An online quality prediction system
By working collaboratively between edge nodes and cloud nodes, and utilizing a variable-length sliding window mechanism and a long short-term memory recurrent neural network, the quality prediction model is dynamically updated, solving the problem of decreased accuracy when the production environment changes, and achieving efficient and accurate quality prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TIANJIN UNIV
- Filing Date
- 2023-03-23
- Publication Date
- 2026-07-31
AI Technical Summary
Existing quality prediction systems suffer from decreased model accuracy when the industrial production environment changes, requiring frequent retraining, which is time-consuming, labor-intensive, and difficult to guarantee accuracy.
The system, which adopts a dynamically updated quality prediction model, works collaboratively between edge nodes and cloud nodes, uses a variable-length sliding window mechanism to update model parameters, and combines a long short-term memory recurrent neural network to adapt to changes in the production environment in real time.
This improved the accuracy of the quality prediction model, reduced the computational load and network traffic required for model updates, and ensured the model's applicability and real-time prediction capabilities in new environments.
Smart Images
Figure CN116245421B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to an online quality prediction system, belonging to the field of industrial product quality prediction. Background Technology
[0002] Online quality prediction is the proactive prediction of product quality on manufacturing production lines. It involves predicting production quality parameters based on known production process parameters. Traditional quality control in manufacturing relies on post-production quality inspection, taking responsive measures only when non-conforming products are detected. By this time, product defects have already occurred, and corrective measures are costly. Quality prediction, however, can predict product quality in advance, allowing for timely correction of process parameters when quality warnings are issued, significantly improving production efficiency and reducing costs. With the development of the Industrial Internet and big data, industrial production processes generate massive amounts of data. Quality prediction is an effective way to utilize this industrial production big data, mining relationships between production and quality parameters from historical production data and applying these relationship models to predict quality in current production.
[0003] Currently, quality prediction mainly relies on machine learning algorithms such as neural networks, using historical data to train models, and then developing quality prediction systems based on these models for application in industrial fields. However, as the industrial production environment changes, the accuracy of the model gradually decreases until the prediction model becomes completely unsuitable for the current production environment. Therefore, it is necessary to frequently collect recently obtained historical data to retrain the model and redeploy it into the system. This method is time-consuming and labor-intensive, and the accuracy of model predictions is difficult to guarantee. Summary of the Invention
[0004] To address the issue of how to improve the accuracy of quality prediction, this invention provides an online quality prediction system that dynamically updates the quality prediction model, thereby improving the overall prediction accuracy of the model.
[0005] The present invention provides an online quality prediction system, the system comprising:
[0006] The data acquisition node unit is used to collect production parameter data and corresponding quality parameter true value data from the production site and send them to the edge node unit.
[0007] The edge node unit is used to preprocess the received production parameter data to obtain the input data of the quality prediction model, input it into the quality prediction model, and output the corresponding quality parameter prediction values. It is also used to send the input data of the quality prediction model and the corresponding true data of the quality parameters to the cloud node unit. It is also used to calculate the error between the true data of the quality parameters corresponding to the input data of the quality prediction model and the predicted value of the quality parameters, obtain the model prediction accuracy, and send it to the cloud node unit.
[0008] The cloud node unit is used to cache the input data of the received quality prediction model and the corresponding true data of the quality parameters, update the parameters of the quality prediction model using the cached data, and determine whether to push the quality prediction model down to the edge node unit for replacement based on the received model prediction accuracy.
[0009] Preferably, it includes multiple edge node units and acquisition node units. Each edge node unit and one acquisition node unit form a group for predicting the quality parameters of a production site. The cloud node unit is used to cache the input data of the quality prediction model of all groups and the corresponding true data of the quality parameters. Based on the prediction accuracy of each group's model, the cached data is used to update the parameters of the quality prediction model of each group. After the parameters are updated, the quality prediction model is sent down to the corresponding edge node unit for replacement.
