production equipment.

[0065] Compared with the prior art, the beneficial effects of the present invention are:

[0066] This automatic water channel parameter generation system based on thermal image analysis and deep learning reduces the number of industrial trial moldings by optimizing the design of the mold cooling water channel positions and the setting of water channel process parameters, and is especially suitable for manufacturing scenarios with strict requirements for mold temperature field control such as precision die casting and injection molding;

[0067] 1. Further, collect the temperature field distribution data of the mold during operation through an infrared thermal imager, and use the simultaneously recorded water channel parameters (including the center coordinates of the water pipe openings, opening sizes, flow rates, and water passing times) as supervision labels to train the deep learning model to automatically learn and predict the optimal water channel parameters from the thermal images;

[0068] 2. Further, the system uses an improved CSPDarknet53 network as the backbone to extract temperature field features and generate the optimal water channel parameters;

[0069] 3. Further, innovatively design a hybrid loss function for jointly optimizing the prediction of water channel positions, sizes, flow rates, and water passing times;

[0070] 4. Further, the method includes steps such as data acquisition and preprocessing, model architecture design, hybrid loss function design, model training, and post-inference processing, and finally outputs an optimized set of water channel parameters. Description of the Drawings

[0071] Figure 1 This is a schematic diagram of the overall system flow of the present invention. Detailed Embodiments

[0072] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0073] Embodiment 1: Please refer to Figure 1 , the present invention provides the following technical solutions: An automatic water channel parameter generation system based on thermal image analysis and deep learning, the system includes the following steps: Step 1, acquire industrial data and perform preprocessing; Step 2, train the network; Step 3, online prediction of the model;

[0074] Step 1 in the system further includes the following steps:

[0075] Step 1.1: Acquire thermal image parameters and water channel parameters, and subsequently associate them with the network target box;

[0076] Step 1.2: Acquire the target box and the associated thermal image annotation data set;

[0077] Step 1.3: Preprocess the acquired thermal image annotation data set:

[0078] Step 2 in the system further includes the following steps:

[0079] Step 2.1: Construct the backbone network structure:

[0080] Step 2.2: Parse the elements in the output:

[0081] Step 2.3: Design the hybrid loss function;

[0082] Step 2.4: Calculate the gradient of the hybrid loss function with respect to the network parameters through the backpropagation algorithm;

[0083] Step 3 in the system further includes the following steps:

[0084] Step 3.1: Acquire accurate temperature field distribution data;

[0085] Step 3.2: Preprocess the collected thermal images according to Step 1.2;

[0086] Step 3.3: Screening and unique indexing of prediction boxes;

[0087] Step 3.4: Calculate parameters based on the output of Step 3.3;

[0088] Step 3.5: Use the calculated position coordinates to determine the exact opening position on the mold.

[0089] In Step 1.1, obtain the thermal image parameters and water channel parameters. The subsequent association with the network target boxes is mainly as follows: Use an infrared thermal imager, fixedly installed directly above the die-casting or injection mold to be measured, ensure that the field of view covers the entire target area, collect the thermal images of the mold surface during operation, select different mold types, different water valve position settings, different cooling parameters, different production batches, etc., and randomly set parameters such as flow rate and water passing time to construct a total of N (N>20000) sets of classical working conditions. Obtain the thermal image data at the time of process stability for each working condition once, denoted as p a (a = 1, 2, …, N), and finally obtain P = {p1, p2, …, p N} pictures, and record the water channel parameters (flow rate and water passing time) at that time through the PLC, and subsequently associate with the grid target boxes.

