Polymer monofilament size and morphology detection and spinning component service life prediction method

Abstract

The present invention discloses a method for detecting the size and morphology of polymer monofilaments and predicting the service life of spinning components. The method obtains the historical data of polymer monofilaments collected by a machine vision system and the incremental data collected in real time, divides the data into a training data set and a test data set, uses a deep convolutional neural network to learn and train, optimizes the extraction of internal features and edge features of the image, obtains a training model, and then uses the test data set to test the training model to obtain the test result. If the test result passes, the result is output, and if it fails, retraining is performed; combined with a deep convolutional neural network CNN-Transformer prediction model, on the one hand, the edge feature extraction of the polymer monofilament image collected by the machine vision system is optimized, and the detection accuracy of the geometric size and surface morphology of the polymer monofilament is improved, and on the other hand, the prediction of the remaining service life of the polymer monofilament spinning component is further realized. The present invention improves the accuracy of the measurement of the line diameter of the polymer monofilament.

Description

Technical Field

[0001] The invention relates to a method for detecting the geometrical size and surface morphology of polymer monofilaments and predicting the remaining service life of a spinning component. Background Art

[0002] In the field of polymer monofilament geometric dimension detection, machine vision is used to measure the diameter of monofilaments to generate a large number of image data sets, and deep learning is applied to image processing. It uses a supervised deep learning neural network, combined with the large-scale image stock generated by machine vision and the continuously generated image increments for training, to automatically extract fine-grained features and combine coarse-grained features to form an abstract visual description; thereby optimizing the geometric dimension accuracy and surface morphology features extracted from the image, automatically classifying polymer monofilaments with out-of-tolerance dimensions distributed at different positions on the image, and predicting the service life of the spinning component based on the changes in the real-time diameter and surface morphology of the polymer monofilament during the spinning process.

[0003] In existing machine vision measurement technology, parallel arranged single wires are imaged in an industrial camera to form an image data set. Due to problems with irradiation light and imaging angle, products with the same wire diameter undergo optical distortion during the imaging process, and arrangement spacing errors and other errors cause measurement errors. Summary of the invention

[0004] The purpose of the present invention is to provide a method for measuring the geometric dimensions of polymer monofilaments and predicting the service life of spinning components based on deep learning, so as to overcome the problems in existing machine vision measurement technology, in which parallel arranged monofilaments are imaged in an industrial camera to form an image data set. Due to the problems of irradiation light and imaging angle, products with the same wire diameter undergo optical distortion during the imaging process, and measurement errors are caused by arrangement spacing errors.

[0005] The technical solution of the present invention is:

[0006] A method for detecting the size and morphology of polymer monofilaments and predicting the service life of spinning components, characterized by: obtaining historical data of polymer monofilaments collected by a machine vision system and incremental data collected in real time, dividing the data into a training data set and a test data set, using a deep convolutional neural network for learning and training, optimizing the extraction of internal features and edge features of the image, obtaining a training model, and then using the test data set to test the training model to obtain a test result, outputting the result if the test result passes, and retraining if the test result fails; combining a deep convolutional neural network CNN and a Transformer prediction model, on the one hand, optimizing the edge feature extraction of polymer monofilament images collected by the machine vision system, improving the detection accuracy of the geometric size and surface morphology of the polymer monofilaments, and on the other hand, further realizing the prediction of the remaining service life of the polymer monofilament spinning component.

