Polylactic acid bio-based fiber dyeing process optimization method and system

Abstract

The present invention relates to the field of neural network technology, and in particular to a method and system for optimizing the dyeing process of polylactic acid bio-based fibers. The method comprises the following steps: obtaining polylactic acid fiber image data and corresponding chemical additive content data; performing self-attention convolution layer calculation on the polylactic acid fiber image data to obtain polylactic acid fiber image convolution layer data; performing multi-dimensional pooling layer processing on the polylactic acid fiber image convolution layer data to obtain polylactic acid fiber image pooling layer data; performing full connection layer calculation on the polylactic acid fiber image pooling layer data to obtain polylactic acid fiber image full connection layer data; constructing and iterating a model according to the full connection layer data of the polylactic acid fiber image and the corresponding chemical additive content data to obtain a polylactic acid bio-based fiber dyeing process optimization model, so as to perform polylactic acid bio-based fiber dyeing process optimization operations. The present invention can improve dyeing uniformity and color fastness, and can also reduce fiber damage during the dyeing process.

Description

Technical Field

[0001] The present invention relates to the field of neural network technology, and in particular to a method and system for optimizing the dyeing process of polylactic acid bio-based fibers. Background Art

[0002] With the development of the textile industry, bio-based fibers, as a new type of environmentally friendly material, have received more and more attention. Polylactic acid (PLA) fiber, as one of them, has become an important material in the textile industry due to its renewable, degradable and excellent mechanical properties. However, PLA fiber faces many technical difficulties in the dyeing process, such as uneven dyeing, low color fastness and damage to the fiber structure during the dyeing process. Therefore, how to optimize the dyeing process of PLA fiber and improve the dyeing quality and stability has become an urgent problem to be solved. Traditional dyeing processes mostly rely on experience and repeated trials, which are inefficient and not universal. Summary of the invention

[0003] In order to solve the above technical problems, the present invention proposes a method and system for optimizing the dyeing process of polylactic acid bio-based fibers to solve at least one of the above technical problems.

[0004] The present application provides a method for optimizing the dyeing process of polylactic acid bio-based fibers, comprising the following steps:

[0005] Step S1: Acquire polylactic acid fiber image data and corresponding chemical additive content data;

[0006] Step S2: performing self-attention convolution layer calculation on the polylactic acid fiber image data to obtain polylactic acid fiber image convolution layer data;

[0007] Step S3: performing multi-dimensional pooling layer processing on the polylactic acid fiber image convolution layer data to obtain polylactic acid fiber image pooling layer data;

[0008] Step S4: performing full connection layer calculation on the pooling layer data of the polylactic acid fiber image to obtain full connection layer data of the polylactic acid fiber image;

[0009] Step S5: construct and iterate a model based on the fully connected layer data of the polylactic acid fiber image and the corresponding chemical additive content data to obtain a polylactic acid bio-based fiber dyeing process optimization model to perform polylactic acid bio-based fiber dyeing process optimization operations.

[0010] The present invention can significantly improve the dyeing quality of polylactic acid bio-based fibers and ensure dyeing uniformity and color fastness through precise feature extraction and optimization model. The optimized dyeing process reduces the number of experiments and the waste of raw materials, thereby reducing production costs. The use of chemical additives and wastewater discharge are reduced, which meets environmental protection requirements and enhances the sustainable development capabilities of enterprises. The optimization model can be adjusted according to different fiber types and dyeing requirements, and has strong adaptability and versatility.

[0011] Preferably, the polylactic acid fiber image data includes first polylactic acid fiber image data and second polylactic acid fiber image data, and step S1 is specifically as follows:

[0012] Capturing images of the polylactic acid bio-based fibers using a high-definition camera to obtain first polylactic acid fiber image data;

[0013] The image of the polylactic acid bio-based fiber is collected by using a high-definition camera and a high-resolution microscope to obtain the second polylactic acid fiber image data;

[0014] Obtain the chemical additive content data corresponding to the polylactic acid fiber image data.

[0015] The macroscopic image data in the present invention provides a wide range of fiber information, which is suitable for the evaluation of the overall dyeing effect; the microscopic image data provides detailed information on the fiber surface, which is suitable for the analysis and optimization of fine structures. Obtaining detailed chemical additive content data and associating it with image data can improve the accuracy of subsequent model construction. The model can more accurately predict the effect of different auxiliary agent formulas on the dyeing effect, thereby optimizing the dyeing process. Through detailed data collection and correlation analysis, the effect of different dyeing parameters on the final dyeing quality can be found, providing a scientific basis and decision-making support for the optimization of the dyeing process. By initially using a high-definition camera for large-scale rapid screening and then using a high-resolution microscope for detailed analysis, the overall cost of image acquisition can be effectively reduced while ensuring the high quality of the data.

[0016] Preferably, step S2 specifically includes:

[0017] Perform multi-scale convolution calculation on the polylactic acid fiber image data to obtain preliminary convolution layer data;

[0018] Perform multi-head self-attention calculation on the preliminary convolutional layer data to obtain the convolutional layer multi-head self-attention data;

[0019] The convolutional layer data of the polylactic acid fiber image is obtained by weighted fusion based on the multi-head self-attention data of the convolutional layer.

[0020] The multi-scale convolution calculation in the present invention can extract features of different scales, improve the feature extraction ability of the model, and ensure the capture of important information. Through the multi-head self-attention mechanism, the model can pay attention to different parts of the image, improve the understanding and processing ability of complex images, especially for detailed and complex polylactic acid fiber images. The introduction of the self-attention mechanism enables the model to capture long-distance dependencies, optimize the quality of feature selection and representation, and provide higher quality convolutional layer data. By weighted fusion of self-attention features of different heads, a variety of feature information is comprehensively considered, and the performance and prediction accuracy of the model are improved. Through precise image feature extraction and representation, the accuracy and consistency of the dyeing process optimization model are improved, ensuring the stability and high quality of the dyeing effect.

[0021] Preferably, step S3 specifically includes:

[0022] Perform multi-scale pooling calculation on the convolution layer data of the polylactic acid fiber image to obtain preliminary polylactic acid fiber image pooling layer data;

[0023] Normalizing the pooling layer features of the preliminary polylactic acid fiber image pooling layer data to obtain image pooling layer normalized data;

[0024] Perform cross-channel pooling fusion on the normalized data of the image pooling layer to obtain the fused data of the image pooling layer;

[0025] The image pooling layer fusion data is dynamically enhanced to obtain the polylactic acid fiber image pooling layer data.

[0026] In the present invention, multi-scale pooling calculation and cross-channel pooling fusion can capture and integrate feature information of different scales and different channels, and improve the model's ability to understand and represent images. Through feature normalization processing, the scale differences between different features are eliminated, the consistency and comparability of the data are improved, and the training effect of the model is optimized. Cross-channel pooling fusion and dynamic feature enhancement enable the model to comprehensively consider a variety of feature information, enhance the performance of the model, and optimize the final dyeing process model. Multi-scale pooling calculation and data dimensionality reduction effectively reduce the amount of calculation, improve the operating efficiency of the model, and ensure efficient data processing and model training. Dynamic feature enhancement enables the model to maintain stable performance under different conditions, improves the robustness and adaptability of the model, and ensures the consistency and stability of the dyeing process.

