PTFE emulsion liquid level cooperative detection method, device and equipment and storage medium

Abstract

The invention discloses a PTFE emulsion liquid level cooperative detection method, device and equipment and a storage medium. The method comprises the steps that liquid level images of a plurality of key point positions on a PTFE emulsion production line are obtained; performing image enhancement and preprocessing on the image data to obtain preprocessed data corresponding to each key point location; inputting the preprocessed data of each key point location into a trunk feature extraction network to obtain a feature map corresponding to each key point location; performing cross-point fusion on the feature maps corresponding to the key points based on a cross-point attention mechanism to obtain cross-point fusion feature maps corresponding to the key points; performing global semantic modeling on the enhanced feature map corresponding to each key point location, and outputting a global semantic feature corresponding to each key point location; and inputting the feature vector corresponding to each key point into a multi-task learning model, and outputting the liquid level state of each key point. The PTFE emulsion liquid level abnormal response capability and the real-time detection level of the system are enhanced.

Description

Technical Field

[0001] This invention relates to the field of PTFE emulsion level detection technology, and in particular to a method, apparatus, equipment and storage medium for synergistic detection of PTFE emulsion level. Background Technology

[0002] Polytetrafluoroethylene (PTFE) is a core chemical product in aerospace material manufacturing, and its emulsion preparation process requires extremely high precision in process parameters. In the PTFE suspension filtration stage, vibrating screen equipment needs to separate coagulated particles through vibration to obtain pure emulsion.

[0003] In the early stages of PTFE emulsion filtration, production relied on manual monitoring. Operators had to be stationed at the equipment in real time, visually judging the liquid level in the vibrating screen and manually adjusting the feed rate. This method was not only labor-intensive but also had significant drawbacks: if operators failed to respond promptly, PTFE emulsion could easily overflow untreated for extended periods, leading to raw material waste, equipment contamination, and even safety hazards. In modern industrial production systems, accurate and real-time liquid level monitoring is crucial for ensuring production efficiency and product quality, and traditional monitoring methods are no longer sufficient to meet practical needs.

[0004] With the development of deep learning technology, many deep learning-based liquid level detection methods have emerged. Most existing deep learning-based liquid level detection methods adopt a single-point modeling strategy, which fails to fully explore the collaborative patterns between key points (such as vibrating screens, emulsion tanks, and waste tanks). Especially when abnormal conditions trigger linkage (such as emulsion tank overflow caused by vibrating screen blockage), single-point prediction cannot accurately identify systemic risks, resulting in detection limitations.

[0005] Therefore, a method, device, equipment, and storage medium for the coordinated detection of PTFE emulsion level were developed to solve the above problems. Summary of the Invention

[0006] This invention proposes a method, device, equipment, and storage medium for the coordinated detection of PTFE emulsion levels, in order to solve the problem that existing single-point prediction cannot accurately identify systemic risks, resulting in detection limitations.

[0007] The present invention achieves the above objectives through the following technical solutions:

[0008] This invention provides a method for synergistic detection of PTFE emulsion liquid level, comprising:

[0009] Acquire image data, which includes liquid level images at several key points on the PTFE emulsion production line;

[0010] The image data is enhanced and preprocessed to obtain preprocessed data corresponding to each key point.

[0011] The preprocessed data of each key point is input into the backbone feature extraction network to obtain the feature map corresponding to each key point.

[0012] Based on the cross-point attention mechanism, the feature maps corresponding to each key point are fused across points to obtain the cross-point fused feature map corresponding to each key point.

[0013] Perform global semantic modeling on the enhanced feature maps corresponding to each key point and output the global semantic features corresponding to each key point.

[0014] The feature vectors corresponding to each key point are input into the multi-task learning model, and the liquid level status of each key point is output. The multi-task learning model includes several independent task heads. Each independent task head is used to classify the liquid level status of the feature vectors corresponding to each key point. Each independent task head is obtained by jointly training all independent task heads through a preset multi-task joint loss function.