[0007] The specific method for optimizing the edge feature extraction of polymer monofilament images collected by a machine vision system is as follows: (1) First, filter the high-frequency features in the image, and use a convolutional neural network to upsample the image. By taking advantage of the convolutional neural network, the traditional fixed parameter filtering operations (such as mean filtering and Gaussian filtering) are replaced by learnable filtering operations, and deep separable convolution is used to learn to filter high-frequency features; at the same time, a dilated convolution is used to expand the receptive field of the convolution result; (2) Quadratic linear interpolation is used to obtain the low-frequency representation of high-level features, and the upsampled low-frequency features are restored to the same scale as the high-level features. Then, the high-level features are connected with the low-frequency features and passed through a convolution kernel of a certain scale to predict the optical flow field. Each binary group of the optical flow field represents the relative offset of the high-level features to match the low-frequency features. The high-level features are deformed according to the offset given by the optical flow field to obtain the final low-frequency internal features. In order to obtain the desired low-frequency internal features, The filtering of the above high-frequency features and the learning of the optical flow field are supervised and constrained. The supervision information uses the semantic segmentation label inside the object. The edge width is set to 2 pixels, and the edge of the complete semantic label is calculated to obtain the edge label of the image and the internal label after removing the edge. The internal label and the probability map predicted by the internal low-frequency feature are used to calculate the cross entropy loss, thereby supervising the learning; (3) The edge feature is extracted by subtracting the high-level feature from the extracted low-frequency internal feature. Since the high-level feature is sampled by the encoder, the high-frequency information contained is incomplete. In order to obtain a more complete edge feature, the network bottom-level features and the high-frequency features obtained above are combined through convolution and fused; at the same time, the category edge of the image is predicted, and then the binary label of the image category edge is calculated based on the true semantic label of the image. The edge probability predicted by the convolution of the edge feature is used to calculate the binary classification cross entropy loss with the edge label to supervise the edge feature extraction.

[0008] Specific method for predicting the remaining service life of polymer monofilament spinning components: (1) Construction of health index (HI): The spinning life of the component is affected by different factors during use, and the degradation rate is constantly changing. At the beginning of use, the performance is relatively stable and the degradation rate is slow. Later, as the use time increases, the degradation rate accelerates and the health level continues to decline; therefore, a quadratic function is used to express the degradation trend of the spinning component life, which is defined as:

[0009]

[0010] t n is the complete life of the spinning assembly, t i is the current time within the complete life cycle of the spinning assembly;

[0011] (2) Construct a spinning component life prediction model: Use the Transformer model to further obtain the global characteristics of health indicators for inference and prediction, thereby realizing the prediction of the remaining useful life (RUL) of the spinning component.

[0012] Machine vision is applied to monofilament wire diameter measurement to generate a large number of image data sets. Deep learning is applied to image processing. It uses different forms of neural networks, combined with the large-scale inventory of visual big data and the continuously generated increments for learning and training, automatically extracts fine-grained features and combines coarse-grained features to form an abstract visual description; thereby optimizing the geometric dimensional accuracy of the polymer monofilaments extracted from the image and capturing the characteristics of surface defects of the polymer monofilaments; then, the various features of the polymer monofilaments obtained through the above extraction are imported into the Transforer prediction model to predict the remaining service life of the polymer monofilament spinning components.

[0013] Advantages of the present invention:

[0014] (1) Extract features from images acquired by machine vision cameras through deep convolutional neural networks, optimize the noise in the data through learning and training, and improve the accuracy of polymer monofilament wire diameter measurement.

[0015] (2) Through deep convolutional neural network self-learning and training, this technical solution can extract the surface defects of polymer monofilaments in the image and classify the types of surface defects of polymer monofilaments, which is convenient for the staff to optimize the production process of polymer monofilaments in the later stage.

[0016] (3) Through the combination of deep convolutional neural network and Transformer prediction model, this technical solution constructs a service life prediction model for polymer monofilament spinning components during the spinning process by collecting data on the dynamic change of wire diameter during the production process of polymer monofilaments, and data on the time, location, and category of surface defects of polymer monofilaments. The service life of this set (batch) of spinning components can be predicted in advance, which facilitates the company to arrange maintenance plans more effectively. The data can be interconnected with the company's APS system, allowing the company to arrange production plans more efficiently and improve the equipment utilization efficiency of the entire company. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0018] Figure 1 This is a flow chart of the image edge feature and internal feature segmentation method based on convolutional neural network.

[0019] Figure 2 This is a schematic diagram of the AlexNet convolutional neural network model structure.

[0020] Figure 3 It is a list of operations, dimensions and parameters of each layer of AlexNet.

[0021] Figure 4 It is a schematic diagram of the component remaining service life prediction process.

[0022] Figure 5 is a graph of the degradation of the health of the spinning assembly.

