Real-time prediction and management and control system and method for cut tobacco production waste raw materials

Abstract

The invention relates to the technical field of tobacco processing, and discloses a real-time prediction and management and control system and method for cut tobacco production waste raw materials, and the method comprises the following steps: S1, data collection: collecting key process production data of each production batch of a cut tobacco production workshop in real time; s2, data preprocessing: carrying out standardization processing on the collected process data and environmental parameters to form a model input vector; s3, model prediction: inputting the preprocessed data into a pre-trained waste raw material prediction model, and outputting a current batch of waste raw material prediction value; s4, deviation analysis and early warning: calculating a deviation value between the predicted value of the waste raw materials and the actually collected output quantity of the waste raw materials, and when the deviation value exceeds a preset threshold value, triggering an early warning signal and outputting early warning information; and S5, optimization and feedback: carrying out iteration on the waste raw material prediction model based on the output early warning information and the prediction result. And the amount of the waste raw materials in each batch can be dynamically and accurately predicted.

Description

Technical Field

[0001] This application relates to the field of tobacco production technology, and to a real-time prediction and control system and method for waste materials in tobacco processing. Background Technology

[0002] As the cigarette manufacturing industry continues to develop towards intelligence and precision, the management of waste materials in tobacco processing workshops has become a key aspect of improving resource utilization and reducing production costs. Currently, tobacco processing workshops generally adopt continuous and assembly-line production models.

[0003] In existing technologies, the management of waste materials largely relies on post-event statistics and human experience judgment, lacking real-time and predictive capabilities. Specifically, existing technologies cannot accurately predict the output of waste materials in real time, resulting in lagging management and control. They also lack an intelligent management and control closed loop that can learn and optimize itself based on real-time data, leading to adjustments in management strategies often relying on delayed human intervention. Summary of the Invention

[0004] This application provides a real-time prediction and control system and method for waste materials in silk production, aiming to solve the above-mentioned problems.

[0005] One approach provides a method for real-time prediction and control of waste materials from silk production, including the following steps:

[0006] S1. Data Acquisition: Real-time acquisition of key process production data for each production batch in the yarn-making workshop;

[0007] S2. Data preprocessing: The collected process data and environmental parameters are standardized to form a model input vector;

[0008] S3. Model Prediction: Input the preprocessed data into the pre-trained waste material prediction model and output the predicted value of the waste material for the current batch;

[0009] S4. Deviation Analysis and Early Warning: Calculate the deviation between the predicted value of the waste raw materials and the actual amount of waste raw materials produced. When the deviation exceeds a preset threshold, trigger an early warning signal and output early warning information and prediction results.

[0010] S5. Optimization and Feedback: Iterate the waste material prediction model based on the early warning information and the prediction results.

[0011] Key process production data include processes such as rehydration, drying, flavoring, and steaming of stems. The silk-making workshop of this application includes a 5T main line, a 2T stem line, and a 3T expanded silk line.

[0012] The key processes of the 5T main line include leaf fiber loosening and rehydration, leaf fiber drying, and blending and flavoring. The key processes of the 2T stem line include at least stem steaming and stem flavoring. The key processes of the 3T expanded yarn line include at least expanded yarn loosening and rehydration and expanded yarn drying. The system collects the cumulative actual amount of inlet material and the average inlet moisture content at each key process point, as well as the cumulative actual amount of outlet material and the average outlet moisture content, and converts them into dry matter.

[0013] After step S2 is completed, the production process is dynamically divided into multiple stable working conditions using an adaptive clustering algorithm (K-means algorithm), and a unique feature subset is selected for each working condition to form an input vector.

[0014] Specifically, the K-means algorithm is used, and the optimal number of working conditions is determined by the silhouette coefficient, while invalid clusters with a sample size lower than a preset threshold are removed.

[0015] Specifically, in step S2, the specific feature selection is performed independently for each working condition. The random forest algorithm is used to calculate the importance of features, and the top k features with a cumulative contribution rate of 95% are selected to form the feature subset for that working condition.

[0016] Specifically, in step S2, the collected process data and environmental parameters need to undergo health screening before standardization. This health screening includes dual filtering of the data based on both the equipment knowledge base and the process knowledge base, eliminating data samples that exceed the reasonable operating range of the equipment and the feasible range of the process.

[0017] In step S3, a waste material prediction model is constructed for each working condition. The waste material prediction model consists of a global model trained based on historical data and a local adaptive model dynamically fine-tuned based on recent sliding window data under the current working condition. The waste material prediction result under the working condition is output through adaptive weight integration.

[0018] Specifically, the adaptive weight integration described in step S3 involves dynamically calculating weights based on the prediction performance of the global and local models on recent data, with the model that performs better being assigned a higher weight.

[0019] Specifically, step S3 also includes an optimal model selection mechanism: the system saves the model with the best historical performance for each working condition, and selects to use the best model or the final trained model based on the model's historical performance during prediction.

[0020] The deviation analysis in step S4 is based on the predicted output and actual input of each process. It calculates the predicted losses at each stage according to the material balance principle, including:

[0021] 2T line loss: S2T_预测 =x 2T_实际 -y 2T_预测

[0022] 3T expansion wire loss: S 3T_预测 =x 3T_实际 -y 3T_预测

[0023] 5T main line moisture regain to yarn drying loss: S 5T_回潮-烘丝入口_预测 =x 5T_回潮_实际 -y 5T_烘丝入口_预测

[0024] 5T main line blending loss:

[0025] S 5T_掺配_预测 =(y 5T_烘丝_预测 +α×y 2T_预测 +β×y 3T_预测 )-y 5T_掺配_预测

[0026] 5T main line blending to fragrance loss: S 5T_掺配-加香_预测 =y 5T_加香_预测 -y 5T_掺配_预测

[0027] And aggregated to obtain the total loss prediction:

[0028] S 总量_预测 =S 5T_掺配-加香_预测 +S 5T_掺配_预测 +S 5T_回潮-烘丝入口_预测 +S 3T_预测 +S 2T_预测

[0029] The relative deviation between the predicted and actual total loss values ​​is calculated. When the deviation exceeds a threshold, a graded early warning is triggered, and abnormal processes are located by analyzing the deviation of each sub-item of loss.

[0030] In step S4, when the deviation value exceeds a preset threshold, a warning signal is triggered, and warning information is output. Specifically, the warning information is a tiered warning, including:

[0031] Level 1 warning: A system-level alarm is triggered when the relative deviation of total loss exceeds the threshold.

[0032] Level 2 warning: Automatically analyze the loss prediction deviation of each process and identify the process with the largest deviation rate as the source of the anomaly;

[0033] Level 3 Early Warning: Generates a detailed early warning report containing abnormal processes, possible causes, and handling suggestions by combining the process knowledge base.

[0034] In step S5, the waste material prediction model is iterated based on the early warning information and prediction results, including:

[0035] When the prediction error increases slightly, online fine-tuning of the local model is triggered;

[0036] When the prediction error continues to worsen, adjust the integration weights of the global model and the local model;

[0037] When performance degrades significantly, it triggers a re-partitioning of operating conditions and model reconstruction.

