Textile fabric winding tension prediction system based on machine learning

Abstract

The invention relates to the technical field of textile industry, in particular to a textile fabric winding tension prediction system based on machine learning, which comprises an environment sensing module, a dynamic data acquisition module, a data preprocessing enhancement module, a deep learning model training module, a self-adaptive control module and a fault diagnosis early warning module. Wherein the environment sensing module is used for monitoring spinning environment data in real time; the dynamic data acquisition module is used for comprehensively monitoring the running state and the physical environment change of the textile machinery; the data preprocessing enhancement module is used for standardizing the data acquired by the acquisition module; and the deep learning model training module is used for carrying out model training on tension change in the textile fabric winding process. According to the invention, real-time and accurate tension control is realized, the production automation and intelligence levels are improved, and fault diagnosis early warning is started, so that the production efficiency and quality of textiles are remarkably improved, and meanwhile, the resource waste and the production cost are reduced.

Description

Technical Field

[0001] This invention relates to the field of textile industry technology, and in particular to a textile fabric take-up tension prediction system based on machine learning. Background Technology

[0002] In the textile industry, precise control of tension during fabric winding is crucial, as tension directly affects the quality of textiles, including but not limited to fabric smoothness, dimensional stability, and the appearance and feel of the final product. Traditional tension control methods mainly rely on operator experience and simple mechanical adjustment systems. These methods often struggle to achieve high-precision tension control when dealing with rapidly changing production conditions and different types of textile materials. Furthermore, existing tension monitoring and adjustment responses are often delayed, failing to respond in real time to unexpected problems during production, leading to low production efficiency and significant material waste.

[0003] With the development of intelligent manufacturing and industrial automation technologies, the textile industry urgently needs a solution that can accurately monitor and adjust fabric take-up tension in real time. Existing technologies have failed to effectively integrate advanced data analysis technologies and machine learning algorithms to optimize the tension control process, especially when facing complex and changing environments and variable material properties during the production process.

[0004] Therefore, there is an urgent need to develop a new textile fabric take-up tension prediction and control system that can use machine learning technology to automatically predict and adjust tension by analyzing real-time collected environmental and mechanical operation data in order to achieve efficient and high-quality production goals. Summary of the Invention

[0005] To achieve the above objectives, the present invention provides a machine learning-based system for predicting the winding tension of textile fabrics.

[0006] The machine learning-based textile fabric winding tension prediction system includes an environmental sensing module, a dynamic data acquisition module, a data preprocessing and enhancement module, a deep learning model training module, an adaptive control module, and a fault diagnosis and early warning module.

[0007] Environmental sensing module: Employs a pre-set environmental sensor network to monitor temperature, humidity, and airflow data of the textile environment in real time;

[0008] Dynamic data acquisition module: Through preset force, speed, vibration and sound sensors, it comprehensively monitors the operating status of textile machinery and changes in the physical environment;

[0009] Data preprocessing enhancement module: applies preset data processing techniques to standardize and clean the data collected by the environmental sensing module and the dynamic data acquisition module, and generates training samples through data enhancement techniques;

[0010] Deep learning model training module: Utilizing a pre-set deep learning algorithm, combined with the characteristics of convolutional neural networks and recurrent neural networks, the module trains a model for tension changes during the winding process of textile fabrics and performs tension prediction.

[0011] Adaptive control module: Based on the prediction results of the deep learning model, it automatically adjusts the operating parameters of the textile machinery to ensure that the tension during the fabric winding process reaches the optimal state;

[0012] Fault diagnosis and early warning module: Combining data from environmental sensing and dynamic data acquisition modules, it analyzes equipment performance and operating status, identifies equipment faults or performance degradation that affect tension control at an early stage, and issues timely warnings to operators to avoid production interruptions or quality problems.

[0013] Furthermore, the environmental sensing module includes a temperature detection unit, a humidity detection unit, and an airflow detection unit; wherein,

[0014] Temperature detection unit: Using thermocouples or thermistors as sensing elements, it is located in the textile production area to monitor and collect ambient temperature data in real time. This temperature detection unit provides temperature data by measuring temperature changes in the air.

[0015] Humidity detection unit: Employs a humidity sensor, including a capacitive or resistive humidity sensor, placed near textile machinery or in the production environment to monitor the relative humidity of the air in real time. This humidity detection unit provides humidity data to the system by measuring the air humidity.

[0016] Airflow detection unit: Employs a wind speed sensor, positioned above the textile production line, to monitor the airflow speed and direction within the production area.

[0017] Furthermore, the dynamic data acquisition module includes a force detection unit, a speed detection unit, a vibration detection unit, and a sound detection unit; wherein,

[0018] Force detection unit: Using strain gauges or pressure sensors, fixed on textile machinery, especially at the positions of the fabric take-up shaft and guide wheel, to monitor fabric tension and mechanical pressure in real time. This force detection unit measures the force acting on the fabric and mechanical parts.

[0019] Speed ​​detection unit: An encoder or speed sensor is installed on the drive shaft of the textile machinery to measure the rotational speed of the mechanical components;

[0020] Vibration detection unit: Using an accelerometer, it is installed on the structural components of the textile machinery to monitor the vibration generated during the operation of the machinery. The vibration data can help identify early signs of mechanical failure or performance degradation.

[0021] Sound detection unit: Using sound sensors or microphones, it is placed in the textile production area to collect the sound generated by the operation of machinery. By analyzing the sound signals, it can identify abnormal or faulty mechanical conditions.

[0022] Furthermore, the data preprocessing enhancement module includes a data standardization unit, a data cleaning unit, and a data enhancement unit; wherein,

[0023] Data standardization unit: Applying preset standardization technology, it processes the temperature, humidity, airflow, force, speed, vibration and sound data collected by the environmental sensing module and the dynamic data acquisition module to ensure that all data are on the same order of magnitude, which facilitates comparison and analysis;

[0024] Data cleaning unit: Utilizes preset data cleaning techniques to optimize the collected data, and ensures data quality and reliability by identifying and handling inconsistencies or errors in the data;

[0025] Data Augmentation Unit: Applying preset data augmentation techniques, including random noise injection and data interpolation, to expand the processed dataset, thereby increasing data diversity and volume, and improving the generalization ability and robustness of deep learning models.

