Adhesive tape production process optimization control method based on industrial internet of things

Abstract

This invention discloses a method for optimizing and controlling the tape production process based on the Industrial Internet of Things (IIoT), belonging to the field of IIoT control technology for tape production. The method involves collecting real-time parameters such as temperature, pressure, speed, thickness, and tension through an IIoT sensor network on the production line to form a multi-dimensional time-series stream. This stream is then cleaned and aligned using multi-source asynchronous data, and features are extracted to obtain a process feature vector. The feature vector is input into a process state discrimination model to obtain state labels, and a set of associated potential process offset patterns is retrieved from a process knowledge graph. Using the offset patterns and a synchronization parameter matrix as input, an improved gradient boosting decision tree is used to predict the future evolution path of process parameters. Based on the deviation between the path and the optimal target interval, the adjustment amount is calculated through a process control quantity mapping network. This method achieves multi-source data synchronization, prior knowledge fusion, and time-series path prediction, improving the accuracy and stability of tape production process control.

Description

Technical Field

[0001] This invention belongs to the field of IoT control technology for tape production, specifically a tape production process optimization and control method based on industrial IoT. Background Technology

[0002] The tape production line relies on industrial IoT sensors to collect process parameters such as temperature, pressure, speed, thickness, and tension. These parameters are collected independently in a multi-dimensional time-series format, with sensor sampling frequencies asynchronous across different processes. Existing solutions only perform simple filtering on the time-series data, construct parameter representations through conventional feature extraction, use a general classification model to determine the process operating status, rely on traditional gradient boosting decision trees for parameter prediction, and calculate control adjustments using empirical formulas or conventional mapping relationships.

[0003] Multi-source asynchronous parameters suffer from timing misalignment and redundant noise, lacking synchronicity and completeness, and process characteristics cannot fully reflect the coupling relationship of multiple parameters. Process state discrimination results are not combined with prior process knowledge, failing to uncover potential deviation patterns. Traditional prediction models only output parameter values ​​at a single moment, failing to reflect continuous time-series evolution trends. Control quantity calculations are disconnected from future parameter deviations, resulting in significant adjustment lag and easy deviation of process parameters from the target range.

[0004] To address the issues of inaccurate timing of multi-source process data, lack of prior offset matching for state discrimination, lack of continuous paths for parameter prediction, and coarse calculation of control quantities, it is necessary to achieve synchronous alignment processing of multi-source data, match potential offset patterns with process knowledge graphs, predict parameter evolution paths by improving models, and calculate precise adjustment quantities based on dedicated networks. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; Therefore, this invention proposes a method for optimizing and controlling the tape production process based on the Industrial Internet of Things, including: Real-time process parameters, including temperature, pressure, speed, thickness, and tension, are collected from an industrial IoT sensor network deployed at each stage of the tape production line, forming a multi-dimensional process time sequence flow. Perform multi-source asynchronous data alignment and cleaning on the multi-dimensional process timeline to generate a synchronous process parameter matrix; Process feature extraction is performed on the synchronous process parameter matrix to obtain a multi-dimensional process feature vector; The multidimensional process feature vector is input into the trained process state discrimination model, and the current process state label is output. Based on the current process status label, a preset process knowledge graph is retrieved to obtain a set of potential process offset patterns associated with the current process status label; Based on the set of potential process offset patterns and the synchronous process parameter matrix, an improved gradient boosting decision tree algorithm is used to predict the future process parameter evolution path. The improved gradient boosting decision tree algorithm is based on the gradient boosting framework and is improved for the task of predicting the process parameter evolution path. Based on the deviation between the future process parameter evolution path and the preset optimal target range of process parameters, a set of process control adjustment quantities are calculated through a process control quantity mapping network.

[0006] Furthermore, multi-source asynchronous data alignment and cleaning are performed on the multi-dimensional process timeline to generate a synchronous process parameter matrix, including: Identify the timestamp sequences of sensor data from different sources in the multidimensional process timeline, using the data source with the highest sampling frequency as the reference time axis; For data sources with other sampling frequencies, a spline interpolation algorithm based on a time window is used to resample the data onto the reference time axis to achieve timestamp alignment. For the data that has been timestamped and aligned, a data cleaning process based on local outlier detection is applied to identify and remove outliers caused by instantaneous sensor anomalies. The cleaned, time-aligned multidimensional process data are arranged in chronological order to form the synchronous process parameter matrix, where each row of the matrix represents a sampling time and each column represents a process parameter.

[0007] Further, process feature extraction is performed on the synchronous process parameter matrix to obtain a multi-dimensional process feature vector, including: In the time domain, the statistical characteristics of each process parameter column in the synchronous process parameter matrix are calculated, including the mean, standard deviation, skewness, and kurtosis. In the frequency domain, a fast Fourier transform is performed on the time series of each process parameter column in the synchronous process parameter matrix to extract the amplitude and frequency of the main spectral components as frequency domain features. At the state transition level, the rate of change of process parameters between adjacent sampling times is calculated, and the number of times the rate of change exceeds the threshold within a fixed time window is counted as dynamic features. The time-domain statistical features, frequency-domain features, and state dynamic features are spliced ​​and combined to form a comprehensive feature set that characterizes the current process operation state, namely the multidimensional process feature vector.

[0008] Furthermore, the multi-dimensional process feature vector is input into the trained process state discrimination model, and the current process state label is output, including: The trained process state discrimination model is a deep neural network classifier that has been trained in advance using historical process data. The multidimensional process feature vector is input into the input layer of the deep neural network classifier; The deep neural network classifier includes multiple fully connected layers and non-linear activation layers. The multi-dimensional process feature vector undergoes layer-by-layer non-linear transformation and feature abstraction through the fully connected layers and non-linear activation layers. In the output layer of the deep neural network classifier, the probability that the multidimensional process feature vector belongs to each preset process state category is calculated by the softmax function; The preset process state category with the highest probability is output as the model's discrimination result, which is the current process state label.