[0038] In one scheme, in step S1, the key process production data includes: basic production data, material cumulative data, moisture parameter data, process control data, and equipment operation data; the environmental parameters include: workshop ambient temperature and workshop ambient humidity.

[0039] In one approach, the waste material prediction model is a BP neural network model, and the structural parameters of the BP neural network model are adjusted using a BO-PSO hybrid algorithm.

[0040] In one embodiment, in step S3, the BO-PSO hybrid optimization algorithm is used to jointly optimize the number of hidden layer neurons, learning rate, and momentum coefficient of the BP neural network model.

[0041] In one embodiment, the number of input layer neurons in the BP neural network model is consistent with the number of process data and environmental parameters, the hidden layer of the BP neural network model is at least one layer, the search space for the number of hidden layer neurons in the BP neural network model is [10, 200], and the number of output layer neurons in the BP neural network model is consistent with the dimension of the output to be predicted.

[0042] In one scheme, in step S4, when the preset threshold is greater than 5%, an early warning is triggered and an early warning message is output.

[0043] In one scheme, the process data and environmental parameters of each production batch in the silk-making workshop include multiple production conditions. After step S2 is completed, the multiple production conditions are classified into corresponding specific conditions by an adaptive clustering algorithm. The input data is projected using the exclusive feature subset corresponding to the condition, and the processed data is input into a pre-trained waste raw material prediction model to output the waste raw material prediction value.

[0044] Specifically, during online prediction, after completing data preprocessing step S2, the real-time production data is first divided into corresponding specific working condition categories using an adaptive clustering algorithm. Then, a subset of features pre-selected for that working condition during training is used to project features onto the input data. Subsequently, the processed data is simultaneously input into both a pre-trained and saved global base model and a local adaptive model specific to that working condition, obtaining global and local model predictions respectively. Based on the weight fusion strategy corresponding to that working condition, and considering the performance of the two models on recent data for that working condition, the system dynamically calculates the fusion weights of the global and local predictions. Finally, based on the obtained fusion weights, a weighted fusion calculation is performed on the global and local model predictions to output the predicted waste material result for that working condition.

[0045] In one approach, after an early warning is triggered, the waste material prediction model locates the corresponding process segment where the deviation occurred and generates an early warning report.

[0046] In one solution, another aspect provides a real-time prediction and control system for waste materials from silk production, used to implement real-time prediction and control methods, including:

[0047] The data acquisition module can acquire process data and environmental parameters from the production execution system and environmental sensors;

[0048] A data preprocessing module, electrically connected to the data acquisition module, is used to clean and standardize the process data and environmental parameters. A prediction model module, electrically connected to the data preprocessing module, receives the preprocessed process data and environmental parameters transmitted by the data preprocessing module and outputs predicted values ​​for waste materials.

[0049] The early warning module is electrically connected to the prediction model module. The early warning module can obtain the deviation value between the predicted value of waste raw materials and the actual amount of waste raw materials produced. When the deviation value exceeds a preset threshold, an early warning signal is output.

[0050] The model optimization module retrains and updates the hyperparameters of the prediction model module based on the newly added process data, environmental parameters, predicted values ​​of waste raw materials, and early warning signals.

[0051] One solution also includes a human-computer interaction interface for visually displaying prediction results, early warning information, historical data tracing results, and model performance indicators.

[0052] Specifically, it also includes a model management module, which is used to maintain a record of the best historical performance model for each operating condition and to intelligently select the best model or the final model when making predictions.

[0053] The beneficial effects of this application are:

[0054] By constructing a dedicated waste material prediction model and inputting standardized real-time process data and environmental parameters, this invention can dynamically and accurately predict the amount of waste materials in each batch, changing the lag of traditional post-event statistics, moving the control point forward, and realizing the transformation from "passive response" to "proactive prediction".

[0055] By setting deviation thresholds and performing real-time comparisons, this invention can trigger early warning signals immediately when production deviations show signs of occurring. This early warning mechanism based on real-time deviation prediction shortens the time from the occurrence of an anomaly to its detection, creating conditions for rapid response and intervention by management personnel, and effectively preventing further expansion of batch losses.

[0056] By feeding back early warning information and prediction results to the model and iteratively optimizing it, this invention constructs a closed-loop intelligent management and control process of "prediction-early warning-feedback-optimization". This process enables the prediction model to continuously adapt to changes in production conditions, achieve self-learning and continuous improvement, thereby maintaining prediction accuracy and management effectiveness in the long term, and ultimately promoting the waste management of the silk-making workshop to a new stage of standardization, visualization and refinement. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is a schematic diagram of the BP neural network model structure in one embodiment of this application;

[0059] Figure 2 This is a diagram of a hybrid prediction model architecture based on dynamic weights in one embodiment of this application;

[0060] Figure 3 This is a schematic diagram of the process of three production lines in one embodiment of this application;

[0061] Figure 4 This is a flowchart of a real-time prediction and control method in one embodiment of this application; Detailed Implementation

[0062] The specific embodiments of this application will be further described in detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate this application, but are not intended to limit the scope of this application. Similarly, the following examples are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0063] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0064] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0065] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection, an electrical connection, or a connection that allows communication between them; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0066] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "over," and "on top" of the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0067] In this invention, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to a specific feature, structure, material, or characteristic described in connection with that embodiment or example, which is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0068] This application proposes improvements and innovations, and presents the following embodiments.

[0069] In some implementations, please refer to Figures 1 to 4 On the one hand, it provides a method for real-time prediction and control of waste materials in silk production, including the following steps:

[0070] S1. Data Acquisition: Real-time acquisition of process data and environmental parameters for each production batch in the yarn-making workshop;

[0071] S2. Data preprocessing: Standardize the collected process data and environmental parameters to form model input vectors;

[0072] S3. Model Prediction: Input the preprocessed data into the pre-trained waste material prediction model and output the predicted value of the waste material for the current batch;

[0073] S4. Deviation Analysis and Early Warning: Calculate the deviation between the predicted value of waste raw materials and the actual amount of waste raw materials produced. When the deviation exceeds the preset threshold, an early warning signal is triggered and early warning information is output.

[0074] S5. Results Output and Feedback: Iterate the waste material prediction model based on the output early warning information and prediction results.

[0075] By constructing a dedicated waste material prediction model and inputting standardized real-time process data and environmental parameters, this invention can dynamically and accurately predict the amount of waste materials in each batch, changing the lag of traditional post-event statistics, moving the control point forward, and realizing the transformation from "passive response" to "proactive prediction".

[0076] By setting deviation thresholds and performing real-time comparisons, this invention can trigger early warning signals immediately when production deviations show signs of occurring. This early warning mechanism based on real-time deviation prediction shortens the time from the occurrence of an anomaly to its detection, creating conditions for rapid response and intervention by management personnel, and effectively preventing further expansion of batch losses.