[0026] Furthermore, the deep learning model training module includes an input unit, a feature extraction unit, a sequence learning unit, and a model training optimization unit; wherein,

[0027] Data input unit: Used to receive the dataset output by the data preprocessing enhancement module as input for deep learning training;

[0028] Feature extraction unit: Uses convolutional neural networks to extract features from input data. It automatically captures spatial features in the data through convolutional layers and extracts them in detail.

[0029] Sequence learning unit: Uses a long short-term memory network to process the data extracted by the feature extraction unit and capture long-term dependencies in the time series;

[0030] Model Training Optimization Unit: Based on convolutional neural networks and long short-term memory networks, the model generates the final tension prediction output through fully connected layers. During training, the backpropagation algorithm and optimizer are used to continuously update the model weights to minimize the difference between the predicted value and the actual tension value.

[0031] Furthermore, the feature extraction unit specifically includes:

[0032] Depthwise separable convolution: Instead of traditional convolution operations, depthwise separable convolution is used. First, spatial convolution is performed on each input channel, then 1x1 convolution is used to fuse the outputs of the depth channels. The specific calculation formula is as follows:

[0033]

[0034] Among them, X k,l,m This represents a pixel value in the input feature map, where k, l represent the spatial location of the pixel, and m represents the channel number of the input feature map; It is a weight value in the depthwise convolution kernel, used in the spatial convolution step of depthwise separable convolution; D k,l,m These are the pixel values ​​in the output feature map after applying depthwise convolution; These are the weights in a 1x1 convolution kernel, used to fuse features extracted through depthwise convolution; Y k,l,n It is a pixel value in the final output feature map, where k,l represent the spatial location of the pixel, and n represents the channel number of the output feature map;

[0035] Attention Mechanism: An attention mechanism is introduced to enhance the network's ability to perceive features. Specifically, spatial attention and channel attention mechanisms are used to automatically learn the important parts of the input data. Spatial attention calculates the importance weight of each position on the feature map, while channel attention evaluates the importance of each feature channel.

[0036] Feature fusion and recalibration: The feature maps processed by the attention step are fused and recalibrated using global average pooling and fully connected layers to improve the discriminative power of the feature representation. Specifically, global average pooling is used to reduce the feature dimension first, then the connected layers learn non-linear combinations between features, and finally the sigmoid activation function is used to output the recalibration weights.

[0037] Multi-scale feature fusion: Feature maps of different scales are fused to capture tension patterns at different levels during the winding of textile fabrics. Specifically, features of different scales are extracted through parallel convolutional paths, and the size of the feature maps is adjusted using convolutional layers to match them. Then, they are fused through element-wise addition or concatenation operations.

[0038] Furthermore, the sequence learning unit includes:

[0039] Time series feature embedding: First, the output of the feature extraction unit is passed through an embedding layer. This embedding layer maps the input data to a high-dimensional space, thereby enhancing the model's ability to learn time series features. The calculation of the embedding layer is expressed as: E = W emb ·X, where X is the input data, W emb is the weight matrix of the embedding layer, and E is the feature representation after embedding;

[0040] Improved LSTM: An improved LSTM is used to process the embedded features. This improved LSTM introduces a gated attention mechanism for finer control of the information flow. The improved LSTM computation formula is as follows:

[0041] f t =σ(Wf · [h t-1 E t ]+b f ),

[0042] i t =σ(W i ·[h t-1 E t ]+b i ),

[0043] o t =σ(W o ·[h t-1 E t ]+b o ),

[0044]

[0045] h t =o t ⊙tanh(C t ), where f t Let i be the activation value of the forget gate at time t; t Let o be the input gate activation value at time t; t The output gate activation value at time t; Let C be the candidate cell state at time t; t h represents the cell state at time t. t The output hidden state at time t; E t W is the input feature representation at time t after time series feature embedding; f W i W o W C and b f ,b i ,b o ,b C ...

[0046] Furthermore, the model training optimization unit includes:

[0047] Define the loss function: First, define the loss function. To quantify the tension value predicted by the model The difference between the actual tension value Y and the mean squared error is used as the loss function, and the specific formula is as follows:

[0048] in, The loss function quantifies the tension value predicted by the model. The difference between the actual tension value Y and the actual tension value; N is the sample size; Y i The actual tension value of the i-th sample. Let be the predicted tension value for the i-th sample;

[0049] Backpropagation algorithm: The backpropagation algorithm is used to calculate the gradient of each weight parameter of the loss function. Specifically, the gradient is calculated layer by layer from the output layer. For each weight W, the gradient is calculated as follows: It can be calculated using the chain rule;

[0050] Weight Update: Using the calculated gradients, the Adam optimizer is used to update the model's weights. The specific formula for weight update is as follows: Among them, W t η represents the model weights at time step t during the training process; η is the learning rate, which controls the size of the gradient descent step size during weight updates. This is the first moment estimate at time step t, i.e., the estimate of the mean of the gradient; This is the second moment estimate at time step t, i.e., the estimate of the gradient variance; ∈ is a small constant used to avoid division by zero;

[0051] Iterative training: Repeatedly execute the backpropagation algorithm and weight updates, and continuously update the model's weights through multiple iterations of training until the loss function converges or the preset number of iterations is reached.

[0052] Furthermore, the adaptive control module includes a prediction result analysis unit, a parameter adjustment decision unit, and an execution unit; wherein,

[0053] Prediction Result Analysis Unit: This unit receives the tension prediction results output by the deep learning model and analyzes the deviation between these predictions and the actual target tension value. By calculating the difference between the predicted and target tensions, it determines whether the operating parameters of the textile machinery need to be adjusted. The formula for calculating the deviation is: Where D represents deviation, Y target It is the target tension value. It is the tension value predicted by the model;

[0054] Parameter Adjustment Decision Unit: Based on the deviation information provided by the prediction result analysis unit, a decision algorithm is used to determine the optimal adjustment scheme for the operating parameters of the textile machinery. By continuously learning the relationship between deviation and parameter adjustment, the adjustment strategy is optimized. The formula for the decision algorithm is: P new =P current +α·ΔP(D), where P new It is the new running parameter value, P current α is the current parameter value, α is the learning rate, and ΔP(D) is the parameter adjustment amount calculated based on the bias D.

