Temperature prediction method for charging and moisture regaining equipment based on time sequence fusion network model

Abstract

The invention discloses a charging and moisture regaining equipment temperature prediction method based on a time sequence fusion network model, and relates to the field of production process control, and the method comprises the steps: collecting time sequence data in a charging and moisture regaining equipment production environment, and carrying out the preprocessing; dividing the data set into a training set, a verification set and a test set, and injecting Gaussian noise into the training set; a prediction model for predicting the outlet temperature is constructed, and the prediction model is a time sequence fusion network model and comprises a residual TCN time sequence convolutional network, an SK-Net multi-scale attention network and a BiLSTM bidirectional circulation network; pre-training the prediction model by using the training set; utilizing the trained prediction model to predict the outlet temperature of the feeding and moisture regaining equipment; and evaluating a prediction result, and if an evaluation index is greater than a threshold value, starting an incremental training process to re-train the prediction model. According to the invention, through a multi-module combined deep learning model, the prediction accuracy of the outlet temperature of the charging and moisture regaining equipment can be improved in a complex and changeable industrial environment.

Description

Technical Field

[0001] The present invention relates to the field of production process control, and particularly to a temperature prediction method for a feeding and humidifying device based on a time series fusion network model. Background Art

[0002] In the cigarette production process, the precise control of the outlet temperature of the feeding and humidifying device is the core guarantee for the uniformity of the tobacco leaf moisture content, the penetration effect of flavors, and the process stability. Traditional PID control relies on the adjustment of empirical parameters and cannot predict the long-delay characteristics of the thermodynamic system, resulting in a too high overshoot; although the predictive control based on the mechanism model (such as the finite element method) can describe the temperature field distribution, it has a high computational complexity, poor real-time performance, and it is difficult to adapt to the attenuation of the heat transfer coefficient caused by equipment aging. The statistical learning model (such as the Auto-Regressive Integrated Moving Average model (ARIMA) + Support Vector Machine (SVM)) relies on artificial feature engineering and has insufficient generalization ability for non-linear working conditions. In recent years, deep learning methods based on the Convolutional Neural Network (CNN) have been introduced into the field of industrial temperature prediction, but the existing solutions still have significant defects in terms of time series modeling ability and industrial scenario adaptability. The implementation solution closest to the technical solution of the present invention is the temperature prediction method based on CNN. This method converts the temperature time series data into a two-dimensional grayscale image (the horizontal axis is time and the vertical axis is the temperature value), extracts local spatio-temporal patterns through a convolutional network, and finally outputs the temperature prediction value through a fully connected layer. Its core lies in using the local perception characteristics of CNN to capture the short-term correlation in temperature fluctuations, but there are significant defects: the image processing destroys the temporal continuity of the original data, resulting in the breakage of the causal chain between the steam valve action and the temperature response; the fixed receptive field of the convolutional kernel cannot model the long-term thermal delay effect, and the long-term prediction trend error is large.

[0003] The existing cut tobacco drying process takes PID control as the core and relies on linear models to describe the heat and mass transfer processes. However, when faced with a dynamic system with high coupling, strong nonlinearity, and significant hysteresis, it is difficult to accurately adjust. The strong correlation between multiple variables such as temperature and humidity leads to interference between control parameters. During the drying stage, the phenomenon of "dry head and dry tail" often occurs, and there are significant deviations in the moisture content, directly affecting the physical properties and combustion quality of the cut tobacco. Production regulation highly depends on the operator's experience and lacks standardized control rules, resulting in obvious quality fluctuations between different batches. The manual response speed cannot match the changes in dynamic working conditions, and it is difficult to transform excellent operation cases into universal strategies, further exacerbating the problems of uneven moisture content distribution and unstable sensory quality. The early linear regression models were limited by the strong nonlinearity, multi-variable coupling, and complex noise interference of production data, resulting in insufficient prediction accuracy. Model updates rely on manual calibration and cannot adapt to fluctuations in raw materials and changes in equipment status, leading to increased energy consumption and limited process efficiency. There are problems such as noise interference, feature redundancy, and dynamic drift in the production process data, and traditional algorithms are difficult to effectively extract key control information. Existing intelligent algorithms have insufficient computational efficiency, are difficult to meet the requirements of real-time control, and miscontrol phenomena occur frequently, still relying on manual calibration and correction.

[0004] In summary, the existing traditional linear models are difficult to depict the nonlinear delay response law between steam pressure and outlet temperature, and the prediction accuracy is low. Summary of the Invention

[0005] In view of the above-mentioned defects of the prior art, the technical problem to be solved by the present invention is that traditional linear models are difficult to depict the nonlinear delay response law between steam pressure and outlet temperature, and the prediction accuracy is low.

[0006] To achieve the above object, the present invention provides a temperature prediction method for a feeding and conditioning equipment based on a time series fusion network model, and the method includes the following steps: S101: Collect time series data in the production environment of the feeding and conditioning equipment, and preprocess the time series data to generate a data set; S103: Divide the data set into a training set, a validation set, and a test set, and inject Gaussian noise into the training set, and the Gaussian noise simulates sensor errors; S105: Construct a prediction model for predicting the outlet temperature of the feeding and conditioning equipment, and the prediction model is configured as a time series fusion network model, and the time series fusion network model includes a residual TCN time series convolutional network, an SK-Net multi-scale attention network, and a BiLSTM bidirectional recurrent network; S107: Pre-train the prediction model using the training set; S109: Use the trained prediction model to predict the outlet temperature of the feeding and conditioning equipment; S111: Evaluate the prediction result of the prediction model. If the evaluation index is greater than the threshold, start the incremental training process to retrain the prediction model.