[0027] Preferably, the image pooling layer fusion data includes the first image pooling layer fusion data and the second image pooling layer fusion data, and the cross-channel pooling fusion is specifically:

[0028] Perform channel decomposition on the normalized data of the image pooling layer to obtain channel decomposition data;

[0029] Information entropy calculation is performed according to the channel decomposition data to obtain channel decomposition information entropy data;

[0030] Determine whether the channel decomposition information entropy data is greater than or equal to a preset channel decomposition information entropy threshold data;

[0031] When it is determined that the channel decomposition information entropy data is greater than or equal to the preset channel decomposition information entropy threshold data, the channel decomposition data is subjected to product pooling fusion to obtain first image pooling layer fusion data;

[0032] When it is determined that the channel decomposition information entropy data is less than the preset channel decomposition information entropy threshold data, weighted average pooling is performed on the channel decomposition data to obtain the second image pooling layer fusion data.

[0033] The channel decomposition in the present invention helps to process the features of different channels independently, ensuring the independence and clarity of the features of each channel. By decomposing the features of different channels, the unique information of each channel can be better analyzed and processed. Through information entropy calculation, the data processing process can be optimized, the features with high information content can be retained, and the quality of feature representation can be improved. According to the information entropy threshold judgment, different fusion strategies are selected to optimize the feature fusion effect. Product pooling fusion can highlight important features, enhance the expressiveness of features, and improve the accuracy of the model. Weighted average pooling can smooth feature representation, reduce the impact of noise, and improve the stability of features.

[0034] Preferably, the channel decomposition data includes first channel decomposition data and second channel decomposition data, and the channel decomposition is specifically:

[0035] Perform convolution channel decomposition on the normalized data of the image pooling layer to obtain the first channel decomposition data;

[0036] Perform clustering channel decomposition on the normalized data of the image pooling layer to obtain the second channel decomposition data.

[0037] In the present invention, through convolution channel decomposition, feature representations of different convolution kernels can be extracted, multiple features in the image can be captured, and the diversity and expression ability of features can be enhanced. Convolution channel decomposition can separate independent feature channels, avoid interference between different features, and improve the independence and clarity of features. Through multi-level convolution operations, features can be refined and enhanced, and more detailed and important feature information can be captured, providing a high-quality data basis for subsequent processing. Through cluster channel decomposition, similar features can be clustered together, reducing data redundancy and optimizing the aggregation and consistency of feature representation. Cluster channel decomposition can effectively simplify the feature processing process, reduce computational complexity, and improve processing efficiency and the overall performance of the model. Cluster analysis can reveal the inherent structure of data, divide features into groups with clear physical meanings, and enhance the structurality and interpretability of features.

[0038] Preferably, the convolution channel decomposition is specifically as follows:

[0039] Performing sparse coding processing on the normalized data of the image pooling layer to obtain sparse coding data of the image pooling layer;

[0040] Perform convolution calculation according to the sparse coding data of the image pooling layer to obtain sparse coding convolution layer data;

[0041] Perform time domain conversion according to the sparse coding convolution layer data to obtain convolution layer time domain conversion data;

[0042] Gradient feature extraction is performed based on the time domain conversion data of the convolution layer to obtain time domain conversion gradient feature data;

[0043] Performing data segmentation on the convolutional layer time domain conversion data according to the time domain conversion gradient feature data to obtain the convolutional layer time domain segmentation data;

[0044] Perform feature reconstruction according to the time domain segmentation data of the convolution layer to obtain feature reconstruction data;

[0045] Channel mapping is performed according to the feature reconstructed data to obtain first channel decomposition data.

[0046] The sparse coding process in the present invention can extract important features from the data, reduce redundant information, and improve the efficiency of feature representation. The convolution operation can effectively identify local features and patterns in the image and improve the feature recognition ability of the model. The time domain conversion can convert the convolution layer data to another dimension, provide feature analysis from different perspectives, and enhance the diversity of features. Gradient feature extraction can capture key changes and edge features in the data and improve the model's ability to recognize important features. By segmenting different feature areas, the independence and clarity of the features are enhanced. Feature reconstruction can reintegrate the segmented features to improve the integrity and consistency of feature representation. Channel mapping can map the reconstructed features to independent channels, enhancing the independence and feature expression capabilities of each channel.

[0047] Preferably, the cluster channel decomposition is specifically as follows:

[0048] Perform feature selection on the normalized data of the image pooling layer to obtain feature selection data of the image pooling layer;

[0049] Perform clustering calculation on the image pooling layer feature selection data to obtain the image pooling layer clustering feature data;

[0050] Extract cluster centers based on cluster feature data of the image pooling layer to obtain cluster center graph data;

[0051] Perform cluster channel mapping according to the cluster center graph data to obtain second channel decomposition data;

[0052] The cluster channel mapping is as follows:

[0053] Perform feature encoding according to the cluster center graph data to obtain encoded feature data;

[0054] Edge features are transferred and calculated according to the encoded feature data to obtain distributed feature data;

[0055] Performing feature correlation aggregation processing on the distributed feature data to obtain feature correlation aggregation data;

[0056] Perform feature mapping on the data according to feature correlation aggregation to obtain feature mapping data;

[0057] Perform feature alignment and fusion according to feature mapping data to obtain feature fusion data;

[0058] Channel mapping is performed based on the feature fusion data to obtain the second channel decomposition data.

[0059] The present invention can screen out the most important features for dyeing effect through feature selection, and improve feature quality. Through clustering, the structure and hierarchy of data are enhanced, which is helpful for further analysis and processing. Cluster center extraction can simplify feature representation, retain only cluster centers, and reduce data volume. The cluster center represents the main mode of the feature, can optimize feature expression, and improve the recognition ability of the model. Through channel mapping, the cluster center feature is mapped to an independent channel to improve the independence of the feature.

[0060] In the present invention, feature encoding and edge computing are used to preprocess data and disperse the computing load, thereby improving the efficiency and real-time performance of data processing. Edge computing reduces data transmission to the central server in local processing and analysis of data, reduces bandwidth occupancy and transmission delay. By analyzing the correlation between features, the feature representation is optimized, and feature combinations that are more meaningful to the task are extracted. The mapped features are more distinguishable in the new feature space, which helps to improve the accuracy and generalization ability of the model. Through feature alignment, the consistency of data from different sources or different stages is ensured, and the deviation between features is eliminated. The channel mapping process ensures the independence and clarity of the data of each channel, and minimizes the interference between different features. By selectively mapping the fused features to the specified channel, the representation of the feature space is optimized, making the distinction between channels more obvious.

[0061] Preferably, step S4 is specifically:

[0062] Perform forward propagation calculation on the pooling layer data of the polylactic acid fiber image to obtain the preliminary image fully connected layer data;

[0063] Perform activation calculation on the preliminary image full connection layer data to obtain the image full connection layer activation calculation data;

[0064] Perform feature combination based on the activation calculation data of the fully connected layer of the image to obtain combined feature data;

[0065] Perform error calculation based on the combined characteristic data and the corresponding chemical additive content data to obtain error calculation data;

[0066] Perform back propagation according to the error calculation data to obtain back propagation data;

[0067] The fully connected layer data of the polylactic acid fiber image pooling layer is calculated according to the back-propagation data to obtain the fully connected layer data of the polylactic acid fiber image.