[0015] Further, the image data undergoes image enhancement and preprocessing, including:

[0016] Perform geometric transformations on the image data;

[0017] Color perturbation is applied to the image data;

[0018] The image data is then subjected to regularization processing.

[0019] Furthermore, the backbone feature extraction network is a ResNet network architecture.

[0020] Furthermore, based on a cross-point attention mechanism, feature maps corresponding to each key point are fused across points, including:

[0021] Flatten the feature maps corresponding to each key point into two-dimensional feature sequences;

[0022] For each key location, construct three matrices: query, key, and value.

[0023] The dot product attention mechanism is used to compute a weighted representation of the two-dimensional feature sequence of each key point with respect to the two-dimensional feature sequences of all other key points.

[0024] Calculate the two-dimensional feature sequence of each key point and the sum of its weighted representation with respect to all other key points to obtain the cross-point fusion feature map corresponding to each key point.

[0025] Furthermore, global semantic modeling is performed on the enhanced feature maps corresponding to each key point, outputting the global semantic features corresponding to each key point, including:

[0026] Flatten the cross-point fusion feature map corresponding to each key point, and divide the flattened feature map into several image blocks;

[0027] Perform a linear transformation on each image patch;

[0028] By inputting all the linearly transformed image patches corresponding to each key point into the Vision Transformer model, the global semantic features corresponding to each key point are obtained.

[0029] Furthermore, the training process for the independent task head is as follows:

[0030] A multi-task learning model is constructed, which includes an independent task head corresponding to each key point. The classifier expression for each independent task head is as follows: ;

[0031] in Represents a probability vector. This represents the weight of the fully connected layer for the classification task of the i-th key point. Z represents the bias term parameter for the classification task of the i-th key point, and Z represents the global semantic feature corresponding to the i-th key point.

[0032] Construct a multi-task joint loss function, which is a weighted sum of all single-task loss functions. The single-task loss function L... i The expression is as follows: ;

[0033] In the formula, N is the batch sample size. Let c be the true label of the nth sample at point i. This represents the predicted probability.

[0034] The multi-task learning model was optimized based on the training samples with the goal of minimizing the joint loss function of the multi-tasks. The network parameters were adjusted using gradient descent and backpropagation. Finally, the model with the highest overall accuracy in the validation set was selected as the optimal model for actual deployment. The training samples included liquid level image samples of several key points on the PTFE emulsion production line labeled with three-class labels, including normal label, overflow label, and occlusion label.

[0035] Further, image data is acquired, including:

[0036] Acquire liquid level images from programmable industrial cameras deployed at key locations in a PTFE emulsion production line, including vibrating screens, emulsion tanks, and waste tanks.

[0037] The acquired liquid level images are processed into frame slices at a uniform frame rate, and each set of images is composed of a time-consistent image tuple for all key locations.

[0038] The present invention also provides a PTFE emulsion level detection device, comprising:

[0039] The acquisition module is used to acquire image data, which includes liquid level images of several key points on the PTFE emulsion production line.

[0040] The preprocessing module is used to perform image enhancement and preprocessing on the image data to obtain preprocessed data corresponding to each key point.

[0041] The feature extraction module is used to input the preprocessed data of each key point into the backbone feature extraction network to obtain the feature map corresponding to each key point.

[0042] The fusion module is used to perform cross-point fusion on the feature maps corresponding to each key point based on the cross-point attention mechanism to obtain the cross-point fused feature map corresponding to each key point.

[0043] The full semantic modeling module is used to perform global semantic modeling on the enhanced feature maps corresponding to each key point and output the global semantic features corresponding to each key point.

[0044] The detection module is used to input the feature vectors corresponding to each key point into a multi-task learning model and output the liquid level status of each key point. The multi-task learning model includes several independent task heads, each of which is used to classify the liquid level status of the feature vectors corresponding to each key point. Each independent task head is obtained by jointly training all independent task heads through a preset multi-task joint loss function.

[0045] This invention also provides a method for synergistic detection of PTFE emulsion level, comprising:

[0046] Memory, used to store computer programs;

[0047] A processor is configured to implement the steps of the PTFE emulsion level co-detection method when executing the computer program.