[0007] Further, in the step S101, the time series data is three-dimensional time series tensor data, including the number of samples, the time step, and the feature dimension. The feature dimension includes the actual value of the outlet temperature, the outlet moisture content, the actual value of the process steam flow rate, the water addition flow rate, and the inlet moisture.

[0008] Further, in the step S101, the preprocessing includes missing data filling, abnormal data filling, and dynamic normalization, where: For the missing data filling, for the data with multiple consecutive missing time steps, linear interpolation is used for filling:

[0009] where, is the missing length, is the time step, is the data at the time step, is the data at the time step, is the data at the time step; For the abnormal data filling, for the continuously abnormal data, call the historical database to retrieve the average value in the recent period under the same working condition to fill the abnormal time period, and add an abnormal flag bit; For the dynamic normalization, perform normalization separately according to the feature dimension:

[0010] where, is the data before normalization, is the data after normalization, is the minimum value of the th dimension feature in the training set, is the th dimension feature in the training set.

[0011] Further, in the step S105, the residual TCN time series convolutional network effectively captures long-term dependencies through causal convolution and dilated convolution, including the following steps: S10511: TCN preprocessing, mapping the input feature dimension to the internal channel number of the TCN through convolution for the preprocessed time series data:

[0012] S10512: Residual block level processing, performing dilated causal convolution on the time series data preprocessed by the TCN to obtain the first layer output tensor:

[0013] S10513: Pass the first layer output tensor through weight normalization and the ReLU activation function:

[0014] S10514: Add the main path output and the original input to obtain the first layer processing result:

[0015] S10514: Loop processing, use the first layer processing result as the input of the next layer, continuously process multiple times to obtain the final result, and complete the channel number expansion of the final result through convolution:

[0016] Among them, is the time series data, , is the time series data, is the batch size, is the time step, is the feature dimension, is the internal channel number of the TCN, is the first layer output tensor, is the first layer processing result, The processing result of the nth layer, is the final result processed by the residual TCN time series convolutional network, is the one-dimensional convolution operation function.

[0017] Furthermore, in the step S105, the SK-Net multi-scale attention network configuration realizes feature extraction and enhances the sensitivity of the model to key regions, including the following steps: S10521: Pass the preprocessed time series data through double-branch convolution to output feature maps of two branches, and process the feature maps through the GELU activation function; S10522: Generate channel attention weights according to the feature maps of the two branches; S10523: Split the channel attention weights into two channel weight vectors, and perform the normalized exponential function operation on the weights of each channel to generate dynamic allocation coefficients; S10524: Broadcast and expand the generated dynamic allocation coefficients along the time step dimension, and perform weighted summation on the double-branch feature maps to obtain the fused enhanced feature map; S10525: Align the dimensions of the preprocessed time series data through convolution and add it to the fused enhanced feature map to obtain the finally output feature map.

[0018] Further, in the step S10522, the generation of the channel attention weight includes the following sub-steps: S105221: Add the feature maps obtained by double-branch convolution element-wise to obtain a fused feature; S105222: Compress the spatial information of the fused feature along the time dimension to generate channel statistics:

[0019] S105223: Compress the channel information of the channel statistics through a fully connected layer, and the activation function of the fully connected layer is the ReLU function; S105224: Restore the processed data to the original number of channels to generate a compact weight vector:

[0020] Among them, is the channel statistic, is the fused feature, is the compact weight vector, is the relu function, is the batch normalization process, is the global average pooling function, is the fully connected layer, is the spatial height of the feature map, is the spatial width of the feature map, 、 are the horizontal and vertical coordinate indices of the spatial position.

[0021] Further, in the step S10523, the generation of the dynamic allocation coefficient is calculated using the following formula:

[0022] Among them, is the channel weight, is the channel weight vector, is the control parameter for dynamically adjusting the channel weight allocation ratio.

[0023] Further, in the step S105, the BiLSTM bidirectional recurrent network reverses the input sequence of the backward LSTM layer in time series and synchronously processes the process data with the forward LSTM. By splicing the double hidden state vectors, the bidirectional causal association is completed to achieve more accurate prediction.

[0024] Further, in the step S105, when the BiLSTM bidirectional recurrent network works, it includes the following steps: S10531: The forward LSTM updates the hidden state and outputs the final sequence in the following manner:

[0025] The final output sequence of the forward LSTM is: ; S10532: The backward LSTM updates the hidden state and outputs the final sequence using the following formula:

[0026] The final output sequence of the backward LSTM is: ; S10533: Concatenate the output results of the forward LSTM and the backward LSTM bidirectionally to obtain the result :

[0027] S10533: Perform inference one more time according to the above steps to obtain

[0028] S10534: Map the obtained output to the prediction target dimension through a fully connected layer to obtain the final prediction result:

[0029] Where: is the concatenated feature, is the predicted future time step, is the variable predicted at each time step, is the time step, is the hidden state of the first layer of the forward LSTM at time step t, is the cell state of the first layer of the forward LSTM at time step t, is the hidden state of the first layer of the backward LSTM at time step t, is the cell state of the first layer of the backward LSTM at time step t, is the bias term of the fully connected layer, is the weight matrix of the fully connected layer.