[0009] Further, based on the current process state label, a preset process knowledge graph is retrieved to obtain a set of potential process offset patterns associated with the current process state label, including: The preset process knowledge graph is stored in a graph structure, where nodes represent different process states or abnormal patterns of process parameters, and edges represent the relationships or causal logic between nodes. Using the current process status label as the starting node, a multi-hop breadth-first traversal is performed in the process knowledge graph; During the traversal, all nodes that characterize abnormal patterns of process parameters and are directly or indirectly connected to the starting node through causal edges are collected to form a relevant node set; The attribute information of each process parameter abnormality mode node is extracted from the relevant node set. The attribute information includes the typical parameter change direction and magnitude triggered by the process parameter abnormality mode node, forming the potential process offset mode set.

[0010] Furthermore, the working principle of the improved gradient boosting decision tree algorithm includes: When constructing the decision tree in each iteration, not only are the original features of the process parameters used, but the set of potential process offset patterns is also encoded as additional guiding features and added to the candidate feature set of the split node. When calculating the optimal split point for a split node, the standard mean square error loss is combined with a process path smoothness regularization term, which penalizes splitting patterns where the predicted values ​​fluctuate drastically between adjacent time points. In the calculation of the output value of the leaf node of the decision tree, not only is the negative gradient direction considered, but also an attraction term based on the historical optimal process path is introduced, so that the output value moves closer to the historical stable process region while following gradient descent. After multiple iterations, the prediction results of each round of decision trees are weighted and accumulated to obtain the final sequence of predicted process parameters, which is the future process parameter evolution path.

[0011] Furthermore, the specific construction method of the process path smoothness regularization term is as follows: Calculate the adjacent differences in the time series of the predicted process parameters of the samples in the left and right child nodes after the sample division at the current split point. The sum of squares of the adjacent differences is calculated for the left and right child nodes respectively, in order to measure the degree of fluctuation in the predicted path caused by the splitting method; The weighted sum of the predicted path fluctuations of the left and right child nodes is used as the smoothness penalty term corresponding to the split point and incorporated into the overall loss calculation of the split point selection. This constrains the decision tree split to tend to produce a smooth process evolution path.

[0012] Furthermore, based on the deviation between the future process parameter evolution path and the preset optimal target range of process parameters, a set of process control adjustment quantities is calculated through a process control quantity mapping network, including: The future process parameter evolution path is compared with the preset optimal target range of process parameters, and the difference sequence between the predicted value of each key process parameter and its corresponding optimal target value at multiple future times is calculated. The difference sequence and the current process status label are input together into the process control quantity mapping network; The process control quantity mapping network is a pre-trained neural network, which learns the complex correspondence between process parameter deviations and process states and various actuator control quantities through multiple nonlinear mappings. The process control quantity mapping network outputs control quantity adjustment values ​​that correspond one-to-one with the actuators of each production line, constituting the process control adjustment quantity.

[0013] Furthermore, the trained process state discrimination model is constructed through the following steps: Collect historical multidimensional process feature vector samples containing multiple process state categories, and label each sample with its corresponding actual process state label to form a training dataset; Construct a deep neural network classifier model, which includes an input layer, at least one fully connected hidden layer, and an output layer connected in sequence, wherein each fully connected hidden layer is followed by a non-linear activation function, and the output layer uses the softmax activation function. The training dataset is divided into a training set and a validation set. The deep neural network classifier model is trained using the training set data. During the training process, the model prediction results are calculated by forward propagation, the error between the prediction results and the true labels is calculated by the cross-entropy loss function, and the model parameters are updated by the backpropagation algorithm and the gradient descent optimizer. During training, the model's performance on non-training data is periodically evaluated using validation set data, and hyperparameters are adjusted and the model is stopped early based on the performance on the validation set to prevent overfitting. Once the model training meets the preset stopping conditions, the model parameters are saved to obtain the trained process state discrimination model.

[0014] Furthermore, the method also includes: The process control adjustment is sent to the corresponding production line actuator to complete the closed-loop optimization control of the tape production process, including: The process control adjustment amount is encapsulated according to a preset communication protocol format to generate a control instruction containing the target actuator identifier, control amount value, and execution time window; The control commands are sent to the corresponding identified production line actuators via the real-time data bus of the Industrial Internet of Things. The production line actuator receives and parses the control command, drives the servo motor, heater, pressure valve or take-up and unwind motor, and adjusts the process parameters to the target position or state indicated by the control quantity value within a specified time window. After the adjustment is completed, the industrial IoT sensor network collects new real-time process parameters again and starts the next round of optimization control cycle.

[0015] Compared with the prior art, the beneficial effects of the present invention are: Multi-source asynchronous data alignment and cleaning of multi-dimensional process time-series streams can eliminate timing misalignments caused by differences in sampling frequencies of different sensors, remove abnormal noise and redundant data from parameters, and form a synchronous process parameter matrix with consistent timing and complete data. Process feature extraction can uncover the coupling correlation between temperature, pressure, speed, thickness, and tension, and construct a multi-dimensional feature vector that can fully characterize the process operating state. The process state discrimination model can output corresponding operating condition labels based on the feature vectors. By retrieving a preset process knowledge graph, the association between state labels and potential process deviation patterns is established, incorporating process mechanisms and historical deviation patterns, clarifying the potential development forms of abnormal operating conditions, distinguishing between regular fluctuations and trend deviations, and overcoming the drawbacks of pure data-driven discrimination without prior process support, thus providing directional constraints for subsequent parameter prediction.

[0016] Using a set of potential process deviation patterns and a synchronous process parameter matrix as dual inputs, an improved gradient boosting decision tree is adapted to time-series evolution prediction tasks, outputting continuous process parameter evolution paths over multiple time periods, fully reflecting the parameter change trends over time. The evolution paths are compared with the optimal target range of process parameters to obtain time-series deviation results. A process control quantity mapping network is used to perform nonlinear mapping calculations on the deviations, outputting process control adjustment quantities that match the time-series changes of the parameters. These adjustment quantities closely align with the future change trends of the parameters, mitigating the lag inherent in traditional control based on real-time single-point parameter calculations, reducing the deviation range between parameters and the target range, stabilizing the operating status of parameters in each process step, optimizing the coordination of multi-parameter collaborative changes, improving the accuracy and real-time performance of process control, and adapting to the dynamic control requirements of continuous conveyor belt production. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the steps of the tape production process optimization and control method based on the Industrial Internet of Things as described in this invention. Figure 2 A flowchart for multi-source asynchronous data alignment and cleaning; Figure 3 A flowchart for outputting status labels for the process status discrimination model. Detailed Implementation