[0068] In the present invention, forward propagation can enhance the representation ability of features and ensure the effective transmission of features in the fully connected layer. The activation calculation introduces nonlinear characteristics, which enables the model to capture complex feature relationships and improve the performance of the model. Feature combination can optimize the combination of different activation calculation data to improve the representation ability of features. Through feature combination, the correlation between features can be enhanced, and the model's understanding and processing capabilities of data can be improved. Error calculation can accurately evaluate the performance of the model and identify the shortcomings of the model. Back propagation can optimize the parameters of the model layer by layer according to the error calculation results to improve the performance of the model. Through the calculation of the fully connected layer after back propagation, the convergence of the model can be achieved to ensure the stability and accuracy of the model.

[0069] Preferably, the present application also provides a polylactic acid bio-based fiber dyeing process optimization system, which is used to execute the polylactic acid bio-based fiber dyeing process optimization method as described above, and the polylactic acid bio-based fiber dyeing process optimization system comprises:

[0070] A polylactic acid fiber data acquisition module is used to obtain polylactic acid fiber image data and corresponding chemical additive content data;

[0071] A polylactic acid fiber image convolution module is used to perform self-attention convolution layer calculation on the polylactic acid fiber image data to obtain polylactic acid fiber image convolution layer data;

[0072] The polylactic acid fiber image pooling module is used to perform multi-dimensional pooling layer processing on the polylactic acid fiber image convolution layer data to obtain the polylactic acid fiber image pooling layer data;

[0073] A polylactic acid fiber image fully connected module is used to perform fully connected layer calculation on the polylactic acid fiber image pooling layer data to obtain the polylactic acid fiber image fully connected layer data;

[0074] The PLA bio-based fiber dyeing process optimization model construction module is used to construct and iterate the model according to the fully connected layer data of the PLA fiber image and the corresponding chemical additive content data, so as to obtain the PLA bio-based fiber dyeing process optimization model for performing the PLA bio-based fiber dyeing process optimization operation.

[0075] The beneficial effects of the present invention are: by obtaining the image data of polylactic acid fibers and the corresponding chemical additive content data through a high-definition camera and a high-resolution microscope, comprehensive collection of multi-source data is achieved. Multi-scale convolution calculation extracts information of different scales, and the self-attention mechanism enhances the performance of the model on long-distance dependent features, thereby capturing more key information. Through multi-scale pooling calculation and pooling layer feature normalization, the data dimension is reduced, and the data consistency and comparability are improved. Forward propagation and activation calculation enhance the expressive power of the model by introducing nonlinear characteristics. Feature combination optimizes feature representation and improves feature correlation. Through multiple iterative training, error calculation and back propagation are performed based on the fully connected layer data and the chemical additive content data, and the model parameters are continuously optimized to achieve model convergence and stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Other features, objects and advantages of the present application will become more apparent by reading the detailed description of non-limiting implementations made with reference to the following drawings:

[0077] Figure 1 A flowchart showing a method for optimizing a dyeing process of polylactic acid bio-based fibers according to an embodiment of the present invention is provided;

[0078] Figure 2 A flow chart showing the steps of a method for collecting data of polylactic acid fibers according to an embodiment is shown;

[0079] Figure 3 A flow chart showing the steps of a polylactic acid fiber image convolution method according to an embodiment is shown;

[0080] Figure 4 A flow chart showing the steps of a polylactic acid fiber image pooling method according to an embodiment is shown;

[0081] Figure 5 A flowchart of the steps of a method for fully connecting polylactic acid fiber images according to an embodiment is shown. DETAILED DESCRIPTION

[0082] The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by technicians in this field without creative work are within the scope of protection of the present invention.

[0083] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor methods and / or microcontroller methods.

[0084] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0085] See also Figures 1 to 5 The present application provides a method for optimizing the dyeing process of polylactic acid bio-based fibers, comprising the following steps:

[0086] Step S1: Acquire polylactic acid fiber image data and corresponding chemical additive content data;

[0087] Specifically, a high-resolution microscope was used to obtain images of polylactic acid fibers at different dyeing stages. The corresponding chemical additive data, including the type, concentration, and usage time of the additive, were collected. The images were preprocessed, such as denoising, contrast enhancement, and standardization. The processed image data was paired with the chemical additive data to establish a data set.

[0088] Step S2: performing self-attention convolution layer calculation on the polylactic acid fiber image data to obtain polylactic acid fiber image convolution layer data;

[0089] Specifically, a convolutional neural network (CNN) is designed, and its structure includes multiple convolutional layers, self-attention mechanism and activation function. The preprocessed polylactic acid fiber image data is input into the first convolutional layer of CNN, and convolution operation is performed to extract primary features. A self-attention mechanism is added after each convolutional layer to highlight the important features in the image through weighted calculation. Multi-layer convolution and self-attention processing are performed to obtain the convolutional layer data of the polylactic acid fiber image.

[0090] Step S3: performing multi-dimensional pooling layer processing on the polylactic acid fiber image convolution layer data to obtain polylactic acid fiber image pooling layer data;

[0091] Specifically, the data output by the convolution layer is input into the multidimensional pooling layer, and the maximum pooling or average pooling method is adopted. The data dimension is reduced by the pooling operation, the computational complexity and the risk of overfitting are reduced, and the main features are retained at the same time. Multi-dimensional pooling processing, such as 2D pooling and 3D pooling, is performed in the pooling layer to further extract important information from the polylactic acid fiber image. The data after the pooling processing is output, that is, the polylactic acid fiber image pooling layer data.

[0092] Step S4: performing full connection layer calculation on the pooling layer data of the polylactic acid fiber image to obtain full connection layer data of the polylactic acid fiber image;

[0093] Specifically, the data output by the pooling layer is expanded into a one-dimensional vector and input into the fully connected layer. A multi-layer fully connected network is designed, each layer contains a large number of neurons, and calculations are performed through weight matrices and activation functions. Forward propagation is performed in the fully connected layer, the output of neurons is calculated layer by layer, and high-level features are extracted. The calculation results of the fully connected layer are output, that is, the fully connected layer data of the polylactic acid fiber image.

[0094] Step S5: construct and iterate a model based on the fully connected layer data of the polylactic acid fiber image and the corresponding chemical additive content data to obtain a polylactic acid bio-based fiber dyeing process optimization model to perform polylactic acid bio-based fiber dyeing process optimization operations.

[0095] Specifically, the polylactic acid fiber image data and the corresponding chemical additive content data are divided to obtain a training data set and a test data set; according to the training data set, training is performed through steps S1-S4 and a preset loss function (square difference loss function) to obtain a preliminary polylactic acid bio-based fiber dyeing process optimization model; according to the test data set, the preliminary polylactic acid bio-based fiber dyeing process optimization model is iteratively trained to obtain a polylactic acid bio-based fiber dyeing process optimization model.

[0096] Specifically, the fully connected layer data of the PLA fiber image and the chemical additive content data are used as input to design a comprehensive optimization model. A multi-layer perceptron (MLP) or other advanced models, such as LSTM and GRU, are used for preliminary model construction. A loss function, such as mean square error (MSE), is defined to evaluate the model performance. The back propagation algorithm is used to optimize the model parameters through continuous iterations to minimize the loss function. Based on the optimized model, the dyeing process parameters of the PLA bio-based fiber are extracted to optimize the actual dyeing process. The optimization model is verified and tested to ensure its good stability and accuracy under different conditions.

[0097] The present invention can significantly improve the dyeing quality of polylactic acid bio-based fibers and ensure dyeing uniformity and color fastness through precise feature extraction and optimization model. The optimized dyeing process reduces the number of experiments and the waste of raw materials, thereby reducing production costs. The use of chemical additives and wastewater discharge are reduced, which meets environmental protection requirements and enhances the sustainable development capabilities of enterprises. The optimization model can be adjusted according to different fiber types and dyeing requirements, and has strong adaptability and versatility.

