Cotton assorting resultant yarn quality prediction method based on mixed deep learning model

Abstract

The invention discloses a cotton assorting resultant yarn quality prediction method based on a hybrid deep learning model, and the method comprises the following steps: S1, inputting raw cotton indexes and technological parameters into a deterministic prediction model, carrying out the multi-scale feature extraction and processing, dynamically adjusting the attention to important features, and outputting a deterministic prediction result; s2, inputting a prediction result output by the deterministic prediction model into an uncertainty prediction model, and performing uncertainty prediction by taking raw cotton indexes and process parameters as conditions; and S3, sampling the uncertainty prediction model for multiple times to obtain a plurality of residual values, and adding the residual values with a result of the certainty prediction model to obtain a final cotton assorting resultant yarn quality prediction result. According to the method, the regression model and the conditional denoising diffusion probability model are combined, so that the prediction result of the whole hybrid deep learning model is closer to the actual production condition, the accuracy and robustness of resultant yarn quality prediction are remarkably improved, and the method is suitable for industrial production scenes with uncertainty and randomness.

Description

Technical Field

[0001] The invention relates to a method for predicting cotton yarn quality based on a hybrid deep learning model, and belongs to the technical field of textiles. Background Art

[0002] In the yarn production process, cotton blending is a key link that affects the quality of yarn. The cotton blending process involves a variety of raw cotton indicators, which have a significant impact on the quality characteristics of the final yarn, such as strength, uniformity and fineness. Traditional cotton blending yarn quality prediction methods mainly rely on statistical analysis and empirical formulas, which usually assume that the relationship between variables is linear. However, in actual production, there is often a complex nonlinear relationship between raw cotton indicators and yarn quality indicators. Therefore, using a simple linear model for prediction usually leads to insufficient accuracy. In addition, the various process parameters in the production process also have an important impact on the quality of the yarn. Therefore, when constructing a yarn quality prediction model, multiple factors of raw cotton indicators and process parameters must be considered simultaneously.

[0003] At the same time, there are some uncertain factors in the production process, such as changes in workshop temperature and humidity, machine maintenance status, and operator skill level. These factors are usually difficult to quantify, resulting in the actual output (yarn quality index) often fluctuating within a certain range under the same input conditions (i.e., the cotton blending scheme and process parameters are exactly the same). Traditional models are difficult to effectively capture these uncertain factors, resulting in large deviations in the prediction results, lack of robustness and reference value.

[0004] Faced with increasingly fierce domestic and international competition, textile enterprises have increasingly stringent requirements on yarn quality. In the efficient and high-quality automated production process, enterprises are in urgent need of more accurate yarn quality prediction models so as to adjust the raw cotton ratio and process parameters in time to reduce production risks. With the development of deep learning technology, prediction models based on deep neural networks have shown excellent performance in various industrial applications. To this end, the present invention provides a method for predicting the quality of cotton yarn based on a hybrid deep learning model, which aims to improve the accuracy and robustness of yarn quality prediction. Summary of the invention

[0005] To solve the above problems, the present invention provides a method for predicting the quality of blended cotton yarn based on a hybrid deep learning model. The present invention constructs two models: the first is a deterministic prediction model (DPN), i.e., a regression model, which aims to predict the yarn quality index based on the raw cotton index and process parameters; the second is an uncertainty prediction model based on the prediction results of the deterministic model, i.e., a conditional denoising diffusion probability model (CDDPM), which further corrects the output of the deterministic model by predicting the residual noise, thereby providing a fluctuation range for each yarn quality index. The method provided by the present invention makes the prediction results of the entire hybrid deep learning model closer to the actual production situation, and improves the accuracy and reliability of the prediction.

[0006] In a first aspect, the present invention provides a method for predicting the quality of blended cotton yarn based on a hybrid deep learning model, comprising the following steps:

[0007] Step S1: Input raw cotton indexes and process parameters into the deterministic prediction model to extract and process multi-scale features, dynamically adjust the focus on important features, and output deterministic prediction results;

[0008] Step S2: input the residual between the prediction result output by the deterministic prediction model and the true value of the training data into the uncertainty prediction model, and use the raw cotton index and process parameters as conditions to perform uncertainty prediction;

[0009] Step S3: by sampling the uncertainty prediction model multiple times, multiple residual values ​​are obtained, and the residual values ​​are added to the results of the deterministic prediction model to obtain the final prediction result of the quality of the blended cotton yarn and its fluctuation range.

[0010] In one embodiment of the present invention, the step S1 specifically includes a regression prediction method;

[0011] The regression prediction method specifically comprises the following steps:

[0012] Step S101: normalizing the raw cotton index and process parameters as input;

[0013] Step S102: obtaining raw cotton index features in the cotton blending process through a multi-scale attention enhanced feature extraction module;

[0014] Step S103: the extracted raw cotton features are fed into the attention-enhanced long short-term memory network module, and the process parameters of each process are sequentially fed into each attention-enhanced long short-term memory network module;

[0015] Step S104: Accumulate the output of each attention-enhanced long short-term memory network module and input it into the gated attention fusion module for fusion, and then add the output of the gated attention fusion module and the output of the last attention-enhanced long short-term memory network module as the final output;

[0016] Step S105: The final output of step S104 is converted into a deterministic prediction result through the FOP module;

[0017] Step S106: Denormalize the model prediction results.