[0055] Execution unit: Based on the decision results of the parameter adjustment decision unit, adjust the operating parameters of the textile machinery, including the winding speed and tension setting, to ensure that the parameters are adjusted in real time to achieve the preset tension value.

[0056] Furthermore, the fault diagnosis and early warning module includes a data integration unit, a feature engineering unit, a fault diagnosis model unit, and an early warning output unit; wherein,

[0057] Data integration unit: First, it integrates the data collected by the environmental sensing module and the dynamic data acquisition module to form a comprehensive dataset;

[0058] Feature Engineering Unit: Based on data integration, feature engineering is used to extract performance indicators (KPIs) as features for fault diagnosis. Performance indicators (KPIs) include equipment operating efficiency, average vibration level, temperature fluctuation rate, and sound spectrum anomaly index features.

[0059] Fault diagnosis model unit: Utilizes a pre-trained deep neural network model to analyze the features extracted by the feature engineering unit, and identifies potential faults and performance degradation signals;

[0060] Early warning output unit: Based on the output of the fault diagnosis model unit, when a potential fault or performance degradation is detected, the early warning output unit will generate corresponding early warning information and promptly convey it to the operator through the system interface or notification mechanism.

[0061] The beneficial effects of this invention are:

[0062] This invention, through real-time collection and analysis of environmental and mechanical operation data, enables the system to automatically predict tension changes during fabric winding and adjust control parameters accordingly. This prediction and feedback mechanism significantly improves the accuracy of tension control, ensures the consistency and reliability of textile quality, thereby reducing material waste and increasing production efficiency.

[0063] This invention, by introducing a machine learning model, can not only handle complex data analysis tasks but also continuously optimize and learn based on real-time data from the production process, automatically adjusting control strategies. This intelligent control reduces the workload of operators, decreases reliance on human experience, and simultaneously improves the stability and predictability of the production process.

[0064] This invention provides strong support for the continuity of textile production and equipment maintenance through fault diagnosis and early warning functions. The system can identify equipment malfunctions or performance degradation that may affect tension control at an early stage and issue timely warnings, enabling the maintenance team to take swift action to avoid production interruptions or quality problems. This early warning mechanism further ensures efficient production operation and stable product quality. Attached Figure Description

[0065] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0066] Figure 1 This is a schematic diagram of a textile fabric winding tension prediction system according to an embodiment of the present invention;

[0067] Figure 2 This is a schematic diagram of the deep learning model training module in an embodiment of the present invention. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0069] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0070] like Figures 1-2 As shown, the machine learning-based textile fabric winding tension prediction system includes an environmental sensing module, a dynamic data acquisition module, a data preprocessing and enhancement module, a deep learning model training module, an adaptive control module, and a fault diagnosis and early warning module; among which,

[0071] Environmental sensing module: It adopts a preset environmental sensor network to monitor the temperature, humidity and airflow data of the textile environment in real time, providing environmental parameters for tension prediction;

[0072] Dynamic data acquisition module: Through preset force, speed, vibration and sound sensors, it comprehensively monitors the operating status of textile machinery and changes in the physical environment;

[0073] Data preprocessing enhancement module: The module applies preset data processing techniques to standardize and clean the data collected by the environmental sensing module and the dynamic data acquisition module, and generates training samples through data augmentation techniques, aiming to improve the model's ability to predict tension changes during the textile process.

[0074] Deep learning model training module: Utilizing a pre-set deep learning algorithm, combined with the characteristics of convolutional neural networks and recurrent neural networks, the module trains a model for tension changes during the winding process of textile fabrics and performs tension prediction.

[0075] Adaptive control module: Based on the prediction results of the deep learning model, it automatically adjusts the operating parameters of the textile machinery, such as the winding speed and tension setting, to ensure that the tension of the fabric during the winding process reaches the optimal state.

[0076] Fault diagnosis and early warning module: Combining data from environmental sensing and dynamic data acquisition modules, it analyzes equipment performance and operating status, identifies equipment faults or performance degradation that affect tension control at an early stage, and issues timely warnings to operators to avoid production interruptions or quality problems.

[0077] The environmental sensing module includes a temperature detection unit, a humidity detection unit, and an airflow detection unit; among which,

[0078] Temperature detection unit: Using thermocouples or thermistors as sensing elements, it is located in the textile production area to monitor and collect ambient temperature data in real time. This temperature detection unit provides temperature data by measuring temperature changes in the air to help predict tension changes that may be affected by temperature fluctuations during the winding process of textile fabric.

[0079] Humidity detection unit: Employs a humidity-sensitive sensor, including a capacitive or resistive humidity sensor, placed near the textile machinery or in the production environment to monitor the relative humidity of the air in real time. This humidity detection unit provides humidity data to the system by measuring the air humidity. Since humidity levels have a direct impact on the tension of textiles, this unit supports more accurate tension prediction.

[0080] Airflow detection unit: Employs a wind speed sensor, positioned above the textile production line, to monitor the airflow speed and direction within the production area. Airflow detection is crucial for understanding and adjusting tension control during the textile fabric winding process, as airflow conditions can affect the fabric's tension and stability.

[0081] Each detection unit is connected to a central processing unit, which is responsible for collecting data from the sensors in each unit and integrating temperature, humidity, and airflow information through a preset data fusion algorithm to form comprehensive environmental data. This data is then transmitted in real time to the data preprocessing and enhancement module to support the tension prediction activities of the deep learning model training module. The design of the environmental sensing module ensures comprehensive monitoring of textile environmental parameters, improving the system's adaptability to environmental changes and the accuracy of tension prediction.