[0077] By feeding back early warning information and prediction results to the model and iteratively optimizing it, this invention constructs a closed-loop intelligent management and control process of "prediction-early warning-feedback-optimization". This process enables the prediction model to continuously adapt to changes in production conditions, achieve self-learning and continuous improvement, thereby maintaining prediction accuracy and management effectiveness in the long term, and ultimately promoting the waste management of the silk-making workshop to a new stage of standardization, visualization and refinement.

[0078] In step S4, when the deviation value exceeds a preset threshold, a warning signal is triggered, and warning information is output. Specifically, the warning information is a tiered warning, including:

[0079] Level 1 warning: A system-level alarm is triggered when the relative deviation of total loss exceeds the threshold.

[0080] Level 2 warning: Automatically analyze the loss prediction deviation of each process and identify the process with the largest deviation rate as the source of the anomaly;

[0081] Level 3 Early Warning: Generates a detailed early warning report containing abnormal processes, possible causes, and handling suggestions by combining the process knowledge base.

[0082] In step S5, the waste material prediction model is iterated based on the early warning information and prediction results, including:

[0083] When the prediction error increases slightly, online fine-tuning of the local model is triggered;

[0084] When the prediction error continues to worsen, adjust the integration weights of the global model and the local model;

[0085] When performance degrades significantly, it triggers a re-partitioning of operating conditions and model reconstruction.

[0086] By combining a hybrid model architecture with an online learning mechanism, this invention integrates historical patterns with real-time dynamics, ensuring both model stability and the ability to adapt to production changes. The system can intelligently select the optimal model based on performance and automatically trigger a self-healing mechanism when performance degradation is detected, achieving a complete intelligent control loop of "prediction-early warning-source tracing-optimization".

[0087] Specifically, the system's feedback optimization mechanism is based on early warning information and prediction results. The system triggers a multi-level adaptive optimization mechanism, which includes:

[0088] Online incremental learning: When the prediction error increases slightly, the system automatically triggers online fine-tuning of the local model. It uses a sliding window mechanism to retain the production data of the most recent N batches and performs rapid fine-tuning for 5 training cycles with a learning rate of 0.01, so as to achieve rapid adaptation to dynamic changes in the production process.

[0089] Dynamic weight adjustment: When the prediction error continues to worsen, the system re-evaluates the prediction performance of the global model and the local model on recent data, and dynamically adjusts the integrated weight allocation of the two based on the mean absolute error index, increasing the weight ratio of the model with higher prediction accuracy.

[0090] Operating condition reconfiguration mechanism: When the model performance degrades significantly, the system determines that the current operating condition division may no longer be applicable, automatically triggers the operating condition reconfiguration process, re-executes cluster analysis, and establishes a new operating condition-model mapping relationship, thereby realizing the system's self-healing capability.

[0091] Best model management: The system maintains a archive of historical best performance models for each working condition, continuously tracks the performance of each model at different production stages, and intelligently selects to use the historical best model or the latest trained model when making predictions, ensuring continuous optimization of prediction accuracy.

[0092] The actual amount of waste material is calculated as follows:

[0093] Using a fixed input ratio production function, the total output coefficient is calculated based on the input coefficient of key processes to obtain the actual amount of waste raw materials, as shown in the following formula.

[0094]

[0095] Where: Q represents output, L and K represent the input of labor and capital respectively, and U and V represent the fixed production technology coefficients of labor and capital respectively. The fixed production technology coefficients of labor and capital are fixed values ​​that have been determined by the process department during the project design or process planning stage.

[0096] The steps for data collection are as follows:

[0097] Key production process data from three main production lines (5T main line, 3T expanded yarn line, and 2T stem yarn line) in the yarn-making workshop were collected via data interfaces to construct a multi-dimensional time-series feature dataset based on historical production batches. The specific data collected includes:

[0098] (1) Basic production data: batch number, brand, production line number, and production timestamp;

[0099] (2) Cumulative material data: cumulative standard value of incoming materials, cumulative allowable deviation of incoming materials, cumulative actual quantity of incoming materials, cumulative offset of incoming materials, cumulative process weight offset of incoming materials, cumulative plant-level standard value of outgoing materials, and cumulative actual quantity of outgoing materials;

[0100] (3) Moisture parameter data: inlet moisture average, inlet moisture standard deviation, inlet moisture coefficient of variation, outlet moisture standard value, outlet moisture cumulative amount, outlet moisture average, outlet moisture standard deviation, outlet moisture coefficient of variation, outlet material moisture qualification rate;

[0101] (4) Process control data: Standard value of inlet hot air temperature, average value of inlet hot air temperature, standard deviation of inlet hot air temperature, coefficient of variation of inlet hot air temperature, set value of circulating hot air temperature, average value of circulating hot air temperature, standard deviation of circulating hot air temperature, qualified rate of circulating hot air temperature, maximum value of inlet material temperature, minimum value of inlet material temperature, average value of inlet material temperature, standard value of outlet material temperature, average value of outlet material temperature, standard deviation of outlet material temperature; standard value of steam injection ratio, cumulative amount of steam injection ratio, average value of steam injection ratio, standard deviation of steam injection ratio, coefficient of variation of steam injection ratio;

[0102] (5) Equipment operation data: hot air fan frequency standard value, hot air fan frequency actual set value, hot air fan frequency actual average value, circulating fan frequency standard value, circulating fan frequency actual set value, circulating fan frequency actual average value, dehumidification damper opening standard value, dehumidification damper opening actual set value, dehumidification damper opening actual average value, circulating damper opening standard value, circulating damper opening actual set value, circulating damper opening actual average value, drum speed / frequency standard value, drum speed / frequency actual set value, drum speed / frequency actual average value;

[0103] (6) Environmental parameters: Environmental monitoring data such as workshop ambient temperature and workshop ambient humidity.

[0104] The data acquisition system interfaces with the production execution system (MES system) to ensure high-frequency acquisition and synchronous transmission of information, constructing a complete dataset of the silk production process, and providing data support for subsequent model training and dynamic prediction.

[0105] The waste material prediction model is trained as follows:

[0106] A prediction model is constructed using the BP neural network algorithm. The model structure is as follows:

[0107] (1) The input layer contains k neurons, corresponding to the 6 types of feature parameters collected;

[0108] (2) The hidden layer can be set to one or more layers according to actual needs;

[0109] (3) The output layer has j neurons, which output the predicted output of each process.

[0110] The structure of a network is determined by its basic processing units and how they are interconnected, and the connection mechanism of these processing units is similar to that of neurons in a nervous system.

[0111] Each neuron is activated by a function that performs a weighted sum of all received inputs to determine whether the neuron is activated. The weighted sum is s. j As shown in the formula below:

[0112]

[0113] Where, x k (k = 1, 2, ..., K) represents multiple input variables of the neuron; w kj b represents the connection weight from the k-th input to the j-th neuron. j (j = 1, 2, ..., J) represents the bias term for the j-th neuron.