[0048] The present invention also provides a storage medium, which is a readable storage medium, and stores a computer program thereon. When the computer program is executed by a processor, it implements the steps of the PTFE emulsion level collaborative detection method.

[0049] The beneficial effects of this invention are as follows:

[0050] This invention proposes a unified framework for data acquisition, feature extraction, feature fusion, and classification prediction in a collaborative detection method, device, equipment, and storage medium for PTFE emulsion levels. By constructing a cross-point attention mechanism, deep interaction of image information from different points is achieved. Furthermore, a multi-task learning model is used to simultaneously optimize the classification task for multiple points, thereby improving the model's ability to represent complex states and its generalization performance. This method is suitable for multi-point level monitoring needs in industrial settings. It enables efficient end-to-end inference in resource-constrained edge computing environments, enhancing the anomaly response capability and real-time detection level of PTFE emulsion levels, and providing reliable technical support for intelligent monitoring and control of PTFE emulsion production lines. Attached Figure Description

[0051] Figure 1 This is a flowchart illustrating the method, apparatus, equipment, and storage medium for synergistic detection of PTFE emulsion level according to this application.

[0052] Figure 2 This is a flowchart illustrating the feature extraction, cross-location feature fusion, and multi-task learning methods in the embodiments of this application.

[0053] Figure 3 This is a schematic diagram of the test results in the embodiments of this application. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0055] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0056] In their practical research, the inventors discovered that single-point modeling strategies only monitor the liquid level at a single point in the emulsion tank or vibrating screen, failing to capture the correlation of features across multiple points in the production process. For example, when the emulsion tank overflows, the vibrating screen is often accompanied by an abnormal phenomenon of excessively high flow rate; the accumulation of solid waste on the vibrating screen plate often indicates a low content of pure emulsion in the material, leading to the overflow of the waste tank. Single-point detection cannot effectively utilize the correlated feature information between images of each point. Training based solely on the features of each independent point often requires a large sample size, increasing training costs. Independent single-point detection requires deploying three models for independent inference, which also places higher demands on hardware equipment.

[0057] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0058] like Figure 1 As shown, a method, apparatus, equipment, and storage medium for synergistic detection of PTFE emulsion level include:

[0059] S1: Acquire image data, which includes liquid level images of several key points on the PTFE emulsion production line;

[0060] In one embodiment, acquiring image data includes:

[0061] Acquire liquid level images from programmable industrial cameras deployed at key locations in a PTFE emulsion production line, including vibrating screens, emulsion tanks, and waste tanks.

[0062] The acquired liquid level images are processed into frame slices at a uniform frame rate, and each set of images is composed of a time-consistent image tuple for all key locations.

[0063] Three programmable industrial cameras are deployed at key locations in the PTFE emulsion production line (vibrating screen, emulsion tank, and waste tank), and images are synchronously acquired using a unified time control signal. The acquired images are processed into frame slices at a uniform frame rate, and each set of images forms a time-consistent image triplet for three points, ensuring that the model can jointly analyze the status of each point at the same time.

[0064] S2: Perform image enhancement and preprocessing on the image data to obtain preprocessed data corresponding to each key point;

[0065] In one embodiment, image enhancement and preprocessing of the image data includes:

[0066] Perform geometric transformations on the image data;

[0067] Color perturbation is applied to the image data;

[0068] The image data is then subjected to regularization processing.

[0069] To enhance the model's robustness to complex on-site environments (such as camera offset, liquid level reflection, and image blur), various image enhancement and regularization techniques are employed. These include geometric transformations (horizontal flipping, translation, rotation, and cropping), color perturbations (hue, saturation, and channel order adjustments), and regularization methods (Gaussian noise addition, random occlusion, batch regularization, etc.) to enhance data distribution diversity and improve training stability and generalization ability.