[0082] The dynamic data acquisition module includes a force detection unit, a velocity detection unit, a vibration detection unit, and a sound detection unit; among which,

[0083] Force detection unit: Using strain gauges or pressure sensors, fixed on textile machinery, especially at the positions of the fabric take-up shaft and guide wheel, it is used to monitor the fabric tension and mechanical pressure in real time. This force detection unit provides the system with direct data on the tension level by measuring the force acting on the fabric and mechanical parts.

[0084] Speed ​​detection unit: Using an encoder or speed sensor, it is installed on the drive shaft of the textile machinery to measure the rotational speed of the mechanical components. Speed ​​data is crucial for understanding and adjusting the fabric tension during the winding process, because changes in winding speed directly affect the tension state.

[0085] Vibration detection unit: Using an accelerometer, it is installed on the structural components of the textile machinery to monitor the vibration generated during the operation of the machinery. Vibration data can help identify early signs of mechanical failure or performance degradation, which may affect the smoothness of the fabric winding process and the accuracy of tension control.

[0086] Sound detection unit: Using sound sensors or microphones, it is placed in the textile production area to collect the sound generated by the operation of machinery. By analyzing the sound signals, it can identify abnormal or faulty mechanical conditions that may affect the quality of textiles and winding tension.

[0087] The dynamic data acquisition module collects data through these detection units, which is then aggregated and preprocessed by a central processing unit. This aggregated data is then transmitted in real time to the data preprocessing and enhancement module to support the tension prediction activities of the deep learning model training module. The design and implementation of the dynamic data acquisition module ensures that the system can comprehensively monitor the operating status of textile machinery and changes in the physical environment, providing accurate and comprehensive real-time data for tension prediction.

[0088] The data preprocessing enhancement module includes a data standardization unit, a data cleaning unit, and a data enhancement unit; among them,

[0089] Data standardization unit: Applying preset standardization techniques, such as Z-score standardization or min-max standardization, to process temperature, humidity, airflow, force, speed, vibration and sound data collected by the environmental sensing module and dynamic data acquisition module, ensuring that all data are on the same order of magnitude, facilitating comparison and analysis, and improving the training efficiency and prediction accuracy of subsequent machine learning models;

[0090] Data cleaning unit: Utilizing preset data cleaning techniques, such as missing value handling, outlier detection and removal, the collected data is optimized. By identifying and handling inconsistencies or errors in the data, the quality and reliability of the data are ensured, providing a solid data foundation for accurate tension prediction.

[0091] Data Augmentation Unit: Applying preset data augmentation techniques, including random noise injection and data interpolation, to expand the processed dataset. By increasing data diversity and scale, it improves the generalization ability and robustness of deep learning models, especially when the amount of data is limited or the data changes complexly, thereby enhancing the predictive performance of the model.

[0092] Random noise injection is a method to increase data diversity by adding randomly generated noise values ​​to the original data. This method is particularly suitable for improving the robustness of models to small perturbations. Assuming the original dataset is represented as X, the data X′ after random noise injection can be calculated using the following formula: X′=X+∈, where ∈ is from a certain distribution (such as the normal distribution). The noise is randomly sampled in the sample, and σ is the standard deviation used to control the noise intensity.

[0093] Data interpolation is a technique for expanding a dataset by creating new data points between existing data points. This method is highly effective for time series data or any data that can be interpolated along numerical dimensions. Linear interpolation is a common method and can be performed using the following formula: Assuming there are two data points X... a and X b and the corresponding target value y aand y b We want to generate a new data point X. c and target value y c , where X c In X a and X b The linear interpolation formula between them is as follows: X c =X a +t(X b -X a ) and y c =y a +t(y b -y a ), where t is a parameter between 0 and 1 used to control X. c In X a and X b The positions between.

[0094] The deep learning model training module includes an input unit, a feature extraction unit, a sequence learning unit, and a model training optimization unit; among which,

[0095] Data Input Unit: This unit receives the datasets output by the data preprocessing and enhancement module. These datasets have been standardized, cleaned, and enhanced to be suitable as input for deep learning training. This unit ensures that the data is fed into the subsequent training process in the correct format and structure.

[0096] Feature extraction unit: Uses convolutional neural networks to extract features from input data. The convolutional layers automatically capture spatial features in the data, such as the tension distribution pattern during the winding process of textile fabrics. This is crucial for understanding tension changes.

[0097] Sequence Learning Unit: The Long Short-Term Memory (RNN) network is used to process the data extracted by the feature extraction unit. The RNN is specifically designed to process sequence data and can capture long-term dependencies in time series, which is crucial for predicting tensions that change over time.

[0098] Model Training Optimization Unit: Based on convolutional neural networks and long short-term memory networks, the model generates the final tension prediction output through fully connected layers. During training, the backpropagation algorithm and optimizer are used to continuously update the model weights to minimize the difference between the predicted value and the actual tension value.

[0099] The feature extraction unit specifically includes:

[0100] Depthwise separable convolution: To improve the computational efficiency of the model and reduce the number of parameters, depthwise separable convolution is used instead of traditional convolution operations. First, spatial convolution is performed on each input channel, and then 1x1 convolution is used to fuse the outputs of the depth channels. The specific calculation formula is as follows:

[0101]

[0102] Among them, X k,l,m X represents a pixel value in the input feature map, where k and l represent the spatial location of the pixel, and m represents the channel number of the input feature map. In the context of predicting the winding tension of textile fabrics, X can represent the value of preprocessed textile machinery operating status data (such as force, speed, vibration, etc.) or environmental data (such as temperature, humidity, etc.) at a specific location and channel. is a weight value in the depthwise convolution kernel, used in the spatial convolution step of depthwise separable convolution. i,j represent positions within the kernel, while m represents the kernel for a specific input channel m. These weights learn how to independently extract spatial features from each input channel; D k,l,m These are the pixel values ​​in the output feature map after applying depthwise convolution, reflecting the spatial feature extraction results based on the input channel m at position k,l; These are the weights in a 1x1 convolution kernel (i.e., a point convolution kernel), used to fuse features extracted through depthwise convolution. Here, m represents the input channels, and n represents the output channels, allowing features from different channels to be combined through point convolution; Y k,l,n Y is a pixel value in the final output feature map, where k, l represent the spatial location of the pixel and n represents the channel number of the output feature map. In the context of tension prediction, Y represents the feature map after spatial feature extraction and channel feature fusion, which provides the foundation for subsequent sequence learning and tension prediction.