[0090] Step 1.2: The main methods for obtaining the target box and the associated thermal image annotation dataset are as follows: Uniformly scale the thermal image collection P to X×X pixels (recommended to be 512×512), convert the original radiation value to a temperature value (unit: °C), and linearly normalize the temperature based on global statistics to [0, 1]; Manually mark the position where the water channel projects onto the thermal image on each image p a ; Specifically, use a rectangular annotation box to mark the position of the water channel, and call this rectangular box the target box. Set the number of target boxes marked on the thermal image to be represented by b (b = 0, 1, 2), and at most B = 3 target boxes can be marked in each grid (it is necessary to ensure that the annotation box is the smallest matrix that can enclose the target when annotating). Divide the image p a into (recommended to be 8×8) grids, and assign a unique coordinate (i, j) to each grid, where i represents the row index (range from 0 to S - 1), and j represents the column index for each target box (the range is also from 0 to S - 1).

[0091] For the target box, according to the situation of the target box, set the following parameters in each grid of the a-th thermal image:

[0092] Confidence c a,i,j,b : If the b-th target box exists in the grid (i, j), then c a,i,j,b = 1, otherwise c a,i,j,b = 0

[0093] Bounding box parameter x a,i,j,b , y a,i,j,b , w a,i,j,b , h a,i,j,b : where x a,i,j,b is the horizontal offset of the center point of the target box relative to the upper left corner of the grid (i, j); y a,i,j,b is the vertical offset of the center point of the target box relative to the upper left corner of the grid (i, j); w a,i,j,b is the width of the target box; h a,i,j,b is the height of the target box;

[0094] Waterway parameter f a,i,j,b , t a,i,j,b : where f a,i,j,b represents the waterway flow rate corresponding to this target box (waterway opening); t a,i,j,b represents the waterway water passing time corresponding to this target box (waterway opening);

[0095] If there is no target in the grid, then this target box c a,i,j,b = 0, and other parameters are also set to 0;

[0096] Repeat the above annotation process to finally obtain the entire dataset:

[0097] where N is the number of thermal images, and B = 3 is the number of predictions for each grid.

[0098] For the dataset, preprocess the thermal image annotation dataset Q; linearly map the bounding box parameters x a,i,j,b , y a,i,j,b , w a,i,j,b , h a,i,j,b to the range (0, 1) based on the size of the thermal image; scale the waterway parameters f a,i,j,b , t a,i,j,b from the maximum and minimum values of the actual process parameters to (0, 1); repeat the above steps to finally obtain the preprocessed dataset where represents the annotation information after preprocessing corresponding to the a-th picture q a ;

[0099] Example 2:

[0100] On the basis of Example 1, it is also disclosed that:

[0101] Step 2.1: Construct the backbone network structure as follows:

[0102] The output shape of the input layer is X×X×1 (X is recommended to be 512), which is equal to the size of the thermal image;

[0103] The first - layer Stem convolution, with a convolution kernel size of 3×3, a stride of 2, 32 output channels, and using the SiLU activation function;

[0104] The second - layer CSP module, with 1 residual unit (including 1×1 and 3×3 convolutions), and the number of channels is compressed to 64;

[0105] The third - layer Transition layer, with a convolution kernel size of 3×3, a stride of 2, and 64 output channels;

[0106] The fourth - layer CSP module, with 1 residual unit (including 1×1 and 3×3 convolution with a dilation rate of 2), and the number of channels is compressed to 128;

[0107] The fifth - layer Transition layer, with a convolution kernel size of 3×3, a stride of 2, and 128 output channels;

[0108] The sixth - layer SPP fast pooling, with 5×5, 9×9, and 13×9 parallel pooling;

[0109] The seventh - layer up - sampling layer, with 2 - fold up - sampling followed by a 1×1 convolution, and 128 output channels;

[0110] The eighth - layer cross - layer feature fusion, concatenating the output of the up - sampling layer and the fourth - layer output;

[0111] The ninth - layer down - sampling layer, with a convolution kernel size of 4×4 and a stride of 4;

[0112] The ninth - layer prediction head, with a 1×1 convolution, and the output is S×S×(B×7) (recommended S = 8, B = 3).