[0070] S3: Input the preprocessed data of each key point into the backbone feature extraction network to obtain the feature map corresponding to each key point;

[0071] In one embodiment, the backbone feature extraction network is a ResNet network architecture. The three images are input into a unified backbone feature extraction network (using the ResNet architecture) to obtain low-level feature maps with spatial hierarchical structure. The network structure employs residual connections and multi-scale feature extraction mechanisms, which not only preserve local edge and texture information but also alleviate the gradient vanishing problem through skip connections, providing a stable semantic representation for the subsequent fusion module.

[0072] S4: Based on the cross-point attention mechanism, the feature maps corresponding to each key point are fused across points to obtain the cross-point fused feature map corresponding to each key point.

[0073] In one embodiment, cross-point fusion of feature maps corresponding to each key point is performed based on a cross-point attention mechanism, including:

[0074] Flatten the feature maps corresponding to each key point into two-dimensional feature sequences;

[0075] For each key location, construct three matrices: query, key, and value.

[0076] The dot product attention mechanism is used to compute a weighted representation of the two-dimensional feature sequence of each key point with respect to the two-dimensional feature sequences of all other key points.

[0077] Calculate the two-dimensional feature sequence of each key point and the sum of its weighted representation with respect to all other key points to obtain the cross-point fusion feature map corresponding to each key point.

[0078] The three feature maps are flattened into a sequence, and the feature response weights between different points are calculated using a cross-point attention mechanism. A query, key, and value matrix is ​​constructed for each point, and feature correlations are calculated based on a dot product attention mechanism. Information from other points is then weighted and aggregated to generate enhanced fused features. This mechanism can explicitly model the physical and logical relationships between points (e.g., abnormal material feeding from a vibrating screen causing emulsion tank overflow).

[0079] Specifically, the image data from the three locations are first processed through a shared ResNet-50 backbone network for feature extraction. Assume the feature maps output from the three input images after passing through the backbone network are as follows: ;

[0080] here, These represent the output feature maps of the vibrating screen, waste bin, and emulsion bin, respectively, where H and W represent the spatial dimensions of the feature map, and D represents the number of channels.

[0081] Next, each feature map is transformed into a two-dimensional feature sequence to enable cross-location information interaction using an attention mechanism:

[0082] ;

[0083] Subsequently, three matrices, Query(Q), Key(K), and Value(V), are constructed for each location to achieve cross-location feature fusion: In the formula, These are learnable linear mapping parameters used to uniformly map the feature space to the attention space. The attention mechanism for point i with respect to point j is defined as follows:

[0084] ;

[0085] In the formula, This is a scaling factor used to normalize the attention score. This normalization mechanism helps stabilize the training process and prevent gradient anomalies. Furthermore, the above formula clearly demonstrates the weight allocation mechanism by which feature i focuses on and fuses feature j.

[0086] For example, when focusing on the characteristics of the emulsion container (C) location, the specific fusion process is as follows:

[0087] ;

[0088] In the above formula The features obtained after the emulsion points are processed by the feature fusion module are enhanced using the above method, and the feature space of the emulsion bucket is enhanced, containing effective information from both the vibrating screen and the waste bucket. Similarly, the enhanced features of the vibrating screen and the waste bucket can be calculated, thereby completing comprehensive cross-point interaction and fusion of features.

[0089] S5: Perform global semantic modeling on the enhanced feature maps corresponding to each key point and output the global semantic features corresponding to each key point.

[0090] In one embodiment, global semantic modeling is performed on the enhanced feature maps corresponding to each key point, and global semantic features corresponding to each key point are output, including:

[0091] Flatten the cross-point fusion feature map corresponding to each key point, and divide the flattened feature map into several image blocks;

[0092] Perform a linear transformation on each image patch;

[0093] By inputting all the linearly transformed image patches corresponding to each key point into the Vision Transformer model, the global semantic features corresponding to each key point are obtained.