[0103] Attention Mechanism: An attention mechanism is introduced to enhance the network's ability to perceive features. Specifically, spatial attention and channel attention are used to automatically learn the important parts of the input data. Spatial attention calculates the importance weight of each position on the feature map, while channel attention evaluates the importance of each feature channel. Spatial attention focuses on where things are important. Let F be the feature map, and the calculation of spatial attention can be expressed by the following formula: S = σ(f agg (Gonv(F))), where Conv is a convolution operation used to reduce dimensionality or change the spatial dimension of the feature map, f aggThe aggregation function (e.g., average pooling and max pooling) is used to generate an attention weight at each location, and σ is the sigmoid activation function that produces an attention map S between the two points. Channel attention focuses on which channel is important. Given a feature map F, channel attention can be obtained by the following formula: C = σ(MLP(GAP(F))), where GAP represents global average pooling, which compresses the spatial information of the feature map into a vector, MLP represents multilayer perceptron, which is used to learn complex interactions between different channels, and σ is the sigmoid function that generates the attention weight C for each channel.

[0104] Feature fusion and recalibration: The feature maps processed by the attention step are fused, and global average pooling and fully connected layers are used to recalibrate the features to improve the discriminative power of the feature representation. Specifically, global average pooling is first used to reduce the feature dimensionality, then the connected layers learn non-linear combinations between features, and finally the sigmoid activation function is used to output recalibration weights. Given the feature maps processed by the attention module, feature fusion and recalibration can be completed through the following steps: F re =σ(FC(GAP(F)))⊙F, where GAP is global average pooling, which compresses the feature map F into a global feature vector; FC is a fully connected layer that learns a non-linear combination of feature vectors; and σ is the sigmoid function, which generates the recalibration weights F for each channel. re ,⊙ represents element-wise multiplication, used to adjust the channel intensity of the original feature map F;

[0105] Multi-scale feature fusion: Feature maps of different scales are fused to capture tension patterns at different levels during the fabric winding process. Specifically, features at different scales are extracted through parallel convolutional paths, and the size of the feature maps is adjusted using convolutional layers for matching. Then, element-wise addition or concatenation operations are used for fusion. Specifically, it is assumed that there are feature maps F1, F2, ..., F from different scales. n Multi-scale feature fusion is represented as:

[0106] F multi =Concat(Conv(F1),Conv(F2),…,Conv(F n )),or

[0107] F multi =Conv(F1)+Conv(F2)+…+Conv(F nIn this context, Conv is a convolution operation used to adjust the number of channels in feature maps of different scales to match them, and Concat represents a feature map concatenation operation that connects feature maps of different scales together to form a richer feature representation. Another approach is to combine feature maps of different scales through element-wise addition to enhance the model's utilization of features at different scales.

[0108] Sequence learning units include:

[0109] Time series feature embedding: To process time series data more effectively, the output of the feature extraction unit is first passed through an embedding layer. This embedding layer maps the input data to a high-dimensional space, thereby enhancing the model's ability to learn time series features. The calculation of the embedding layer is expressed as: E = W emb ·X, where X is the input data, W emb is the weight matrix of the embedding layer, and E is the feature representation after embedding;

[0110] Improved LSTM: An improved LSTM is used to process the embedded features. This improved LSTM introduces a gated attention mechanism for finer control of the information flow. The improved LSTM computation formula is as follows:

[0111] f t =σ(M f ·[h t-1 E t ]+b f ),

[0112] i t =σ(W i ·[h t-1 E t ]+b i ),

[0113] o t =σ(W o ·[h t-1 E t ]+b o ),

[0114]

[0115] h t =o t ⊙tanh(C t ), where f t The forget gate activation value at time t determines the cell state C at the previous time step. t-1 How much information will be forgotten or retained? Forget gates help models manage long-term dependent information by controlling the forgetting mechanism of information flow; t The input gate activation value at time t determines the input E at the current time.t How much information (feature embedding) will be updated to the cell state C? t The input gate enables the model to add new and important information to the cell state; t The output gate activation value at time t determines the current cell state C. t How much information in the output will affect the hidden state h? t The output gate controls which information will be used for prediction or passed to the next layer; Let E be the candidate cell state at time t, determined by the current input E. t and the previous hidden state h t-1 Together, the candidate cell state is determined to contain new information that may be added to the current cell state; C t The cell state at time t is the "memory" part of the model, which integrates the forgetting gate f. t Decision and input gate i t The new information updates maintain the long-term state information of the model; h t The output hidden state at time t is the output of the LSTM unit and can be used for subsequent processing or as a feature for the final predicted tension. The hidden state depends on the output gate o. t and current cell state C t The information is calculated and carries the output information of the model at the current moment; E t This is the input feature representation at time t after time series feature embedding, which may include environmental parameters of the textile process (such as temperature and humidity) and mechanical operating parameters (such as speed and force); W f W i Q o W C and b f ,b i ,b o ,b C , are the weight matrices and bias terms for the forget gate, input gate, output gate, and candidate cell state, respectively. These parameters are automatically adjusted during the learning process and are used to control the behavior of the gating mechanism and the update of the cell state; σ is the sigmoid activation function, which is used to compress the output of the gating unit to between 0 and 1, representing the proportion of information that passes through;

[0116] With the improved LSTM described above, the model can more effectively capture long-term dependencies in time series data, especially dynamic tension changes during the winding process of textile fabrics. The final LSTM output can be used for subsequent fully connected layers to generate accurate tension predictions.