[0018] In one embodiment of the present invention, step S101 specifically includes:

[0019] The input raw cotton index is The process parameters are The yarn quality index is Where m represents the number of samples, d c ,d p and d y They represent the dimensions of raw cotton index, process parameters and final yarn quality index respectively; the normalized input of raw cotton index, process parameters and final yarn quality index is expressed as: c =Norm(x c ), x′ p =Norm(x p ) and y′=Norm(y), where Norm(.) represents the normalization operation, and the specific formula is:

[0020]

[0021] Among them, x represents the original data, μ represents the mean of the original data, and σ represents the standard deviation of the original data;

[0022] The multi-scale attention-enhanced feature extraction module in step S102 includes 4 convolutional layers, 3 C2F layers, 1 SPPF layer and 1 CBAM attention mechanism layer, and the input is the normalized raw cotton index x′ c ; The convolution layer includes a 1D convolution layer, a BN layer, and a ReLU activation function. The convolution kernel size of the convolution layer is 3 and the step size is 2; the number of input and output feature channels of the C2F layer is the output channel of the previous convolution layer, and a residual connection is used; the number of input and output feature channels of the SPPF layer is the output feature channel number of the previous C2F; specifically described by the following formula:

[0023] F c =F CBAM (F SPPF (F C2f (F Conv(F C2f (F Conv (F C2f (F Conv (F Conv (x′ c )))))))))

[0024] The step S103 specifically includes:

[0025] Add the process features one by one, which can be described by the following formula:

[0026]

[0027] in, represents the features processed by the i-th group of attention-enhanced long short-term memory network modules;

[0028] The step S104 specifically includes:

[0029] The output features of the four groups of attention-enhanced long short-term memory network modules are fused through the gated attention fusion module, and then the output of the gated attention fusion module is added to the output of the last attention-enhanced long short-term memory network module to obtain the global fusion feature, which is specifically described by the following formula:

[0030]

[0031] Among them, F g is the global fusion feature, F σ Represents the Sigmoid activation function, F CBAM represents the CBAM attention mechanism, n represents the number of groups of AEL modules;

[0032] The step S106 specifically includes:

[0033] The predicted final yarn quality index Perform denormalization, Denorm(.) represents the denormalization operation, and the specific formula is:

[0034] Denorm(x)=x*σ+μ

[0035] Among them, x represents the normalized data, μ represents the mean of the original data, and σ represents the standard deviation of the original data.

[0036] In one embodiment of the present invention, the attention-enhanced long short-term memory network module in step S103 is normalized into one group with one LSTM, one multi-head attention mechanism and one layer, and one group of attention-enhanced long short-term memory network modules receives a process parameter in a main process, wherein the output of the multi-scale attention-enhanced feature extraction is first converted into a data dimension by one LSTM module when entering the first group of attention-enhanced long short-term memory network modules;

[0037] The gated attention fusion module in step S104 includes: 1 linear layer, 1 Sigmoid layer and 1 CBAM attention mechanism; wherein the input dimension and output dimension of the linear layer are consistent;

[0038] The FOP module in step S105 includes: 1 convolution module, 1 C2F module, 1 one-dimensional maximum pooling layer and 2 linear layers; wherein the convolution layer is used to reduce the number of channels, C2F reduces the amount of calculation by reducing the number of channels, the maximum pooling layer is used to reduce the feature dimension, and finally the multi-dimensional features are converted into deterministic prediction results through the linear layer.

[0039] In one embodiment of the present invention, step S2 specifically includes the following steps:

[0040] Step S201: During the training process, residual noise is generated through a diffusion process; the results are predicted by a deterministic prediction model The residual Δy0 is obtained by comparing it with the true value y of the training data, which is input into the diffusion process of the uncertainty prediction model, and Gaussian noise is gradually added to generate the final residual noise Δy N ; The diffusion process starts from Δy0, and Gaussian noise is gradually added after 100 time steps, finally making Δy N Close to the standard Gaussian distribution;

[0041] Step S202: During the training process, the denoising network is used to predict the noise added in each step of the denoising process; the final residual noise Δy after denoising is N As input, the raw cotton index and process parameters are sent to the denoising network as conditions. Through the convolution layer, linear layer and back diffusion process, the noise is gradually predicted and removed, and finally the residual Δy0 is restored; in the inference process, the randomly generated Gaussian noise is used as input, and the raw cotton index and process parameters are sent to the denoising network as conditions. After 100 time steps of step-by-step denoising, the predicted residual is finally sampled.

[0042] Step S203: During the inference process, the prediction residual obtained by sampling in step S202 is Prediction results with deterministic prediction model By adding them together, the yarn quality prediction result Y with uncertainty characterization is obtained, ensuring that the prediction result contains both certainty and uncertainty information.

[0043] In one embodiment of the present invention, step S201 specifically includes:

[0044] Preliminary forecast output from a deterministic forecast model The residual value Δy0 is obtained from the true value y of the corresponding training data. The diffusion process of the uncertainty prediction model is used to gradually add Gaussian noise to the residual Δy0 to generate the final residual noise Δy N ;

[0045] The forward diffusion process of the uncertainty prediction model is as follows:

[0046]

[0047] where Δy N , α N and ∈ N are the samples at time step N, the parameters controlling the noise intensity, and the Gaussian noise added at time step N;

[0048] The forward diffusion process is deduced as follows:

[0049]

[0050] The final expression is:

[0051]

[0052] where Δy N is a linear combination of the original residual Δy0 and Gaussian noise ∈; as N increases, Gradually decreases, causing the noise component in the data to gradually increase until it finally approaches the standard Gaussian distribution;

[0053] The step S202 specifically includes:

[0054] During the training process, the denoising network predicts the noise added in each step of the denoising process; the final residual noise Δy N As input, the raw cotton index and process parameters are sent to the denoising network as conditions. The noise is gradually predicted and removed through the convolution layer, linear layer and reverse diffusion process, and finally the residual Δy0 is restored. The reverse denoising process is to gradually N At the beginning, until Δy0, the noise is completely removed;

[0055] During the inference process, the denoising network receives randomly generated Gaussian noise as input, takes raw cotton indicators and process parameters as conditions, and performs step-by-step denoising for 100 time steps to finally sample the prediction residuals.