[0113] Step 2.2: Parse the elements in the output. Specifically, for each grid (i, j) and each of the 7 parameters of each bounding box b, they are respectively set to where is the predicted confidence; are the predicted bounding - box parameters; are the predicted water - channel parameters;

[0114] Step 2.3: Design of the hybrid loss function: The hybrid loss function is used to supervise the prediction accuracy of the model for the water - channel position, size, and parameters. The specific formula is as follows:

[0115]

[0116] where S is the grid size 8, B is the number of bounding boxes predicted for each grid cell 3, N is the number of images, a is the image index, i, j are the grid indices, b is the bounding - box index, c a,i,j,b represents the true confidence (1 means there is a target, 0 means there is no target); x a,i,j,b , y a,i,j,b , wa,i,j,b , h a,i,j,b is the true bounding box parameter; f a,i,j,b , t a,i,j,b is the true water channel parameter (flow rate and time), λ conf , λ loc , λ param is a hyperparameter, and it is recommended to be set to 6, 3, 1;

[0117] Step 2.4: Calculate the gradient of the hybrid loss function with respect to the network parameters through the backpropagation algorithm, and use the AdamW optimizer (initial learning rate 3e-4, weight decay 0.05) to update the network weights and bias terms. The training is set with a batch size of 32 and a total of 300 iterations until the loss converges.

[0118] Example Three:

[0119] Based on Example Two, it is also disclosed that:

[0120] Step 3.1: Obtain accurate temperature field distribution data. Specifically, during the actual mold manufacturing process, when designing or optimizing the water channel openings, use an infrared thermal imager to capture the thermal image of the mold in real time, ensuring that the installation position and field of view angle of the infrared thermal imager are consistent with those in the training stage and covering the entire target area to obtain accurate temperature field distribution data;

[0121] Step 3.2: Preprocess the captured thermal image according to Step 1.2, including scaling to X×X pixels (recommended to be 512×512), temperature normalization, etc., to obtain the preprocessed thermal image p new , and input p new into the trained deep learning model to obtain the predicted output whose shape is S×S×(B×7). Parse the parameters according to the method provided in Step 2.2, predict B bounding boxes for each grid (i, j), and each bounding box contains 7 parameters:

[0122] Step 3.3: Screening of predicted boxes and unique indexing is specifically as follows:

[0123] Set the confidence threshold τ = 0.5, and only retain the predicted boxes. Sort all the predicted boxes in descending order of confidence, and use the non-maximum suppression algorithm to screen the boxes. The traditional algorithm uses the intersection over union (IoU). The calculation method of the intersection over union in the present invention is improved as follows:

[0124] For two predicted boxes m and n:

[0125]

[0126] Among them, IoU(m,n) measures the spatial overlap degree of the bounding boxes, and the specific formula is where Area(m∩n) represents the area of the overlapping part of the two bounding boxes, and Area(m∪n) represents the total area covered by the two bounding boxes; is the calculation formula for the flow difference, where f max -f min is the difference between the maximum and minimum values of the flow in the training data; set α = 0.7, β = 0.3, and finally, through the improved non-maximum suppression algorithm, after processing, we get K k ={(xx k , yy k , ww k , hh k , ff k , tt k )} where k represents the unique index of the retained prediction box.

[0127] Step 3.4: According to the output of Step 3.3, calculate the parameters, specifically:

[0128] According to the output of 3.3, for the corresponding responsible grid (i k , j k ), calculate the following parameters:

[0129] Coordinates of the upper left corner of the grid:

[0130] Coordinates of the center point relative to the image:

[0131] where W and H are the width and height of the actual mold;

[0132] Dimensions: w′ k = ww k ×W, h′ k = hh k ×H

[0133] Flow and time: f′ k = f min + ff k ×(f max - f min ), t′ k = t min + tt k ×(t max - t min )

[0134] where f max , f min , t max , t minThe maximum and minimum flow rates and times in the training data;

[0135] Step 3.5: Use the calculated position coordinates to determine the exact hole-opening positions on the mold, specifically:

[0136] Use the calculated position coordinates to determine the exact hole-opening positions on the mold; use the calculated opening sizes to determine the sizes of the water-channel openings, and submit this position and size information to the mold designers or processors for manufacturing a new mold or modifying an existing mold;

[0137] Use the calculated flow rate values to set the flow control valves or water pumps for the corresponding water channels, and use the calculated water-passing time values to set the opening durations of the solenoid valves for the corresponding water channels in the control program (such as PLC) of the production equipment.