[0023] Figure 6 It is a flow chart for obtaining health indicators of the spinning component life degradation process.

[0024] Figure 7 It is a schematic diagram of the structure of the spinning component service life prediction model. DETAILED DESCRIPTION

[0025] A method for polymer monofilament size and morphology detection and spinning component life prediction uses convolutional neural networks on image segmentation tasks and can complete supervised end-to-end training on images of any scale. The remaining service life of the spinning component is predicted by combining the deep convolutional neural network with the Transforer prediction model to understand the target features in the image.

[0026] Since polymer monofilaments have various colors, transparent, translucent, opaque, large diameter scale span, uneven arrangement spacing or parallel wires, etc., the polymer wire diameter boundary on the image captured by the camera will be unclear, making it difficult to extract edge features in the software, resulting in inaccurate measurement of the polymer wire diameter. Since the category edge pixels of the image are more difficult to distinguish, while the internal pixels are easier to distinguish, and the internal pixel features of the object extracted by the neural network can be well clustered, the better the clustering effect, the better the segmentation effect. Therefore, the technical solution proposes an image feature decoupling algorithm combined with edge information, which is used for the network to more clearly model and constrain the relevant information of the image edge during the image feature extraction process. The image feature decoupling algorithm decouples the small-scale high-level features output by the original segmentation network into edge features and internal features through different information supervision. The edge features are supervised by the category edge information calculated by the true value label, and the decoupling of the internal features learns more consistent internal information of the segmentation target through optical flow transformation. Since edge information is used for supervision in the feature extraction process, the edge features learned by the network are clearer and the internal feature consistency is stronger. At present, the segmentation network uses upsampling operations to extract high-level features many times during the encoding process. The final segmentation result is calculated by the small-scale high-level features through quadratic linear interpolation or deconvolution upsampling to obtain a segmentation mask with the same scale as the original image. This method will make it difficult to restore high-frequency information. The classification confidence of the final segmentation mask at the edge of the category is low, and the edge contour line transition is smooth. Therefore, for the difficult-to-distinguish area at the edge of the category in the rough segmentation result obtained by the network for the first time, a multi-layer perceptron structure (MLP) is used for pixel-level correction, and the multi-layer features obtained by the original segmentation network encoding are input during the correction. This not only solves the problem of high-frequency information loss of the image caused by upsampling, but also solves the problem of low accuracy of smoothing and edge classification results caused by upsampling operations. At the same time, the number of parameters of the multi-layer perceptron structure used for correction is small, and the calculation amount of this scheme is much smaller than that of the network structure that performs step-by-step upsampling through the decoder.

[0027] The image data set used for edge feature extraction of polymer monofilaments in images of the technical solution comes from the polymer monofilament measurement images previously captured by the camera and stored in the system memory, as well as the incremental images captured by the camera in real time during the real-time production process as the data set. The image is encoded in the convolutional neural network, and the high-frequency information and low-frequency information of the image are extracted respectively through Fourier transform, and modeled into high-frequency features and low-frequency features (i.e., category edge features and internal features). The high-frequency features and low-frequency features are constrained to increase the network's attention to the edges of the image, and the optical flow network is used to increase the consistency of non-edge internal pixel features. The high-frequency features and low-frequency features are then restored to obtain the complete original image, with less loss of edge features of the original image, stronger consistency of internal features, clearer semantics, and reduced difficulty of subsequent network segmentation.