[0114] Neuron output y j By using the activation function f(·) on s j It is obtained by performing a nonlinear transformation.

[0115] y j =fs j

[0116] The choice of activation function depends on the specific requirements:

[0117] The hidden layer uses the ReLU function: f(·) = max(0,s) to avoid the gradient vanishing problem.

[0118] The output layer uses a linear activation function to ensure that the output range is unrestricted and adapts to the numerical range requirements of yield prediction.

[0119] The implementation of a backpropagation (BP) network consists of two steps: forward propagation and backward propagation. In the forward propagation stage, the input data is processed by weighted summations and activation functions of neurons in each layer, ultimately yielding a predicted value at the output layer. In the backward propagation stage, based on the error E between the predicted and actual values, the network weights and thresholds are adjusted layer by layer through backpropagation of the error gradient to reduce the error between the predicted and target values, gradually bringing the error to the required accuracy. The computational error for a single learning sample is shown in the following formula.

[0120] E(n) = [y(n) - Y(n)] 2

[0121] Where n = 1, 2, 3, ..., N (N is the total number of samples); E(n) is the error between the predicted value and the actual value of the nth sample; y(n) is the calculated predicted value of the nth sample; Y(n) is the actual output value of the nth sample.

[0122] The training objective function of a BP neural network is based on the mean squared error of all samples, and the calculation formula is as follows:

[0123]

[0124] Among them, E MSE This represents the mean squared error between the predicted and actual values ​​for all samples.

[0125] The PSO optimization process is as follows:

[0126] Particle swarm optimization (PSO) is based on the idea of ​​cooperative search by swarms, simulating the movement of particles in the search space to find the optimal solution. The position x of each particle... i This represents a combination of hyperparameters, where velocity v is... i Indicates the update direction, updated according to the following formula:

[0127] Speed ​​updates:

[0128]

[0129] Location update:

[0130]

[0131] Where ω is the inertia weight, balancing local and global search; c1 and c2 are learning factors (generally ranging from 1.5 to 2.0); r1 and r2 are random numbers with mean distribution, ranging from [0,1]; p i is the historical best position of particle i; g is the global best position of all particles.

[0132] The BO optimization process is as follows:

[0133] Bayesian optimization simulates the objective function by constructing a surrogate model (such as a Gaussian process), as shown below:

[0134] In each iteration, new sampling points are selected based on the Expected Improvement (EI) criterion, as shown below:

[0135] EI(x)=(μ(x)-f best -ξ)Φ(Z)+σ(x)φ(Z)

[0136]

[0137] Where μ(x) and σ(x) represent the predicted mean and standard deviation of the surrogate model; f best ξ is the current best observation; ξ is the exploration factor (a small positive number); Φ(Z) and φ(Z) are the cumulative distribution function and probability density function of the standard normal distribution.

[0138] BO-PSO hybrid optimization strategy:

[0139] This method uses PSO for global search to quickly explore the hyperparameter space, while Bayesian optimization is used to refine the local search after particle convergence, thereby improving the model's performance stability and robustness.

[0140] Both global search methods for PSO and BO optimization processes are available on the market.

[0141] Optimization objectives and constraints:

[0142] The final optimization objective is to minimize the full sample mean squared error (MSE), calculated as follows:

[0143]

[0144] Where θ represents the hyperparameter vector to be optimized; y(n; θ) is the predicted value of the neural network based on the current parameters on the nth sample; constraints include: number of neurons ∈ [10, 200], learning rate ∈ [0.0001, 0.1], etc.

[0145] The system inputs key production process data for the current batch into a pre-trained neural network prediction system in real time. The system first automatically identifies the production batch type (large or small batch) based on the cumulative actual amount of incoming materials, and then classifies the identified batch data into specific operating conditions. It then uses a subset of features pre-selected for that specific operating condition during training to project features onto the input data. The processed data is simultaneously input into a pre-trained and saved global base model and a local adaptive model specific to that operating condition, obtaining global and local model predictions respectively. Based on the weighted integration strategy corresponding to that operating condition, and considering the performance of the two models on recent data for that condition, the system dynamically calculates the fusion weights of the global and local predictions. Finally, based on the obtained fusion weights, the global and local model predictions are weighted and fused to output the predicted waste material result for that operating condition.

[0146] The layered calculation for losses is as follows:

[0147] The silk-making workshop has three production lines operating as follows:

[0148] The expanded tobacco produced by the 3T expanded tobacco line and the stem tobacco produced by the 2T stem line are mixed with the main leaf tobacco in the blending process of the 5T main line according to the production process standard ratio. Then, the tobacco is flavored in the flavoring process of the 5T leaf tobacco main line to form the finished tobacco.

[0149] 2T wire loss prediction value S 2T_预测 Calculated using the following formula:

[0150] S 2T_预测 =x 2T_实际 -y 2T_预测

[0151] Among them, y 2T_纯丝预测 Forecast output of 2T stem line; x 2T_实际 This represents the actual amount of raw materials input for a 2T wire rod.

[0152] Predicted loss value S of 3T expansion wire 3T_预测 Calculated using the following formula:

[0153] S 3T_预测 =x 3T_实际 -y 3T_预测

[0154] Among them, y 3T_预测 Forecast output of 3T expanded yarn; x 3T_实际 This represents the actual amount of raw materials input for 3T expansion yarn.

[0155] Predicted loss value S from the 5T main line re-moistening process to the yarn drying process 5T_回潮-烘丝_预测 Calculated using the following formula:

[0156] S 5T_回潮-烘丝入口_预测 =x 5T_回潮_实际 -y 5T_烘丝_预测

[0157] Among them, y 5T_烘丝_预测 The predicted output value for the 5T main wire drying process; x 5T_回潮_实际 This represents the actual amount of raw materials input for the loosening and rehydration process of the 5T main line.

[0158] Loss prediction value S of the 5T main line blending process 5T_掺配_预测 Calculated using the following formula:

[0159] S 5T_掺配_预测 =(y 5T_烘丝_预测 +α×y 2T_预测 +β×y 3T_预测 )-y 5T_掺配_预测 Among them, y 5T_掺配_预测 α represents the predicted output of the 5T main wire blending process; α represents the blending ratio of 2T stem wire, and β represents the blending ratio of 3T expanded wire. The values ​​are based on the process production standards.

[0160] Predicted loss value S from blending process to flavoring process on the 5T main line 5T_掺配-加香_预测 Calculated using the following formula:

[0161] S 5T_掺配-加香_预测 =y 5T_加香_预测 -y 5T_掺配_预测

[0162] Among them, y 5T_加香_预测 The predicted output value for the fragrance addition process on the 5T main production line.

[0163] In summary, the total predicted loss for the three lines is S.总量_预测 .

[0164] S 总量_预测 =S 5T_掺配-加香_预测 +S 5T_掺配_预测 +S 5T_回潮-烘丝入口_预测 +S 3T_预测 +S 2T_预测 ;

[0165] By applying the principles of layered loss calculation and material balance, this invention achieves precise location of abnormal processes from total loss early warning to specific abnormal processes, and constructs a multi-level traceability mechanism of "total loss anomaly → sub-item loss analysis → process-level location", providing production managers with accurate abnormal diagnosis basis.

