Industrial control parameter intelligent optimization recommendation method based on convolutional neural network
By deploying a lightweight convolutional neural network model at the edge, multi-source time-series data is converted into operating condition images, and online fine-tuning is performed using a feedback mechanism. This solves the problems of reliance on expert experience and poor real-time performance in existing technologies, and enables real-time optimization and self-optimization of industrial control parameters.
Patent Information
- Application Number
- CN202511168457.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-20
- Publication Date
- 2025-11-25
AI Technical Summary
Existing technologies in industrial control rely on expert experience, lack systematicity, have poor real-time performance, weak model generalization ability, cannot handle high-dimensional and complex input information, lack feedback mechanisms, and cannot self-optimize.
A lightweight convolutional neural network model is used to respond in real time at the edge, converting multi-source time-series data into working condition images, and performing online fine-tuning through a feedback mechanism to achieve model self-optimization.
It enables real-time optimization and recommendation of industrial control parameters, improves the model's perception and generalization ability to complex working conditions, and realizes the model's self-optimization and long-term adaptation.
Smart Images
Figure CN121008545A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of industrial automation control, and particularly relates to an intelligent optimization recommendation method for industrial control parameters based on a convolutional neural network. BACKGROUND
[0002] Parameter intelligent control has been applied in multiple industries. In chemical production, optimal reaction parameters can be recommended based on multi-source real-time data of a reaction kettle. In a metallurgical scene, adaptive control parameters are given for variables such as molten steel temperature and equipment load. In a manufacturing assembly line, dynamic adjustment suggestions are provided for motor speed, conveyor belt speed and the like to achieve optimization of the production process.
[0003] However, the prior art has the following disadvantages: it relies on expert experience, parameter setting lacks systematicness; it is mostly offline calculation, poor in real-time performance; the model has weak generalization ability and relies on artificial feature design; it cannot process high-dimensional complex input information; and it lacks a feedback mechanism, so the model cannot be self-optimized.
[0004] Therefore, a new method is urgently needed. SUMMARY
[0005] The application aims to provide an intelligent optimization recommendation method for industrial control parameters based on a convolutional neural network, which realizes edge real-time response through a lightweight model, converts multi-source time series data into working condition images, triggers online fine-tuning of the model with actual effect error, and realizes continuous self-optimization.
[0006] To achieve the above-mentioned purpose, the application provides an intelligent optimization recommendation method for industrial control parameters based on a convolutional neural network, comprising the following steps:
[0007] S1, real-time collection and storage of multi-source working condition data to obtain structured time series data;
[0008] S2, preprocessing of the structured time series data in S1, and conversion of the preprocessed structured time series data to generate working condition images;
[0009] S3, construction of a lightweight model;
[0010] S4, supervised training of the model in S3 based on the working condition images in S2 to obtain a trained model and training history records;
[0011] S5, deployment of the trained model in S4 to an edge, input of the working condition images in S2 as real-time parameters into the recommendation model to obtain recommended parameters;
[0012] S6, record the actual production effect after the recommended parameter in S5 is executed, calculate the error between the actual effect and the predicted effect, mark the sample containing the working condition image, the recommended parameter and the actual effect as the to-be-learned sample and deliver the to-be-learned sample to S4 to assist the model iteration, and meanwhile, replace the model in S5 with the fine-tuned optimization model.
[0013] Preferably, the real-time multi-source working condition data is collected in S1, and the multi-dimensional data is collected in parallel through an industrial internet of things gateway at a fixed sampling interval, and the multi-dimensional data includes control parameters, sensor variables, device states, product quality indicators and external environment information.
[0014] Preferably, the data preprocessing in S2 is specifically: the mean ± 3 times standard deviation method is used to remove outliers, the wavelet transform is used to decompose the signal to remove high-frequency noise, and then the standardization method is used to unify the dimension, and the data is mapped to the [-1, 1] interval.