[0094] After completing the cross-location feature fusion, the transfer and subsequent processing of the fused features are crucial. Specifically, the fused features need to undergo appropriate transformations and structural processing before being fed into the Vision Transformer (ViT) for global semantic modeling. The detailed steps are as follows: Through a cross-location attention mechanism, the features of the vibrating screen, waste bin, and emulsion bin are used to calculate corresponding enhanced feature maps. Let the fused feature map for each location be:

[0095]

[0096] in, These are the feature maps of the vibrating screen, waste bin, and emulsion bin, respectively, enhanced by the attention mechanism. To adapt to the input format of the ViT model, the feature map at each point needs to be flattened and divided into several smaller blocks. Assume each feature map... The space dimensions are Then we divide the feature map into The size of each small block is P, where P represents the size of the block. And then convert it into a one-dimensional vector through a flattening operation. Let the k-th block at point i be... Each small block can then be represented in the following way:

[0097] ;

[0098] Then, it is necessary to process each A linear transformation can be performed on the image patch to compress the dimension to D.

[0099] ;

[0100] In the formula, The initial input includes category information and location code. It is category information. Indicates the first The feature map of each point is flattened into a vector, and E is the input embedding. It is a positional encoding matrix. Is The layer output undergoes layer normalization and multi-head self-attention. Z is... The final output of the layer, the feature vector Z of ViT, will be used for subsequent multi-task learning tasks to predict the liquid level status (normal, overflow, occlusion, etc.) of the vibrating screen, waste tank, and emulsion tank.

[0101] S6: Input the feature vectors corresponding to each key point into the multi-task learning model and output the liquid level status of each key point. The multi-task learning model includes several independent task heads. Each independent task head is used to classify the liquid level status of the feature vectors corresponding to each key point. Each independent task head is obtained by jointly training all independent task heads through a preset multi-task joint loss function.

[0102] In one embodiment, the training process for the independent task head is as follows:

[0103] A multi-task learning model is constructed, which includes an independent task head corresponding to each key point. The classifier expression for each independent task head is as follows:

[0104] ;

[0105] in Represents a probability vector. This represents the weight of the fully connected layer for the classification task of the i-th key point. Z represents the bias term parameter for the classification task of the i-th key point, and Z represents the global semantic feature corresponding to the i-th key point.

[0106] Construct a multi-task joint loss function, which is a weighted sum of all single-task loss functions. The single-task loss function L... i The expression is as follows:

[0107] ;

[0108] In the formula, N is the batch sample size. Let c be the true label of the nth sample at point i. This represents the predicted probability.

[0109] The fused features from the three locations are input into their respective independent classification heads to predict the liquid level status at each location (e.g., normal, overflow, occlusion). A weighted cross-entropy multi-task loss function is designed for joint training, allowing each task to retain its own discriminative power while sharing underlying features. The loss weights between tasks can be dynamically adjusted based on the risk level of the location or data imbalance, thus ensuring training stability and optimization efficiency. To achieve accurate prediction of liquid level status at multiple locations simultaneously, this embodiment constructs a multi-task learning framework based on shared underlying features and multiple independent task heads. The aforementioned cross-location fused features... The Vision Transformer (ViT) network generates a global semantic feature representation Z, which is simultaneously input into three independent task heads, each corresponding to a state classification task at a given location.

[0110] The classifier structures for the three independent tasks are defined as follows:

[0111] ;

[0112] In the formula, The weights and bias parameters of the fully connected layer are defined for the classification task at the i-th location. Each classification task outputs a three-class probability vector (e.g., normal, overflow, occlusion). During training, a joint loss function is designed to simultaneously optimize the classification performance of the three tasks. :

[0113] ; where each single-task loss function L i Using the definition of cross-entropy loss:

[0114] In the formula, N is the batch sample size. Let c be the ground truth label of the nth sample at point i. Represents the predicted probability. The coefficients in the loss function. This is used to adjust the optimization intensity between tasks to reflect the importance of different tasks or sample imbalance. For example, changes in the liquid level in an emulsion tank directly affect production safety, so a higher weight can be set to emphasize the priority of this task's optimization. The rationale for the joint loss function design lies in: the shared global semantic feature representation Z among the three tasks allows for implicit transfer and sharing of effective discriminative information between tasks; the independent classifier structure for each task allows the model to adaptively learn for the unique features of each point; and through multi-task joint optimization, it significantly reduces the overfitting problem that may occur in single-task models and improves the recognition accuracy of classes with fewer samples.