[0166] The deviation analysis and early warning are explained as follows:

[0167] The system automatically triggers intelligent early warnings by calculating the relative deviation between the predicted and actual total losses in real time. The formula for calculating the relative deviation is as follows:

[0168]

[0169] The warning threshold Δ is set at 5% based on multiple loss test data from three production lines (5T main line, 3T expanded yarn line, and 2T stem yarn line) in the yarn making workshop, taking into account the actual historical loss (actual historical loss = actual input - actual output), production load fluctuations, and target control accuracy requirements.

[0170] The system adopts a three-level progressive early warning mechanism:

[0171] Level 1 Early Warning (System-level Monitoring): When the relative deviation of total loss exceeds the 5% threshold, a system-level alarm is immediately triggered to notify production management personnel to pay attention to the abnormal overall loss situation.

[0172] Level 2 Early Warning (Process-Level Traceability): The system automatically analyzes the loss prediction deviation of each sub-process. By comparing the actual and predicted deviation rates of each process in the 2T filament, 3T expansion wire, and 5T main line, the system accurately identifies the process with the largest deviation rate as the main source of anomaly.

[0173] Level 3 Early Warning (Intelligent Diagnosis and Handling): Combining the built-in process knowledge base, the system automatically generates detailed early warning reports containing key information such as timestamps, production batches, raw material grades, production line numbers, and abnormal process segments. Based on historical anomaly cases and process rules, it intelligently recommends possible causes of anomalies and corresponding handling suggestions, providing decision support for managers.

[0174] The reverse analysis steps of the system are as follows:

[0175] The system not only sets a threshold (such as 5%) for the total deviation, but also sets an internal threshold (which can be 5%, or a more finely set value) for each sub-deviation (the deviation corresponding to the amount of waste for each production line or even a key process).

[0176] The system checks which component's relative deviation first exceeds its internal threshold. This process that first "crosses the line" is identified as the source of the anomaly. This method can quickly pinpoint the stage where abnormal fluctuations first occur.

[0177] The prediction results include information such as the predicted value of waste materials.

[0178] The purpose of standardization is to transform the raw, messy, and inconsistent production data from the data acquisition module into a clean, complete, and uniformly scaled dataset for subsequent neural network models to learn and predict effectively. Without this step, problems in the raw data will directly lead to inaccurate model predictions, training failures, or even complete ineffectiveness.

[0179] Standardization processing includes data cleaning, data standardization, and normalization. Standardization processing is a mature and existing data processing method.

[0180] Data cleaning primarily addresses errors, anomalies, and inconsistencies in data.

[0181] Missing value handling: Due to sensor malfunction, transmission interruption, or manual entry omissions, some data points may be empty (NULL or NaN). Fill them with the mean or median of the same feature column (such as "workshop temperature") under the same historical period or normal production batch.

[0182] Outlier handling: There may be outliers in the data that do not conform to the production logic, such as extremely high instantaneous temperatures or negative input quantities. These may be caused by sensor false alarms or recording errors.

[0183] Based on process knowledge, set reasonable upper and lower limits for each parameter (e.g., "moisture parameter data" must be between 5% and 30%), and values ​​outside the range are considered abnormal.

[0184] Z-Score normalization is used to scale input features of different dimensions to the same scale. Box plots (IQR method) or Z-Score (standard deviation method) are used to identify values ​​that deviate significantly from the data center. Identified outliers are treated as missing values ​​and imputed; or the record is directly removed.

[0185] Inconsistent data correction: Data logic contradictions, such as "output dry matter mass" being greater than "input dry matter mass", or inconsistent record of the same batch of grade information in different systems.

[0186] Logical validation is performed by writing business rule scripts, and then corrected after verification with the production management system (MES) or operators.

[0187] Standardization primarily addresses the issue of inconsistent data dimensions and scales.

[0188] The Z-Score normalization formula and the Min-Max normalization formula are used to scale input features of different dimensions to the same scale. Both the Z-Score normalization formula and the Min-Max normalization formula are commonly used formulas for data processing in the market.

[0189] Through a systematic data cleaning and standardization process, the problems of high noise, inconsistent dimensions, and logical errors commonly found in industrial field data were effectively solved. This step transforms the raw "dirty data" into a high-quality, uniformly scaled dataset, laying a reliable data foundation for the subsequent training of the PB neural network model. This fundamentally ensures the accuracy, stability, and reliability of the prediction model, avoiding model training failure or prediction distortion caused by data quality issues.

[0190] In one embodiment, in step S1, the key production process data includes: basic production data, material cumulative data, moisture parameter data, process control data, and equipment operation data; the environmental parameters include: workshop ambient temperature and workshop ambient humidity.

[0191] Specifically, key production process data can be obtained through the system provided by the corresponding production line.

[0192] By explicitly using key parameters such as basic production data, cumulative material data, moisture parameters, process control data, equipment operation data, and environmental parameters as model inputs, a clear physical correlation and process logic between the input features and the output of waste raw materials is ensured. This feature selection makes the patterns learned by the model more consistent with actual production, significantly improving the rationality and interpretability of the prediction results, and providing accurate decision-making basis for precise control.

[0193] In one embodiment, the waste material prediction model is a BP neural network model, and the structural parameters of the BP neural network model are adjusted by a BO-PSO hybrid algorithm.

[0194] A hybrid optimization algorithm, BO-PSO, is employed to automatically tune the structural parameters of a backpropagation (BP) neural network, overcoming the blindness, inefficiency, and subjectivity of traditional manual trial-and-error parameter tuning. This hybrid strategy combines the global fast search capability of PSO with the local fine-grained search capability of BO, enabling it to find the optimal combination of model hyperparameters with higher efficiency and less computational resources. This ensures that the final prediction model possesses excellent prediction accuracy and strong generalization ability.

[0195] The structural parameters of a BP neural network model include: the number of hidden layers, the number of neurons in the input layer, the number of neurons in the hidden layer, the learning rate, and the momentum coefficient.

[0196] The waste material prediction model is a BP neural network model trained based on historical production data, and its model structure parameters are optimized using the BO-PSO hybrid optimization algorithm.

[0197] The BP neural network is a multilayer feedforward neural network, short for "backpropagation neural network". It is one of the most widely used and mature neural network models (existing models).

[0198] In one embodiment, in step S3, the BO-PSO hybrid optimization algorithm is used to jointly optimize the number of hidden layer neurons, learning rate, and momentum coefficient of the BP neural network model.

[0199] The BO-PSO hybrid optimization algorithm is used to collaboratively optimize key hyperparameters of the BP neural network in the hybrid prediction model, including the number of hidden layer neurons, learning rate, and momentum coefficient. In this hybrid optimization framework, Bayesian optimization (BO) is responsible for efficient exploration in the global parameter space to quickly locate potential high-quality regions; while particle swarm optimization (PSO) performs fine-tuning within the high-quality regions identified by Bayesian optimization, obtaining the optimal hyperparameter combination through swarm intelligence search.