[0015] Preferably, the recommended parameter obtained in S5 is specifically: the real-time working condition image is expanded in the batch dimension, input into the deployed model for forward propagation inference, and the recommended parameter is output after the batch dimension is squeezed.
[0016] Preferably, the to-be-learned sample marked in S6 is specifically: an error threshold is set, and when the actual error exceeds the error threshold, the sample is marked as the to-be-learned sample.
[0017] Preferably, S6 further includes fine-tuning the model by using the small batch gradient descent method, and when updating the parameters, the bottom convolution layer parameters are fixed, and only the top weight is updated.
[0018] The application also provides an industrial control parameter intelligent optimization recommendation system based on a convolutional neural network, comprising:
[0019] A multi-source working condition data real-time collection and storage module is used for collecting and storing multi-source working condition data in real time to obtain structured time series data.
[0020] A working condition image conversion module is connected with the multi-source working condition data real-time collection and storage module, and is used for preprocessing the structured time series data and converting the preprocessed structured time series data to generate working condition images.
[0021] A model construction module is connected with the working condition image conversion module, and is used for constructing a model.
[0022] A model supervised training module is connected with the model construction module, and is used for supervising and training the model based on the working condition image to obtain a trained model and a training history record.
[0023] An edge end deployment and real-time parameter recommendation module, connected with the model supervised training module, is configured to deploy the trained model to the edge end and input the working condition image as a real-time parameter to obtain a recommended parameter.
[0024] A feedback data acquisition and model online optimization module, connected with the edge end deployment and real-time parameter recommendation module, is configured to record the actual production effect after the recommended parameter is executed in S5, calculate the error between the actual effect and the predicted effect, mark the sample containing the working condition image, the recommended parameter and the actual effect as a to-be-learned sample and deliver the to-be-learned sample to the model supervised training module to assist the model iteration, and replace the model in the edge end deployment and real-time parameter recommendation module with the fine-tuned optimized model.
[0025] Therefore, the industrial control parameter intelligent optimization and recommendation method based on the convolutional neural network has the following beneficial effects compared with the prior art.
[0026] (1) The light-weight CNN model is designed and deployed to the edge end, the real-time parameter recommendation is realized through the optimized inference process, the dynamic working condition demand is met, the working condition rules are automatically learned by the CNN, the manual feature design is replaced, the dependence on the expert experience is eliminated, the multi-source time series data is converted into the working condition image, the advantage of the CNN in processing high-dimensional information is utilized, and the perception and generalization ability of the model to the complex working condition are improved.
[0027] (2) The feedback mechanism is constructed, the model online fine-tuning is triggered through the error between the actual effect and the predicted effect, and the model self-optimization and long-term adaptation are realized.
[0028] The technical solutions of the embodiments of the present application will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS
[0029] Figure 1 The flowchart of the industrial control parameter intelligent optimization and recommendation method based on the convolutional neural network is shown in the figure. DETAILED DESCRIPTION
[0030] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application. Unless otherwise defined, the technical terms or scientific terms used in the present application should be understood as the usual meanings understood by those of ordinary skill in the art.
[0031] Embodiment one
[0032] As Figure 1 shown, an industrial control parameter intelligent optimization recommendation method based on a convolutional neural network of the present application comprises the following steps:
[0033] S1, real-time acquisition and storage of multi-source working condition data to obtain structured time series data;
[0034] In this step, through the industrial protocol interface PLC control system, DCS system and sensor network, multi-dimensional data is collected in parallel at a fixed sampling interval, including control parameters, sensor variables, device status, product quality indicators and external environment information. The collected data is stored as structured time series data in chronological order;
[0035] S2, pre-processing of the structured time series data in S1, and converting the pre-processed structured time series data to generate working condition images;
[0036] In this step, the mean ± 3 times standard deviation method is used to remove outliers, wavelet transform is used to decompose the signal to remove high-frequency noise, and then a standardization method is used to unify the dimension, such as Z-score normalization or Min-Max normalization, to map the data to the [-1, 1] interval.