[0018] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] The present invention provides a method for optimizing and controlling the tape production process based on the Industrial Internet of Things (IIoT). The overall implementation scheme is as follows: Real-time process parameters, including temperature, pressure, speed, thickness, and tension, are collected from an IIoT sensor network deployed at each stage of the tape production line, forming a multi-dimensional process time-series stream. Multi-source asynchronous data alignment and cleaning are performed on the multi-dimensional process time-series stream to generate a synchronous process parameter matrix. Process feature extraction is performed on the synchronous process parameter matrix to obtain a multi-dimensional process feature vector. The multi-dimensional process feature vector is input into a trained process state discrimination model, outputting a current process state label. Based on the current process state label, a preset process knowledge graph is retrieved to obtain a set of potential process offset patterns associated with the current process state label. Based on the set of potential process offset patterns and the synchronous process parameter matrix, an improved gradient boosting decision tree algorithm is used to predict the future process parameter evolution path. The improved gradient boosting decision tree algorithm is based on the gradient boosting framework and is improved for the process parameter evolution path prediction task. Based on the deviation between the future process parameter evolution path and the preset optimal target range of process parameters, a set of process control adjustment quantities is calculated through a process control quantity mapping network.

[0020] In one embodiment of the present invention, the specific process of performing multi-source asynchronous data alignment and cleaning on a multi-dimensional process timeline to generate a synchronous process parameter matrix includes: (See reference) Figure 2 This method identifies the timestamp sequences of sensor data from different sources in a multidimensional process timeline, using the data source with the highest sampling frequency as the baseline time axis. For data sources with other sampling frequencies, a spline interpolation algorithm based on a time window is used to resample the data onto the baseline time axis, achieving timestamp alignment. For the timestamp-aligned data, a data cleaning process based on local outlier detection is applied to identify and remove outliers caused by transient sensor anomalies. The cleaned, time-aligned multidimensional process data is then arranged chronologically to form a synchronized process parameter matrix, where each row represents a sampling time and each column represents a process parameter.

[0021] The specific process of extracting process features from the synchronous process parameter matrix to obtain a multidimensional process feature vector includes: In the time domain, calculating the statistical features of each process parameter column in the synchronous process parameter matrix, including mean, standard deviation, skewness, and kurtosis. In the frequency domain, performing a Fast Fourier Transform on the time series of each process parameter column in the synchronous process parameter matrix to extract the amplitude and frequency of the main spectral components as frequency domain features. At the state transition level, calculating the rate of change of process parameters between adjacent sampling times and counting the number of times the rate of change exceeds a threshold within a fixed time window as dynamic features. The time domain statistical features, frequency domain features, and state dynamic features are then concatenated and combined to form a comprehensive feature set representing the current process operating state, i.e., the multidimensional process feature vector.

[0022] In practical implementation, the tape production line includes multiple processes such as coating, drying, and winding. An industrial IoT sensor network is deployed to collect real-time process parameters. For example, temperature and pressure sensors are installed at the coating head, speed sensors at the transfer rollers, thickness sensors at the thickness gauge, and tension sensors at the tension rollers. The sampling interval for the temperature and pressure sensors is 0.5 seconds, for the speed sensors it is 1 second, and for the thickness and tension sensors it is 2 seconds. The data streams from different sensors together constitute a multi-dimensional process timeline containing temperature, pressure, speed, thickness, and tension. In some embodiments, multi-source asynchronous data alignment and cleaning are performed on the multi-dimensional process timeline to generate a synchronous process parameter matrix. The timestamp sequences of sensor data from different sources in the multi-dimensional process timeline are identified, and the temperature sensor data stream is determined to have the highest sampling frequency of 0.5 seconds. The timestamp sequence of the temperature sensor data stream is then set as the reference time axis. For the speed sensor data stream, a spline interpolation algorithm based on a time window is adopted. At each time point of the reference time axis, a cubic spline function is constructed based on the data points of the original timestamp sequence of the speed sensor to calculate the speed value corresponding to that reference time, thereby realizing the resampling and timestamp alignment of the data on the reference time axis.

[0023] For the data streams from the thickness and tension sensors, a spline interpolation algorithm based on a time window is also used to resample the data sampled every 2 seconds onto a reference time axis with a 0.5-second interval. In specific implementation, after timestamp alignment, a data cleaning process based on local outlier detection is applied to calculate the local outlier score of the aligned multi-dimensional process parameter vector at each time point. Outliers with scores exceeding a preset threshold are identified and removed. These outliers correspond to noise data generated by instantaneous sensor anomalies. After removing outliers, the cleaned, time-aligned temperature, pressure, speed, thickness, and tension data are arranged in chronological order to form a synchronized process parameter matrix. Each row of the synchronized process parameter matrix represents a sampling time at a 0.5-second interval, and each column represents the specific value of the temperature, pressure, speed, thickness, and tension process parameters.

[0024] In practical implementation, process feature extraction is performed on the synchronous process parameter matrix to obtain a multi-dimensional process feature vector. In the time domain, the statistical characteristics of each process parameter column in the synchronous process parameter matrix are calculated within a set time window. For a window containing N sampling times, the mean, standard deviation, skewness, and kurtosis of the temperature parameter column are calculated; similarly, the mean, standard deviation, skewness, and kurtosis of the pressure, velocity, thickness, and tension parameter columns are calculated. In the frequency domain, a Fast Fourier Transform is performed on the time series of each process parameter column in the synchronous process parameter matrix within the set time window. The top K spectral components with the highest amplitudes are extracted from the transformed spectrum, and the amplitude values ​​and corresponding frequency values ​​of these K components are recorded as frequency domain features. In practical implementation, a dimensionless metric formula for evaluating the significance of spectral components is defined as follows: ; in: It is the amplitude of the k-th spectral component. It is the reference value for the signal amplitude of this process parameter within the current time window. It is the frequency corresponding to the k-th spectral component. This serves as the system's reference frequency benchmark. At the state transition level, the rate of change of each process parameter between adjacent sampling times in the synchronous process parameter matrix is ​​calculated. The cumulative number of times the absolute value of the rate of change of each process parameter exceeds its preset threshold within a fixed time window is counted, and this cumulative number is used as the dynamic characteristic of that process parameter. Optionally, the mean, standard deviation, skewness, and kurtosis (four features) of each process parameter obtained in the time domain, and K amplitude features, K frequency features, and K formula-based features of each process parameter obtained in the frequency domain are used. The calculated saliency features and one dynamic feature for each process parameter obtained at the state transition level are sequentially concatenated. For a production line containing five process parameters, this ultimately constitutes a dimension... The comprehensive feature set is the multidimensional process feature vector.