[0030] Further, when pre-training the prediction model in the step S107, the following loss function is used to calculate the Huber loss:

[0031] Where, is the Huber loss, is the predicted value, is the actual value, is the threshold coefficient of the Huber loss.

[0032] In a preferred embodiment of the present invention, compared with the prior art, the present invention has the following beneficial effects: 1. The present invention directly processes the original time series data instead of image conversion, avoiding the control lag caused by time series breakage in the CNN method, and can maintain time series continuity, thus realizing short-time closed-loop response; 2. The present invention adopts bidirectional time series modeling, and the forward-backward chain reasoning of BiLSTM accurately traces the root cause of anomalies, greatly reducing the fault location time; 3. The present invention effectively captures the long-term causal chain between various factors and temperature response through the dilated causal convolution of TCN, realizes long-term thermal delay modeling, and compared with the CNN image method, the long-term prediction error is greatly reduced; 4. The present invention adopts the dual-branch dynamic weight allocation of SK-Net to simultaneously extract second-level fluctuations (such as sudden changes in steam flow) and minute-level trends (such as slow increase in humidity), realizes multi-scale feature fusion, and greatly improves the anomaly detection accuracy under the condition of sudden changes in raw material humidity; 5. The present invention adopts a dynamic attention mechanism, and the channel weights of SK-Net can adaptively adjust the feature attention, making the prediction error of the model stable within a very small range under different raw material batches.

[0033] The following will further illustrate the concept, specific structure and technical effects of the present invention with reference to the accompanying drawings, so as to fully understand the purpose, features and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is a schematic diagram of the steps of the prediction method according to a preferred embodiment of the present invention; Figure 2 is a schematic diagram of the overall architecture of the prediction model according to a preferred embodiment of the present invention; Figure 3 is a schematic diagram of the overall structure of the SK-Net multi-scale attention network according to a preferred embodiment of the present invention; Figure 4 is a schematic diagram of the overall structure of the BiLSTM bidirectional recurrent network according to a preferred embodiment of the present invention; Figure 5 is a schematic diagram of the overall structure of the residual TCN temporal convolutional network according to a preferred embodiment of the present invention; Figure 6 is a schematic diagram of the dilated causal convolution structure of the residual TCN temporal convolutional network according to a preferred embodiment of the present invention; Figure 7 It is a schematic diagram of the prediction result of the prediction model of a preferred embodiment of the present invention. Detailed implementation manners

[0035] The following introduces multiple preferred embodiments of the present invention with reference to the accompanying drawings of the specification to make its technical content clearer and easier to understand. The present invention can be embodied in many different forms of embodiments, and the protection scope of the present invention is not limited to the embodiments mentioned in the text.

[0036] In the drawings, components with the same structure are denoted by the same numerical labels, and components with similar structures or functions everywhere are denoted by similar numerical labels. The size and thickness of each component shown in the drawings are arbitrarily shown, and the present invention does not limit the size and thickness of each component. To make the illustration clearer, the thickness of some parts in the drawings is appropriately exaggerated.

[0037] Aiming at the fact that traditional linear models are difficult to depict the non-linear delay response law between steam pressure and outlet temperature, the present invention adopts an expanded causal convolution TCN module (Temporal Convolutional Network, TCN), expands the model's temporal receptive field through a hierarchical expansion structure, and captures long-term delay correlation features on the premise of strictly maintaining causality, so as to solve the problem of control inaccuracy caused by signal transmission lag in time-varying systems; aiming at the multi-parameter high-dimensional coupling characteristics such as temperature and humidity, steam flow, etc., a channel-spatial dual attention mechanism is designed. The channel attention dynamically identifies the influence weights of key process parameters, and the spatial deformable convolution adaptively adjusts the local feature extraction mode of sensor data, effectively separating the hidden interference between multi-variables, and realizing accurate modeling of the heterogeneous temperature field distribution in the drying oven; by constructing a bidirectional long short-term memory network symmetric coding architecture, through fusing the forward process time series and the reverse causal correlation, a parameter dynamic compensation mechanism is established; combined with the hierarchical downsampling characteristics of the residual TCN, double filtering of noise suppression and drift correction is realized during the feature extraction process, enhancing the adaptability of the model to complex working conditions. An online incremental learning framework is constructed based on the residual connection and the sliding time window mechanism, enabling the model to dynamically adjust parameters according to real-time process data. When changes in raw material characteristics or equipment status are detected, it automatically triggers the progressive optimization of the control strategy, breaking through the limitations of the response lag of the traditional manual calibration mode.

[0038] As Figure 1 shown, a method for predicting the temperature of a feeding and humidifying device based on a time series fusion network model provided by an embodiment of the present invention includes the following steps: S101: Collect time series data in the production environment of the feeding and humidifying device, and preprocess the time series data to generate a data set.

[0039] The collected time series data is three-dimensional time series tensor data, including the number of samples, time steps, and feature dimensions. Among them, the feature dimensions include multiple relevant parameters in the production environment of the feeding and conditioning equipment, including the actual value of the outlet temperature, the outlet moisture content, the actual value of the process steam flow rate, the water addition flow rate, and the inlet moisture content.

[0040] For the collected time series data, preprocessing is required to generate relevant data sets. In this embodiment, the preprocessing includes missing data filling, abnormal data filling, and dynamic normalization, where: 1) Missing data filling: For data with multiple consecutive missing time steps, linear interpolation is used to fill:

[0041] Among them, is the missing length, is the time step, is the data at time step, is the data at time step, is the data at time step.