[0138] The content not described in detail in this specification belongs to the prior art well-known to those skilled in the art.

[0139] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. An automatic generation system for waterway parameters based on thermal image analysis and deep learning, characterized in that The system includes the following steps: Step 1, obtain industrial data and perform preprocessing; Step 2, train the network; Step 3, perform online prediction of the model; Step 1 in the system further includes the following steps: Step 1.1: Obtain the thermal image parameters and water channel parameters, and subsequently associate them with the network target boxes; Step 1.2: Obtain the target boxes and the associated thermal image annotation datasets; Step 1.3: Perform preprocessing on the obtained thermal image annotation datasets: Step 2 in the system further includes the following steps: Step 2.1: Construct the backbone network structure: Step 2.2: Parse the elements in the output: Step 2.3: Design the hybrid loss function; Step 2.4: Calculate the gradient of the hybrid loss function with respect to the network parameters through the backpropagation algorithm; Step 3 in the system further includes the following steps: Step 3.1: Obtain accurate temperature field distribution data; Step 3.2: Preprocess the collected thermal images according to Step 1.2; Step 3.3: Screening and unique indexing of the prediction boxes; Step 3.4: Calculate the parameters according to the output of Step 3.3; Step 3.5: Use the calculated position coordinates to determine the precise opening positions on the mold.

2. The automatic generation system for waterway parameters based on thermal image analysis and deep learning according to claim 1, characterized in that: In step 1.1, the thermal image parameters and water channel parameters are obtained. The subsequent association with the network target box mainly includes: using an infrared thermal imager, fixedly installed directly above the die casting or injection molding die to be measured, ensuring that the field of view covers all target areas, collecting the thermal images of the die surface during operation, selecting different die types, different water valve position settings, different cooling parameters, different production batches, etc., and randomly setting parameters such as flow rate and water passing time to construct a total of N (N>20000) groups of classical working conditions. Each time a working condition is obtained, the thermal image data at process stability is recorded as p a (a = 1, 2, …, N), and finally P = {p1, p2, …, p N} pictures are obtained, and the water channel parameters (flow rate and water passing time) at that time are recorded through the PLC, and then associated with the grid target box.

3. The automatic waterway parameter generation system based on thermal image analysis and deep learning according to claim 1, characterized in that: Step 1.2: Obtaining the target boxes and the associated thermal image annotation datasets mainly involves: uniformly scaling the thermal image collection P to X×X pixels (recommended to be 512×512), converting the original radiation value to a temperature value (unit: °C), and linearly normalizing the temperature to [0,1] based on the global statistics; Manually mark the position of the water channel on each image p a after it is projected onto the thermal image; specifically, use a rectangular annotation box to mark the position of the water channel, and call this rectangular box the target box. Let the number of target boxes marked on the thermal image be represented by b (b = 0, 1, 2), and each grid can mark at most B = 3 target boxes (when annotating, it is necessary to ensure that the annotation box is the smallest matrix that can enclose the target). Divide the image p a into S×S (recommended to be 8×8) grids, and assign a unique coordinate (i, j) to each grid, where i represents the row index (ranging from 0 to S - 1), and j represents the column index for each target box (also ranging from 0 to S - 1).