[0025] In some embodiments, the process feature extraction process is continuously cyclical. Each time a new synchronous process parameter matrix sliding window is obtained, the aforementioned feature calculation and splicing operations at the time domain, frequency domain, and state transition levels are repeated. It can be understood that the multi-dimensional process feature vector comprehensively represents the overall operating state of the production process within the current time window. Optionally, in frequency domain feature extraction, the number K of the main spectral components can be preset according to the complexity of the actual process signal; for example, in vibration analysis of conveyor belt production, K can be set to 3. Reference amplitude It can be set to the average amplitude of the process parameter signal during historical steady-state operation, with a reference frequency. It can be set to half of the system sampling frequency.

[0026] In one embodiment of the present invention, the specific process of inputting a multi-dimensional process feature vector into a trained process state discrimination model and outputting the current process state label includes: (See below) Figure 3 The trained process state discrimination model is a deep neural network classifier pre-trained using historical process data. Multidimensional process feature vectors are input into the input layer of the deep neural network classifier. The deep neural network classifier contains multiple fully connected layers and non-linear activation layers, and the multidimensional process feature vectors undergo layer-by-layer non-linear transformation and feature abstraction through these layers. In the output layer of the deep neural network classifier, the probability of the multidimensional process feature vectors belonging to each preset process state category is calculated using the softmax function. The preset process state category with the highest probability is output as the model's discrimination result, which is the current process state label.

[0027] The trained process state discrimination model is constructed through the following steps: First, historical multi-dimensional process feature vector samples containing various process state categories are collected, and each sample is labeled with its corresponding actual process state label to form a training dataset. Second, a deep neural network classifier model is constructed, consisting of a sequentially connected input layer, at least one fully connected hidden layer, and an output layer. Each fully connected hidden layer is followed by a non-linear activation function, and the output layer uses the softmax activation function. Third, the training dataset is divided into a training set and a validation set. The deep neural network classifier model is trained using the training set data. During training, the model's prediction results are calculated using forward propagation, the error between the prediction results and the true labels is calculated using the cross-entropy loss function, and the model parameters are updated using the backpropagation algorithm and gradient descent optimizer. Fourth, during training, the model's performance on untrained data is periodically evaluated using the validation set data. Hyperparameter adjustments and early model stopping are performed based on the performance on the validation set to prevent overfitting. Finally, when the model training meets the preset stopping conditions, the model parameters are saved, resulting in the trained process state discrimination model.

[0028] In practice, the trained process state discrimination model is a deep neural network classifier pre-trained using historical process data. On the tape production line, multi-dimensional process feature vectors are input to the input layer of the trained process state discrimination model in real time. The dimension of the multi-dimensional process feature vectors is consistent with the number of neurons in the input layer of the deep neural network classifier. The deep neural network classifier contains multiple fully connected layers and non-linear activation layers. After the multi-dimensional process feature vectors undergo a linear transformation in the first fully connected layer, they are non-linearly mapped through a non-linear activation function to generate the first layer of abstract features. The first layer of abstract features continues to be passed to subsequent fully connected layers and non-linear activation layers for layer-by-layer non-linear transformation and feature abstraction. In the output layer of the deep neural network classifier, the probability of the multi-dimensional process feature vectors belonging to various preset process state categories is calculated using the softmax function. The preset process state categories include "stable operation," "slight temperature deviation," "abnormal tension," and "uneven winding," among others. The preset process state category with the highest probability is output as the model's discrimination result, which is the current process state label. For example, "abnormal tension" is output as the current process state label.

[0029] In some embodiments, the trained process state discrimination model is constructed through the following steps: collecting historical multidimensional process feature vector samples containing multiple process state categories, and labeling each sample with its corresponding actual process state label to form a training dataset. The historical multidimensional process feature vector samples in the training dataset are derived from historical operating data of the conveyor belt production line. This data is obtained by performing timestamp alignment and data cleaning on the raw process parameter sequences collected by historical sensors, and calculating time-domain statistical features, frequency-domain spectral features, and state transition dynamic features. The process state label for each sample is confirmed by domain engineers based on production records and product quality reports for the corresponding time period. A deep neural network classifier model is constructed, comprising a sequentially connected input layer, at least one fully connected hidden layer, and an output layer. Each fully connected hidden layer is followed by a non-linear activation function, and the output layer uses the softmax activation function. Optionally, the non-linear activation function can be the ReLU function. The number of neurons in the input layer of the deep neural network classifier model is set to the dimension value of the multidimensional process feature vector, and the number of neurons in the output layer is set to the total number of preset process state categories.

[0030] In practice, the training dataset is divided into a training set and a validation set. The deep neural network classifier model is trained using the training set data. During training, the model's prediction results are calculated using forward propagation, and the error between the prediction results and the true labels is calculated using the cross-entropy loss function. The calculation method is as follows: ; in: This indicates the number of samples in a training batch. This indicates the total number of process status categories. It is an indicator function, when the sample The true category is The value is 1 if the condition is met, and 0 otherwise. This indicates that the deep neural network classifier model predicts samples. Category The probability of this is understood. It can be understood that the model parameters are updated through backpropagation and gradient descent optimizer. The gradient descent optimizer adjusts the weight and bias parameters of the fully connected layers in the deep neural network classifier model based on the gradient calculated by the cross-entropy loss function. In some embodiments, during training, the model's performance on non-training data is periodically evaluated using validation set data, and hyperparameter adjustments and early model stopping are performed based on the performance on the validation set to prevent overfitting. When the model training meets a preset stopping condition, the model parameters are saved, resulting in a trained process state discrimination model. The preset stopping condition may be that the classification accuracy on the validation set no longer improves for several consecutive evaluation cycles, or that the total number of training rounds reaches a preset upper limit. Optionally, hyperparameter adjustments include adjusting the learning rate, the number of neurons in the fully connected hidden layers, and the training batch size. Early model stopping is triggered when the validation set loss function value begins to rise, and the model parameters are rolled back to the state where the validation set performance is optimal for saving. It can be understood that the finally saved trained process state discrimination model is deployed in the real-time control system of the conveyor belt production line to continuously receive real-time multi-dimensional process feature vectors and output the current process state label.