[0042] 2) Abnormal data filling: For continuously abnormal data, the historical database is called to retrieve the average value in the recent period under the same working conditions to fill the abnormal time period, and an abnormal flag bit is added.

[0043] 3) Dynamic normalization: For the collected time series data, normalization is performed separately according to the feature dimensions:

[0044] Among them, is the data before normalization, is the data after normalization, is the minimum value of the d-th feature in the training set, is the maximum value of the d-th feature in the training set.

[0045] In this embodiment, when performing the normalization process, parameters such as temperature and humidity are mapped to the range of [0, 1].

[0046] S103: Divide the data set into a training set, a validation set, and a test set, and inject Gaussian noise into the training set. The Gaussian noise simulates sensor errors.

[0047] The data set is divided into a training set, a validation set, and a test set. The training set is used to train the prediction model, the validation set is used to verify the trained prediction model, and the test set is used to test the trained prediction model.

[0048] To simulate sensor errors and improve the robustness of the model, Gaussian noise (σ = 0.01) is injected into the training set.

[0049] S105: Construct a prediction model for predicting the outlet temperature of the feeding and humidifying equipment. The prediction model is configured as a temporal fusion network model, and the temporal fusion network model includes a residual TCN temporal convolutional network, a multi-scale attention network (Selective Kernel Networks, SK-Net), and a BiLSTM bidirectional recurrent network.

[0050] In this embodiment, the prediction model for predicting the outlet temperature of the feeding and humidifying equipment adopts a temporal fusion network architecture. The prediction model uses a cascaded fusion model including residual TCN (temporal convolution) + SK-Net (multi-scale attention) + BiLSTM (bidirectional recurrent network). The dilated causal convolutional TCN network expands the temporal receptive field of the model through a hierarchical dilation structure, captures long-term delay correlation features while strictly maintaining causality, and solves the problem of control inaccuracy caused by signal transmission lag in time-varying systems; for the multi-parameter high-dimensional coupling characteristics such as temperature and humidity, steam flow, etc., the SK-Net multi-scale attention network is adopted, and a channel-spatial dual attention mechanism is designed. The channel attention dynamically identifies the influence weights of key process parameters, and the spatial deformable convolution adaptively adjusts the local feature extraction mode of sensor data, effectively separating the hidden interference between multi-variables and realizing the accurate modeling of the heterogeneous temperature field distribution in the drying oven; by constructing a bidirectional BiLSTM symmetric encoding architecture, through fusing the forward process time series and the reverse causal association, a parameter dynamic compensation mechanism is established; combined with the hierarchical downsampling characteristics of residual TCN, double filtering of noise suppression and drift correction is realized during the feature extraction process, enhancing the adaptability of the model to complex working conditions.

[0051] 1) The residual TCN temporal convolutional network effectively captures long-term dependencies through causal convolution and dilated convolution, including the following steps: S10511: TCN preprocessing, mapping the input feature dimension to the internal channel number of TCN through convolution for the preprocessed time series data:

[0052] S10512: Residual block hierarchical processing, processing the time series data preprocessed by TCN through dilated causal convolution to obtain the first-layer output tensor:

[0053] S10513: Pass the first-layer output tensor through weight normalization and the ReLU activation function:

[0054] S10514: Add the main path output to the original input to obtain the first-layer processing result:

[0055] S10514: Perform loop processing. Use the first-layer processing result as the input for the next layer. After continuous processing multiple times, obtain the final result, and complete the channel number expansion for the final result through convolution:

[0056] Among them, is time series data, , is time series data, is the batch size, is the time step, is the feature dimension, is the number of internal channels of the TCN, is the output tensor of the first layer, is the first-layer processing result, the processing result of the nth layer, is the final result processed by the residual TCN time series convolutional network, is the one-dimensional convolution operation function.

[0057] 2) The SK-Net multi-scale attention network configuration realizes feature extraction and enhances the sensitivity of the model to key regions, including the following steps: S10521: Process the preprocessed time series data through double-branch convolution, output the feature maps of the two branches, and process the feature maps through the GELU activation function; S10522: Generate channel attention weights based on the feature maps of the two branches; S10523: Split the channel attention weights into two channel weight vectors, and perform the normalized exponential function operation on the weights of each channel to generate dynamic allocation coefficients; S10524: Broadcast and expand the generated dynamic allocation coefficients along the time step dimension, and perform weighted summation on the double-branch feature maps to obtain the fused enhanced feature map; S10525: Align the dimensions of the preprocessed time series data through convolution and add it to the fused enhanced feature map to obtain the finally output feature map.

[0058] In this embodiment, step S10522 for generating channel attention weights includes the following sub-steps: S105221: Add the feature maps obtained by double-branch convolution element-wise to obtain the fused feature; S105222: Compress the spatial information of the fusion feature along the time dimension to generate channel statistics:

[0059] S105223: Compress the channel information of the channel statistics through a fully connected layer, and the activation function of the fully connected layer is the ReLU function; S105224: Restore the processed data to the original number of channels to generate a compact weight vector:

[0060] where, is the channel statistics, is the fusion feature, is the compact weight vector, is the relu function, is the batch normalization process, is the global average pooling function, is the fully connected layer, is the spatial height of the feature map, is the spatial width of the feature map, 、 are the horizontal and vertical coordinate indices of the spatial position.