[0115] During the training phase, the model parameters are minimized by the joint loss function. Optimization was performed using gradient descent and backpropagation to adjust network parameters, ensuring the model's robustness and generalization performance. Finally, the model with the highest overall accuracy on the validation set was selected as the optimal model for practical deployment. During the inference phase, the model receives synchronized image data from three locations and outputs classification results for all three locations in a single forward propagation. This real-time, multi-task classification process meets the rapid response requirements of industrial production sites.

[0116] The multi-task learning model was optimized based on the training samples with the goal of minimizing the joint loss function of the multi-tasks. The network parameters were adjusted using gradient descent and backpropagation. Finally, the model with the highest overall accuracy in the validation set was selected as the optimal model for actual deployment. The training samples included liquid level image samples of several key points on the PTFE emulsion production line labeled with three-class labels, including normal label, overflow label, and occlusion label.

[0117] The overall model supports end-to-end training, uses the Adam optimizer for gradient updates, and employs a learning rate decay mechanism and cross-validation during training to avoid overfitting. During the inference phase, a set of three-point images is input, and the three-point states can be output synchronously with only one forward propagation, meeting the deployment requirements of high real-time performance and low computational cost in industrial settings.

[0118] This invention also provides a PTFE emulsion liquid level collaborative detection device, comprising: an acquisition module for acquiring image data, the image data including liquid level images of several key points on a PTFE emulsion production line; a preprocessing module for performing image enhancement and preprocessing on the image data to obtain preprocessed data corresponding to each key point; a feature extraction module for inputting the preprocessed data of each key point into a backbone feature extraction network to obtain feature maps corresponding to each key point; and a fusion module for performing cross-point attention mechanism on the feature maps corresponding to each key point. Cross-point fusion is performed to obtain cross-point fused feature maps corresponding to each key point; a full semantic modeling module is used to perform global semantic modeling on the enhanced feature maps corresponding to each key point and output the global semantic features corresponding to each key point; a detection module is used to input the feature vectors corresponding to each key point into a multi-task learning model and output the liquid level status of each key point. The multi-task learning model includes several independent task heads, each of which is used to classify the liquid level status of the feature vectors corresponding to each key point. Each independent task head is obtained by jointly training all independent task heads through a preset multi-task joint loss function.

[0119] This invention also provides a method for synergistic detection of PTFE emulsion level, comprising:

[0120] Memory, used to store computer programs;

[0121] A processor is configured to implement the steps of the PTFE emulsion level co-detection method when executing the computer program.

[0122] The present invention also provides a storage medium, which is a readable storage medium, and stores a computer program thereon. When the computer program is executed by a processor, it implements the steps of the PTFE emulsion level collaborative detection method.

[0123] Compared with traditional liquid level detection methods, the PTFE emulsion liquid level collaborative detection method, device, equipment and storage medium proposed in this invention have significant advantages in terms of accuracy, robustness and system deployment efficiency. Existing methods often rely on single-point images or physical sensors for detection, which suffers from problems such as failing to capture the linkage between points, poor adaptability to complex working conditions, and high deployment costs due to model redundancy. This invention, however, introduces a cross-point attention mechanism to achieve deep fusion and causal modeling of features between multiple key points such as vibrating screens, emulsion tanks, and waste bins, effectively improving the accuracy and interpretability of anomaly detection. Combined with a multi-task learning framework, it simultaneously completes multi-point state prediction within the same model, significantly reducing parameter redundancy and computational overhead compared to the traditional "multi-model independent inference" approach, improving inference efficiency, and adapting to the edge computing needs of industrial sites. Furthermore, by integrating the local perception capabilities of ResNet and the global modeling capabilities of ViT, the model can simultaneously identify minute changes in liquid level boundaries and long-range dependencies between points, exhibiting more robust performance in scenarios with complex backgrounds, fluctuating lighting, and partial occlusion. Through various data augmentation strategies designed for industrial scenarios, the model's adaptability to changes in on-site imaging conditions is further enhanced. In summary, this invention significantly improves the feasibility of engineering deployment while enhancing detection performance, providing a more reliable and efficient solution for intelligent liquid level monitoring in the production of PTFE emulsions and other high-purity liquids.