[0056] The reverse denoising process of the uncertainty prediction model is as follows:

[0057]

[0058] where ∈ θ (Δy N ,N,c) is the noise prediction value learned by the denoising network for a given Gaussian noise Δy N , time step N and condition c to predict the noise;

[0059] During the training process, the inputs to the denoising network are the residual noise Δy N , time step N and conditional parameters, the conditional parameters are raw cotton indicators and process parameters; the denoising process of the uncertainty prediction model has 5 residual blocks, each residual block contains 3 convolutional layers and 1 hole convolutional layer, and finally the outputs of each residual block are accumulated, and the noise is predicted through 1 convolutional layer

[0060] For the input residual noise, time step and conditional parameters, in order to keep the dimensions consistent, the dimension size is converted through mapping. The specific mapping definition is as follows:

[0061] Input projection: x = ReLU(W i *Δy N +b i )

[0062] Time step embedding: N emb =W t *F emb (N)+b i

[0063] Input conditional projection: c = W c *inputs+b i

[0064] The above three inputs are fused in the residual block, and the features of the fused content are extracted, which is described in the following formula:

[0065] x conbined =F DiatedConv (F Conv (d emb )+x)+c

[0066] x left ,x right =F split(x conbined )

[0067] x left ,x right =F split (F σ (x left )*F tanh (x right ))

[0068] output = x + x left

[0069] skip=F ReLU (F Conv (x right ))

[0070] where F DiatedConv represents the dilated convolution, F Conv represents the convolutional layer, F split Indicates separation by channel dimension, F σ Represents the Sigmoid activation function, F tanh represents the hyperbolic tangent activation function, F ReLU ReLU activation function.

[0071] The output of the current residual block is used as the input of the next residual block, and the skip of each residual block is accumulated and then used to predict the noise through the linear layer. The specific description is as follows:

[0072]

[0073] F proj (x) = F ReLU (F Conv (F Linear (x)))

[0074] The step S203 specifically includes:

[0075] During the inference process, the final residual value is predicted by the denoising network Then compare the initial prediction results of the deterministic prediction model Adding them together, the yarn quality prediction result with uncertainty representation is obtained, ensuring that the prediction result contains both certainty and uncertainty information;

[0076] The final prediction results are shown as follows:

[0077]

[0078] in and are the preliminary prediction results of the deterministic prediction model and the residual values ​​of the reverse denoising prediction of the uncertainty prediction model, respectively. x and p represent the input raw cotton parameters and process parameters, respectively.

[0079] In one embodiment of the present invention, the specific sampling in step S3 is represented as follows:

[0080] s={s 1 ,s 2 ,…,s m ,s M}

[0081] s min =min{s}

[0082] s max =max{s}

[0083] in

[0084] In a second aspect, the present invention provides a cotton yarn quality prediction system based on a hybrid deep learning model, comprising:

[0085] The deterministic prediction module is used to input raw cotton indicators and process parameters into the deterministic prediction model for multi-scale feature extraction and processing, dynamically adjust the focus on important features, and output deterministic prediction results;

[0086] The uncertainty prediction module is used to input the residual between the prediction result output by the deterministic prediction model and the true value of the training data into the uncertainty prediction model, and use the raw cotton index and process parameters as conditions to perform uncertainty prediction;

[0087] The cotton yarn quality prediction module is used to obtain multiple residual values ​​by sampling the uncertainty prediction model multiple times, and add them to the results of the deterministic prediction model to obtain the final cotton yarn quality prediction result and its fluctuation range.

[0088] In a third aspect, the present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the steps of a method for predicting the quality of cotton yarn based on a hybrid deep learning model are implemented.

[0089] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a method for predicting the quality of cotton yarn based on a hybrid deep learning model.

[0090] The beneficial effects of the present invention are:

[0091] The present invention proposes a method for predicting the quality of blended cotton yarn based on a hybrid deep learning model, which combines a regression model and a conditional denoising diffusion probability model, and efficiently models the complex nonlinear relationship in the blended cotton yarn process, thereby significantly improving the accuracy and robustness of the yarn quality prediction. In the regression network, the present invention adopts a lightweight deep learning model, obtains the raw cotton index features through a multi-scale feature extraction module, and uses the multi-head attention mechanism and long short-term memory (LSTM) module in the Transformer to extract the time series features of the process parameters. These two features are then effectively fused to generate a deterministic prediction result. In the conditional denoising diffusion probability model, the present invention uses the raw cotton index and process parameters as conditions to learn the random and uncertain factors in the yarn process, and provide a fluctuation range for each yarn quality index. This design makes the prediction result more robust and suitable for industrial production scenarios with uncertainty and randomness. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings required to be used in the embodiments or the prior art description are briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on the provided drawings without creative work.

[0093] Figure 1 is a schematic diagram of the model network architecture of the present invention;

[0094] Figure 2 It is a network architecture diagram of the deterministic prediction model in the present invention;

[0095] Figure 3 is a schematic diagram of the structure of the residual block in the present invention;

[0096] Figure 4 is an example diagram of a multi-scale attention enhanced feature extraction (MAFE) module in the deterministic prediction model of the present invention;

[0097] Figure 5 It is an example diagram of an attention-enhanced long short-term memory network (AEL) module and a gated attention fusion (GAF) module in the deterministic prediction model of the present invention;

[0098] Figure 6 is an example diagram of a convolutional block attention module (CBAM) in the deterministic prediction model of the present invention;

[0099] Figure 7 This is an example diagram of the FOP (Feature-To-Outcome Prediction) module in the deterministic prediction model of the present invention. DETAILED DESCRIPTION

[0100] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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 ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0101] The present invention provides a method for predicting the quality of blended cotton yarn based on a hybrid deep learning model, which aims to accurately predict various quality indicators of the yarn by extracting deep features of raw cotton indicators and process parameters, and further correct the output of the deterministic model by predicting residual noise, providing a fluctuation range for each yarn quality indicator, so that the prediction result is closer to the actual production situation.

[0102] It should be noted that Figure 1 It is a schematic diagram of the model network architecture of the present invention, showing the structure of each layer of the overall system and their interconnection relationship; Figure 2 It is a network architecture diagram of the deterministic prediction model in the present invention, which describes the overall structure of the specific model from the input of raw cotton index and process parameters to the prediction of yarn quality; Figure 3 It is a schematic diagram of the structure of the residual block in the present invention, showing the convolution operation, dilated convolution, and the interaction between conditional and diffusion embedding of each residual block in the denoising process of the conditional denoising diffusion probability model; Figure 4 is an example diagram of a multi-scale attention enhanced feature extraction (MAFE) module in the deterministic prediction model of the present invention; Figure 5 It is an example diagram of an attention-enhanced long short-term memory network (AEL) module and a gated attention fusion (GAF) module in the deterministic prediction model of the present invention; Figure 6 is an example diagram of a convolutional block attention module (CBAM) in the deterministic prediction model of the present invention; Figure 7 This is an example diagram of the FOP (Feature-To-Outcome Prediction) module in the deterministic prediction model of the present invention.