[0028] The specific approach is as follows: (1) First filter the high-frequency features in the image, use a convolutional neural network to upsample the image, and use the advantages of convolutional neural networks to replace traditional fixed parameter filtering operations (such as mean filtering and Gaussian filtering) with learnable filtering operations, and learn to filter high-frequency features through deep separable convolution. At the same time, use hollow convolution to expand the receptive field of the convolution. (2) Use quadratic linear interpolation to obtain the low-frequency representation of high-level features, restore the upsampled low-frequency features to the same scale as the high-level features, and then connect the high-level features with the low-frequency features and pass them through a convolution kernel of a certain scale to predict the optical flow field. Each binary group of the optical flow field represents the relative offset of the high-level feature to match the low-frequency feature. The high-level feature is deformed according to the offset given by the optical flow field to obtain the final low-frequency internal features. In order to obtain the desired low-frequency internal features, we also need to supervise the filtering of the above high-frequency features and the learning of the optical flow field. The supervision information uses the semantic segmentation label inside the object, sets the edge width to 2 pixels, calculates the edge of the complete semantic label, obtains the edge label of the image and the internal label after removing the edge, and calculates the cross entropy loss between the internal label and the probability map predicted by the internal low-frequency feature, so as to supervise the learning. (3) The extraction of edge features can be obtained by subtracting the high-level features from the extracted low-frequency internal features. Since the high-level features are sampled by the encoder, the high-frequency information contained is incomplete. In order to obtain more complete edge features, we combine the underlying network features and the high-frequency features obtained above through convolution and fuse them; at the same time, the category edge of the image is predicted, and then the binary label of the image category edge is calculated based on the real semantic label of the image. The edge probability predicted by the convolution of the edge feature and the edge label are used to calculate the binary cross entropy loss, so as to supervise the edge feature extraction.

[0029] The segmentation skeleton network used is the AlexNet convolutional neural network (CNN), and its structural diagram is as follows Figure 2; It contains an input layer, 5 convolutional layers and three fully connected layers. In the figure, conv represents convolution operation, ReLU represents rectified linear unit, pool represents pooling operation, norm represents local response normalization, dropout represents dropout output operation, IP represents full connection, and softmax represents soft maximum function. The AlexNet convolutional neural network (CNN) is used to learn and train the internal features and edge feature segmentation of the image, so that the boundaries and defect contours of the polymer monofilament wire diameter on the image are clearer, making the polymer monofilament wire diameter measurement more accurate.

[0030] (1) The first layer is the input layer, and the input is the original image.

[0031] (2) The second layer is a convolution layer, which uses 96 large 11×11×3 convolution kernels. These convolution kernels are divided into two groups (48 convolution kernels in each group). The input layer is convolved with a step size of 4 pixels to obtain two sets of convolution results. Then, the ReLU activation function is used on the convolution results to obtain activation results. Then, the two sets of activation results are overlapped with the maximum pooling with a window size of 3×x3 and a step size of 2 pixels to obtain the pooling results of the two groups. Finally, the local response normalization operation is performed on the pooling results to obtain the normalized results of the two groups.

[0032] (3) The third layer is a convolution layer, which uses 256 convolution kernels of size 27×27×48. These convolution kernels are divided into two groups (128 convolution kernels in each group). The normalized results of the second layer are convolved with a step size of 1 pixel to obtain two sets of convolution results. Then, the ReLU activation function is used on the convolution results to obtain two sets of activation results. Next, the two sets of activation results are subjected to overlapping maximum pooling with a window size of 3×3 and a step size of 2 pixels to obtain two sets of pooling results. Finally, the local response normalization operation is performed on the pooling results to obtain two sets of normalized results.

[0033] (4) The fourth layer is a convolution layer, which uses 384 convolution kernels of size 13×13×256 and performs convolution operation on the normalized result of the third layer with a step size of 1 pixel to obtain two sets of convolution results. Then, the ReLU activation function is used on the convolution result to obtain the activation result.

[0034] (5) The fifth layer is a convolution layer, which uses 384 convolution kernels of size 13×13×192. These convolution kernels are divided into two groups (192 convolution kernels in each group), and the activation results of the fourth layer are convolved with a step size of 1 pixel to obtain two groups of convolution results. Then, the ReLU activation function is used on the convolution results to obtain the activation results.

[0035] (6) The sixth layer is a convolution layer, which uses 256 convolution kernels of size 13×13×192. These convolution kernels are divided into two groups (128 convolution kernels in each group). The activation results of the fifth layer are convolved with a step size of 1 pixel to obtain two groups of convolution results. Then, the ReLU activation function is used on the convolution results to obtain activation results. Next, the two groups of activation results are overlapped and max pooled with a window size of 3×3 and a step size of 2 pixels to obtain two groups of pooling results.