[0117] The model training optimization unit includes:

[0118] Define the loss function: First, define the loss function. To quantify the tension value predicted by the model The difference between the actual tension value Y and the mean squared error is used as the loss function, and the specific formula is as follows:

[0119] in, The loss function quantifies the tension value predicted by the model. The difference between the actual tension value Y and the actual tension value Y, in the context of fabric winding, represents the accuracy of the model in predicting fabric tension; N is the number of samples, representing the total number of fabric winding events used to train the model in the training dataset; Y i This represents the actual tension value of the i-th sample, i.e., the tension value actually measured during the fabric winding process. This is the predicted tension value for the i-th sample, i.e., the tension value predicted by the model based on the input features;

[0120] Backpropagation algorithm: The backpropagation algorithm is used to calculate the gradient of each weight parameter of the loss function. Specifically, the gradient is calculated layer by layer from the output layer. For each weight W, the gradient is calculated as follows: It can be calculated using the chain rule;

[0121] Weight Update: Using the calculated gradients, the Adam optimizer is used to update the model's weights. The specific formula for weight update is as follows: Among them, W t Let t be the model weights at time step t during the training process, representing the parameter values ​​of the deep learning model at this moment; η is the learning rate, which controls the size of the gradient descent step size during weight updates and is a key hyperparameter in the optimization process. It is the first moment estimate at time step t, that is, the estimate of the mean of the gradient. It is one of the mechanisms used in the Adam optimizer to adjust the step size of each weight update. is the second moment estimate at time step t, i.e., the estimate of the gradient variance. In the Adam optimizer, it helps adjust the learning rate of each parameter to make the training process more stable; ∈ is a small constant used to avoid division by zero and maintain numerical stability. In practical applications, ∈ is usually set to a very small positive value.

[0122] Iterative training: Repeatedly execute the backpropagation algorithm and weight updates, and continuously update the model's weights through multiple iterations of training until the loss function converges or the preset number of iterations is reached.

[0123] The adaptive control module includes a prediction result analysis unit, a parameter adjustment decision unit, and an execution unit; among which,

[0124] Prediction Result Analysis Unit: This unit receives the tension prediction results output by the deep learning model and analyzes the deviation between these predictions and the actual target tension value. By calculating the difference between the predicted and target tensions, it determines whether the operating parameters of the textile machinery need to be adjusted. The formula for calculating the deviation is: Where D represents deviation, Y target It is the target tension value. It is the tension value predicted by the model;

[0125] Parameter Adjustment Decision Unit: Based on the deviation information provided by the prediction result analysis unit, a decision algorithm is used to determine the optimal adjustment scheme for the operating parameters of the textile machinery. By continuously learning the relationship between deviation and parameter adjustment, the adjustment strategy is optimized. The formula for the decision algorithm is: P new =P current +α·ΔP(D), where P new It is the new running parameter value, P current α is the current parameter value, α is the learning rate, and ΔP(D) is the parameter adjustment amount calculated based on the bias D.

[0126] Execution unit: Based on the decision results of the parameter adjustment decision unit, it adjusts the operating parameters of the textile machinery, including the winding speed and tension settings, to ensure that the parameters are adjusted in real time to achieve the preset tension value, thereby optimizing the quality and production efficiency of textiles.

[0127] The fault diagnosis and early warning module includes a data integration unit, a feature engineering unit, a fault diagnosis model unit, and an early warning output unit; among which,

[0128] Data integration unit: First, it integrates the data collected by the environmental sensing module and the dynamic data acquisition module to form a comprehensive dataset, including temperature, humidity and airflow data obtained from the environmental sensing module, and force, speed, vibration and sound data obtained from the dynamic data acquisition module.

[0129] Feature Engineering Unit: Based on data integration, feature engineering extracts performance indicators (KPIs) as features for fault diagnosis. These KPIs include equipment operating efficiency, average vibration level, temperature fluctuation rate, and sound spectrum anomaly index features, enabling the model to more accurately capture patterns related to equipment failure and performance degradation. The specific calculation formula is as follows:

[0130] Equipment operating efficiency:

[0131] Actual production is the number of products actually produced within a given time period, while theoretical production is the number of products that the equipment is expected to produce under optimal conditions.

[0132] Average Vibration Level:

[0133] Where Vibrationi is the vibration intensity value measured in the i-th measurement, and N is the number of measurements;

[0134] Temperature Fluctuation Rate:

[0135] Where max(Temp) and min(Temp) are the highest and lowest temperature values ​​during the monitoring period, respectively. It is the average temperature;

[0136] Sound Spectral Anomaly Index:

[0137] Anomaly sound =∑|S actual -S normal |, where S actual This is the current sound spectrum data, S normal It is the baseline sound spectrum data under normal operating conditions;

[0138] Fault diagnosis model unit: This unit uses a pre-trained deep neural network model to analyze the features extracted by the feature engineering unit, identifying potential faults and performance degradation signals. Specifically, let the input feature vector be X, and the model output fault probability be Y.

[0139] X = [Efficiency, Vibration] avg Fluctuation temp Anomaly sound ];

[0140] The specific steps of feature engineering extraction in deep neural network model analysis are as follows:

[0141] Feature input: The input layer receives the KPI feature vector and sets it as X;

[0142] Hidden layer calculation: For each hidden layer l, the output of node j Calculated in the following way:

[0143] in, and These are the weights and bias parameters for the l-th layer, used to transfer and transform data between layers of the neural network; σ is the output of the (l-1)th layer and the input of the lth layer; σ is the activation function used to introduce nonlinearity, enabling the model to learn and simulate complex function mappings. Commonly used activation functions in the hidden layers include ReLU, while in the output layer, the sigmoid function used for binary classification problems can convert the output into a fault probability between 0 and 1.

[0144] Output layer: The output Y of the last layer represents the probability of a fault. It is calculated from the output of the last hidden layer using a softmax or sigmoid function and is used for classification problems. The specific formula is as follows:

[0145] Y is the output of the model, representing the probability of failure. Based on this probability, it can be determined whether the textile machinery may have a failure or performance degradation. In this process, by continuously adjusting the weights w and bias b, the model learns how to identify whether the equipment is operating normally based on the input KPIs. Once the model identifies a potential failure or performance degradation signal, it will issue an early warning in time, so that the operator can take corresponding measures, such as maintenance or adjustment, to prevent production interruption or quality problems.