[0098] Preferably, the polylactic acid fiber image data includes first polylactic acid fiber image data and second polylactic acid fiber image data, and step S1 is specifically as follows:

[0099] Capturing images of the polylactic acid bio-based fibers using a high-definition camera to obtain first polylactic acid fiber image data;

[0100] Specifically, a high-definition camera was used to photograph the polylactic acid fiber samples, ensuring the camera's resolution and focus clarity to capture the fiber's macroscopic structure.

[0101] The image of the polylactic acid bio-based fiber is collected by using a high-definition camera and a high-resolution microscope to obtain the second polylactic acid fiber image data;

[0102] Specifically, a high-definition camera and a high-resolution microscope are used to take microscopic photos of the polylactic acid fibers to obtain the detailed structure of the fibers. The microscope should have sufficient magnification to capture the tiny features on the fiber surface.

[0103] Obtain the chemical additive content data corresponding to the polylactic acid fiber image data.

[0104] Specifically, record the type, concentration, and usage time of chemical additives used in the dyeing process. Use standardized recording forms or data collection systems to ensure the accuracy and completeness of the data. Associate the chemical additive data with the corresponding image data. Each set of image data should have corresponding additive information. This can be achieved through a unified numbering system or database management system.

[0105] The macroscopic image data in the present invention provides a wide range of fiber information, which is suitable for the evaluation of the overall dyeing effect; the microscopic image data provides detailed information on the fiber surface, which is suitable for the analysis and optimization of fine structures. Obtaining detailed chemical additive content data and associating it with image data can improve the accuracy of subsequent model construction. The model can more accurately predict the effect of different auxiliary agent formulas on the dyeing effect, thereby optimizing the dyeing process. Through detailed data collection and correlation analysis, the effect of different dyeing parameters on the final dyeing quality can be found, providing a scientific basis and decision-making support for the optimization of the dyeing process. By initially using a high-definition camera for large-scale rapid screening and then using a high-resolution microscope for detailed analysis, the overall cost of image acquisition can be effectively reduced while ensuring the high quality of the data.

[0106] Preferably, step S2 specifically includes:

[0107] Perform multi-scale convolution calculation on the polylactic acid fiber image data to obtain preliminary convolution layer data;

[0108] Specifically, a convolutional neural network (CNN) is constructed by presetting parameters, which includes multiple convolution kernels of different scales (2 x 2, 3 x 3, 5 x 5 or other convolution kernels) to extract features at different levels. For each polylactic acid fiber image data, convolution operations are performed using convolution kernels of different sizes.

[0109] Perform multi-head self-attention calculation on the preliminary convolutional layer data to obtain the convolutional layer multi-head self-attention data;

[0110] Specifically, the multi-head self-attention mechanism in Transformer is used to perform self-attention calculation on the preliminary convolutional layer data. Each convolution output is calculated by multi-head self-attention to capture important features between different channels.

[0111] The convolutional layer data of the polylactic acid fiber image is obtained by weighted fusion based on the multi-head self-attention data of the convolutional layer.

[0112] Specifically, the data obtained by multi-head self-attention calculations are weighted and fused to generate the final convolutional layer features. The data from multiple heads are fused into a comprehensive feature representation using weighted average or other weighted methods.

[0113] The multi-scale convolution calculation in the present invention can extract features of different scales, improve the feature extraction ability of the model, and ensure the capture of important information. Through the multi-head self-attention mechanism, the model can pay attention to different parts of the image, improve the understanding and processing ability of complex images, especially for detailed and complex polylactic acid fiber images. The introduction of the self-attention mechanism enables the model to capture long-distance dependencies, optimize the quality of feature selection and representation, and provide higher quality convolutional layer data. By weighted fusion of self-attention features of different heads, a variety of feature information is comprehensively considered, and the performance and prediction accuracy of the model are improved. Through precise image feature extraction and representation, the accuracy and consistency of the dyeing process optimization model are improved, ensuring the stability and high quality of the dyeing effect.

[0114] Preferably, step S3 specifically includes:

[0115] Perform multi-scale pooling calculation on the convolution layer data of the polylactic acid fiber image to obtain preliminary polylactic acid fiber image pooling layer data;

[0116] Specifically, pooling windows of different sizes are used to perform multi-scale pooling operations on the convolutional layer data. Maximum pooling and average pooling methods are used to extract features.

[0117] Normalizing the pooling layer features of the preliminary polylactic acid fiber image pooling layer data to obtain image pooling layer normalized data;

[0118] Specifically, the pooling layer data is normalized to eliminate feature scale differences and improve model training results. Use batch normalization or layer normalization methods.

[0119] Perform cross-channel pooling fusion on the normalized data of the image pooling layer to obtain the fused data of the image pooling layer;

[0120] Specifically, the normalized data is pooled and fused across channels to integrate the features of each channel, using weighted average or other fusion methods.

[0121] The image pooling layer fusion data is dynamically enhanced to obtain the polylactic acid fiber image pooling layer data.

[0122] Specifically, dynamic feature enhancement is performed on the fused data to highlight important features. Feature enhancement is performed using learnable parameters or adaptive methods.

[0123] Input layer: receives the fused data from the pooling layer. Attention mechanism layer: generates attention weights using convolution operations and activation functions. Learnable weighting mechanism: uses learnable parameters to weight features. Output layer: outputs enhanced feature data.

[0124] Receives pooling layer fusion data of shape (batch_size / batch data size, channels / channels, height / height, width / width). Generates attention weights using 1x1 or 2x2 convolution operations. Uses sigmoid activation function to limit weight values ​​between 0 and 1. Scales features using learnable parameters. Learnable parameters are initialized to all 1 and optimized through training process. Outputs enhanced feature data with the same shape as input.

[0125] Use the learnable parameter scale to scale the features. The scale parameter shape is (1, channels, 1,1), which means that each channel is scaled separately. The scale parameter is initialized to all 1s and optimized through the training process.

[0126] In the present invention, multi-scale pooling calculation and cross-channel pooling fusion can capture and integrate feature information of different scales and different channels, and improve the model's ability to understand and represent images. Through feature normalization processing, the scale differences between different features are eliminated, the consistency and comparability of the data are improved, and the training effect of the model is optimized. Cross-channel pooling fusion and dynamic feature enhancement enable the model to comprehensively consider a variety of feature information, enhance the performance of the model, and optimize the final dyeing process model. Multi-scale pooling calculation and data dimensionality reduction effectively reduce the amount of calculation, improve the operating efficiency of the model, and ensure efficient data processing and model training. Dynamic feature enhancement enables the model to maintain stable performance under different conditions, improves the robustness and adaptability of the model, and ensures the consistency and stability of the dyeing process.

[0127] Preferably, the image pooling layer fusion data includes the first image pooling layer fusion data and the second image pooling layer fusion data, and the cross-channel pooling fusion is specifically:

[0128] Perform channel decomposition on the normalized data of the image pooling layer to obtain channel decomposition data;

[0129] Specifically, the normalized data is decomposed by channel to obtain independent data for each channel. The shape of the normalized pooling layer data is (batch_size, channels, height, width). For example, batch_size=8, channels=256, height=32, width=32. Decompose the input data into channels independent tensors by channel dimension (dim=1). Split the input tensor along the specified dimension and return a tuple containing the split tensors.