[0103] The present invention uses the actual production data of a cotton spinning mill, which contains 2900 samples, each of which includes 25 input features and 9 output features. The input features consist of 18 raw cotton indicators and 7 process parameters in four main processes, and the output features are 9 quality indicators of the yarn. All data are preprocessed, and the processing steps include standardization and missing value filling.

[0104] like Figure 1 and Figure 2As shown, the present invention provides an overall architecture of a method for predicting the quality of cotton yarn based on a hybrid deep learning model. The network model is divided into two parts: a deterministic prediction model (DPN) and an uncertainty prediction model (i.e., a conditional denoising diffusion probability model CDDPM), which respectively realize deterministic prediction and uncertainty prediction.

[0105] A method for predicting the quality of blended cotton yarn based on a hybrid deep learning model comprises the following steps:

[0106] Step S1: Input raw cotton indexes and process parameters into DPN for multi-scale feature extraction and processing, dynamically adjust the focus on important features, and output deterministic prediction results;

[0107] In some embodiments, the step S1 specifically includes a regression prediction method;

[0108] The regression prediction method specifically comprises the following steps:

[0109] Step S101: normalizing the raw cotton index and process parameters as input;

[0110] Step S102: Obtaining raw cotton index features in the cotton blending process through a multi-scale attention enhanced feature extraction (MAFE) module;

[0111] Step S103: the extracted raw cotton features are fed into an attention-enhanced long short-term memory network (AEL) module, and the process parameters of each process are sequentially fed into each AEL module;

[0112] Step S104: The output of each AEL module is accumulated and input into the gated attention fusion (GAF) module for fusion, and then the output of GAF is added to the output of the last AEL module as the final output;

[0113] Step S105: The final output of step S104 is converted into a deterministic prediction result through the FOP (Feature-To-Outcome Prediction) module;

[0114] Step S106: Denormalize the model prediction results.

[0115] Furthermore, the step S101 specifically includes:

[0116] The input raw cotton index is The process parameters are The yarn quality index is Where m represents the number of samples, d c ,d p and d yThey represent the dimensions of raw cotton index, process parameters and final yarn quality index respectively. The normalized input of raw cotton index, process parameters and final yarn quality index can be expressed as: c =Norm(x c ), x′ p =Norm(x p ) and y′=Norm(y), where Norm(.) represents the normalization operation, and the specific formula is:

[0117]

[0118] Among them, x represents the original data, μ represents the mean of the original data, and σ represents the standard deviation of the original data.

[0119] Furthermore, if Figure 4 As shown, the multi-scale attention enhanced feature extraction (MAFE) module in step S102 is specifically composed of 4 convolutional layers Conv (composed of 1D convolutional layer, BN (BatchNorm) layer and ReLU activation function), 3 C2F (Cross Stage Partial Connections) layers, 1 SPPF (Spatial Pyramid Pooling-Fast) layer and 1 CBAM attention mechanism layer, and the input is the naturalized raw cotton index x′ c The convolution kernel size of the convolution layer is 3, and the stride is 2; the number of input and output feature channels of the C2F layer is the output channel of the previous convolution layer, and residual connection is used; the number of input and output feature channels of the SPPF layer is the output feature channel of the previous C2F. It can be described in detail by the following formula:

[0120] F c =F CBAM (F SPPF (F C2f (F Conv (F C2f (F Conv (F C2f (F Conv (F Conv (x′ c )))))))))

[0121] Furthermore, if Figure 5As shown, the attention enhanced long short-term memory network (AEL) module in step S103 is composed of 1 LSTM, 1 multi-head attention mechanism (Multi-Head Attention) and 1 layer normalization (LayerNorm) as 1 group, with a total of four groups (the number of groups is determined by the number of main processes), and 1 group of AEL receives the process parameters in a main process, wherein the output of the MAFE module first converts the data dimension through 1 LSTM module when entering the first group of AEL;

[0122] The step S103 specifically includes:

[0123] Add the process features one by one, which can be described by the following formula:

[0124]

[0125] in, represents the features after processing by the i-th group of AEL modules;

[0126] Furthermore, if Figure 5 As shown, the gated attention fusion (GAF) module in step S104 consists of a linear layer (Linear), a Sigmoid layer and a CBAM attention mechanism (such as Figure 6 The input dimension and output dimension of the linear layer are consistent;

[0127] The step S104 specifically includes:

[0128] The output features of the four groups of AEL modules are fused through the GAF module, and then the output of the GAF module is added to the output of the last AEL module to obtain the global fusion feature, which is specifically described by the following formula:

[0129]

[0130] Among them, F g is the global fusion feature, F σ Represents the Sigmoid activation function, F CBAM represents the CBAM attention mechanism, n represents the number of groups of AEL modules;

[0131] Furthermore, if Figure 7 As shown, the FOP module in step S105 is composed of 1 convolution module (Conv), 1 C2F module, 1 one-dimensional maximum pooling layer (Maxpool1d) and 2 linear layers (Linear); wherein the convolution layer is used to reduce the number of channels, C2F reduces the amount of calculation by reducing the number of channels, the maximum pooling layer is used to reduce the feature dimension, and finally the multi-dimensional features are converted into deterministic prediction results through the linear layer.

[0132] Furthermore, the step S106 specifically includes:

[0133] The predicted final yarn quality index Perform denormalization, Denorm(.) represents the denormalization operation, and the specific formula is:

[0134] Denorm(x)=x*σ+μ

[0135] Among them, x represents the normalized data, μ represents the mean of the original data, and σ represents the standard deviation of the original data.

[0136] In the prior art, usually only the raw cotton index features are extracted while the process parameters are ignored, or the raw cotton index and the process parameters are directly spliced ​​as the input of the model. The method provided by the present invention obtains the raw cotton index features through a multi-scale feature extraction module, extracts the process parameter time series features through an LSTM module (in the actual production process, each process has a time series relationship, and the next process will start after one process ends), and then effectively merges these two features.