[0031] In one embodiment of the present invention, the specific process of retrieving a preset process knowledge graph based on the current process state label to obtain a set of potential process offset patterns associated with the current process state label includes: the preset process knowledge graph is stored in a graph structure, where nodes represent different process states or abnormal process parameter patterns, and edges represent the relationships or causal logic between nodes. Using the current process state label as the starting node, a multi-hop breadth-first traversal is performed in the process knowledge graph. During the traversal, all nodes representing abnormal process parameter patterns that are directly or indirectly connected to the starting node through causal edges are collected to form a set of related nodes. Attribute information of each abnormal process parameter pattern node is extracted from the set of related nodes. The attribute information includes the typical parameter change direction and magnitude triggered by the abnormal process parameter pattern node, forming a set of potential process offset patterns.

[0032] In practical implementation, the pre-defined process knowledge graph is stored in a graph structure. Nodes in the process knowledge graph represent different process states or abnormal process parameter patterns, while edges represent the relationships or causal logic between nodes. For example, in the process knowledge graph, "stable operation" is a process state node, and "low drying temperature" is a process parameter abnormality pattern node. A directed edge exists between these two nodes, and the edge's attribute can be labeled "may lead to". The construction of the process knowledge graph is based on expert knowledge and historical event analysis in the tape production field. The node set includes various states and patterns such as "stable operation," "abnormal tension," "uneven coating," "abnormal oven temperature gradient," and "winding pressure fluctuation." The edge set defines the causal, accompanying, or evolutionary relationships between these states and patterns.

[0033] In some embodiments, a preset process knowledge graph is retrieved based on the current process state label output from the process state discrimination model. The current process state label is, for example, "tension anomaly". Using "tension anomaly" as the starting node, a multi-hop breadth-first traversal is performed in the process knowledge graph. The number of hops can be set based on prior knowledge, for example, a maximum hop count of 3. The traversal process begins with the "tension anomaly" node and visits all nodes directly connected to it via causal edges. These nodes constitute the first-hop related nodes, such as "uneven winding" and "thickness fluctuation" nodes. Using the first-hop related nodes as new starting points, nodes directly connected to these nodes via causal edges that have not yet been visited are visited. These nodes constitute the second-hop related nodes; for example, the "thickness fluctuation" node may be associated with the "coating head pressure anomaly" node. The traversal process continues until the set maximum number of hops is reached.

[0034] In practical implementation, during the traversal process, all nodes representing abnormal process parameter patterns that are directly or indirectly connected to the starting node through causal edges are collected, forming a relevant node set. In the process knowledge graph, node types can be divided into "process state" and "process parameter abnormal pattern." During traversal, only nodes of type "process parameter abnormal pattern" are collected. For example, starting from the "tension abnormality" node, after a maximum of 3 hops, multiple process parameter abnormal pattern nodes may be collected, including "winding motor speed fluctuation," "substrate serpentine movement," "sudden drop in oven outlet temperature," and "coating roller eccentricity." These nodes together constitute a relevant node set related to the current "tension abnormality" state. It can be understood that in the knowledge graph design, the edge weight can represent the association strength, and nodes can be sorted or filtered according to the edge weight during traversal and node collection.

[0035] Attribute information for each process parameter anomaly pattern node is extracted from the relevant node set. Each process parameter anomaly pattern node in the process knowledge graph has predefined structured attribute information. This attribute information includes the typical parameter change direction and magnitude triggered by the process parameter anomaly pattern node. For example, the attribute information for the "sudden drop in oven outlet temperature" node defines the affected process parameter as "drying temperature," the typical change direction as "negative," and the typical change range as "[-8°C, -3°C]." This type of attribute information is extracted from all collected process parameter anomaly pattern nodes to form a potential process offset pattern set. The potential process offset pattern set is a list containing multiple entries. Each entry records a potential process parameter offset pattern that may be triggered by the current process state, including the name of the affected parameter, the change direction, and the estimated change range.

[0036] Optionally, during knowledge graph traversal, the correlation between anomaly patterns of certain process parameters and the current process status label can be calculated by combining the relationship type and weight of the edges. Defining the correlation strength: ; in: Indicates the initial process state node. This indicates the node representing the abnormal mode of the process parameter to be evaluated. Indicates from node To the node The set of all paths Indicates a path The number of jumps, Representing a path The product of the weights of all edges. This can be understood as being based on the degree of association. The patterns in the potential process offset pattern set can be sorted to provide a priority reference for subsequent prediction steps.

[0037] For clarity, refer to Table 1, which lists some process parameter anomaly pattern nodes and their attribute information collected and extracted from the process knowledge graph, starting with the aforementioned "tension anomaly": Table 1. Set of Potential Process Offset Patterns: In some embodiments, the generation of the potential process offset pattern set is dynamic, changing with the current process state label. The content of the process knowledge graph can be updated and maintained offline, and the retrieval results of the potential process offset pattern set can be made more comprehensive and accurate by injecting new expert experience or mining new pattern nodes and related edges from historical anomalies.

[0038] In one embodiment of the present invention, the improved gradient boosting decision tree algorithm works as follows: When constructing the decision tree in each iteration, not only are the original features of the process parameters used, but also the set of potential process offset patterns is encoded as additional guiding features and added to the candidate feature set of the splitting node. When calculating the optimal splitting point, the standard mean squared error loss is fused with a process path smoothness regularization term, which penalizes splitting patterns where the predicted values ​​fluctuate drastically between adjacent time points. In the calculation of the output value of the leaf nodes of the decision tree, not only the negative gradient direction is considered, but also an attraction term based on the historical optimal process path is introduced, so that the output value, while following gradient descent, moves closer to the historically stable process region. After multiple iterations, the prediction results of each round of the decision tree are weighted and accumulated to obtain the final sequence of predicted process parameters, i.e., the future evolution path of the process parameters.