[0036] (7) The 7th layer is a fully connected layer, which uses 4096 neurons and is divided into two groups (2048 convolution neurons in each group). The pooling result of the 6th layer is fully connected, and then the ReLU activation function is used on the fully connected processing result to obtain the activation result. Then, the dropout operation with a probability of 0.5 is used on the activation result to obtain the dropout result.

[0037] (8) The 8th layer is a fully connected layer, which uses 4096 neurons and is divided into two groups (2048 convolution neurons in each group). The dropout result of the 7th layer is fully connected, and then the ReLU activation function is used on the fully connected processing result to obtain the activation result. Then, the dropout operation with a probability of 0.5 is used on the activation result to obtain the dropout result.

[0038] (9) The last layer is a 1000-way softmax output layer, which is used to generate a label distribution covering 1000 categories.

[0039] The cross entropy function is used as the objective function for image constraint internal feature extraction, edge feature extraction, and the final decoupled fusion features to evaluate the difference between the model prediction and the image label. The loss function for evaluating internal feature extraction Calculate the cross entropy loss for all non-edge pixel classification. Evaluate the loss function for edge feature extraction Calculate the binary cross entropy loss of all pixels on the edge. Evaluate the loss function of the final decoupled fusion feature Calculate the multi-classification cross entropy loss for all pixels. The cross entropy loss formula is:

[0040]

[0041] Where x represents the final prediction result of the network, x[class] represents the network's prediction value of the true label category at that location, and k is the total number of semantic label categories. In this technical solution, there are only two types of image semantic labels (pixels belong to edge pixels or non-edge pixels). We define two hyperparameters λ1 and λ2, which are the weights of internal feature extraction loss and edge feature loss, respectively. In this solution, 1 and 5 are selected. The final loss function is as follows:

[0042]

[0043] This technical solution uses edge F as an evaluation index for evaluating edge segmentation effects in semantic segmentation. The calculation method is to calculate the F score of the predicted result along the edge of the predicted segmentation mask, give an edge width threshold, and measure the classification accuracy of edge pixels with a specified width. This technical solution uses a width threshold range of 3 pixels, and its formula is defined as follows:

[0044]

[0045] Where P b represents the precision of edge pixels, and Rb represents the recall rate of edge pixels.

[0046] Based on the above technology, the diameter of polymer monofilaments is accurately measured. We can construct the health index of the spinning assembly during the spinning cycle by measuring the diameter change and surface defects of the polymer monofilaments during the production process, so as to predict the remaining service life of the spinning assembly. Figure 4 .

[0047] Construction of health index (HI). The spinning life of the component is affected by different factors during use, and its degradation rate is constantly changing. At the beginning of use, its performance is relatively stable and the degradation rate is slow. Later, as the use time increases, its degradation rate accelerates and the health level continues to decline. Therefore, we use a quadratic function to express the degradation trend of the spinning component life, which is defined as:

[0048]

[0049] t n is the complete life of the spinning assembly, t i is the current time within the complete life cycle of the spinning assembly.

[0050] Acquisition of health indicators: ① The original image data of polymer monofilament diameter change, surface defects, etc. obtained by image processing is used as the input of the convolutional neural network. The data is divided into two parts, one as a training set and the other as a test set. The labels of the training set are as follows: Figure 5 ②. The above input data has been extracted by the convolutional neural network to extract the required features. Here we only need to use the error back propagation to minimize the loss function to train the model.

[0051] ③. Construct health indicators. Input the test set obtained in step ① into the convolutional neural network trained in step ②. The network output is the health indicator.

[0052] Construct a spinning component life prediction model. This technical solution uses the Transformer model to further obtain the global characteristics of health indicators for inference and prediction, thereby realizing the prediction of the remaining useful life (RUL) of the spinning component. The overall structure of the model consists of an encoding component and a decoding component. The encoding component is composed of 4 stacked encoders, each of which contains three ProbSparse sub-layers, which are connected by Dilated-SKConv convolution modules. The structure of each encoder is the same, but the weight parameters they use are different. The decoding component is also composed of 4 layers of decoders stacked together. The structure of each decoder is basically the same as that of the encoder. The difference is that there is an additional encoding-decoding self-attention layer in the middle of the decoder to help the decoder focus on the relevant parts of the input features. Its structure is as follows Figure 7 .