[0146] Early warning output unit: Based on the output of the fault diagnosis model unit, when a potential fault or performance degradation is detected, the early warning output unit will generate corresponding early warning information and promptly convey it to the operator through the system interface or notification mechanism. This step ensures that the operator can quickly take necessary measures, such as adjusting equipment parameters or carrying out maintenance, to avoid production interruption or textile quality problems.

[0147] Through the collaborative work of the above units, the fault diagnosis and early warning module can effectively utilize the data provided by the environmental sensing and dynamic data acquisition modules to comprehensively analyze the operating status of the equipment, identify and warn of equipment failures or performance degradation that may affect the tension control of textile winding, thereby ensuring the continuity of the production process and the quality of textiles.

[0148] This invention is intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A machine learning-based textile fabric take-up tension prediction system, characterized in that, It includes an environmental sensing module, a dynamic data acquisition module, a data preprocessing and enhancement module, a deep learning model training module, an adaptive control module, and a fault diagnosis and early warning module; among which, Environmental sensing module: Employs a pre-set environmental sensor network to monitor temperature, humidity, and airflow data of the textile environment in real time; Dynamic data acquisition module: Through preset force, speed, vibration and sound sensors, it comprehensively monitors the operating status of textile machinery and changes in the physical environment; Data preprocessing enhancement module: applies preset data processing techniques to standardize and clean the data collected by the environmental sensing module and the dynamic data acquisition module, and generates training samples through data enhancement techniques; Deep learning model training module: Utilizing a pre-set deep learning algorithm, combined with the characteristics of convolutional neural networks and recurrent neural networks, the module trains a model for tension changes during the winding process of textile fabrics and performs tension prediction. Adaptive control module: Based on the prediction results of the deep learning model, it automatically adjusts the operating parameters of the textile machinery to ensure that the tension during the fabric winding process reaches the optimal state; Fault diagnosis and early warning module: Combining data from environmental sensing and dynamic data acquisition modules, it analyzes equipment performance and operating status, identifies equipment faults or performance degradation that affect tension control at an early stage, and issues timely warnings to operators to avoid production interruptions or quality problems.

2. The textile fabric take-up tension prediction system based on machine learning according to claim 1, characterized in that, The environmental sensing module includes a temperature detection unit, a humidity detection unit, and an airflow detection unit; wherein... Temperature detection unit: Using thermocouples or thermistors as sensing elements, it is located in the textile production area to monitor and collect ambient temperature data in real time. This temperature detection unit provides temperature data by measuring temperature changes in the air. Humidity detection unit: Employs a humidity sensor, including a capacitive or resistive humidity sensor, placed near textile machinery or in the production environment to monitor the relative humidity of the air in real time. This humidity detection unit provides humidity data to the system by measuring the air humidity. Airflow detection unit: Employs a wind speed sensor, positioned above the textile production line, to monitor the airflow speed and direction within the production area.

3. The textile fabric take-up tension prediction system based on machine learning according to claim 2, characterized in that, The dynamic data acquisition module includes a force detection unit, a speed detection unit, a vibration detection unit, and a sound detection unit; wherein, Force detection unit: Using strain gauges or pressure sensors, fixed on textile machinery, especially at the positions of the fabric take-up shaft and guide wheel, to monitor fabric tension and mechanical pressure in real time. This force detection unit measures the force acting on the fabric and mechanical parts. Speed ​​detection unit: An encoder or speed sensor is installed on the drive shaft of the textile machinery to measure the rotational speed of the mechanical components; Vibration detection unit: Using an accelerometer, it is installed on the structural components of the textile machinery to monitor the vibration generated during the operation of the machinery. The vibration data can help identify early signs of mechanical failure or performance degradation. Sound detection unit: Using sound sensors or microphones, it is placed in the textile production area to collect the sound generated by the operation of machinery. By analyzing the sound signals, it can identify abnormal or faulty mechanical conditions.

4. The textile fabric take-up tension prediction system based on machine learning according to claim 3, characterized in that, The data preprocessing enhancement module includes a data standardization unit, a data cleaning unit, and a data enhancement unit; wherein... Data standardization unit: Applying preset standardization technology, it processes the temperature, humidity, airflow, force, speed, vibration and sound data collected by the environmental sensing module and the dynamic data acquisition module to ensure that all data are on the same order of magnitude, which facilitates comparison and analysis; Data cleaning unit: Utilizes preset data cleaning techniques to optimize the collected data, and ensures data quality and reliability by identifying and handling inconsistencies or errors in the data; Data Augmentation Unit: Applying preset data augmentation techniques, including random noise injection and data interpolation, to expand the processed dataset, thereby increasing data diversity and volume, and improving the generalization ability and robustness of deep learning models.

5. The textile fabric take-up tension prediction system based on machine learning according to claim 4, characterized in that, The deep learning model training module includes an input unit, a feature extraction unit, a sequence learning unit, and a model training optimization unit; wherein... Data input unit: Used to receive the dataset output by the data preprocessing enhancement module as input for deep learning training; Feature extraction unit: Uses convolutional neural networks to extract features from input data. It automatically captures spatial features in the data through convolutional layers and extracts them in detail. Sequence learning unit: It uses a long short-term memory network to process the data extracted by the feature extraction unit and capture long-term dependencies in the time series. Model Training Optimization Unit: Based on convolutional neural networks and long short-term memory networks, the model generates the final tension prediction output through fully connected layers. During training, the backpropagation algorithm and optimizer are used to continuously update the model weights to minimize the difference between the predicted value and the actual tension value.