[0010] Preferably, the cloud node unit updates the parameters of the quality prediction model by selecting cached data using a sliding window based on the received model prediction accuracy:
[0011] The sliding window consists of a variable length portion L c and fixed length part L m It consists of two parts:
[0012] L = L c +L m
[0013] The size of the variable length depends on the received model prediction accuracy. If the model prediction accuracy is within a high threshold range, the variable length portion L is increased. c If the model's prediction accuracy is within a low threshold range, then reduce the variable length portion L. c Discarding some historical data.
[0014] As a preferred option, the variable length portion L c for:
[0015]
[0016] The model prediction accuracy is the root mean square error R, L cmax For the variable length part L c The maximum value of R1 and R2 are thresholds; k represents the coefficient, k<0.
[0017] As a preferred option, the quality prediction model employs a long short-term memory recurrent neural network.
[0018] Preferably, the cloud node unit determines whether to push the quality prediction model down to the edge node unit for replacement based on the received model prediction accuracy, specifically including:
[0019] If the model prediction accuracy is within a preset threshold range where the accuracy decreases slightly, the latest quality prediction model will be deployed to the edge node unit for replacement after the quality prediction model parameters are updated.
[0020] If the model's prediction accuracy falls within a preset threshold range indicating a decrease in severity, the currently updated quality prediction model will be immediately deployed to the edge node unit for replacement.
[0021] Preferably, the system also includes a visualization terminal for real-time display of quality prediction results.
[0022] As a preferred option, the cloud node unit is also equipped with a historical database to store the inputs and outputs of the quality prediction model;
[0023] The visualization terminal is also used to trace historical data by accessing historical databases.
[0024] As a preferred option, the visualization terminal is also used to manage each node unit and the quality prediction model.
[0025] Preferably, preprocessing includes handling missing values and normalization.
[0026] The beneficial effects of this invention are that it fully utilizes the computing and storage capabilities of massive cloud data and the advantages of low latency and fast response of edge hosts. This ensures timely prediction of product quality on the production line, enables dynamic updates to the quality prediction model, improves the accuracy of the model, and expands data storage capabilities. The variable-length sliding window mechanism ensures that the quality prediction model can be dynamically updated using the latest data based on its adaptability to the production environment, discarding some low-value data from the production process. This reduces the computational load of model update training and improves the prediction accuracy of the updated model in the new production environment. The visualization terminal of this invention allows staff to view online quality prediction curves on-site, as well as remotely view and analyze historical data, manage various nodes, and manage user permissions. Attached Figure Description
[0027] Figure 1 This is an overall architecture diagram of the present invention;
[0028] Figure 2 This is a diagram showing the specific work content of each node;
[0029] Figure 3 This is a schematic diagram of a variable-length sliding window;
[0030] Figure 4 This is a schematic diagram of the prediction model decentralization strategy;
[0031] Among them, 1-cloud node unit, 2-sliding window, 3-edge node unit, 4-visual terminal, and 5-collection node unit. Detailed Implementation
[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.
[0033] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.
[0034] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.
[0035] like Figure 1 and Figure 2 As shown, the online quality prediction system of this embodiment includes:
[0036] The data acquisition node unit is used to collect production parameter data and corresponding quality parameter true value data from the production site and send them to the edge node unit.
[0037] In this embodiment, the data acquisition node unit is deployed at the industrial production site. It uses sensors to collect production process parameters, environmental parameters, and other parameters that affect production quality in real time, while recording the corresponding product quality parameters. There is a one-to-one correspondence between a set of production parameters and one or a set of quality parameters. The data acquisition node unit marks the corresponding production parameters and quality parameters with the same flag and sends them to the edge node unit in sequence. The communication between the data acquisition node unit and the edge node unit can be wired or wireless, depending on the situation.
[0038] Before building a quality prediction model, sufficient historical data needs to be collected for model training. The collected data should be arranged in chronological order and the order should not be disrupted. During the online prediction phase, the data collection node sends the production parameters and corresponding quality parameters to the edge node in sequence for online quality prediction and calculation of the model prediction accuracy.