[0061] In this embodiment, the dynamic allocation coefficient generated in step S10523 is calculated using the following formula:

[0062] where, is the channel weight, is the channel weight vector, is the control parameter for dynamically adjusting the channel weight allocation ratio.

[0063] 3) The BiLSTM bidirectional recurrent network reverses the input sequence of the backward LSTM layer in time sequence and synchronously processes the process data with the forward LSTM. By splicing the dual hidden state vectors, bidirectional causal association is completed to achieve more accurate prediction.

[0064] The BiLSTM bidirectional recurrent network includes the following steps when working: S10531: The forward LSTM updates the hidden state and outputs the final sequence in the following way:

[0065] The final output sequence of the forward LSTM is: ; S10532: The backward LSTM updates the hidden state and outputs the final sequence using the following formula:

[0066] The final output sequence of the backward LSTM is: ; S10533: Bi-directionally concatenate the output results of the forward LSTM and the backward LSTM to obtain the result :

[0067] S10533: Perform inference one more time according to the above steps to obtain

[0068] S10534: Map the obtained output to the prediction target dimension through a fully connected layer to obtain the final prediction result:

[0069] Where: is the concatenated feature, is the predicted future time step, is the variable predicted at each time step, is the time step, is the hidden state of the first layer of the forward LSTM at time step t, is the cell state of the first layer of the forward LSTM at time step t, is the hidden state of the first layer of the backward LSTM at time step t, is the cell state of the first layer of the backward LSTM at time step t, is the bias term of the fully connected layer, is the weight matrix of the fully connected layer.

[0070] S107: Pre-train the prediction model using the training set.

[0071] When pre-training the prediction model, the following loss function is used to calculate the Huber loss:

[0072] Where, is the Huber loss, is the predicted value, is the actual value, is the threshold coefficient of the Huber loss.

[0073] S109: Use the trained prediction model to predict the outlet temperature of the feeding and humidifying equipment.

[0074] S111: Evaluate the prediction results of the prediction model. If the evaluation index is greater than the threshold, start the incremental training process to retrain the prediction model.

[0075] Compared with the prior art, the temperature prediction method for the feeding and humidifying equipment based on the time series fusion network model provided by the embodiments of the present invention has the following characteristics: 1. Maintaining temporal continuity: The present invention directly processes the original time series data instead of the converted image data, avoiding the control lag caused by temporal breaks in the CNN method of the prior art and enabling short-term closed-loop response. 2. Bidirectional time series modeling: The present invention uses the forward-backward chain reasoning of BiLSTM to accurately trace the root cause of anomalies, greatly reducing the fault location time. 3. Long-term thermal delay modeling: The present invention effectively captures the long-term causal chain between various factors and temperature response through the dilated causal convolution of TCN. Compared with the CNN image method, the long-term prediction error is greatly reduced. 4. Multi-scale feature fusion: The SK-Net dual-branch dynamic weight allocation designed in the present invention simultaneously extracts second-level fluctuations (such as sudden changes in steam flow) and minute-level trends (such as slow increase in humidity). Under the condition of sudden changes in raw material humidity, the anomaly detection accuracy is greatly improved. 5. Dynamic attention mechanism: The channel weights in the SK-Net of the present invention can adaptively adjust the feature attention, making the prediction error of the model stable within a very small range under different raw material batches.

[0076] The following will describe the present invention in detail in combination with the preferred embodiments of the present invention.

[0077] The present invention provides a method for predicting the outlet temperature of the feeding and humidifying equipment based on a hybrid network model of residual TCN and BiLSTM. The overall architecture of the model is as Figure 2 shown. By combining multiple deep learning technology modules, the ability to extract features and model temporal dependencies from the operation state data of the feeding and humidifying equipment is improved. Among them, the SK-Net attention mechanism enhances the sensitivity of the model to key regions. The residual TCN time series convolutional network effectively captures long-term dependencies through causal convolution and dilated convolution. The BiLSTM network further improves the capture accuracy of temporal dependencies through the combination of forward and backward LSTM. Through this deep learning model combining multiple modules, in a complex and changing industrial environment, the present invention has high prediction accuracy and robustness for the outlet temperature of the feeding and humidifying equipment, effectively ensuring the quality stability of tobacco leaf processing.

[0078] The prediction method provided in this embodiment includes the following steps: Step 1: Construct a sliding window for input data.

[0079] When collecting production data in the production process of the feeding and humidifying equipment, define a window length of 60 time steps (step size 1 step) and construct a three-dimensional time series tensor , where: is the number of samples; is the time step size, is the feature dimension. In this embodiment, , and the features include parameters such as the actual value of the outlet temperature, the outlet moisture content, the actual value of the process steam flow rate, the water addition flow rate, and the inlet moisture content.

[0080] Step 2: Data repair strategy For data with continuous missing values ≤ 3 time steps, use linear interpolation to fill:

[0081] where, is the missing length.

[0082] If the data anomaly persists for > 5 minutes, call the historical database to retrieve the average value in the most recent 30 days under the same working conditions , and use to fill the abnormal segment and add an abnormal flag bit.

[0083] Step 3: Dynamic normalization.

[0084] Perform Min - Max normalization separately by feature dimension:

[0085] where, is the minimum / maximum value of the th - dimensional feature in the training set; map parameters such as temperature and humidity to [0, 1].