6. The textile fabric take-up tension prediction system based on machine learning according to claim 5, characterized in that, The feature extraction unit specifically includes: Depthwise separable convolution: Instead of traditional convolution operations, depthwise separable convolution is used. First, spatial convolution is performed on each input channel, then 1x1 convolution is used to fuse the outputs of the depth channels. The specific calculation formula is as follows: Among them, X k,l,m This represents a pixel value in the input feature map, where k, l represent the spatial location of the pixel, and m represents the channel number of the input feature map; It is a weight value in the depthwise convolution kernel, used in the spatial convolution step of depthwise separable convolution; D k,l,m These are the pixel values ​​in the output feature map after applying depthwise convolution; These are the weights in a 1x1 convolution kernel, used to fuse features extracted through depthwise convolution; Y k,l,n It is a pixel value in the final output feature map, where k,l represent the spatial location of the pixel, and n represents the channel number of the output feature map; Attention Mechanism: An attention mechanism is introduced to enhance the network's ability to perceive features. Specifically, spatial attention and channel attention mechanisms are used to automatically learn the important parts of the input data. Spatial attention calculates the importance weight of each position on the feature map, while channel attention evaluates the importance of each feature channel. Feature fusion and recalibration: The feature maps processed by the attention step are fused and recalibrated using global average pooling and fully connected layers to improve the discriminative power of the feature representation. Specifically, global average pooling is used to reduce the feature dimension first, then the connected layers learn non-linear combinations between features, and finally the sigmoid activation function is used to output the recalibration weights. Multi-scale feature fusion: Feature maps of different scales are fused to capture tension patterns at different levels during the winding of textile fabrics. Specifically, features of different scales are extracted through parallel convolutional paths, and the size of the feature maps is adjusted using convolutional layers to match them. Then, they are fused through element-wise addition or concatenation operations.

7. The textile fabric take-up tension prediction system based on machine learning according to claim 6, characterized in that, The sequence learning unit includes: Time series feature embedding: First, the output of the feature extraction unit is passed through an embedding layer. This embedding layer maps the input data to a high-dimensional space, thereby enhancing the model's ability to learn time series features. The calculation of the embedding layer is expressed as: E = W emb ·X, where X is the input data, W emb is the weight matrix of the embedding layer, and E is the feature representation after embedding; Improved LSTM: An improved LSTM is used to process the embedded features. This improved LSTM introduces a gated attention mechanism for finer control of the information flow. The improved LSTM computation formula is as follows: f t =σ(W f ·[h t-1 ,E t ]+b f ), I t =σ(W i ·[h t-1 ,E t ]+b i ), the t =σ(W o ·[h t-1 ,E t ]+b o ), h t =o t ⊙tanh(C t ), where f t Let i be the activation value of the forget gate at time t; t Let be the input gate activation value at time t; o t The output gate activation value at time t; C represents the candidate cell state at time t; t h represents the cell state at time t. t The output hidden state at time t; E t W is the input feature representation at time t after time series feature embedding; f W i W o W C and b f b i b o b C ...

8. The textile fabric take-up tension prediction system based on machine learning according to claim 7, characterized in that, The model training optimization unit includes: Define the loss function: First, define the loss function. To quantify the tension value predicted by the model The difference between the actual tension value Y and the mean squared error is used as the loss function, and the specific formula is as follows: in, The loss function quantifies the tension value predicted by the model. The difference between the actual tension value Y and the actual tension value; N is the sample size; Y i The actual tension value of the i-th sample. Let be the predicted tension value for the i-th sample; Backpropagation algorithm: The backpropagation algorithm is used to calculate the gradient of each weight parameter of the loss function. Specifically, the gradient is calculated layer by layer from the output layer. For each weight W, the gradient is calculated as follows: It can be calculated using the chain rule; Weight Update: Using the calculated gradients, the Adam optimizer is used to update the model's weights. The specific formula for weight update is as follows: Among them, W t η represents the model weights at time step t during the training process; η is the learning rate, which controls the size of the gradient descent step size during weight updates. This is the first moment estimate at time step t, i.e., the estimate of the mean of the gradient; This is the second moment estimate at time step t, i.e., the estimate of the gradient variance; ∈ is a small constant used to avoid division by zero; Iterative training: Repeatedly execute the backpropagation algorithm and weight updates, and continuously update the model's weights through multiple iterations of training until the loss function converges or the preset number of iterations is reached.

9. The textile fabric take-up tension prediction system based on machine learning according to claim 8, characterized in that, The adaptive control module includes a prediction result analysis unit, a parameter adjustment decision unit, and an execution unit; wherein... Prediction Result Analysis Unit: This unit receives the tension prediction results output by the deep learning model and analyzes the deviation between these predictions and the actual target tension value. By calculating the difference between the predicted and target tensions, it determines whether the operating parameters of the textile machinery need to be adjusted. The formula for calculating the deviation is: Where D represents deviation, Y target It is the target tension value. It is the tension value predicted by the model; Parameter Adjustment Decision Unit: Based on the deviation information provided by the prediction result analysis unit, a decision algorithm is used to determine the optimal adjustment scheme for the operating parameters of the textile machinery. By continuously learning the relationship between deviation and parameter adjustment, the adjustment strategy is optimized. The formula for the decision algorithm is: P new =P current +α·ΔP(D), where P new It is the new running parameter value, P current α is the current parameter value, α is the learning rate, and ΔP(D) is the parameter adjustment amount calculated based on the bias D. Execution unit: Based on the decision results of the parameter adjustment decision unit, adjust the operating parameters of the textile machinery, including the winding speed and tension setting, to ensure that the parameters are adjusted in real time to achieve the preset tension value.

10. The textile fabric take-up tension prediction system based on machine learning according to claim 9, characterized in that, The fault diagnosis and early warning module includes a data integration unit, a feature engineering unit, a fault diagnosis model unit, and an early warning output unit; wherein... Data integration unit: First, it integrates the data collected by the environmental sensing module and the dynamic data acquisition module to form a comprehensive dataset; Feature Engineering Unit: Based on data integration, feature engineering is used to extract performance indicators (KPIs) as features for fault diagnosis. Performance indicators (KPIs) include equipment operating efficiency, average vibration level, temperature fluctuation rate, and sound spectrum anomaly index features. Fault diagnosis model unit: Utilizes a pre-trained deep neural network model to analyze the features extracted by the feature engineering unit, and identifies potential faults and performance degradation signals; Early warning output unit: Based on the output of the fault diagnosis model unit, when a potential fault or performance degradation is detected, the early warning output unit will generate corresponding early warning information and promptly convey it to the operator through the system interface or notification mechanism.

Images