[0039] The specific construction method of the process path smoothness regularization term is as follows: At the current split point, after sample partitioning, calculate the adjacent differences in the time series of the predicted process parameters of the samples in the left and right child nodes. Calculate the sum of squares of the adjacent differences for each child node to measure the degree of fluctuation in the predicted path caused by the splitting method. The weighted sum of the predicted path fluctuations of the left and right child nodes is used as the smoothness penalty term corresponding to the split point, and incorporated into the overall loss calculation for split point selection. This constrains the decision tree splitting to tend towards producing a smooth process evolution path.

[0040] In its implementation, the improved gradient boosting decision tree algorithm is based on the gradient boosting framework and is improved for the task of predicting the evolution path of process parameters. Its working principle includes that when building the decision tree in each iteration, not only the original features of the process parameters are used, but also the set of potential process offset patterns is encoded as additional guiding features and added to the candidate feature set of the split node. The set of potential process offset patterns includes patterns such as "fluctuation of winding motor speed" and "sudden drop in oven outlet temperature" and their typical change directions and magnitudes. The encoding process generates numerical features for each pattern that represent its matching degree or influence intensity with the current process state. These guiding features, together with the original time-series features and statistical features, constitute the candidate feature set of the split node. In some embodiments, when the improved gradient boosting decision tree algorithm calculates the optimal split point for a split node, it integrates the standard mean squared error loss with a process path smoothness regularization term. The process path smoothness regularization term penalizes splits where the predicted values ​​fluctuate drastically between adjacent time points. The specific construction of the process path smoothness regularization term measures the degree of fluctuation in the predicted path by calculating the sum of the squares of the adjacent differences in the predicted process parameters of the samples in the left and right child nodes under the current split point. The degree of fluctuation of the left and right child nodes is weighted and summed with the sample size as the weight, which serves as the smoothness penalty term corresponding to the split point. Finally, the selection of the split point is determined based on the weighted sum of the standard mean squared error loss and the smoothness penalty term, thereby constraining the decision tree split to tend to produce a smooth process evolution path.

[0041] In its implementation, the improved gradient boosting decision tree algorithm, in calculating the output value of the leaf nodes of the decision tree, not only considers the negative gradient direction but also introduces an attraction term based on the historical optimal process path. The historical optimal process path is extracted from high-quality batch data in the history of tape production, representing a stable and ideal sequence of process parameters. For all samples falling into a certain leaf node, the output value of the leaf node... It consists of two parts: ; in: This indicates the output value of the leaf node, whose dimensions are consistent with the process parameters to be predicted; It is the learning rate, a dimensionless hyperparameter; It represents the sum of the first-order gradients of all samples in the leaf node under the current model prediction, and its dimension is the reciprocal of the dimension of the target value. This represents the sum of the second-order gradients of all samples in the leaf node under the current model prediction, and its dimension is the square of the reciprocal of the dimension of the target value; therefore The dimensions of the calculation result are consistent with the dimensions of the target value; It is a hyperparameter that controls the weight of the attraction term, and it is dimensionless; This is the attraction term, calculated as the average of the historical best process path values ​​corresponding to the samples in this leaf node. Its dimensions are consistent with the target value dimensions (e.g., °C). The dimensions of both terms on the right side of the formula are the same as those on the left side. The units are consistent, and this calculation ensures that the output value follows the negative gradient direction. While updating, we will also move towards historically stable process areas. After multiple iterations, the prediction results of each round of decision trees are weighted and summed to obtain the final sequence of predicted process parameters, i.e., the future evolution path of process parameters. To clearly illustrate the role of the process path smoothness regularization term in node splitting, refer to Table 2, which shows a simplified calculation example containing 6 samples, each containing a prediction sequence of length 3, and a candidate split point dividing the sample into left child nodes (samples 1, 2, 3) and right child nodes (samples 4, 5, 6): Table 2. Calculation table of process path smoothness regularization term: Understandably, when selecting split points, the algorithm calculates a weighted sum of the standard mean squared error loss corresponding to all candidate features and split points, along with the process path smoothness regularization term calculated in the example above, and selects the split point that minimizes the overall loss. Minimal features and split points. In some embodiments, the weighting coefficients between the standard mean squared error loss and the process path smoothness regularization term. and the weight of the attraction term in the leaf node output calculation. All of these can be adjusted and optimized using a validation set. The improved gradient boosting decision tree algorithm, through the above mechanism, utilizes the guidance of a set of potential process offset patterns and integrates path smoothness with historical optimal attractiveness to ultimately output a future process parameter evolution path that better conforms to process physical constraints.

[0042] In one embodiment of the present invention, the specific process of calculating a set of process control adjustment quantities through a process control quantity mapping network based on the deviation between the future process parameter evolution path and the preset optimal target range of process parameters includes: comparing the future process parameter evolution path with the preset optimal target range of process parameters, and calculating the difference sequence between the predicted values ​​of each key process parameter at multiple future time points and their corresponding optimal target values. The difference sequence and the current process state label are then input into the process control quantity mapping network. The process control quantity mapping network is a pre-trained neural network that internally learns the complex correspondence between process parameter deviations and process states and the control quantities of various actuators through multiple layers of nonlinear mapping. The process control quantity mapping network outputs control quantity adjustment values ​​that correspond one-to-one with the actuators of each production line, constituting the process control adjustment quantities.

[0043] The method also includes sending process control adjustments to the corresponding production line actuators to complete closed-loop optimization control of the tape production process. This includes: encapsulating the process control adjustments according to a preset communication protocol format to generate control commands containing the target actuator identifier, control value, and execution time window. The control commands are then sent to the corresponding identified production line actuators via the industrial IoT real-time data bus. The production line actuators receive and parse the control commands, driving servo motors, heaters, pressure valves, or take-up / unwind motors to adjust the process parameters to the target position or state indicated by the control value within the specified time window. After adjustment, the industrial IoT sensor network collects new real-time process parameters again, initiating the next round of optimization control cycle.

[0044] In practical implementation, based on the deviation between the future evolution path of process parameters and the preset optimal target range of process parameters, a set of process control adjustment quantities is calculated through a process control quantity mapping network. The process involves comparing the future evolution path of process parameters with the preset optimal target range of process parameters, and calculating the sequence of differences between the predicted values ​​of each key process parameter at multiple future time points and their corresponding optimal target values. The preset optimal target range of process parameters is determined based on the tape production process standards and historical high-quality production data. For example, for the temperature parameter in the drying process, the optimal target range can be set as follows: The optimal target value can be set as the midpoint of the interval, 152.5°C. The future process parameter evolution path gives a sequence of predicted drying temperatures over a period of time. The difference between the predicted temperature at each moment in the prediction sequence and 152.5°C is calculated to form a difference sequence for drying temperature. This operation is performed on all key process parameters that need to be controlled to obtain a set containing multiple process parameter difference sequences.