[0124] like Figure 2As shown, a multi-point liquid level detection dataset was constructed based on a real industrial scenario of a PTFE emulsion production line to verify the effectiveness of the cross-point feature fusion and multi-task learning method proposed in this invention. Liquid level data was collected from three key points: a vibrating screen, an emulsion tank, and a waste tank. Images were preprocessed to 224*224 pixels at a frame rate of 25 frames per second, with a total of over 30,000 samples collected. Three-class labels (normal, overflow, occlusion) were obtained through joint annotation by manual and sensor data. In terms of model design, a shared ResNet-50 backbone network was used to extract features from the three-point images. The ResNet-50 backbone network includes several CNN layers, and a Cross Attention mechanism was used to achieve feature interaction between points. Subsequently, a Vision Transformer structure was fused to further model global dependencies. To compare the performance advantages of the method of this invention, we tested the detection accuracy of three fusion methods (feature concatenation, weighted summation, and cross-attention) and various model structures (ResNet, ViT, ViT+ResNet, DensNet121, and mobilenet_v3_small) at various locations. Referring to Tables 1 and 2, the results show that the Cross Attention fusion method combined with the ViT structure significantly outperforms other methods: taking the vibrating screen location as an example, the accuracy of the cross-attention method reached 92.84%, an improvement of more than 13 percentage points compared to the feature concatenation method; at the emulsion bucket location, the accuracy increased from 65.2% for ResNet to 89.7%, effectively enhancing the model's responsiveness to complex state changes. Furthermore, the inference latency was controlled within 100ms, meeting the real-time requirements of industrial applications. Figure 3 As shown in the experiment, the multi-task learning structure significantly reduces the computational resource consumption of the model, such as... Figure 3 As shown, this reduces inference overhead by approximately 67% compared to a three-single-point model deployment scheme. Figure 3 As shown, by merging the three ResNetV2 backbones using a multi-task framework, low-level feature extraction can be completed with only one forward pass, compressing the overall computational cost from 51 GFLOPs to 17 GFLOPs. The main saving is in the heaviest convolutional backbone (-18.6 GFLOPs, accounting for 90% of the total reduction). This not only reduces inference latency by two-thirds but also significantly reduces GPU memory and power consumption. The above results demonstrate that the collaborative detection method proposed in this invention significantly outperforms traditional schemes in terms of accuracy, efficiency, and engineering adaptability, and has the potential for widespread application in practical chemical production scenarios.

[0125] Table 1

[0126]

[0127] Table 2

[0128]

[0129] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for synergistic detection of PTFE emulsion level, characterized in that, include: Acquire image data, which includes liquid level images at several key points on the PTFE emulsion production line; The image data is enhanced and preprocessed to obtain preprocessed data corresponding to each key point. The preprocessed data of each key point is input into the backbone feature extraction network to obtain the feature map corresponding to each key point. Based on the cross-point attention mechanism, the feature maps corresponding to each key point are fused across points to obtain the cross-point fused feature map corresponding to each key point. Perform global semantic modeling on the enhanced feature maps corresponding to each key point and output the global semantic features corresponding to each key point. The feature vectors corresponding to each key point are input into the multi-task learning model, and the liquid level status of each key point is output. The multi-task learning model includes several independent task heads. Each independent task head is used to classify the liquid level status of the feature vectors corresponding to each key point. Each independent task head is obtained by jointly training all independent task heads through a preset multi-task joint loss function.

2. The method for synergistic detection of PTFE emulsion level according to claim 1, characterized in that, Image enhancement and preprocessing of the image data include: Perform geometric transformations on the image data; Color perturbation is applied to the image data; The image data is then subjected to regularization processing.