[0200] The BO-PSO hybrid algorithm combines the global search capability of Particle Swarm Optimization (PSO) with the efficient local search capability of Bayesian Optimization (BO). Its optimization objective is to find an optimal set of hyperparameters (such as [number of neurons, learning rate, momentum coefficient]) such that the prediction error (mean squared error, MSE) is minimized when the BP neural network is trained with these parameters.

[0201] Specifically, a detailed explanation of PSO global exploration is as follows:

[0202] 1. Initialize the particle swarm: Randomly generate a swarm of "particles". The position x of each particle is... i It's a combination of hyperparameters, such as [number of neurons = 50, learning rate = 0.01, momentum = 0.9]. Each particle also has a flight velocity v. i .

[0203] 2. Evaluate the particles: For each particle (i.e., each set of hyperparameters), initialize and train a backpropagation neural network, and then calculate its performance (e.g., MSE) on the validation set. This MSE value is the fitness of the particle.

[0204] 3. Update individual and global optimum: Each particle remembers the best position p it found. iThe entire particle swarm also remembers the optimal position g found by all the particles.

[0205] 4. Update particle state: According to formulas (7) and (8), each particle updates its velocity and position by combining its historical best pi and global best g, thereby moving and exploring in the hyperparameter space.

[0206] 5. Function: The PSO algorithm performs a broad global search at this stage, quickly scanning different regions of the hyperparameter space to locate potential "basins".

[0207] The detailed explanation of BO local mining is as follows:

[0208] 1. Switching condition: When the PSO search converges (e.g., the overall movement of the particle swarm becomes very small, or after a certain number of iterations), the algorithm switches to the BO stage.

[0209] 2. Constructing a surrogate model: The BO algorithm treats all evaluated {hyperparameter combinations -> MSE} from the previous PSO stage and this stage as observation data, and uses a probabilistic model (usually a Gaussian process GP) to fit this unknown "hyperparameter -> performance" function. This GP model can provide the performance prediction mean μ(x) and uncertainty σ(x) at any untested point x.

[0210] 3. Selecting the Next Evaluation Point: BO uses an acquisition function (e.g., Equation 10, to determine the next most worthwhile combination of hyperparameters to evaluate) to improve EI. The EI function balances "exploitation" and "exploration":

[0211] 4. Sample in regions where the model performs well in prediction (where μ(x) is small).

[0212] 5. Sample in regions of high uncertainty (where σ(x) is large) to reduce model uncertainty.

[0213] 6. Evaluation and Update: Evaluate the BP network with this selected combination of hyperparameters, then add this new data point to the observation set and update the surrogate model.

[0214] 7. Function: The BO stage performs a fine local search in the potential area found by PSO, which can find the precise optimal point with fewer evaluations.

[0215] 8. Final output: After the entire hybrid optimization process is completed, the output is the set of hyperparameters that minimizes the MSE of the BP neural network validation set throughout the entire search history.

[0216] In one embodiment, the number of input layer neurons in the BP neural network model is consistent with the number of process data and environmental parameters, the hidden layer of the BP neural network model is at least one layer, the search space for the number of hidden layer neurons in the BP neural network model is [10,200], and the number of output layer neurons in the BP neural network model is consistent with the dimension of the output to be predicted.

[0217] By strictly mapping the number of neurons in the input, hidden, and output layers to the actual problem (number of input features and number of prediction targets), and by setting a reasonable search space for hidden layer neurons, scientific guidance is provided for model structure design. This design ensures that the network has sufficient ability to fit complex nonlinear relationships, while effectively controlling model complexity and avoiding underfitting or overfitting caused by inappropriate structure, making the model more practical for engineering applications while maintaining performance.

[0218] Specifically, in the silk-making workshop, the relationship between the output of waste raw materials and parameters such as "basic production data, cumulative material data, moisture parameter data, process control data, equipment operation data, and environmental parameters" can never be a simple linear relationship.

[0219] The impact of moisture parameters on losses may not increase proportionally; it may have a dramatic effect within a certain humidity range, while leveling off in another range. This complex, tortuous, and non-uniform trend relationship is called a "nonlinear relationship."

[0220] If the performance (such as validation set MSE) of the first hidden layer still does not meet the requirements after optimization by the BO-PSO hybrid algorithm, then it is upgraded to a second hidden layer.

[0221] Adding a third or more layers should only be considered when the model performs poorly on the training set (underfitting) and cannot be improved by increasing the number of neurons in a single layer or extending the training time.

[0222] Specifically, the number of neurons in the output layer of the BP neural network model is consistent with the dimension of the output to be predicted. That is, the BP neural network model can not only predict the total waste, but also simultaneously and independently predict the waste of each production line, and even key processes, namely the loss of the 2T stem line, the loss of the 3T expanded yarn line, the loss from the 5T main line's rehydration to drying, the loss of the 5T main line's blending process, and the loss from the 5T main line's blending process to the flavoring process. Therefore, the dimension of the output to be predicted is five, and the corresponding number of neurons in the output layer of the BP neural network model is also five. The first neuron outputs S. 2T_预测 The value of . The second neuron outputs S. 3T_预测 The value of . The third neuron outputs S. 5T_烘丝入口_预测 The value of S. The fourth neuron outputs S. 5T_掺配_预测 The value of the fifth neuron is S.5T_加香_预测 .

[0223] The model can predict the output of all targets with a single forward propagation, eliminating the need to train a separate model for each target.

[0224] These output metrics are inherently correlated (e.g., losses in the preceding process affect those in the following process). By having a single model learn all the targets simultaneously, it can capture these intrinsic physical or logical connections and may predict more accurately than by training multiple independent models.

[0225] For the subsequent early warning and traceability modules, the predicted values ​​of all key processes can be obtained at once, and a comprehensive deviation analysis can be performed immediately to quickly locate which specific production line or process has an anomaly.

[0226] In one embodiment, in step S4, when the preset threshold exceeds 5%, an early warning is triggered and an early warning message is output. Setting the early warning threshold explicitly to 5% of the relative deviation provides a clear, unified, and operable quantitative standard for anomaly detection. This threshold is based on a comprehensive setting of historical loss data and production control accuracy, effectively capturing meaningful abnormal fluctuations and avoiding false alarms about normal production disturbances, while ensuring the sensitivity of the early warning and preventing missed alarms. This achieves the best balance between cost control and ensuring control effectiveness.

[0227] Specifically, the preset threshold is calculated as |(actual value - predicted value) / actual value| × 100%.

[0228] In one embodiment, the key process production data of each production batch in the silk-making workshop includes multiple production conditions. After step S2 is completed, the multiple production conditions are classified into corresponding specific conditions by an adaptive clustering algorithm. The input data is projected using the exclusive feature subset corresponding to the condition, and the processed data is input into a pre-trained waste raw material prediction model to output the waste raw material prediction value.