[0045] In practical implementation, the difference sequence and the current process status label are input into the process control quantity mapping network, which is a pre-trained neural network. The difference sequence is organized in vector form. For example, if the drying temperature, coating pressure, and winding tension are predicted for the next 5 moments, the difference sequence vector contains 15 elements. The current process status label is also converted into a vector after being thermally encoded. The difference sequence vector and the process status label encoded vector are concatenated at the input layer to form the complete input vector of the process control quantity mapping network. Internally, the process control quantity mapping network learns the complex correspondence between process parameter deviations and process statuses and the control quantities of various actuators through multiple layers of nonlinear mapping. The output of the process control quantity mapping network corresponds one-to-one with the control quantity adjustment values ​​of each production line actuator, constituting the process control adjustment quantity. The process control adjustment quantity is represented in vector form, for example, the adjustment quantity for the heater power, coating pump pressure valve opening, and winding motor speed. The internal mapping relationship of the process control quantity mapping network can be described as follows: ; in: This represents the output vector of process control adjustments. This indicates that the difference sequence vector With process status label encoding vector The input vector after concatenation in dimensions and These are the weight matrix and bias vector of the first fully connected layer in the process control quantity mapping network. Represents a non-linear activation function. and These are the weight matrix and bias vector of the second fully connected layer in the process control quantity mapping network. This represents the activation function of the output layer. The process control quantity mapping network is learned through training. , , , These parameters enable the mapping of appropriate actuator control adjustments when a specific sequence of process parameter differences and process status are input.

[0046] The method also includes sending process control adjustments to the corresponding production line actuators to complete closed-loop optimization control of the tape production process. This includes encapsulating the process control adjustments according to a preset communication protocol format to generate control instructions containing the target actuator identifier, control quantity value, and execution time window. In some embodiments, the preset communication protocol adopts the OPCUA protocol, and the control instructions include the node identifier of the target actuator in the OPCUA address space, the control quantity value, and the expected time window for the adjustment to take effect. The control instructions are sent to the corresponding identified production line actuators via the real-time data bus of the Industrial Internet of Things (IIoT). The real-time data bus can adopt the Industrial Ethernet protocol to ensure the real-time and deterministic nature of instruction transmission. The production line actuators receive and parse the control instructions, driving servo motors, heaters, pressure valves, or take-up / unwind motors to adjust the process parameters to the target position or state indicated by the control quantity value within a specified time window. For example, if the control instruction contains the target actuator identifier "Heater_Zone2_Power", the control quantity value "+2.5%", and the execution time window "next control cycle", the heater in zone two of the production line, upon receiving the instruction, will increase its power by 2.5% in the next control cycle.

[0047] After adjustment, the industrial IoT sensor network collects new real-time process parameters again, initiating the next round of optimization control loop. The new real-time process parameters are collected and form a new multi-dimensional process time sequence stream. The system repeatedly executes the complete process from data alignment and cleaning to generating new control commands, achieving continuous closed-loop optimization control. It can be understood that the process control quantity mapping network needs to be pre-trained. The training data comes from the process parameter deviation sequence recorded in historical production processes, the corresponding process state at each moment, and the final control adjustment amount. The network is trained through supervised learning to learn the mapping relationship from deviations and states to control quantities. Optionally, the structure of the process control quantity mapping network is not limited to two layers; the number of hidden layers can be increased according to the complexity of the actual control problem, and a nonlinear activation function can be used. You can choose the ReLU function, the output layer activation function. You can choose between a linear function or a sigmoid function based on the range of the control quantity.

[0048] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for optimizing and controlling the tape production process based on the Industrial Internet of Things, characterized in that, The method includes: Real-time process parameters, including temperature, pressure, speed, thickness, and tension, are collected from an industrial IoT sensor network deployed at each stage of the tape production line, forming a multi-dimensional process time sequence flow. Perform multi-source asynchronous data alignment and cleaning on the multi-dimensional process timeline to generate a synchronous process parameter matrix; Process feature extraction is performed on the synchronous process parameter matrix to obtain a multi-dimensional process feature vector; The multidimensional process feature vector is input into the trained process state discrimination model, and the current process state label is output. Based on the current process status label, a preset process knowledge graph is retrieved to obtain a set of potential process offset patterns associated with the current process status label; Based on the set of potential process offset patterns and the synchronous process parameter matrix, an improved gradient boosting decision tree algorithm is used to predict the future process parameter evolution path. The improved gradient boosting decision tree algorithm is based on the gradient boosting framework and is improved for the task of predicting the process parameter evolution path. Based on the deviation between the future process parameter evolution path and the preset optimal target range of process parameters, a set of process control adjustment quantities are calculated through a process control quantity mapping network.

2. The method for optimizing and controlling the tape production process based on the Industrial Internet of Things as described in claim 1, characterized in that, Perform multi-source asynchronous data alignment and cleaning on the multi-dimensional process timeline to generate a synchronous process parameter matrix, including: Identify the timestamp sequences of sensor data from different sources in the multidimensional process timeline, using the data source with the highest sampling frequency as the reference time axis; For data sources with other sampling frequencies, a spline interpolation algorithm based on a time window is used to resample the data onto the reference time axis to achieve timestamp alignment. For the data that has been timestamped and aligned, a data cleaning process based on local outlier detection is applied to identify and remove outliers caused by instantaneous sensor anomalies. The cleaned, time-aligned multidimensional process data are arranged in chronological order to form the synchronous process parameter matrix, where each row of the matrix represents a sampling time and each column represents a process parameter.

3. The method for optimizing and controlling the tape production process based on the Industrial Internet of Things as described in claim 1, characterized in that, Process feature extraction is performed on the synchronous process parameter matrix to obtain a multi-dimensional process feature vector, including: In the time domain, the statistical characteristics of each process parameter column in the synchronous process parameter matrix are calculated, including the mean, standard deviation, skewness, and kurtosis. In the frequency domain, a fast Fourier transform is performed on the time series of each process parameter column in the synchronous process parameter matrix to extract the amplitude and frequency of the main spectral components as frequency domain features. At the state transition level, the rate of change of process parameters between adjacent sampling times is calculated, and the number of times the rate of change exceeds the threshold within a fixed time window is counted as dynamic features. The time-domain statistical features, frequency-domain features, and state dynamic features are spliced ​​and combined to form a comprehensive feature set that characterizes the current process operation state, namely the multidimensional process feature vector.