3. The method for synergistic detection of PTFE emulsion level according to claim 1, characterized in that, The backbone feature extraction network is a ResNet network architecture.

4. The method for synergistic detection of PTFE emulsion level according to claim 1, characterized in that, Based on a cross-point attention mechanism, feature maps corresponding to each key point are fused across points, including: Flatten the feature maps corresponding to each key point into two-dimensional feature sequences; For each key location, construct three matrices: query, key, and value. The dot product attention mechanism is used to compute a weighted representation of the two-dimensional feature sequence of each key point with respect to the two-dimensional feature sequences of all other key points. Calculate the two-dimensional feature sequence of each key point and the sum of its weighted representation with respect to all other key points to obtain the cross-point fusion feature map corresponding to each key point.

5. The method for synergistic detection of PTFE emulsion level according to claim 1, characterized in that, Global semantic modeling is performed on the enhanced feature maps corresponding to each key point, and the global semantic features corresponding to each key point are output, including: Flatten the cross-point fusion feature map corresponding to each key point, and divide the flattened feature map into several image blocks; Perform a linear transformation on each image patch; By inputting all the linearly transformed image patches corresponding to each key point into the Vision Transformer model, the global semantic features corresponding to each key point are obtained.

6. The method for synergistic detection of PTFE emulsion level according to claim 1, characterized in that, The training process for an independent task head is as follows: A multi-task learning model is constructed, which includes an independent task head corresponding to each key point. The classifier expression for each independent task head is as follows: ； in Represents a probability vector. This represents the weight of the fully connected layer for the classification task of the i-th key point. Z represents the bias term parameter for the classification task of the i-th key point, and Z represents the global semantic feature corresponding to the i-th key point. Construct a multi-task joint loss function, which is a weighted sum of all single-task loss functions. The single-task loss function L... i The expression is as follows: ， In the formula, N is the batch sample size. Let c be the true label of the nth sample at point i. Indicates the predicted probability; The multi-task learning model was optimized based on the training samples with the goal of minimizing the joint loss function of the multi-tasks. The network parameters were adjusted using gradient descent and backpropagation. Finally, the model with the highest overall accuracy in the validation set was selected as the optimal model for actual deployment. The training samples included liquid level image samples of several key points on the PTFE emulsion production line labeled with three-class labels, including normal label, overflow label, and occlusion label.

7. The method for synergistic detection of PTFE emulsion level according to claim 1, characterized in that, Acquire image data, including: Acquire liquid level images from programmable industrial cameras deployed at key locations in a PTFE emulsion production line, including vibrating screens, emulsion tanks, and waste tanks. The acquired liquid level images are processed into frame slices at a uniform frame rate, and each set of images is composed of a time-consistent image tuple for all key locations.

8. A PTFE emulsion level detection device, characterized in that, include: The acquisition module is used to acquire image data, which includes liquid level images of several key points on the PTFE emulsion production line. The preprocessing module is used to perform image enhancement and preprocessing on the image data to obtain preprocessed data corresponding to each key point. The feature extraction module is used to input the preprocessed data of each key point into the backbone feature extraction network to obtain the feature map corresponding to each key point. The fusion module is used to perform cross-point fusion on the feature maps corresponding to each key point based on the cross-point attention mechanism to obtain the cross-point fused feature map corresponding to each key point. The full semantic modeling module is used to perform global semantic modeling on the enhanced feature maps corresponding to each key point and output the global semantic features corresponding to each key point. The detection module is used to input the feature vectors corresponding to each key point into a multi-task learning model and output the liquid level status of each key point. The multi-task learning model includes several independent task heads, each of which is used to classify the liquid level status of the feature vectors corresponding to each key point. Each independent task head is obtained by jointly training all independent task heads through a preset multi-task joint loss function.

9. A method for synergistic detection of PTFE emulsion level, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the PTFE emulsion level co-detection method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium is a readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the PTFE emulsion level co-detection method as described in any one of claims 1 to 7.

Images