[0053] In order to integrate the multi-level features (broken filaments, thick and thin filaments, surface roughness, missing parts, indentations, spots, etc.) of the remaining life of the spinning assembly reflected by the polymer monofilaments and improve the prediction accuracy, this model integrates the output of a convolutional neural network in the output part of each Dilated-SKConv convolution module.

[0054] This technical solution first pre-processes the original data and inputs it into the deep convolutional neural network for self-learning training to improve the accuracy of polymer monofilament diameter measurement, and statistically classifies the changes in polymer monofilament diameter and surface defects; uses these data for training to construct the health index (HI) of the spinning assembly, and then uses the health index as the input of the Transformer prediction model for learning and training, and the model output is the remaining service life of the spinning assembly.

Claims

1. A method for detecting the size and morphology of polymer monofilaments and predicting the service life of spinning components, characterized by: The historical data of polymer monofilaments collected by the machine vision system and the incremental data collected in real time are obtained, and the data are divided into a training data set and a test data set. The deep convolutional neural network CNN is used for learning and training, and the extraction of internal features and edge features of the image is optimized to obtain a training model. The training model is then tested using the test data set to obtain the test results. If the test result passes, the result is output. If it fails, retraining is performed. The deep convolutional neural network CNN and the Transformer life prediction model are combined to optimize the edge feature extraction of the polymer monofilament image collected by the machine vision system and improve the detection accuracy of the geometric size and surface morphology of the polymer monofilament. On the other hand, the remaining service life of the polymer monofilament spinning component is further predicted. The image feature decoupling algorithm combined with edge information is used to model and constrain the information of image edges during the image feature extraction process. The image feature decoupling algorithm decouples the small-scale high-level features output by the original segmentation network into edge features and internal features through different information supervision. The edge features are supervised by the category edge information calculated by the true value label, while the decoupling of the internal features learns the internal information of the segmentation target through optical flow transformation. The multi-layer perceptron structure is used to perform pixel-level correction on the difficult-to-distinguish areas at the category edge in the rough segmentation results initially obtained by the network, and the multi-layer features encoded by the original segmentation network are input during the correction. The image data set used for edge feature extraction of polymer monofilaments in images comes from the polymer monofilament measurement images collected by the camera in the past and stored in the system memory, as well as the incremental images collected by the camera in real time during the real-time production process. The images are encoded in the convolutional neural network and modeled into category edge features and internal features. The category edge features and internal features are constrained to increase the network's attention to the image edges. The optical flow network is then used to increase the consistency of non-edge internal pixel features, and then the category edge features and internal features are restored to obtain the complete original image. The specific method for optimizing the edge feature extraction of polymer monofilament images collected by the machine vision system is as follows: the image is input into the segmentation skeleton network, and the segmentation skeleton network outputs high-level features and low-level features, and then: (1) the high-frequency features in the image are first filtered out, and the image is upsampled using a convolutional neural network. By taking advantage of the convolutional neural network, the traditional fixed parameter filtering operation is replaced by a learnable filtering operation, and the high-frequency features are learned to be filtered through deep separable convolution; at the same time, a dilated convolution is used to expand the receptive field of the convolution; (2) the low-frequency representation of the high-level features is obtained by using quadratic linear interpolation, and the upsampled low-frequency features are restored to the same scale as the high-level features. Then, the high-level features are connected with the low-frequency features and passed through the convolution kernel together to predict the optical flow field. Each binary group in the field represents the relative offset of the high-level feature to match the low-frequency feature. The high-level feature is deformed according to the offset given by the optical flow field to obtain the final low-frequency internal feature. In order to obtain the desired low-frequency internal feature, the filtering of the high-frequency feature and the learning of the optical flow field are supervised and constrained. The supervision information uses the semantic segmentation label inside the object, sets the edge width to 2 pixels, calculates the edge of the complete semantic label, obtains the edge label of the image and the internal label after removing the edge, and calculates the cross entropy loss of the probability map