[0137] Existing models are based only on deterministic predictions, that is, only one certain output is generated for each input. This approach ignores the inevitable randomness and uncertainty factors in the actual production environment. Therefore, when faced with a complex and changing production environment, existing models often show limitations and cannot flexibly respond to fluctuations in output under the same input. In order to solve this problem, the present invention introduces uncertainty prediction based on the conditional denoising diffusion probability model (CDDPM). This method can more realistically simulate the complex interactions of multiple factors in the production process and fully consider the random disturbances in the production environment. The model constructed based on this method not only improves the prediction accuracy of yarn quality, but also quantifies the uncertainty of the prediction, making it more robust under different production conditions.

[0138] Step S2: Input the residual between the prediction result output by DPN and the true value of the training data into the uncertainty prediction model (i.e., conditional denoising diffusion probability model CDDPM), and take the raw cotton index and process parameters as conditions to perform uncertainty prediction;

[0139] In some embodiments, step S2 specifically includes the following steps:

[0140] Step S201: During the training process, residual noise is generated through a diffusion process. The residual Δy0 is obtained by comparing it with the true value y of the training data, and then input into the diffusion process of CDDPM (Conditional Denoising Diffusion Probabilistic Model), and Gaussian noise is gradually added to generate the final residual noise Δy N The diffusion process starts from Δy0, and Gaussian noise is gradually added after 100 time steps, eventually making Δy N Close to the standard Gaussian distribution;

[0141] Step S202: During the training process, the denoising network is used to predict the noise added in each step of the denoising process. The final residual noise Δy after denoising is N As input, the raw cotton index and process parameters are sent to the denoising network as conditions. Through the convolution layer, linear layer and back diffusion process, the noise is gradually predicted and removed, and the residual Δy0 is finally restored. In the inference process, the randomly generated Gaussian noise is used as input, and the raw cotton index and process parameters are sent to the denoising network as conditions. After 100 time steps of step-by-step denoising, the predicted residual can be sampled.

[0142] Step S203: During the inference process, the prediction residual obtained by sampling in step S202 is Compared with DPN prediction results By adding them together, the yarn quality prediction result Y with uncertainty characterization is obtained, ensuring that the prediction result contains both certainty and uncertainty information.

[0143] Furthermore, if Figure 1 As shown, the step S201 specifically includes:

[0144] Initial prediction output by DPN The residual value Δy0 is obtained by the true value y of the corresponding training data. The diffusion process of CDDPM is used to gradually add Gaussian noise to the residual Δy0 to generate the final residual noise Δy N In this embodiment, Δy is set to N Close to the standard Gaussian distribution;

[0145] The forward diffusion process of CDDPM is as follows:

[0146]

[0147] where Δy N , α N and ∈ N are the samples at time step N, a parameter that controls the noise intensity (usually decreases with time), and the Gaussian noise added at time step N (from a standard normal distribution N(0,I)).

[0148] The forward diffusion process can be deduced as follows:

[0149]

[0150] Finally it can be expressed as:

[0151]

[0152] where Δy N is a linear combination of the original residual Δy0 and Gaussian noise ∈. As N increases, Gradually decreases, causing the noise component in the data to gradually increase until it finally approaches the standard Gaussian distribution. From this, it can be seen that there is no need to gradually add noise during the noise addition process, and Δy can be directly obtained through one-step calculation. N , which can significantly improve the computational efficiency.

[0153] Furthermore, the step S202 specifically includes:

[0154] During the training process, the denoising network predicts the noise added in each step of the denoising process. The final residual noise Δy N As input, the raw cotton index and process parameters are sent to the denoising network as conditions. The noise is gradually predicted and removed through the convolution layer, linear layer and reverse diffusion process, and finally the residual Δy0 is restored. The reverse denoising process is to gradually N At the beginning, until Δy0, the noise is completely removed;

[0155] During the inference process, the denoising network receives randomly generated Gaussian noise as input, takes raw cotton indicators and process parameters as conditions, and after 100 time steps of stepwise denoising, the prediction residual can be finally sampled.

[0156] The reverse denoising process of CDDPM is as follows:

[0157]

[0158] where ∈ θ (Δy N ,N,c) is the noise prediction value learned by the denoising network for a given Gaussian noise Δy N , time step N and condition c to predict the noise.

[0159] like Figure 3 As shown, during the training process, the inputs in the denoising network are the residual noise Δy N, time step N and conditional parameters, where the conditional parameters are raw cotton indicators and process parameters; the denoising process of CDDPM has 5 residual blocks, each of which mainly contains 3 convolutional layers and 1 hole convolutional layer. Finally, the outputs of each residual block are accumulated, and the noise is predicted through 1 convolutional layer (including 1 ReLU activation function and 1 linear layer).

[0160] For the input residual noise, time step and conditional parameters, in order to keep the dimensions consistent, the dimension size can be converted by mapping. The specific mapping definition is as follows:

[0161] Input projection: x = ReLU(W i *Δy N +b i )

[0162] Time step embedding: N emb =W t *F emb (N)+b i

[0163] Input conditional projection: c = W c *inputs+b i

[0164] The above three inputs are fused in the residual block, and the features of the fused content are extracted, which is described in the following formula:

[0165] x conbined =F DiatedConv (F Conv (d emb )+x)+c

[0166] x left ,x right =F split (x conbined )

[0167] x left, x right =F split (F σ (x left )*F tanh (x right ))

[0168] output = x + x left

[0169] skip=F ReLU (F Conv (x right ))

[0170] where F DiatedConvrepresents the dilated convolution, F Conv represents the convolutional layer, F split Indicates separation by channel dimension, F σ Represents the Sigmoid activation function, F tanh represents the hyperbolic tangent activation function, F ReLU ReLU activation function.