[0086] Step 4: Dataset division and data augmentation.

[0087] Divide the first time series data into the training set, the middle data into the validation set; and the last data into the test set. Inject Gaussian noise into the training set to simulate sensor errors and improve the robustness of the model.

[0088] Step 5: Multi - scale attention feature selection The overall structure of the SK - Net multi - scale attention network is as Figure 3 shown.

[0089] Step 5.1: Split, dual - branch convolutional feature extraction For the time series data pre - processed in Step 4 ( is the batch size, is the time step, and is the feature dimension) are processed through two branches respectively: 1) Branch 1: Adopt a convolutional kernel (number of channels 64, stride 1, padding 1), and output the feature map ; 2) Branch 2: Adopt a convolutional kernel (number of channels 64, stride 1, padding 2), and output the feature map ; Enhance the non-linear expression ability of the outputs of both branches through the GELU activation function.

[0090] Step 5.2: Fusion, generation of channel attention weights Add the obtained in step 5.1 and element-wise to obtain the fused feature ; Compress the spatial information of along the time dimension T to generate the channel statistic , and the calculation formula is as follows:

[0091] Compress the channel information of the statistic through a fully connected layer (output dimension 16). The activation function of the fully connected layer is ReLU, and then restore the processed data to the original number of channels 64 to generate a compact weight vector . The formula is as follows:

[0092] where, is the relu function, and B represents batch normalization processing.

[0093] Step 5.3: Selection, dynamic feature weight allocation Split the obtained in step 5.2 into two channel weight vectors , which respectively correspond to the channel weights of the 3×3 and 5×5 branches. Perform the normalized exponential function operation (Softmax) on the weights of each channel to generate the dynamic allocation coefficient. The formula is as follows:

[0094] where, is the allocation weight for each channel.

[0095] Broadcast and expand along the time step dimension T to , and perform weighted summation on the feature maps of both branches:

[0096] Among them, ⊙ represents channel-wise multiplication.

[0097] Then the output is obtained, and the fused enhanced feature map is obtained:

[0098] The time-series data preprocessed in step 4 is passed through convolution (output channels 64, without activation function) to align the dimensions, obtaining ; Then add them:

[0099] The final output feature map is obtained: .

[0100] Step 6: Residual TCN convolution The structure of the residual TCN time-series convolutional network is as Figure 5 shown.

[0101] Step 6.1: TCN preprocessing.

[0102] The time-series data preprocessed in step 4 is passed through convolution to map the input feature dimension from to the internal channel number of the TCN , and the formula is as follows:

[0103] Step 6.2: Residual block hierarchical processing The result obtained in step 6.1 is processed by dilated causal convolution (the structure of the dilated causal convolution is as Figure 6 shown), and the parameters of the dilated causal convolution are set as (kernel size , dilation factor , output channels 64; left padding: ).

[0104] Finally, the output tensor of the first layer is obtained:

[0105] Then the obtained result is passed through weight normalization and the ReLU activation function:

[0106] Finally, the output of the main path is added to the original input to obtain the result of the first layer processing:

[0107] Step 6.3: Loop processing Use the result obtained in Step 6.2 as the input for the next layer, i.e.: , and then process it continuously three times to obtain the final result .

[0108] Then, through convolution, expand the number of channels from 64 to 128 to adapt to the input requirements of the downstream module:

[0109] Step 7: Feature concatenation Concatenate the results from Step 5 and Step 6 .

[0110] Where: is the batch size, is the time step, are the output channels of TCN and SK-Net respectively. After concatenation, the dimension is .

[0111] Then expand the concatenated features along the time step dimension, and the input feature dimension for each time step is 192.

[0112] The final input sequence Where .

[0113] Step 8: BiLSTM bidirectional recurrent network processing The overall structure of the BiLSTM bidirectional recurrent network is as Figure 4 shown.

[0114] Step 8.1: Bidirectional LSTM processing The input of the forward LSTM is the result processed in Step 7.1, i.e. ; The formula for updating the hidden state is as follows:

[0115] The final output sequence after passing through the hidden state is: .

[0116] The input of the backward LSTM is the reverse time step ; The formula for updating the hidden state is as follows:

[0117] The final output sequence after passing through the hidden state is: .

[0118] Finally, the results of the forward LSTM and the backward LSTM are concatenated bidirectionally to obtain the result

[0119]

[0120] Similarly, perform another inference according to the above steps to obtain

[0121] Step 8.2: Prediction head (fully connected layer) Take the output of Step 8.1 , and map it to the prediction target dimension through a fully connected layer to obtain the final prediction result , and the formula is as follows:

[0122] Among them, is the predicted future time step, is the variable predicted at each time step.

[0123] Step 9: Model training Step 9.1: Forward calculation: Input the batch data of Step 4 After being processed by Step 5, Step 6, Step 7, and Step 8, output the predicted value , and calculate the Huber loss .

[0124] The formula of the Huber loss function is as follows:

[0125] Step 9.2: Backpropagation The loss gradient Backpropagates from the prediction head to the BiLSTM Feature concatenation layer TCN / SK-Net Input layer, and the gradients of TCN and SK-Net are jointly updated with weights through the chain rule.

[0126] First, calculate the total gradient norm ; If , scale the gradient proportionally: .

[0127] The weight is updated by the Adam optimizer to update the parameters, and at the same time, weight decay is applied ( .