4. The automatic waterway parameter generation system based on thermal image analysis and deep learning according to claim 3, characterized in that: For the target boxes, according to the situation of the target boxes, set the following parameters in each grid of the a-th thermal image: Confidence level c a,i,j,b If there is a b-th bounding box within the grid (i, j), then c a,i,j,b = 1, otherwise c a,i,j,b = 0. Bounding box parameters x a,i,j,b , y a,i,j,b , w a,i,j,b , h a,i,j,b : where x a,i,j,b is the horizontal offset of the center point of the bounding box relative to the upper left corner of the grid (i, j); y a,i,j,b is the vertical offset of the center point of the bounding box relative to the upper left corner of the grid (i, j); w a,i,j,b is the width of the bounding box; h a,i,j,b is the height of the bounding box; Waterway parameter f a,i,j,b , t a,i,j,b : where f a,i,j,b represents the waterway flow rate corresponding to the target box (waterway opening); t a,i,j,b represents the water passing time of the waterway corresponding to the target box (waterway opening); If there is no target in the grid, the target box c a,i,j,b = 0, and other parameters are also set to 0; Repeat the above annotation process, and finally obtain the entire dataset: Where N is the number of thermograms, and B = 3 is the number of predictions for each grid.

5. The automatic waterway parameter generation system based on thermal image analysis and deep learning according to claim 4, characterized in that: For the dataset, preprocess the thermal image annotation dataset Q; linearly map the bounding box parameters x a,i,j,b , y a,i,j,b , w a,i,j,b , h a,i,j,b to the range (0, 1) based on the thermal image size; scale the waterway parameters f a,i,j,b , t a,i,j,b from the maximum and minimum values of the actual process parameters to (0, 1); Repeat the above steps to finally obtain the preprocessed dataset wherein represents the q-th image of the a-th image a corresponding annotation information after preprocessing 6. The automatic generation system of waterway parameters based on thermal image analysis and deep learning according to claim 1, characterized in that: Step 2.1: The backbone network structure is constructed as follows: The output shape of the input layer is X×X×1 (X is recommended to be 512), which is equal to the size of the thermal image; The first layer is the Stem convolution, with a convolution kernel size of 3×3, a stride of 2, 32 output channels, and the SiLU activation function is used; The second layer is the CSP module, with 1 residual unit (including 1×1 and 3×3 convolutions), and the channels are compressed to 64; The third layer is the Transition layer, with a convolution kernel size of 3×3, a stride of 2, and 64 output channels; The fourth layer is the CSP module, with 1 residual unit (including 1×1 and 3×3 convolutions with a dilation rate of 2), and the channels are compressed to 128; The fifth layer is the Transition layer, with a convolution kernel size of 3×3, a stride of 2, and 128 output channels; The sixth layer is the SPP fast pooling, with 5×5, 9×9, and 13×9 parallel pooling; The seventh layer is the upsampling layer, with 2-fold upsampling followed by a 1×1 convolution, and 128 output channels; The eighth layer is the cross-layer feature fusion, splicing the output of the upsampling layer and the fourth layer; The ninth layer is the downsampling layer, with a convolution kernel size of 4×4 and a stride of 4; The ninth layer is the prediction head, with a 1×1 convolution, and the output is S×S×(B×7) (recommended S = 8, B = 3) 7. The automatic waterway parameter generation system based on thermal image analysis and deep learning according to claim 1, characterized in that: Step 2.2: Parse the elements in the output. Specifically, for each grid (i, j) and each of the 7 parameters of each bounding box b, set them to where is the predicted confidence; are the predicted bounding box parameters; are the predicted waterway parameters; 2.3: Design of the hybrid loss function: The hybrid loss function is used to supervise the prediction accuracy of the model for the water channel positions, sizes, and parameters. The specific formula is as follows: where S is the grid size of 8, B is the number of bounding boxes predicted per grid cell of 3, N is the number of images, a is the image index, i, j are the grid indices, b is the bounding box index, c a,i,j,b represents the ground truth confidence (1 means there is an object, 0 means there is no object); x a,i,j,b , y a,i,j,b , w a,i,j,b , h a,i,j,b are the ground truth bounding box parameters; f a,i,j,b , t a,i,j,b are the ground truth watercourse parameters (flow rate and time), λ conf , λ loc , λ param are hyperparameters, and are recommended to be set to 6, 3, 1; Step 2.4: Calculate the gradients of the hybrid loss function with respect to the network parameters through the backpropagation algorithm, and use the AdamW optimizer (initial learning rate 3e-4, weight decay 0.05) to update the network weights and bias terms. The training is set with a batch size of 32 and a total of 300 iterations until the loss converges.