[0171] The output of the current residual block is used as the input of the next residual block, and the skip of each residual block is accumulated and then predicted through the linear layer to predict the noise The specific description is as follows:

[0172]

[0173] F proj (x) = F ReLU (F Conv (F Linear (x)))

[0174] Furthermore, the step S203 specifically includes:

[0175] During the inference process, the final residual value is predicted by the denoising network Compared with the initial prediction results of DPN By adding them together, the yarn quality prediction result with uncertainty representation is obtained, ensuring that the prediction result contains both certainty and uncertainty information;

[0176] The final prediction results are shown as follows:

[0177]

[0178] in and They are the preliminary prediction results of DPN and the residual values ​​of the reverse denoising prediction of CDDPM, respectively. x and p represent the input raw cotton parameters and process parameters, respectively.

[0179] Step S3: By sampling the uncertainty prediction model multiple times, multiple residual values ​​are obtained, and added to the results of the deterministic prediction model to obtain the final prediction range and a deterministic prediction for reference by users.

[0180] In this example, the maximum number of sampling times is set to 20 times. Each sampling result is compared with the preliminary prediction result of DPN. If there is a sample greater than or less than the preliminary prediction result in N sampling times, this is used as the maximum and minimum values. If there is no sample greater than or less than the preliminary prediction result after the maximum number of sampling times, the preliminary prediction result is set to the maximum or minimum value.

[0181] The specific sampling is as follows:

[0182] s={s 1 ,s 2 ,…,s m ,s M}

[0183] s min =min{s}

[0184] s max =max{s}

[0185] in

[0186] Table 1 shows the mean square error (MSE) and determination coefficient (R) of the deterministic prediction results on the actual production data set of a cotton spinning mill. 2 ), where the bold value represents the result with the best prediction performance, the underlined value represents the result with the second best prediction performance, SOTA represents the method with the best prediction effect in the industry, Ours represents the prediction method proposed by us, and the method of the present invention is represented by DPN(Our).

[0187] Table 1 shows the MSE and R of the deterministic prediction results on the actual production data set of a cotton spinning mill. 2

[0188]

[0189] The HybridConvLSTM method proposed in the present invention is an improvement based on the CNN-LSTM model, that is, based on the feature extraction module in the original CNN-LSTM, the traditional convolutional neural network (CNN) is replaced by a multi-scale attention enhanced feature extraction module (MAFE). HybridConvGRU is a further improvement based on the HybridConvLSTM method, replacing the LSTM module in the HybridConvLSTM with a gated recurrent unit (GRU).

[0190] By comparing with the benchmark model, the proposed deterministic prediction model (DPN) shows the best performance in the task of cotton yarn quality prediction, with the highest MSE and R 2 The indicators are significantly better than other models.

[0191] Table 2 shows the results of ablation experiments on different module combinations on a real production dataset of a cotton mill. Each column in the experiment corresponds to a different module, including MAFE, AEL, GAF, LSTM, GRU, and FOP. The prediction methods used in the experiment include HybridConvLSTM, HybridConvGRU, and DPN. The check mark (√) in the table indicates that this module is included in the method.

[0192] Table 2 shows the module ablation experiment results on the actual production data set of a cotton spinning mill.

[0193]

[0194] The ablation experiment results verify the effectiveness of the AEL module and GAF ​​module in the present invention in time series data modeling, and also prove the ability of the MAFE module in feature extraction.

[0195] In summary, the present invention provides a method for predicting the quality of blended cotton yarn based on a hybrid deep learning model, including a deterministic prediction model (DPN) and an uncertainty prediction model (CDDPM). In the deterministic prediction model, the present invention adopts a lightweight deep learning model, the raw cotton features are processed by a convolution module to generate a multi-scale feature map; the process parameters are processed by a multi-head attention module and an LSTM module in a Transformer to extract their sequence features. In this process, an attention mechanism (CBAM) is used to increase the focus on key features, thereby enhancing the model's ability to capture complex relationships. The extracted raw cotton features and process parameter features are then feature fused through a cross-attention mechanism. This mechanism can capture the interactive information between raw cotton characteristics and process parameters, thereby better predicting yarn quality indicators. The fused features are input into a prediction network for further processing. The prediction network adopts a multi-head self-attention mechanism to ensure that complex feature associations are modeled, and finally outputs the yarn quality prediction results through a fully connected layer. In order to avoid the singularity of deterministic prediction, the present invention proposes an uncertainty prediction model, using CDDPM (conditional denoising diffusion probability model) to correct the deterministic prediction value of the first stage by predicting residual noise, thereby generating a final prediction result with uncertainty. The denoising network is embedded with raw cotton and process parameters as conditions, and the predicted residual value represents the fluctuation caused by randomness and uncertainty in the cotton blending process. This step provides uncertainty analysis for yarn quality prediction, making the prediction more comprehensive and suitable for industrial scenarios with high uncertainty and randomness.

[0196] The cotton blending-yarn quality prediction system of the present invention does not adopt a large pre-training model, but ensures high efficiency and low computing cost in the reasoning process through a lightweight architecture design. At the same time, the architecture has good scalability and can be applied to the prediction of cotton blending yarn quality under different raw cotton types and production processes. The method is suitable for the prediction of yarn quality in the textile industry. Through the introduction of multi-scale feature extraction and cross-attention mechanism, the accuracy and stability of the prediction can be effectively improved, helping enterprises to optimize production processes and raw material selection.

[0197] In addition, the present invention provides a cotton yarn quality prediction system based on a hybrid deep learning model, comprising:

[0198] The deterministic prediction module is used to input raw cotton indicators and process parameters into the deterministic prediction model for multi-scale feature extraction and processing, dynamically adjust the focus on important features, and output deterministic prediction results;

[0199] The uncertainty prediction module is used to input the residual between the prediction result output by the deterministic prediction model and the true value of the training data into the uncertainty prediction model, and use the raw cotton index and process parameters as conditions to perform uncertainty prediction;

[0200] The cotton yarn quality prediction module is used to obtain multiple residual values ​​by sampling the uncertainty prediction model multiple times, and add them to the results of the deterministic prediction model to obtain the final cotton yarn quality prediction result and its fluctuation range.