[0128] Step 9.3: The forward propagation in Step 9.1 and the backward propagation in Step 9.2 constitute a training process. After 100 rounds of training, a trained model is finally obtained.

[0129] Step 10: Working stage After being preprocessed in Step 3, real-time data (signals such as temperature and steam flow) is fed into the model trained in Step 9 in 32 batches per minute for prediction. The prediction results are as Figure 7 shown.

[0130] The prediction results are transmitted to the PLC controller within <50ms. Based on the predicted values, the PLC calculates the valve opening (such as from 45% → 52%) through the incremental PID algorithm to suppress the impact of sudden changes in raw material humidity.

[0131] When the prediction error is continuously greater than 1% for 3 consecutive times, it automatically switches to the fuzzy control mode to avoid the out-of-control accumulation of the integral term.

[0132] Step 10: Shutdown and evaluation After the equipment is shut down, the system automatically generates evaluation indicators, such as: 1) RMSE: Root mean square error, the comprehensive deviation intensity between the predicted value and the measured value in the laboratory; 2) MAE: Mean absolute error, the average absolute value of the prediction error; 3) : Coefficient of determination, the explanatory ability of the model for temperature fluctuations.

[0133] If the RMSE is greater than the threshold, the system calls the data for the past 30 days and starts the incremental training process to retrain the model. If the indicators are qualified, a heat map of feature weights is generated to prompt the engineer to check the sensor calibration status; at the same time, the PID parameters are adjusted according to the attention weights to maintain the stable operation of the equipment.

[0134] In the process of cigarette production, the precise control of the outlet temperature of the feeding and conditioning equipment is the core guarantee for the uniformity of tobacco leaf moisture content, the penetration effect of flavorings, and process stability. Traditional PID control relies on the adjustment of empirical parameters and cannot predict the long-delay characteristics of the thermodynamic system, resulting in a too high overshoot. Although predictive control based on a mechanism model (such as the finite element method) can describe the temperature field distribution, it has a high computational complexity, poor real-time performance, and is difficult to adapt to the attenuation of the heat transfer coefficient caused by equipment aging. Statistical learning models (such as ARIMA + SVM) rely on artificial feature engineering and have insufficient generalization ability for non-linear working conditions. In recent years, deep learning methods based on convolutional neural networks (CNNs) have been introduced into the field of industrial temperature prediction, but existing solutions still have significant defects in terms of time series modeling ability and industrial scenario adaptability. For the temperature prediction method based on CNN, this method converts temperature time series data into two-dimensional grayscale images (the horizontal axis is time and the vertical axis is the temperature value), extracts local spatio-temporal patterns through a convolutional network, and finally outputs the temperature prediction value through a fully connected layer. Due to the local perception characteristics of CNN, short-term correlations in temperature fluctuations are captured, but there are significant defects: the image processing destroys the temporal continuity of the original data, resulting in the breakage of the causal chain between steam valve actions and temperature responses; the fixed receptive field of the convolutional kernel cannot model the long-term heat delay effect, and the long-term prediction trend error is relatively large. Through a time series fusion network architecture, the present invention realizes a cascaded fusion model of residual TCN time series convolution + SK-Net multi-scale attention + BiLSTM bidirectional recurrent network. Through the BiLSTM bidirectional recurrent network, the input sequence of the backward LSTM layer is time-reversed so that it processes process data synchronously with the forward LSTM. By splicing the double hidden state vectors, two-way causal association is completed to achieve more accurate prediction; and through a dynamic data repair and enhancement strategy, a hierarchical repair mechanism of short-term missing linear interpolation and long-term abnormal historical mean filling is adopted, and Gaussian noise is injected to enhance the anti-interference ability of the data. In a complex and changeable industrial environment, the method has high prediction accuracy and robustness for the outlet temperature of the feeding and conditioning equipment.

[0135] The preferred specific embodiments of the present invention have been described in detail above. It should be understood that those of ordinary skill in the art can make many modifications and variations based on the concept of the present invention without creative work. Therefore, all technical solutions that can be obtained by those skilled in the art in the technical field based on the concept of the present invention through logical analysis, reasoning, or limited experiments on the basis of the prior art should fall within the protection scope determined by the claims.

Claims

1. A temperature prediction method for a feeding and humidifying equipment based on a time series fusion network model, characterized in that, The method includes the following steps: S101: Collect time series data in the production environment of the feeding and humidifying equipment, and preprocess the time series data to generate a data set; S103: Divide the data set into a training set, a validation set, and a test set, and inject Gaussian noise into the training set, where the Gaussian noise simulates sensor errors; S105: Construct a prediction model for predicting the outlet temperature of the feeding and humidifying equipment, where the prediction model is configured as a temporal fusion network model, and the temporal fusion network model includes a residual TCN temporal convolutional network, an SK-Net multi-scale attention network, and a BiLSTM bidirectional recurrent network; S107: Pre-train the prediction model using the training set; S109: Use the trained prediction model to predict the outlet temperature of the feeding and humidifying equipment; S111: Evaluate the prediction results of the prediction model. If the evaluation index is greater than the threshold, start an incremental training process to re-train the prediction model.

2. The method according to claim 1, wherein In the step S101, the time series data is three-dimensional temporal tensor data, including the number of samples, the time step, and the feature dimension. The feature dimension includes the actual value of the outlet temperature, the outlet moisture content, the actual value of the process steam flow, the water addition flow, and the inlet moisture.