[0130] Information entropy calculation is performed according to the channel decomposition data to obtain channel decomposition information entropy data;

[0131] Specifically, the information entropy is calculated for each decomposed channel data to quantify the complexity of the data. Each channel data is processed and flattened into a two-dimensional matrix. The flattened data is normalized and the probability of each pixel value is calculated. The frequency of each pixel value is counted using the histogram method and then divided by the total number of pixels to obtain the probability distribution. The information entropy is calculated based on the probability distribution. For each channel data, the information entropy of each sample is calculated and then the average is calculated to obtain the average information entropy of the channel.

[0132] Determine whether the channel decomposition information entropy data is greater than or equal to a preset channel decomposition information entropy threshold data;

[0133] Specifically, an information entropy threshold is set and the entropy of each channel is compared.

[0134] When it is determined that the channel decomposition information entropy data is greater than or equal to the preset channel decomposition information entropy threshold data, the channel decomposition data is subjected to product pooling fusion to obtain first image pooling layer fusion data;

[0135] Specifically, for channels that meet the threshold, perform product pooling operations and multiply the data of each channel to obtain the fusion result. Initialize a tensor to store the result of product pooling. During initialization, the shape of the tensor is the same as the shape of each channel data, and all elements are set to 1. Traverse the channel data that meet the threshold and multiply the data of each channel with the initialized result tensor element by element. The final fusion result tensor contains the result of element-by-element multiplication of all channel data that meet the threshold.

[0136] When it is determined that the channel decomposition information entropy data is less than the preset channel decomposition information entropy threshold data, weighted average pooling is performed on the channel decomposition data to obtain the second image pooling layer fusion data.

[0137] Specifically, for channels that do not meet the threshold, perform weighted average pooling. Prepare a list of channel data that do not meet the threshold, and the shape of each channel data is (batch_size, height, width). Prepare a corresponding list of weighted coefficients, each coefficient corresponding to a channel data. Initialize a tensor weighted_sum to store the sum of the weighted channel data. During initialization, the shape of the tensor is the same as the shape of each channel data, and all elements are set to 0. Traverse the list of channel data that do not meet the threshold, multiply the data of each channel by the corresponding preset weight, and then accumulate it to weighted_sum. Divide weighted_sum by the sum of all weights to obtain the fusion result of weighted average pooling.

[0138] The channel decomposition in the present invention helps to process the features of different channels independently, ensuring the independence and clarity of the features of each channel. By decomposing the features of different channels, the unique information of each channel can be better analyzed and processed. Through information entropy calculation, the data processing process can be optimized, the features with high information content can be retained, and the quality of feature representation can be improved. According to the information entropy threshold judgment, different fusion strategies are selected to optimize the feature fusion effect. Product pooling fusion can highlight important features, enhance the expressiveness of features, and improve the accuracy of the model. Weighted average pooling can smooth feature representation, reduce the impact of noise, and improve the stability of features.

[0139] Preferably, the channel decomposition data includes first channel decomposition data and second channel decomposition data, and the channel decomposition is specifically:

[0140] Perform convolution channel decomposition on the normalized data of the image pooling layer to obtain the first channel decomposition data;

[0141] Specifically, the convolution operation is used to perform channel decomposition on the normalized data of the image pooling layer, and different features are extracted through different convolution kernels. Two convolution kernels of different sizes are used to perform convolution operations on the normalized data of the image pooling layer to extract features of different scales. The output of the convolution operation is used as the first channel decomposition data.

[0142] Perform clustering channel decomposition on the normalized data of the image pooling layer to obtain the second channel decomposition data.

[0143] Specifically, a clustering algorithm (such as K-means) is used to decompose the channels and cluster the channels according to feature similarity. The normalized data of the image pooling layer is flattened into a two-dimensional matrix, and the data of each channel is used as a feature vector. The flattened data is clustered using the K-means clustering algorithm, and each channel is assigned to a different cluster. A mask is generated based on the clustering results, and the original data is decomposed according to the mask to obtain the channel decomposition data of each cluster. These clustered data are used as the second channel decomposition data.

[0144] In the present invention, through convolution channel decomposition, feature representations of different convolution kernels can be extracted, multiple features in the image can be captured, and the diversity and expression ability of features can be enhanced. Convolution channel decomposition can separate independent feature channels, avoid interference between different features, and improve the independence and clarity of features. Through multi-level convolution operations, features can be refined and enhanced, and more detailed and important feature information can be captured, providing a high-quality data basis for subsequent processing. Through cluster channel decomposition, similar features can be clustered together, reducing data redundancy and optimizing the aggregation and consistency of feature representation. Cluster channel decomposition can effectively simplify the feature processing process, reduce computational complexity, and improve processing efficiency and the overall performance of the model. Cluster analysis can reveal the inherent structure of data, divide features into groups with clear physical meanings, and enhance the structurality and interpretability of features.

[0145] Preferably, the convolution channel decomposition is specifically as follows:

[0146] Performing sparse coding processing on the normalized data of the image pooling layer to obtain sparse coding data of the image pooling layer;

[0147] Specifically, sparse coding technology is used to process image data and extract sparse features. Sparse autoencoders or other sparse representation methods are used.

[0148] Perform convolution calculation according to the sparse coding data of the image pooling layer to obtain sparse coding convolution layer data;

[0149] Specifically, convolution operation is used to extract features from sparse coded data. Define multiple convolution layers, each of which contains several convolution kernels. For example, the size of the convolution kernel in the first layer is 3x3, and the number is 64; the size of the convolution kernel in the second layer is 3x3, and the number is 128. Input the sparse coded data into the first convolution layer and perform convolution operation. The convolution operation includes sliding window operation on the input data, and each operation calculates the inner product of the convolution kernel and the input data, and adds a bias. Apply activation function (such as ReLU) to the result of convolution operation for nonlinear transformation. The output of the first convolution operation is a feature map with a shape of (batch_size, 64, output_height_1, output_width_1). Input the feature map output by the first convolution layer into the second convolution layer, and repeat the above convolution operation. The output of the second convolution operation is a higher-level feature map with a shape of (batch_size, 128, output_height_2, output_width_2).

[0150] Perform time domain conversion according to the sparse coding convolution layer data to obtain convolution layer time domain conversion data;

[0151] Specifically, the convolution layer data is transformed into the time domain, such as Fourier transform, etc. A two-dimensional Fourier transform is performed on each convolution layer data / feature map to convert the data from the time domain to the frequency domain, such as using a fast Fourier transform (FFT) algorithm.

[0152] The feature map conv_output output by the convolution layer has a shape of (8, 128, 32, 32), contains 8 samples, each sample has 128 feature maps, and each feature map is 32x32 in size. A two-dimensional Fourier transform is performed on each feature map. The Fourier transform converts each feature map from the spatial domain to the frequency domain. The Fourier transform can be efficiently calculated using the Fast Fourier Transform (FFT) algorithm. The result of the Fourier transform is a complex number, with each element containing a real part and an imaginary part. The shape of the result is the same as the input, which is (8,128, 32, 32).

[0153] Gradient feature extraction is performed based on the time domain conversion data of the convolution layer to obtain time domain conversion gradient feature data;

[0154] Specifically, the gradient calculation method is used to extract features from the time domain conversion data. Input the time domain conversion data time_domain_data, the shape is (8, 128, 32,32). Use the Sobel operator to calculate the gradient of each feature map. The Sobel operator calculates the gradient in the x and y directions respectively. The calculation result is two gradient maps with the same shape as the input, representing the gradients in the x and y directions respectively. The gradient magnitude is calculated based on the gradients in the x and y directions. The gradient magnitude is calculated using the square root of the sum of squares. The shape of the obtained gradient magnitude map is (8, 128, 32, 32), which is the same as the input shape.