[0201] The specific limitations and beneficial effects of the above-mentioned cotton yarn quality prediction system based on a hybrid deep learning model can be found in the above-mentioned limitations on a cotton yarn quality prediction method based on a hybrid deep learning model, which will not be repeated here. The above-mentioned modules can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0202] In addition, the present invention also provides a computer device, which may include a processor, a memory, a network interface and a database connected through a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the processor executes the steps of a method for predicting the quality of cotton yarn based on a hybrid deep learning model as in any of the above embodiments.

[0203] The working process, working details and technical effects of the computer equipment provided in this embodiment can be found in the above embodiment of a method for predicting the quality of cotton yarn based on a hybrid deep learning model, which will not be described in detail here.

[0204] In addition, the present invention also provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the steps of a method for predicting the quality of cotton yarn based on a hybrid deep learning model as in any of the above embodiments are implemented. The computer-readable storage medium refers to a carrier for storing data, which may include but is not limited to a floppy disk, an optical disk, a hard disk, a flash memory, a USB flash drive and / or a memory stick, etc., and the computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices.

[0205] The working process, working details and technical effects of the computer-readable storage medium provided in this embodiment can be found in the above embodiment of a method for predicting the quality of cotton yarn based on a hybrid deep learning model, which will not be repeated here.

[0206] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0207] Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent replacements for some of the technical features therein. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting the quality of blended cotton yarn based on a hybrid deep learning model, characterized in that: The following steps are involved: Step S1: Input raw cotton indexes and process parameters into the deterministic prediction model to extract and process multi-scale features, dynamically adjust the focus on important features, and output deterministic prediction results; Step S2: input the residual between the prediction result output by the deterministic prediction model and the true value of the training data into the uncertainty prediction model, and use the raw cotton index and process parameters as conditions to perform uncertainty prediction; Step S3: by sampling the uncertainty prediction model multiple times, multiple residual values ​​are obtained, and the residual values ​​are added to the results of the deterministic prediction model to obtain the final prediction result of the quality of the blended cotton yarn and its fluctuation range.

2. The method for predicting the quality of blended cotton yarn based on a hybrid deep learning model according to claim 1, characterized in that: The step S1 specifically includes a regression prediction method; The regression prediction method specifically comprises the following steps: Step S101: normalizing the raw cotton index and process parameters as input; Step S102: obtaining raw cotton index features in the cotton blending process through a multi-scale attention enhanced feature extraction module; Step S103: the extracted raw cotton features are fed into the attention-enhanced long short-term memory network module, and the process parameters of each process are sequentially fed into each attention-enhanced long short-term memory network module; Step S104: Accumulate the output of each attention-enhanced long short-term memory network module and input it into the gated attention fusion module for fusion, and then add the output of the gated attention fusion module and the output of the last attention-enhanced long short-term memory network module as the final output; Step S105: The final output of step S104 is converted into a deterministic prediction result through the FOP module; Step S106: Denormalize the model prediction results.

3. The method for predicting the quality of blended cotton yarn based on a hybrid deep learning model according to claim 2, characterized in that: The step S101 specifically includes: The input raw cotton index is The process parameters are The yarn quality index is Where m represents the number of samples, d c ,d p and d y They represent the dimensions of raw cotton index, process parameters and final yarn quality index respectively; the normalized input of raw cotton index, process parameters and final yarn quality index is expressed as: c =Norm(x c ), x′ p =Norm(x p ) and y′=Norm(y), where Norm(.) represents the normalization operation, and the specific formula is: Among them, x represents the original data, μ represents the mean of the original data, and σ represents the standard deviation of the original data; The multi-scale attention-enhanced feature extraction module in step S102 includes 4 convolutional layers, 3 C2F layers, 1 SPPF layer and 1 CBAM attention mechanism layer, and the input is the normalized raw cotton index x′ c ; The convolution layer includes a 1D convolution layer, a BN layer, and a ReLU activation function. The convolution kernel size of the convolution layer is 3 and the step size is 2; the number of input and output feature channels of the C2F layer is the output channel of the previous convolution layer, and a residual connection is used; the number of input and output feature channels of the SPPF layer is the output feature channel number of the previous C2F; specifically described by the following formula: F c =F CBAM (F SPPF (F C2f (F Conv (F C2f (F Conv (F C2f (F Conv (F Conv (x′ c ))))))))) The step S103 specifically includes: Add the process features one by one, which can be described by the following formula: in, represents the features processed by the i-th group of attention-enhanced long short-term memory network modules; The step S104 specifically includes: The output features of the four groups of attention-enhanced long short-term memory network modules are fused through the gated attention fusion module, and then the output of the gated attention fusion module is added to the output of the last attention-enhanced long short-term memory network module to obtain the global fusion feature, which is specifically described by the following formula: Among them, F g is the global fusion feature, F σ Represents the Sigmoid activation function, F CBAM represents the CBAM attention mechanism, n represents the number of groups of AEL modules; The step S106 specifically includes: The predicted final yarn quality index Perform denormalization, Denorm(.) represents the denormalization operation, and the specific formula is: Denorm(x)=x*σ+μ Among them, x represents the normalized data, μ represents the mean of the original data, and σ represents the standard deviation of the original data.

4. The method for predicting the quality of blended cotton yarn based on a hybrid deep learning model according to claim 3, characterized in that: The attention-enhanced long short-term memory network module in step S103 is normalized into one group with one LSTM, one multi-head attention mechanism and one layer, and one group of attention-enhanced long short-term memory network modules receives a process parameter in a main process, wherein the output of the multi-scale attention-enhanced feature extraction is first converted into a data dimension by one LSTM module when entering the first group of attention-enhanced long short-term memory network modules; The gated attention fusion module in step S104 includes: 1 linear layer, 1 Sigmoid layer and 1 CBAM attention mechanism; wherein the input dimension and output dimension of the linear layer are consistent; The FOP module in step S105 includes: 1 convolution module, 1 C2F module, 1 one-dimensional maximum pooling layer and 2 linear layers; wherein the convolution layer is used to reduce the number of channels, C2F reduces the amount of calculation by reducing the number of channels, the maximum pooling layer is used to reduce the feature dimension, and finally the multi-dimensional features are converted into deterministic prediction results through the linear layer.