obtained by predicting the internal label and the internal low-frequency feature, thereby supervising the learning. (3) Extraction of edge features: subtracting the high-level features from the extracted low-frequency internal features and connecting them with the low-level features to obtain the edge features. The category edge of the image is predicted, and then the binary label of the image category edge is calculated based on the real semantic label of the image. The edge feature is predicted by convolution, and the edge probability and edge label are used to calculate the binary cross entropy loss to supervise the edge feature extraction. The segmentation skeleton network used is the AlexNet convolutional neural network (CNN), which includes an input layer, five convolutional layers, and three fully connected layers. The AlexNet convolutional neural network (CNN) is used to learn the internal features of the training image and the segmentation of the edge features, and the output is the wire diameter change and surface defect of the polymer monofilament during the production process. The cross entropy function is used as the objective function for image constraint internal feature extraction, edge feature extraction, and the final decoupled fusion features to evaluate the difference between the model prediction and the image label; the loss function for evaluating internal feature extraction is Calculate the classification cross entropy loss of all non-edge pixels; evaluate the loss function of edge feature extraction Calculate the binary cross entropy loss of all pixels on the edge; evaluate the loss function of the final decoupled fusion feature Calculate the multi-classification cross entropy loss for all pixels; the cross entropy loss formula is: Where x represents the final prediction result of the network, x[class] represents the network's predicted value of the true label category at that location, and k is the total number of categories of semantic labels; there are only two categories of image semantic labels, pixels belonging to edge pixels or non-edge pixels, and two hyperparameters λ1 and λ2 are defined, which are the weights of internal feature extraction loss and edge feature loss, respectively, and are selected as 1 and 5; the final loss function is as follows: Edge F is used as an evaluation index for evaluating edge segmentation effect in semantic segmentation. The calculation method is to calculate the F score of the predicted result along the edge of the predicted segmentation mask, give an edge width threshold, and measure the classification accuracy of edge pixels with a specified width. The width threshold range is 3 pixels, and its formula is defined as follows: Where P b Indicates the accuracy of edge pixels, R b Represents the recall rate of pixels at the edge; By measuring the diameter changes and surface defects of the polymer monofilaments obtained during the production process, the health index of the spinning assembly during the spinning cycle is constructed to predict the remaining service life of the spinning assembly. The health index is then used as the input of the Transformer prediction model for learning and training, and the model output is the remaining service life of the spinning assembly. Construction of health index: The spinning life of the component is affected by different factors during use, and its degradation rate is constantly changing. When it is first used, its performance is relatively stable and the degradation rate is slow. Later, as the use time increases, its degradation rate accelerates and the health level continues to decline; therefore, a quadratic function is used to express the degradation trend of the spinning component life, which is defined as: t n is the complete life of the spinning assembly, t i is the current time within the complete life cycle of the spinning assembly; Acquisition of health indicators: ① The original image data of polymer monofilament diameter changes and surface defects obtained by image processing is used as the input of the convolutional neural network, and the data is divided into two parts, one as a training set and the other as a test set; ② The above input data has been extracted by the convolutional neural network to extract the required features, and here only the error back propagation minimization loss function needs to be used to train the model; ③Construct health indicators, input the test set obtained in step ① into the convolutional neural network trained in step ②, and the network output is the health indicator; Constructing a spinning component life prediction model: The Transformer model is used to further obtain the global features of health indicators for inference and prediction; thereby realizing the prediction of the remaining service life of the spinning component; the overall structure of the model consists of an encoding component-decoding component, in which the encoding component is composed of 4 encoders stacked, each encoder contains three ProbSparse sub-layers, connected by Dilated-SKConv convolution modules, and the structure of each encoder is the same, but the weight parameters they use are different; the decoding component is also composed of 4 layers of decoders stacked, and the structure of each decoder is basically the same as that of the encoder, except that there is an additional encoding-decoding self-attention layer in the middle of the decoder to help the decoder focus on the relevant parts of the input features; In order to fuse the multi-level features of the remaining life of the spinning assembly reflected by the polymer monofilament and improve the prediction accuracy, the output of a convolutional neural network is fused to the output of each Dilated-SKConv convolution module.

Images