3. The method according to claim 2, wherein In the step S101, the preprocessing includes missing data filling, abnormal data filling, and dynamic normalization, where: For the missing data filling, for data with multiple consecutive missing time steps, linear interpolation is used for filling: Among them, is the missing length, is the time step, is the data of the time step, is the data of the time step, is the data of the time step; For the abnormal data filling, for persistent abnormal data, the historical database is called to retrieve the average value in the recent period under the same working conditions to fill the abnormal time period, and an abnormal flag bit is added; For the dynamic normalization, normalization is performed separately according to the feature dimension: in, is the data before normalization, is the normalized data, For the The minimum value of the dimension feature in the training set, For the The maximum value of the dimension feature in the training set.

4. The method according to claim 3, characterized in that, In the step S105, the residual TCN temporal convolutional network effectively captures long-term dependencies through causal convolution and dilated convolution, including the following steps: S10511: TCN preprocessing, mapping the input feature dimension to the internal channels of the TCN through convolution for the preprocessed time series data; S10512: Residual block level processing, processing the time series data after TCN preprocessing through dilated causal convolution to obtain the first layer output tensor; S10513: Pass the first layer output tensor through weight normalization and the ReLU activation function; S10514: Add the main path output to the original input to obtain the first layer processing result; S10514: Loop processing, use the first layer processing result as the input of the next layer, continuously process multiple times to obtain the final result, and complete the channel number expansion of the final result through convolution; Among them, is time series data, , is time series data, is the batch size, is the time step, is the feature dimension, is the number of internal channels of the TCN, is the input data of the TCN, is the output tensor of the first layer, is the result of passing the output of the first layer through the ReLU activation function, is the processing result of the first layer, the processing result of the nth layer, is the final result processed by the residual TCN temporal convolutional network, is the one-dimensional convolution operation function.

5. The method according to claim 4, wherein In the step S105, the SK-Net multi-scale attention network is configured to achieve feature extraction and enhance the sensitivity of the model to key regions, including the following steps: S10521: Process the preprocessed time series data through double-branch convolution, output feature maps of two branches, and process the feature maps through the GELU activation function; S10522: Generate channel attention weights based on the feature maps of the two branches; S10523: Split the channel attention weights into two channel weight vectors, and perform a normalized exponential function operation on the weights of each channel to generate dynamic allocation coefficients; S10524: Broadcast and expand the generated dynamic allocation coefficients along the time step dimension, and perform weighted summation on the dual-branch feature maps to obtain a fused enhanced feature map; S10525: Align the dimensions of the preprocessed time series data through convolution, and add it to the fused enhanced feature map to obtain the finally output feature map.

6. The method according to claim 5, wherein In the step S10522, the generation of the channel attention weights includes the following sub-steps: S105221: Add the feature maps obtained by dual-branch convolution element-wise to obtain a fused feature; S105222: Compress the spatial information of the fused feature along the time dimension to generate channel statistics; S105223: Compress the channel information of the channel statistics through a fully connected layer, and the activation function of the fully connected layer is the ReLU function; S105224: Restore the processed data to the original number of channels to generate a compact weight vector; Among them, is the channel statistic, is the fused feature, is the compact weight vector, is the relu function, is the batch normalization process, is the global average pooling function, is the fully connected layer, is the spatial height of the feature map, is the spatial width of the feature map, 、 are the horizontal and vertical coordinate indices of the spatial position.

7. The method according to claim 6, wherein In the step S10523, the dynamic allocation coefficients are calculated using the following formula: Among them, is the channel weight, is the channel weight vector, is the control parameter for dynamically adjusting the channel weight allocation ratio.

8. The method according to claim 7, wherein In the step S105, the BiLSTM bidirectional recurrent network reverses the time sequence of the input sequence of the backward LSTM layer, and synchronously processes the process data with the forward LSTM. By splicing the dual hidden state vectors, bidirectional causal association is completed to achieve more accurate prediction.

9. The method according to claim 8, characterized in that In the step S105, the BiLSTM bidirectional recurrent network includes the following steps when working: S10531: The forward LSTM updates the hidden state and outputs the final sequence in the following way: The final output sequence of the forward LSTM is as follows: ; S10532: The backward LSTM updates the hidden state and outputs the final sequence using the following formula: The final output sequence of the backward LSTM is as follows: ; S10533: Bi-directionally splice the output results of the forward LSTM and the backward LSTM to obtain a result : S10533: Perform one more inference according to the above steps to obtain S10534: Map the obtained output to the predicted target dimension through a fully connected layer to obtain the final prediction result: Wherein: is the forward LSTM, is the backward LSTM. is the concatenated feature, is to predict the future time step, is the variable predicted at each time step, is the time step, is the hidden state of the first layer of the forward LSTM at time step t, is the cell state of the first layer of the forward LSTM at time step t, is the hidden state of the first layer of the backward LSTM at time step t, is the cell state of the first layer of the backward LSTM at time step t, is the bidirectional concatenation result of the first layer of BiLSTM, is the bidirectional concatenation result of the second layer of BiLSTM. is the bias term of the fully connected layer, is the weight matrix of the fully connected layer.

10. The method according to claim 9, wherein When pre-training the prediction model in the step S107, the following loss function is used to calculate the Huber loss: Among them, is the Huber loss, is the predicted value, is the actual value, is the threshold coefficient of the Huber loss.

Images