5. The method for predicting the quality of blended cotton yarn based on a hybrid deep learning model according to claim 4, characterized in that: The step S2 specifically includes the following steps: Step S201: During the training process, residual noise is generated through a diffusion process; the results are predicted by a deterministic prediction model The residual Δy0 is obtained by comparing it with the true value y of the training data, which is input into the diffusion process of the uncertainty prediction model, and Gaussian noise is gradually added to generate the final residual noise Δy N ; The diffusion process starts from Δy0, and Gaussian noise is gradually added after 100 time steps, finally making Δy N Close to the standard Gaussian distribution; Step S202: During the training process, the denoising network is used to predict the noise added in each step of the denoising process; the final residual noise Δy after denoising is N As input, the raw cotton index and process parameters are sent to the denoising network as conditions. Through the convolution layer, linear layer and back diffusion process, the noise is gradually predicted and removed, and finally the residual Δy0 is restored; in the inference process, the randomly generated Gaussian noise is used as input, and the raw cotton index and process parameters are sent to the denoising network as conditions. After 100 time steps of step-by-step denoising, the predicted residual is finally sampled. Step S203: During the inference process, the prediction residual obtained by sampling in step S202 is Prediction results with deterministic prediction model By adding them together, the yarn quality prediction result Y with uncertainty characterization is obtained, ensuring that the prediction result contains both certainty and uncertainty information.

6. The method for predicting the quality of blended cotton yarn based on a hybrid deep learning model according to claim 5, characterized in that: The step S201 specifically includes: Preliminary forecast output from a deterministic forecast model The residual value Δy0 is obtained from the true value y of the corresponding training data. The diffusion process of the uncertainty prediction model is used to gradually add Gaussian noise to the residual Δy0 to generate the final residual noise Δy N ; The forward diffusion process of the uncertainty prediction model is as follows: where Δy N , α N and ∈ N are the samples at time step,N,, the parameter controlling the noise intensity, and the Gaussian noise added at time step,N,, respectively; The forward diffusion process is deduced as follows: The final expression is: where Δy N is a linear combination of the original residual Δy0 and Gaussian noise ∈; as N increases, Gradually decreases, causing the noise component in the data to gradually increase until it finally approaches the standard Gaussian distribution; The step S202 specifically includes: During the training process, the denoising network predicts the noise added in each step of the denoising process; the final residual noise Δy N As input, the raw cotton index and process parameters are sent to the denoising network as conditions. The noise is gradually predicted and removed through the convolution layer, linear layer and reverse diffusion process, and the residual Δy0 is finally restored. The reverse denoising process starts from ΔyN step by step to Δy0, and the noise is completely removed. During the inference process, the denoising network receives randomly generated Gaussian noise as input, takes raw cotton indicators and process parameters as conditions, and performs step-by-step denoising for 100 time steps to finally sample the prediction residuals. The reverse denoising process of the uncertainty prediction model is as follows: where ∈ θ (Δy N , N, c) is the noise prediction value learned by the denoising network for a given Gaussian noise Δy N , time step N and condition c to predict the noise; During the training process, the inputs to the denoising network are the residual noise Δy N , time step N and conditional parameters, the conditional parameters are raw cotton indicators and process parameters; the denoising process of the uncertainty prediction model has 5 residual blocks, each residual block contains 3 convolutional layers and 1 hole convolutional layer, and finally the outputs of each residual block are accumulated, and the noise is predicted through 1 convolutional layer For the input residual noise, time step and conditional parameters, in order to keep the dimensions consistent, the dimension size is converted through mapping. The specific mapping definition is as follows: Input projection: x = ReLU(W i *Δy N +b i ) Time step embedding: N emb =W t *F emb (N)+b i Input conditional projection: c = W c *inputs+b i The above three inputs are fused in the residual block, and the features of the fused content are extracted, which is described in the following formula: x conbined =F DiatedConv (F Conv (d emb )+x)+c x left ,x right =P split (x conbined ) x left ,x right =F split (F σ (x left )*F tanh (x right )) output=x+x left skip=F ReLU (F Conv (x right )) where F DiatedConv represents the dilated convolution, F Conv represents the convolutional layer, F split Indicates separation by channel dimension, F σ Represents the Sigmoid activation function, F tanh represents the hyperbolic tangent activation function, F ReLU ReLU activation function. The output of the current residual block is used as the input of the next residual block, and the skip of each residual block is accumulated and then used to predict the noise through the linear layer. The specific description is as follows: F proj (x)=F ReLU (F Conv (F Linear (x))) The step S203 specifically includes: During the inference process, the final residual value is predicted by the denoising network Then compare the initial prediction results of the deterministic prediction model Adding them together, the yarn quality prediction result with uncertainty representation is obtained, ensuring that the prediction result contains both certainty and uncertainty information; The final prediction results are shown as follows: in and are the preliminary prediction results of the deterministic prediction model and the residual values ​​of the reverse denoising prediction of the uncertainty prediction model, respectively. x and p represent the input raw cotton parameters and process parameters, respectively.

7. The method for predicting the quality of blended cotton yarn based on a hybrid deep learning model according to claim 6, characterized in that: The specific sampling in step S3 is as follows: s={s 1 ,s 2 ,...,s m ,s M } s min =min{s} s max =max{s} in 8. A cotton yarn quality prediction system based on a hybrid deep learning model, characterized in that: include: The deterministic prediction module is used to input raw cotton indicators and process parameters into the deterministic prediction model for multi-scale feature extraction and processing, dynamically adjust the focus on important features, and output deterministic prediction results; The uncertainty prediction module is used to input the residual between the prediction result output by the deterministic prediction model and the true value of the training data into the uncertainty prediction model, and use the raw cotton index and process parameters as conditions to perform uncertainty prediction; The cotton yarn quality prediction module is used to obtain multiple residual values ​​by sampling the uncertainty prediction model multiple times, and add them to the results of the deterministic prediction model to obtain the final cotton yarn quality prediction result and its fluctuation range.

9. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of a method for predicting cotton yarn quality based on a hybrid deep learning model as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of a method for predicting cotton yarn quality based on a hybrid deep learning model as described in any one of claims 1 to 7 are implemented.

Images