[0037] Select a fixed time window, arrange the multi-channel time series data in a "variable-time" matrix, each row represents a variable, and each column represents the standardized value of the variable at a certain sampling time. Convert to a two-dimensional "working condition image", which can be in the form of a single-channel grayscale image or a multi-channel image;
[0038] S3, model construction;
[0039] In this step, the CNN model uses a 3x3 convolution kernel, replaces the traditional convolution with a depth separable convolution, and extracts local features from the working condition image;
[0040] Use the ReLU function to enhance non-linear expression;
[0041] Use MaxPool pooling (take the maximum value in the region) to reduce the dimensionality of the feature map after convolution and retain key features;
[0042] Flatten the pooled feature map into a one-dimensional vector and combine high-order features through a weight matrix;
[0043] For continuous parameter recommendation, use a linear activation function, and for parameter interval classification, use a Softmax activation function. At the same time, reduce the computational complexity by simplifying the number of convolution layers and using INT8 quantization compression technology to ensure that the single inference time is controlled within milliseconds, meeting the real-time requirements of industry;
[0044] S4, based on the working condition image in S2, the model in S3 is supervised training, and a trained model and a training history record are obtained;
[0045] In this step, the pre-processed working condition image is used as the input sample, the corresponding historical optimal control parameter is used as the label, the continuous parameter is directly used as the regression label, and the parameter interval is classified and labeled;
[0046] Adam or SGD optimization algorithm is used, and the loss function is selected according to the task type - mean square error for continuous value recommendation, and cross entropy for classification task; 20% of the data is divided into a validation set to prevent overfitting;
[0047] The input sample and the label are iteratively trained for 50 rounds, the network parameters are updated by the optimization algorithm to minimize the loss function, and a trained CNN model and a training history record are output;
[0048] S5, the trained model in S4 is deployed to the edge, the working condition image in S2 is used as a real-time parameter input recommendation model, and a recommended parameter is obtained;
[0049] In this step, the trained model is deployed on the industrial edge device, and data interaction channels are established with the DCS / PLC / SCADA system through industrial standard protocols such as OPCUA and Modbus, and real-time access to current working condition data is realized;
[0050] The received current working condition data is preprocessed and image converted repeatedly according to S2, and real-time working condition images are generated;
[0051] The real-time working condition image is expanded in batch dimension, input into the deployed CNN model for forward propagation inference, and the recommended parameter is output after the batch dimension is squeezed, and is fed back to the execution unit through the control system interface;
[0052] S6, record the actual production effect after the recommended parameter in S5 is executed; and calculate the error between the actual effect and the predicted effect; the sample containing the working condition image, the recommended parameter and the actual effect is marked as a to-be-learned sample and is transmitted to S4 to assist model iteration; at the same time, the fine-tuned optimization model is used to replace the model in S5;
[0053] In this step, the actual production indicators such as energy consumption, product qualification rate and yield after the recommended parameter is executed are recorded through sensors and monitoring systems, the "optimal parameter" is recalibrated, and the error between the actual effect and the model prediction expectation is calculated;
[0054] Set an error threshold, when the actual error exceeds the threshold, mark the sample as "to-be-learned sample"; in this embodiment, the error threshold is 5%;
[0055] Periodically aggregate the 'to-be-learned samples' to form an incremental training set, fine-tune the model using the small batch gradient descent method, fix the parameters of the bottom convolutional layer, only update the weights of the top fully connected layer, and use a low learning rate and a small number of rounds of training to avoid covering historical knowledge;
[0056] The fine-tuned model automatically replaces the old model deployed at the edge, realizes online evolution of the model, and improves long-term adaptability.
[0057] Therefore, the industrial control parameter intelligent optimization recommendation method based on the convolutional neural network realizes real-time response at the edge through the lightweight model, converts multi-source time series data into working condition images, triggers online fine-tuning of the model according to actual effect errors relying on a feedback mechanism, and realizes continuous self-optimization.