4. The method for optimizing and controlling the tape production process based on the Industrial Internet of Things as described in claim 1, characterized in that, The multidimensional process feature vector is input into the trained process state discrimination model, and the current process state label is output, including: The trained process state discrimination model is a deep neural network classifier that has been trained in advance using historical process data. The multidimensional process feature vector is input into the input layer of the deep neural network classifier; The deep neural network classifier includes multiple fully connected layers and non-linear activation layers. The multi-dimensional process feature vector undergoes layer-by-layer non-linear transformation and feature abstraction through the fully connected layers and non-linear activation layers. In the output layer of the deep neural network classifier, the probability that the multidimensional process feature vector belongs to each preset process state category is calculated by the softmax function; The preset process state category with the highest probability is output as the model's discrimination result, which is the current process state label.

5. The method for optimizing and controlling the tape production process based on the Industrial Internet of Things as described in claim 1, characterized in that, Based on the current process state label, a preset process knowledge graph is retrieved to obtain a set of potential process offset patterns associated with the current process state label, including: The preset process knowledge graph is stored in a graph structure, where nodes represent different process states or abnormal patterns of process parameters, and edges represent the relationships or causal logic between nodes. Using the current process status label as the starting node, a multi-hop breadth-first traversal is performed in the process knowledge graph; During the traversal, all nodes that characterize abnormal patterns of process parameters and are directly or indirectly connected to the starting node through causal edges are collected to form a relevant node set; The attribute information of each process parameter abnormality mode node is extracted from the relevant node set. The attribute information includes the typical parameter change direction and magnitude triggered by the process parameter abnormality mode node, forming the potential process offset mode set.

6. The method for optimizing and controlling the tape production process based on the Industrial Internet of Things as described in claim 1, characterized in that, The working principle of the improved gradient boosting decision tree algorithm includes: When constructing the decision tree in each iteration, not only are the original features of the process parameters used, but the set of potential process offset patterns is also encoded as additional guiding features and added to the candidate feature set of the split node. When calculating the optimal split point for a split node, the standard mean square error loss is combined with a process path smoothness regularization term, which penalizes splitting patterns where the predicted values ​​fluctuate drastically between adjacent time points. In the calculation of the output value of the leaf node of the decision tree, not only is the negative gradient direction considered, but also an attraction term based on the historical optimal process path is introduced, so that the output value moves closer to the historical stable process region while following gradient descent. After multiple iterations, the prediction results of each round of decision trees are weighted and accumulated to obtain the final sequence of predicted process parameters, which is the future process parameter evolution path.

7. The method for optimizing and controlling the tape production process based on the Industrial Internet of Things as described in claim 6, characterized in that, The specific construction method of the process path smoothness regularization term is as follows: Calculate the adjacent differences in the time series of the predicted process parameters of the samples in the left and right child nodes after the sample division at the current split point. The sum of squares of the adjacent differences is calculated for the left and right child nodes respectively, in order to measure the degree of fluctuation in the predicted path caused by the splitting method; The weighted sum of the predicted path fluctuations of the left and right child nodes is used as the smoothness penalty term corresponding to the split point and incorporated into the overall loss calculation of the split point selection. This constrains the decision tree split to tend to produce a smooth process evolution path.

8. The method for optimizing and controlling the tape production process based on the Industrial Internet of Things as described in claim 1, characterized in that, Based on the deviation between the future process parameter evolution path and the preset optimal target range of process parameters, a set of process control adjustment quantities is calculated through a process control quantity mapping network, including: The future process parameter evolution path is compared with the preset optimal target range of process parameters, and the difference sequence between the predicted value of each key process parameter and its corresponding optimal target value at multiple future times is calculated. The difference sequence and the current process status label are input together into the process control quantity mapping network; The process control quantity mapping network is a pre-trained neural network, which learns the complex correspondence between process parameter deviations and process states and various actuator control quantities through multiple nonlinear mappings. The process control quantity mapping network outputs control quantity adjustment values ​​that correspond one-to-one with the actuators of each production line, constituting the process control adjustment quantity.

9. The method for optimizing and controlling the tape production process based on the Industrial Internet of Things as described in claim 1, characterized in that, The trained process state discrimination model is constructed through the following steps: Collect historical multidimensional process feature vector samples containing multiple process state categories, and label each sample with its corresponding actual process state label to form a training dataset; Construct a deep neural network classifier model, which includes an input layer, at least one fully connected hidden layer, and an output layer connected in sequence, wherein each fully connected hidden layer is followed by a non-linear activation function, and the output layer uses the softmax activation function. The training dataset is divided into a training set and a validation set. The deep neural network classifier model is trained using the training set data. During the training process, the model prediction results are calculated by forward propagation, the error between the prediction results and the true labels is calculated by the cross-entropy loss function, and the model parameters are updated by the backpropagation algorithm and the gradient descent optimizer. During training, the model's performance on non-training data is periodically evaluated using validation set data, and hyperparameters are adjusted and the model is stopped early based on the performance on the validation set to prevent overfitting. Once the model training meets the preset stopping conditions, the model parameters are saved to obtain the trained process state discrimination model.

10. The method for optimizing and controlling the tape production process based on the Industrial Internet of Things according to claim 1, characterized in that, The method further includes: The process control adjustment is sent to the corresponding production line actuator to complete the closed-loop optimization control of the tape production process, including: The process control adjustment amount is encapsulated according to a preset communication protocol format to generate a control instruction containing the target actuator identifier, control amount value, and execution time window; The control commands are sent to the corresponding identified production line actuators via the real-time data bus of the Industrial Internet of Things. The production line actuator receives and parses the control command, drives the servo motor, heater, pressure valve or take-up and unwind motor, and adjusts the process parameters to the target position or state indicated by the control quantity value within a specified time window. After the adjustment is completed, the industrial IoT sensor network collects new real-time process parameters again and starts the next round of optimization control cycle.

Images