[0155] Performing data segmentation on the convolutional layer time domain conversion data according to the time domain conversion gradient feature data to obtain the convolutional layer time domain segmentation data;

[0156] Specifically, data segmentation is performed based on gradient features to extract salient regions. The threshold of the gradient magnitude map is set according to the specific task and used to segment the data. The threshold can be a fixed value or dynamically determined based on the statistical information of the gradient magnitude map. A threshold is applied to the gradient magnitude map, and regions greater than the threshold are marked as salient regions. A mask is generated based on the threshold, and the mask shape is the same as the input data. The salient regions in the mask are marked as 1, and the remaining regions are marked as 0. The original data is segmented using the mask to extract the salient regions.

[0157] Perform feature reconstruction according to the time domain segmentation data of the convolution layer to obtain feature reconstruction data;

[0158] Specifically, the segmented data is feature reconstructed to restore important information. The segmented data segmented_data has a shape of (8, 128, 32, 32) and contains 8 samples, each sample has 128 feature maps, and each feature map is 32x32 in size. An inverse transform operation is performed on each segmented feature map to restore important information. The frequency domain information is converted back to the time domain using an inverse Fourier transform (inverse FFT). The inverse Fourier transform restores the frequency domain features previously obtained by the Fourier transform to the original time domain features. The result of the inverse Fourier transform is the restored feature map. The shape of the restored feature map is (8, 128, 32, 32), which is the same as the input shape.

[0159] Channel mapping is performed according to the feature reconstructed data to obtain first channel decomposition data.

[0160] Specifically, channel mapping is performed on the reconstructed feature data to generate first channel decomposition data. The data of each channel is mapped using a linear transformation. Channel mapping can be implemented using a fully connected layer. After flattening the data of each channel, it is mapped through a fully connected layer and the result is restored to its original shape.

[0161] The sparse coding process in the present invention can extract important features from the data, reduce redundant information, and improve the efficiency of feature representation. The convolution operation can effectively identify local features and patterns in the image and improve the feature recognition ability of the model. The time domain conversion can convert the convolution layer data to another dimension, provide feature analysis from different perspectives, and enhance the diversity of features. Gradient feature extraction can capture key changes and edge features in the data and improve the model's ability to recognize important features. By segmenting different feature areas, the independence and clarity of the features are enhanced. Feature reconstruction can reintegrate the segmented features to improve the integrity and consistency of feature representation. Channel mapping can map the reconstructed features to independent channels, enhancing the independence and feature expression capabilities of each channel.

[0162] Preferably, the cluster channel decomposition is specifically as follows:

[0163] Perform feature selection on the normalized data of the image pooling layer to obtain feature selection data of the image pooling layer;

[0164] Specifically, feature selection methods (such as principal component analysis (PCA)) are used to extract important features from the normalized data.

[0165] Perform clustering calculation on the image pooling layer feature selection data to obtain the image pooling layer clustering feature data;

[0166] Specifically, a clustering algorithm (such as K-means) is used to cluster the feature selection data.

[0167] Extract cluster centers based on cluster feature data of the image pooling layer to obtain cluster center graph data;

[0168] Specifically, extract the central features of each cluster according to the clustering results. The cluster feature data clustered_features, with a shape of (8, 128, 32,32), contains 8 samples, each sample has feature data of 128 channels, and each feature map is 32x32 in size. The cluster label data cluster_labels, with a shape of (8, 32, 32), represents the cluster label of each position. Assume that num_clusters = 10. Initialize the cluster_centers tensor with a shape of (8,10, 128) to store the central features of each cluster. For each sample i and each cluster j: Generate a Boolean mask mask = (cluster_labels[i]== j). Use the mask to extract the features of the corresponding position in clustered_features[i]. Calculate the mean of the extracted features center = masked_features.mean(dim=(1, 2)). Store the mean in cluster_centers[i, j]. The calculated cluster centers are stored in cluster_centers. The final shape of cluster_centers is (8, 10, 128).

[0169] Perform cluster channel mapping according to the cluster center graph data to obtain second channel decomposition data;

[0170] Specifically, use the cluster center features for channel mapping to generate the second channel decomposition data. Ensure that the input cluster center graph data and cluster label data have the correct shape. Assume that the shape of the cluster center graph data cluster_centers is (8, 10, 128) and the shape of the cluster label data cluster_labels is (8,32, 32). Initialize a tensor mapped_channels to store the mapped channel data. The shape of the mapping result is (batch_size, channels, height, width). Traverse each sample and each position, and map the cluster center features to the corresponding positions according to the cluster labels. Use the cluster label data to determine the cluster center features of each position and map them to the corresponding channel data. The result of channel mapping is new feature data with the same shape as the input.

[0171] The cluster channel mapping is as follows:

[0172] Perform feature encoding according to the cluster center graph data to obtain encoded feature data;

[0173] Specifically, the cluster center graph data is standardized to reduce the magnitude differences between features. PCA is used to reduce the dimension of the standardized data, extract the main components, and retain key information. The reduced-dimensional data is encoded and converted, such as using One-Hot Encoding to encode the category data.

[0174] Edge features are transferred and calculated according to the encoded feature data to obtain distributed feature data;

[0175] Specifically, data and computing tasks are distributed to edge computing devices, such as edge servers or smart devices. Feature calculations are performed on edge devices, including local aggregation and feature enhancement.

[0176] Performing feature correlation aggregation processing on the distributed feature data to obtain feature correlation aggregation data;

[0177] Specifically, the correlation coefficient between the distributed feature data, such as the Pearson correlation coefficient, is calculated, and the features are weighted and aggregated based on the correlation to form a comprehensive feature representation.

[0178] Perform feature mapping on the data according to feature correlation aggregation to obtain feature mapping data;

[0179] Specifically, an appropriate mapping function, such as linear mapping or nonlinear mapping, is selected according to the statistical properties of the feature, and the selected mapping function is applied to transform the feature correlation aggregation data into a new feature space.

[0180] Perform feature alignment and fusion according to feature mapping data to obtain feature fusion data;

[0181] Specifically, feature mapping data from different sources are aligned to eliminate deviations and inconsistencies between features, and the aligned features are processed comprehensively using weighted average or other fusion methods.

[0182] Specifically, the feature mapping data includes original feature data and alignment target feature data, and the two are aligned and fused to obtain feature fusion data, the original feature data is the feature mapping data selected in sequence, and the alignment target feature data is the feature mapping data other than the original feature data; wherein the feature alignment is calculated by the feature alignment index calculation formula to obtain the feature alignment index for alignment, and the feature alignment index calculation formula is specifically:

[0183] ;

[0184] For feature fusion data, For the The original feature data, is the feature mapping order item data, The influence intensity data for the feature map (generated by calculation via polynomial or exponential), For the The feature attribute data corresponding to the alignment target feature data, For the Align target feature data, is auxiliary feature data (for nonlinear functions such as trigonometric functions to make adjustments), is the preset dimension-upgraded data corresponding to the feature mapping data, is the feature alignment weight factor data, is the feature phase angle data (the phase angle of the original feature data and the aligned target feature data in the mapping space, used to capture the phase difference between features).

[0185] Channel mapping is performed based on the feature fusion data to obtain the second channel decomposition data.