[0058] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application but not to limit it, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that: it can still modify or equivalently replace the technical solutions of the present application, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.
Claims
1. A method for intelligent optimization and recommendation of industrial control parameters based on convolutional neural networks, characterized in that, Includes the following steps: S1. Real-time acquisition and storage of multi-source operating condition data to obtain structured time series data; S2. Preprocess the structured time series data in S1, and convert the preprocessed structured time series data into working condition images; S3. Model building; S4. Based on the working condition images in S2, perform supervised training on the model in S3 to obtain the trained model and training history. S5. Deploy the model trained in S4 to the edge, and use the working condition images in S2 as real-time parameter inputs to obtain recommended parameters; S6. Record the actual production effect after the recommended parameters in S5 are implemented, and calculate the error between the actual effect and the predicted effect. The samples containing working condition images, recommended parameters, and actual effects are marked as samples to be learned and passed to S4 to assist model iteration; at the same time, the model in S5 is replaced with the fine-tuned optimized model.
2. The intelligent optimization and recommendation method for industrial control parameters based on convolutional neural networks according to claim 1, characterized in that, The S1 collects multi-source operating condition data in real time. Through the industrial IoT gateway, it collects multi-dimensional data in parallel at fixed sampling intervals. The multi-dimensional data includes control parameters, sensor variables, equipment status, product quality indicators, and external environmental information.
3. The intelligent optimization and recommendation method for industrial control parameters based on convolutional neural networks according to claim 1, characterized in that, The data preprocessing in S2 is as follows: outliers are removed by using the mean ± 3 standard deviation method, high-frequency noise is removed by decomposing the signal through wavelet transform, and the dimensions are unified by standardization method to map the data to the interval [-1,1].
4. The intelligent optimization and recommendation method for industrial control parameters based on convolutional neural networks according to claim 1, characterized in that, The recommended parameters are obtained in S5 by expanding the batch dimension of the real-time working condition image, inputting it into the deployed model for forward propagation inference, and then outputting the recommended parameters after compressing the batch dimension.
5. The intelligent optimization and recommendation method for industrial control parameters based on convolutional neural networks according to claim 1, characterized in that, In S6, the specific steps for marking a sample as a learning sample are as follows: set an error threshold, and when the actual error exceeds the error threshold, mark the sample as a learning sample.
6. The intelligent optimization and recommendation method for industrial control parameters based on convolutional neural networks according to claim 1, characterized in that, S6 also includes fine-tuning the model using mini-batch gradient descent; and when updating parameters, the parameters of the bottom convolutional layers are fixed, and only the weights of the top layer are updated.
7. An intelligent optimization and recommendation system for industrial control parameters based on convolutional neural networks, applied to the intelligent optimization and recommendation method for industrial control parameters based on convolutional neural networks as described in any one of claims 1-6, characterized in that, include: The multi-source operating condition data real-time acquisition and storage module is used to acquire and store multi-source operating condition data in real time to obtain structured time series data; The working condition image conversion module is connected to the multi-source working condition data real-time acquisition and storage module, and is used to preprocess the structured time series data and convert the preprocessed structured time series data into working condition images; The model building module, connected to the working condition image conversion module, is used to build the model; The model supervised training module is connected to the model building module. The model is trained under supervision using the base condition images to obtain the trained model and the training history. An edge deployment and real-time parameter recommendation module is connected to the model supervised training module. It is used to deploy the trained model to the edge and use the working condition image as a real-time parameter input to obtain recommended parameters. The feedback data acquisition and online model optimization module is connected to the edge deployment and real-time parameter recommendation module. It is used to record the actual production effect after the recommended parameters are executed and to calculate the error between the actual effect and the predicted effect. Samples containing working condition images, recommended parameters, and actual effects are labeled as samples to be learned and passed to the model supervised training module to assist model iteration; at the same time, the models in the edge deployment and real-time parameter recommendation modules are replaced with the fine-tuned optimized models.