[0229] Specifically, the tobacco processing workshop produces different cigarettes according to market demand. Therefore, the size of the tobacco leaves and the required parameters for rehydration and drying are different. Each production of one type of cigarette is considered a production condition. Each production batch in the tobacco processing workshop produces multiple types of cigarettes. Therefore, the production data of the key processes in each production batch includes multiple production conditions.

[0230] During online prediction, after completing data preprocessing step S2, the real-time production data is first divided into corresponding specific working condition categories using an adaptive clustering algorithm. Then, a subset of features pre-selected for this working condition during training is used to project features onto the input data. Subsequently, the processed data is simultaneously input into both a pre-trained and saved global base model and a local adaptive model specific to this working condition, obtaining global and local model predictions respectively. Based on the weight fusion strategy corresponding to this working condition, and considering the performance of the two models on recent data for that working condition, the system dynamically calculates the fusion weights of the global and local predictions. Finally, based on the obtained fusion weights, a weighted fusion calculation is performed on the global and local model predictions to output the predicted waste material result for this working condition.

[0231] Explanation of "projecting features onto the input data using a subset of features pre-selected for this specific working condition during training":

[0232] For each operating condition, the system uses feature importance analysis to select a specific subset of features. This subset contains the most relevant and effective features for predicting waste materials under that operating condition.

[0233] Example: For condition A (high-end grade), its specific feature subset might be {raw material moisture content, drying temperature, ambient humidity}. For condition B (low-end grade), its specific feature subset might be {feed quantity, production line speed, workshop temperature}.

[0234] Perform the "projection" operation:

[0235] When a new production batch of data arrives, the system first determines which operating condition it belongs to (e.g., operating condition A).

[0236] Then, the system extracts only the data corresponding to the specific feature subset belonging to working condition A from the complete data vector of that batch.

[0237] This operation of "extracting and retaining a specific subset from the full set of features" is called "projection".

[0238] To put it figuratively, it means "projecting" or "mapping" a high-dimensional data point that may contain irrelevant features into a lower-dimensional subspace composed of key features.

[0239] The steps for dividing working conditions and selecting features are as follows:

[0240] The production process is dynamically divided into multiple stable operating conditions using an adaptive clustering algorithm. First, based on the key feature of the cumulative actual amount of incoming materials, large-batch and small-batch production patterns are automatically identified and separated. Then, within each batch type, a K-means clustering algorithm is used for refined operating condition division. The optimal number of operating conditions is determined by the silhouette coefficient, while invalid clusters with fewer than 10 samples are removed.

[0241] A unique feature subset is selected for each working condition. Feature selection is performed independently for each working condition. The random forest algorithm is used to calculate feature importance, and 100 decision trees are constructed using the TreeBagger function. The Out-of-Bag (OOB) error variation function is used to calculate the importance score of each input feature. The feature importance scores of all output variables are averaged to obtain the comprehensive importance score. The input features are sorted in descending order of comprehensive importance score, and the cumulative contribution rate is calculated. The top k features with a cumulative contribution rate of 95% are selected to form the feature subset for that working condition, where k ranges from 5 to 30 features.

[0242] Specifically, the training steps for the waste material prediction model are as follows:

[0243] A hybrid prediction model is constructed for each working condition. This model consists of a global base model trained on historical data and a local adaptive model dynamically fine-tuned based on online sliding window data. Both the global and local models employ a backpropagation (BP) neural network architecture. An adaptive weight ensemble strategy is used to output the predicted output values ​​for each process. This strategy dynamically allocates weights based on the prediction performance of the global and local models on recent data windows, using mean absolute error (MAE) as the core performance evaluation metric. Models with higher prediction accuracy are assigned higher ensemble weights.

[0244] 1. Model Architecture Design

[0245] (1) Global Basic Model

[0246] Function Description: The global model G is trained based on historical data from all operating conditions to extract the general patterns of the overall production process.

[0247] Purpose: This model is used to provide a stable predictive benchmark for all operating conditions, reflecting long-term trends and systemic patterns.

[0248] Model structure: Input layer (n features) → 2 hidden layers (ReLU activation) → Output layer (linear)

[0249] formula:

[0250]

[0251] in, Let G(·) be the predicted output of the global model in batch t, ​​and let G(·) be the mapping function of the global BP neural network structure. For the input feature matrix that includes all operating conditions in the batch, θ G The parameter set for the global model is updated as follows:

[0252]

[0253] Where, η G The learning rate for the global model. Let L(·) be the gradient of the loss function with respect to the parameters. This is the output feature matrix that includes all operating conditions in the batch.

[0254] Training method: Offline batch training + periodic retraining

[0255] (2) Local Adaptive Model

[0256] Functional description: Each specific working condition has a local model L, which focuses on the dynamic changes and short-term characteristics under that working condition.

[0257] Function: To capture recent dynamics and minor changes in the production process, enabling real-time fine-tuning of the global model.

[0258] Model structure: Consistent with the global model, but updated more frequently (fine-tuning rate n) L >n G ).

[0259] formula:

[0260]

[0261] in, L is the predicted output of the local model for batch t under the current operating conditions. c (·) represents the mapping function of the local BP neural network structure. Let θ be the input feature matrix for the batch under the current operating conditions. L,C The parameter set for the local model under the current operating conditions is as follows:

[0262] Using sliding window sample data Update according to the current operating conditions:

[0263]

[0264] Where: n L To improve local learning efficiency.

[0265] Online learning trigger mechanism: The online learning mechanism is dynamically triggered based on real-time monitoring indicators.

[0266]

[0267] Where, δ thr D is the error trigger threshold. limit This is a drift constraint for operating condition variables, used to measure the difference between the current operating condition input feature distribution and its historical reference distribution. This represents the model confidence level. The calculation is performed using the standardized Euclidean distance method, as shown in the following formula:

[0268]

[0269] Where n is the number of features. This represents the mean value of the i-th feature under the current batch operating condition C. This is a historical average. This is the historical reference standard deviation.

[0270] 2. Model Collaboration and Weight Allocation Mechanism

[0271] The final prediction output of the hybrid model It is the global model prediction output. and local model prediction output Weighted sum:

[0272]

[0273] in, and For adaptive weights of the global and local models, satisfying

[0274] By combining the prediction results of the global model and the local model, weights are dynamically allocated:

[0275]

[0276] Use exponential smoothing to avoid weight jitter:

[0277]

[0278] Where α is the smoothing coefficient, α∈[0.7,0.9]; The newly calculated weights for the global model.

[0279] The system also integrates an optimal model selection mechanism: it independently maintains a historical best-performing model archive for each operating condition, recording the performance data of each model at different production stages and the corresponding optimal weight configuration. During actual prediction, the system intelligently selects to use either the historical best model or the latest trained model based on the characteristics of the current production operating condition and historical performance data, ensuring an optimal balance between prediction accuracy and adaptability to the operating condition.