[0186] Specifically, a new feature channel is reconstructed based on the fused feature data to ensure the independence and validity of each channel feature, and the reconstructed feature is mapped to the specified channel to form the second channel decomposition data.

[0187] Determine mapping rules, such as feature strength, category, or purpose, based on the characteristics of feature fusion data and application requirements. Assign reconstructed feature data to different channels and map according to predefined rules. For each reconstructed feature, apply mapping rules to assign it to the corresponding channel. Ensure that the features in each channel are independent and consistent and meet the requirements of the application. Generate second channel decomposition data.

[0188] The present invention can screen out the most important features for dyeing effect through feature selection, and improve feature quality. Through clustering, the structure and hierarchy of data are enhanced, which is helpful for further analysis and processing. Cluster center extraction can simplify feature representation, retain only cluster centers, and reduce data volume. The cluster center represents the main mode of the feature, can optimize feature expression, and improve the recognition ability of the model. Through channel mapping, the cluster center feature is mapped to an independent channel to improve the independence of the feature.

[0189] In the present invention, feature encoding and edge computing are used to preprocess data and disperse the computing load, thereby improving the efficiency and real-time performance of data processing. Edge computing reduces data transmission to the central server in local processing and analysis of data, reduces bandwidth occupancy and transmission delay. By analyzing the correlation between features, the feature representation is optimized, and feature combinations that are more meaningful to the task are extracted. The mapped features are more distinguishable in the new feature space, which helps to improve the accuracy and generalization ability of the model. Through feature alignment, the consistency of data from different sources or different stages is ensured, and the deviation between features is eliminated. The channel mapping process ensures the independence and clarity of the data of each channel, and minimizes the interference between different features. By selectively mapping the fused features to the specified channel, the representation of the feature space is optimized, making the distinction between channels more obvious.

[0190] Preferably, step S4 is specifically:

[0191] Perform forward propagation calculation on the pooling layer data of the polylactic acid fiber image to obtain the preliminary image fully connected layer data;

[0192] Specifically, a fully connected layer is used to perform forward propagation calculations on the pooling layer data. Input the pooling layer data pooled_output, with a shape of (8, 128, 32, 32). Flatten the data of each sample into a one-dimensional vector. The flattened shape is (8, 128 * 32 * 32), i.e. (8, 131072). Define a fully connected layer to map the flattened data to a new feature space. The input dimension of the fully connected layer is 131072 (i.e. 128 * 32 * 32), and the output dimension is assumed to be 1024. Input the flattened data into the fully connected layer for forward propagation calculations. The forward propagation calculation obtains the preliminary image fully connected layer data with a shape of (8, 1024).

[0193] Perform activation calculation on the preliminary image full connection layer data to obtain the image full connection layer activation calculation data;

[0194] Specifically, an activation function (such as ReLU) is used to perform activation calculation on the output of the fully connected layer.

[0195] Perform feature combination based on the activation calculation data of the fully connected layer of the image to obtain combined feature data;

[0196] Specifically, the activated data is subjected to feature combination, such as combining through multiple fully connected layers.

[0197] Perform error calculation based on the combined characteristic data and the corresponding chemical additive content data to obtain error calculation data;

[0198] Specifically, the error between the combined features and the chemical additive content data is calculated, such as using the mean square error (MSE).

[0199] Define multiple fully connected layers to further combine the activation data and extract more features. Use two fully connected layers, with the output dimension of the first layer being 512 and the output dimension of the second layer being 256. Input the activation data into the first fully connected layer for forward propagation calculation. Perform activation calculation on the output, such as using the ReLU activation function. Input the activated data into the second fully connected layer for forward propagation calculation. Perform activation calculation on the output to obtain the combined feature data. Concatenate the combined feature data with the chemical additive content data to form a new feature vector. Define an error calculation function, such as the mean square error (MSE). Use the error calculation function to calculate the error between the combined feature data and the target data to obtain the error calculation data.

[0200] Perform back propagation according to the error calculation data to obtain back propagation data;

[0201] Specifically, the back-propagation algorithm is used to calculate the gradient to update the network parameters.

[0202] Make sure the error (loss) is calculated. The loss function is mean squared error (MSE). Calculate the error between the predicted value and the true value to get the loss value. The loss value is used to guide the backpropagation to calculate the gradient. Initialize the gradient of the loss value to 1. Start from the output layer and calculate the gradient backward layer by layer. Use the chain rule to calculate the gradient of each layer. Use gradient descent to update the weights and bias of each layer.

[0203] The fully connected layer data of the polylactic acid fiber image pooling layer is calculated according to the back-propagation data to obtain the fully connected layer data of the polylactic acid fiber image.

[0204] Specifically, the fully connected layer parameters are updated according to the back-propagation results, and a new forward propagation calculation is performed.

[0205] In the present invention, forward propagation can enhance the representation ability of features and ensure the effective transmission of features in the fully connected layer. The activation calculation introduces nonlinear characteristics, which enables the model to capture complex feature relationships and improve the performance of the model. Feature combination can optimize the combination of different activation calculation data to improve the representation ability of features. Through feature combination, the correlation between features can be enhanced, and the model's understanding and processing capabilities of data can be improved. Error calculation can accurately evaluate the performance of the model and identify the shortcomings of the model. Back propagation can optimize the parameters of the model layer by layer according to the error calculation results to improve the performance of the model. Through the calculation of the fully connected layer after back propagation, the convergence of the model can be achieved to ensure the stability and accuracy of the model.

[0206] Preferably, the present application also provides a polylactic acid bio-based fiber dyeing process optimization system, which is used to execute the polylactic acid bio-based fiber dyeing process optimization method as described above, and the polylactic acid bio-based fiber dyeing process optimization system comprises:

[0207] A polylactic acid fiber data acquisition module is used to obtain polylactic acid fiber image data and corresponding chemical additive content data;

[0208] A polylactic acid fiber image convolution module is used to perform self-attention convolution layer calculation on the polylactic acid fiber image data to obtain polylactic acid fiber image convolution layer data;

[0209] The polylactic acid fiber image pooling module is used to perform multi-dimensional pooling layer processing on the polylactic acid fiber image convolution layer data to obtain the polylactic acid fiber image pooling layer data;

[0210] A polylactic acid fiber image fully connected module is used to perform fully connected layer calculation on the polylactic acid fiber image pooling layer data to obtain the polylactic acid fiber image fully connected layer data;

[0211] The PLA bio-based fiber dyeing process optimization model construction module is used to construct and iterate the model according to the fully connected layer data of the PLA fiber image and the corresponding chemical additive content data, so as to obtain the PLA bio-based fiber dyeing process optimization model for performing the PLA bio-based fiber dyeing process optimization operation.

[0212] The beneficial effects of the present invention are: by obtaining the image data of polylactic acid fibers and the corresponding chemical additive content data through a high-definition camera and a high-resolution microscope, comprehensive collection of multi-source data is achieved. Multi-scale convolution calculation extracts information of different scales, and the self-attention mechanism enhances the performance of the model on long-distance dependent features, thereby capturing more key information. Through multi-scale pooling calculation and pooling layer feature normalization, the data dimension is reduced, and the data consistency and comparability are improved. Forward propagation and activation calculation enhance the expressive power of the model by introducing nonlinear characteristics. Feature combination optimizes feature representation and improves feature correlation. Through multiple iterative training, error calculation and back propagation are performed based on the fully connected layer data and the chemical additive content data, and the model parameters are continuously optimized to achieve model convergence and stability.