8. The automatic generation system for waterway parameters based on thermal image analysis and deep learning according to claim 1, characterized in that: Step 3.1: Obtain accurate temperature field distribution data, specifically: during the actual mold manufacturing process, when the water channel openings need to be designed or optimized, use an infrared thermal imager to capture the thermal image of the mold in real time, ensuring that the installation position and field of view angle of the infrared thermal imager are consistent with those in the training stage and covering the entire target area to obtain accurate temperature field distribution data; Step 3.2: Preprocess the captured thermal image according to Step 1.2, including scaling it to X×X pixels (recommended to be 512×512), temperature normalization, etc., to obtain the preprocessed thermal image p new , and input p new into the trained deep learning model to obtain the predicted output whose shape is S×S×(B×7). Parse out the parameters according to the method provided in Step 2.

2. For each grid (i, j), B bounding boxes are predicted, and each bounding box contains 7 parameters:

9. The automatic waterway parameter generation system based on thermal image analysis and deep learning according to claim 1, wherein: Step 3.3: The screening and unique indexing of the prediction boxes are specifically as follows: Set the confidence threshold τ = 0.5, and only retain the predicted bounding boxes. Sort all the predicted bounding boxes in descending order of confidence, and use the non-maximum suppression algorithm to screen the boxes. The traditional algorithm uses the intersection over union (IoU). The present invention improves the calculation method of the intersection over union as follows: For two prediction boxes m and n: Among them, IoU(m,n) measures the spatial overlap degree of the bounding boxes, and the specific formula is Among them, Area(m∩n) represents the area of the overlapping part of the two bounding boxes, and Area(m∪n) represents the total area covered by the two bounding boxes; is the calculation formula for the flow difference, where f max -f min is the difference between the maximum and minimum values of the flow in the training data; Set α = 0.7, β = 0.3, and finally, through the improved non-maximum suppression algorithm, K is obtained after processing k ={(xx k ,yy k ,ww k ,hh k ,ff k ,tt k )} where k represents the unique index of the retained prediction box.

10. The automatic waterway parameter generation system based on thermal image analysis and deep learning according to claim 1, characterized in that: Step 3.4: According to the output of Step 3.3, calculate the parameters, specifically: According to the output in 3.3, calculate the following parameters based on the corresponding responsible grid (i k ,j k ): Grid upper left corner coordinates: The coordinates of the center point relative to the image: where W and H are the width and height of the actual mold; Size: w′ k = ww k × W, h′ k = hh k × H Flow rate and time: f′ k = f min + ff k × (f max - f min ), t′ k = t min + tt k × (t max - t min ) where f max , f min , t max , t min are the maximum and minimum flow rates and times in the training data; Step 3.5: Use the calculated position coordinates to determine the precise opening positions on the mold, specifically: Use the calculated position coordinates to determine the precise opening positions on the mold; use the calculated opening sizes to determine the sizes of the water channel openings, and hand over this position and size information to the mold design or processing personnel for manufacturing new molds or modifying existing molds; Use the calculated flow rate values to set the flow control valves or water pumps for the corresponding water circuits, and use the calculated water passing time values to set the opening duration of the corresponding water circuit solenoid valves in the control program (such as PLC) of the production equipment.

Images