[0280] In one embodiment, after an early warning is triggered, the waste material prediction model locates the corresponding process segment where the deviation occurred and generates an early warning report.

[0281] Specifically, after an alert is triggered, the system uses reverse analysis of the prediction model to locate the key production line and specific process segment that caused the abnormal deviation, and generates an alert report that includes the batch number, brand, production line identifier, abnormal process, and timestamp.

[0282] After an alert is triggered, the system can quickly pinpoint the specific abnormal production line and process segment through reverse analysis using a model, and automatically generate an alert report containing key information. This represents a leap from "overall alarm" to "precise positioning," freeing managers from tedious root cause investigations, significantly shortening problem response time, enabling targeted interventions, and significantly improving the efficiency of handling production anomalies and the speed of management loop execution.

[0283] Specifically, in step S3, the waste raw material prediction model establishes sub-models for the 5T main line, 2T stem line and 3T expanded yarn line in the yarn making workshop, or sets corresponding output nodes in a unified model, so as to predict the output of waste raw materials by production line, batch and process.

[0284] Specifically, for the 5T main yarn, the predicted loss values ​​are further distinguished between the re-moistening to drying process, the blending process, and the blending to fragrance process. Based on the blending process ratio, the impact of the input of 2T stem yarn and 3T expanded yarn on the total loss is comprehensively calculated.

[0285] In one embodiment, another aspect provides a real-time prediction and control system for waste materials from silk production, used to implement a real-time prediction and control method, including:

[0286] The data acquisition module can acquire process data and environmental parameters from the production execution system and environmental sensors;

[0287] The data preprocessing module is electrically connected to the data acquisition module. The data preprocessing module can clean and standardize process data and environmental parameters.

[0288] The prediction model module is electrically connected to the data preprocessing module. The prediction model module can receive the preprocessed process data and environmental parameters transmitted by the data preprocessing module and output the predicted value of waste raw materials.

[0289] The early warning module is electrically connected to the prediction model module. The early warning module can obtain the deviation value between the predicted value of waste raw materials and the actual amount of waste raw materials produced. When the deviation value exceeds the preset threshold, an early warning signal is output.

[0290] The model optimization module retrains and updates the hyperparameters of the prediction model module based on the newly added process data, environmental parameters, predicted values ​​of waste raw materials, and early warning signals.

[0291] The Production Execution System (MES) is a dedicated system for the production line and is an existing system. Environmental sensors include those for collecting data on moisture content at the inlet and outlet of each process step, as well as ambient temperature and humidity in the workshop.

[0292] By materializing the method into a complete system encompassing modules such as data acquisition, preprocessing, prediction, early warning, and optimization, the process of predicting and managing waste materials has been automated, systematized, and integrated. Each module performs its specific function and works collaboratively, forming a highly efficient intelligent control loop. This not only significantly reduces the intensity of manual operations and the probability of human error but also enables the invention to be rapidly implemented as an independent, deployable solution within enterprises, enhancing the technology's practical value and scalability.

[0293] In one embodiment, a human-computer interaction interface is also included for visually displaying prediction results, early warning information, historical data tracing results, and model performance indicators.

[0294] By setting up a human-computer interaction interface, complex model prediction results, early warning information, and performance data are presented to managers in a visual and intuitive way. This greatly lowers the barrier to entry for the technology, enabling production managers without a professional algorithm background to easily understand and use the system's output for decision-making. This enhances the system's usability and interactivity, thereby promoting the practical application and effective empowerment of intelligent manufacturing technology at the workshop level.

[0295] The above are merely optional embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application. Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for real-time prediction and management of waste raw materials in a tobacco manufacturing process, characterized in that, Comprise the following steps: S1. Data acquisition: real-time acquisition of key process data of each production batch in the silk making workshop; S2. Data preprocessing: standardizing the collected process data and environmental parameters to form a model input vector; S3. Model prediction: input the preprocessed data into the pre-trained waste material prediction model to output the waste material prediction value of the current batch; S4. Deviation analysis and early warning: calculate the deviation value between the waste material prediction value and the actual waste material output, and when the deviation value exceeds the preset threshold, trigger an early warning signal and output early warning information and prediction results; S5. Optimization and feedback: based on the early warning information and the prediction results, the waste material prediction model is iterated.

2. The real-time prediction and control method of claim 1, wherein, In step S1, the key process data includes: basic production data, material cumulative data, moisture parameter data, process control data and equipment operation data; the environmental parameters include: workshop environmental temperature, workshop environmental humidity.

3. The real-time prediction and control method of claim 2, wherein, The waste material prediction model is a BP neural network model, and the structure parameters of the BP neural network model are adjusted by a BO-PSO hybrid algorithm.

4. The real-time prediction and control method of claim 3, wherein, In step S3, the BO-PSO hybrid optimization algorithm is used to jointly optimize the number of hidden layer neurons, learning rate and momentum coefficient of the BP neural network model.

5. The real-time prediction and control method of claim 4, wherein, The number of input layer neurons of the BP neural network model is consistent with the number of process data and environmental parameters, the hidden layer of the BP neural network model is at least one layer, the search space of the number of hidden layer neurons of the BP neural network model is [10, 200], and the number of output layer neurons of the BP neural network model is consistent with the dimension of the output to be predicted.

6. The real-time prediction and control method according to any one of claims 1-5, wherein, In step S4, when the preset threshold is greater than 5%, an early warning is triggered and early warning information is output.

7. The real-time prediction and control method of claim 6, wherein, The key process data of each production batch in the silk making workshop includes multiple production conditions, after step S2 is executed, the multiple production conditions are classified into corresponding specific conditions by an adaptive clustering algorithm, the input data is projected using the exclusive feature subset corresponding to each specific condition, the processed data is input into the pre-trained waste material prediction model, and the waste material prediction value is output.

8. The real-time prediction and control method of claim 6, wherein, After triggering the early warning, the waste material prediction model locates the corresponding process section that causes the deviation and generates an early warning report.

9. A real-time prediction and control system for tobacco primary material in a tobacco production waste, for implementing the real-time prediction and control method according to any one of claims 1-8, characterized in that, Comprise: A data acquisition module that can obtain process data and environmental parameters from a production execution system and an environmental sensor; A data preprocessing module electrically connected to the data acquisition module, which can clean and standardize the process data and environmental parameters; A prediction model module electrically connected to the data preprocessing module, which can receive the preprocessed process data and environmental parameters transmitted by the data preprocessing module and output a waste material prediction value; An early warning module is electrically connected with the prediction model module, the early warning module can acquire a deviation value between the waste raw material prediction value and the actual collected waste raw material output, and outputs an early warning signal when the deviation value exceeds a preset threshold value; A model optimization module re-trains and updates hyperparameters of the prediction model module according to the newly added process data and the environment parameters, the waste raw material prediction value and the early warning signal.

10. The real-time predictive and governance system of claim 9, wherein, A man-machine interaction interface is further included for visual display of the prediction result, early warning information, historical data tracing result and model performance index.

Images