[0213] Therefore, from any point of view, the embodiments should be regarded as illustrative and non-restrictive, and the scope of the present invention is limited by the attached application documents rather than the above description, and it is intended that all changes falling within the meaning and scope of equivalent elements of the application documents are included in the present invention.

[0214] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A method for optimizing the dyeing process of polylactic acid bio-based fibers, characterized in that: The following steps are involved: Step S1: Acquire polylactic acid fiber image data and corresponding chemical additive content data; Step S2: performing self-attention convolution layer calculation on the polylactic acid fiber image data to obtain polylactic acid fiber image convolution layer data; Step S3: performing multi-dimensional pooling layer processing on the polylactic acid fiber image convolution layer data to obtain polylactic acid fiber image pooling layer data; Step S4: performing full connection layer calculation on the pooling layer data of the polylactic acid fiber image to obtain full connection layer data of the polylactic acid fiber image; Step S5: constructing and iterating a model according to the fully connected layer data of the polylactic acid fiber image and the corresponding chemical additive content data to obtain a polylactic acid bio-based fiber dyeing process optimization model to perform a polylactic acid bio-based fiber dyeing process optimization operation; Step S3 is specifically as follows: Perform multi-scale pooling calculation on the convolution layer data of the polylactic acid fiber image to obtain preliminary polylactic acid fiber image pooling layer data; Normalizing the pooling layer features of the preliminary polylactic acid fiber image pooling layer data to obtain image pooling layer normalized data; Perform cross-channel pooling fusion on the normalized data of the image pooling layer to obtain the fused data of the image pooling layer; Perform dynamic feature enhancement on the image pooling layer fusion data to obtain the polylactic acid fiber image pooling layer data; The image pooling layer fusion data includes the first image pooling layer fusion data and the second image pooling layer fusion data. The cross-channel pooling fusion is specifically as follows: Perform channel decomposition on the normalized data of the image pooling layer to obtain channel decomposition data; Information entropy calculation is performed according to the channel decomposition data to obtain channel decomposition information entropy data; Determine whether the channel decomposition information entropy data is greater than or equal to a preset channel decomposition information entropy threshold data; When it is determined that the channel decomposition information entropy data is greater than or equal to the preset channel decomposition information entropy threshold data, the channel decomposition data is subjected to product pooling fusion to obtain first image pooling layer fusion data; When it is determined that the channel decomposition information entropy data is less than the preset channel decomposition information entropy threshold data, weighted average pooling is performed on the channel decomposition data to obtain the second image pooling layer fusion data.

2. The method according to claim 1, characterized in that The polylactic acid fiber image data includes first polylactic acid fiber image data and second polylactic acid fiber image data, and step S1 is specifically as follows: Capturing images of the polylactic acid bio-based fibers using a high-definition camera to obtain first polylactic acid fiber image data; The image of the polylactic acid bio-based fiber is collected by using a high-definition camera and a high-resolution microscope to obtain the second polylactic acid fiber image data; Obtain the chemical additive content data corresponding to the polylactic acid fiber image data.

3. The method according to claim 1, characterized in that Step S2 is specifically as follows: Perform multi-scale convolution calculation on the polylactic acid fiber image data to obtain preliminary convolution layer data; Perform multi-head self-attention calculation on the preliminary convolutional layer data to obtain the convolutional layer multi-head self-attention data; The convolutional layer data of the polylactic acid fiber image is obtained by weighted fusion based on the multi-head self-attention data of the convolutional layer.

4. The method according to claim 1, characterized in that: The channel decomposition data includes first channel decomposition data and second channel decomposition data, and the channel decomposition is specifically: Perform convolution channel decomposition on the normalized data of the image pooling layer to obtain the first channel decomposition data; Perform clustering channel decomposition on the normalized data of the image pooling layer to obtain the second channel decomposition data.

5. The method according to claim 4, characterized in that The convolution channel decomposition is specifically as follows: Performing sparse coding processing on the normalized data of the image pooling layer to obtain sparse coding data of the image pooling layer; Perform convolution calculation according to the sparse coding data of the image pooling layer to obtain sparse coding convolution layer data; Perform time domain conversion according to the sparse coding convolution layer data to obtain convolution layer time domain conversion data; Gradient feature extraction is performed based on the time domain conversion data of the convolution layer to obtain time domain conversion gradient feature data; Performing data segmentation on the convolutional layer time domain conversion data according to the time domain conversion gradient feature data to obtain the convolutional layer time domain segmentation data; Perform feature reconstruction according to the time domain segmentation data of the convolution layer to obtain feature reconstruction data; Channel mapping is performed according to the feature reconstructed data to obtain first channel decomposition data.

6. The method according to claim 4, characterized in that The cluster channel decomposition is specifically as follows: Perform feature selection on the normalized data of the image pooling layer to obtain feature selection data of the image pooling layer; Perform clustering calculation on the image pooling layer feature selection data to obtain the image pooling layer clustering feature data; Extract cluster centers based on cluster feature data of the image pooling layer to obtain cluster center graph data; Perform cluster channel mapping according to the cluster center graph data to obtain second channel decomposition data; The cluster channel mapping is as follows: Perform feature encoding according to the cluster center graph data to obtain encoded feature data; Edge features are transferred and calculated according to the encoded feature data to obtain distributed feature data; Performing feature correlation aggregation processing on the distributed feature data to obtain feature correlation aggregation data; Perform feature mapping on the data according to feature correlation aggregation to obtain feature mapping data; Perform feature alignment and fusion according to feature mapping data to obtain feature fusion data; Channel mapping is performed based on the feature fusion data to obtain the second channel decomposition data.

7. The method according to claim 1, characterized in that Step S4 is specifically as follows: Perform forward propagation calculation on the pooling layer data of the polylactic acid fiber image to obtain the preliminary image fully connected layer data; Perform activation calculation on the preliminary image full connection layer data to obtain the image full connection layer activation calculation data; Perform feature combination based on the activation calculation data of the fully connected layer of the image to obtain combined feature data; Perform error calculation based on the combined characteristic data and the corresponding chemical additive content data to obtain error calculation data; Perform back propagation according to the error calculation data to obtain back propagation data; The fully connected layer data of the polylactic acid fiber image pooling layer is calculated according to the back-propagation data to obtain the fully connected layer data of the polylactic acid fiber image.

8. A polylactic acid bio-based fiber dyeing process optimization system, characterized in that: Used to perform the polylactic acid bio-based fiber dyeing process optimization method according to claim 1, the polylactic acid bio-based fiber dyeing process optimization system comprises: A polylactic acid fiber data acquisition module is used to obtain polylactic acid fiber image data and corresponding chemical additive content data; A polylactic acid fiber image convolution module is used to perform self-attention convolution layer calculation on the polylactic acid fiber image data to obtain polylactic acid fiber image convolution layer data; The polylactic acid fiber image pooling module is used to perform multi-dimensional pooling layer processing on the polylactic acid fiber image convolution layer data to obtain the polylactic acid fiber image pooling layer data; A polylactic acid fiber image fully connected module is used to perform fully connected layer calculation on the polylactic acid fiber image pooling layer data to obtain the polylactic acid fiber image fully connected layer data; The PLA bio-based fiber dyeing process optimization model construction module is used to construct and iterate the model according to the fully connected layer data of the PLA fiber image and the corresponding chemical additive content data, so as to obtain the PLA bio-based fiber dyeing process optimization model for performing the PLA bio-based fiber dyeing process optimization operation.

Images