[0039] The edge node unit is used to preprocess the received production parameter data to obtain the input data of the quality prediction model, input it into the quality prediction model, and output the corresponding quality parameter prediction values. It is also used to send the input data of the quality prediction model and the corresponding true data of the quality parameters to the cloud node unit. It is also used to calculate the error between the true data of the quality parameters corresponding to the input data of the quality prediction model and the predicted value of the quality parameters, obtain the model prediction accuracy, and send it to the cloud node unit.
[0040] In this embodiment, the edge node unit is deployed close to the data source, resulting in a shorter data transmission link and lower data transmission latency. This allows for real-time calculation and display of data. However, its computing and data storage capabilities are limited, making it unsuitable for tasks involving large-scale computations or storing massive amounts of data. Therefore, the edge node unit is responsible for the forward propagation calculation of the model, performing real-time quality prediction and data preprocessing. The edge node unit preprocesses the collected raw data, including handling missing values and normalization. It uses the preprocessed production parameter data as input to the quality prediction model to predict production quality in real time. Simultaneously, it calculates the model prediction accuracy of the current quality prediction model based on the predicted quality value and the actual quality value transmitted by the data acquisition node. The calculated model prediction accuracy is then transmitted to the cloud node.
[0041] In this embodiment, the RMSE (Root Mean Square Error) is used as the metric to evaluate the prediction accuracy of the model. The edge node unit transmits the RMSE of the most recent N sets of data to the cloud node unit.
[0042] The cloud node unit is used to cache the input data of the received quality prediction model and the corresponding true data of the quality parameters, update the parameters of the quality prediction model using the cached data, and determine whether to push the quality prediction model down to the edge node unit for replacement based on the received model prediction accuracy.
[0043] In this embodiment, the cloud node unit is located in the cloud and connected to various edge nodes via network links. The cloud node has strong computing and data storage capabilities, enabling the computation and storage of massive amounts of data. However, due to network transmission limitations, it experiences higher latency compared to the work site, preventing real-time data processing. The cloud node unit receives data from the edge nodes, caches the data, and uses the cached data to update and iterate the quality prediction model in the cloud in real time. When the prediction accuracy of the edge node's prediction model decreases, the latest quality prediction model is promptly deployed to the edge node to adapt to changes in the production environment and improve the accuracy of model predictions. Simultaneously, a historical database is established on the cloud node to facilitate the traceability of historical data.
[0044] In a preferred embodiment, the system of this implementation may further include multiple edge node units and acquisition node units. Each edge node unit and one acquisition node unit form a group for predicting the quality parameters of a production site. For example, each production line or production workshop may be configured with one group. The cloud node unit is used to cache the input data of the quality prediction model of all groups and the corresponding true value data of the quality parameters. Based on the prediction accuracy of each group's model, the cached data is used to update the parameters of the quality prediction model of each group. After the parameters are updated, the quality prediction model is sent down to the corresponding edge node unit for replacement.
[0045] In a preferred embodiment, the cloud node unit updates the parameters of the quality prediction model by selecting cached data using a sliding window based on the received model prediction accuracy.
[0046] Figure 3 This is a schematic diagram of a variable-length sliding window deployed on a cloud node. The sliding window mechanism is applied to the cloud node's data cache. Each time, the cloud node retrieves data from the sliding window to update and train the quality prediction model, obtaining the latest model. The sliding window consists of a variable-length portion L. c and fixed length part L m It consists of two parts:
[0047] L = L c +L m
[0048] The size of the variable length depends on the received model prediction accuracy. If the model prediction accuracy is within a high threshold range, the variable length portion L is increased. c If the model's prediction accuracy is within a low threshold range, then reduce the variable length portion L. c Discarding some historical data.
[0049] The fixed-length portion represents the minimum length of the entire sliding window. Its size must be carefully considered. If the length is too small, outliers in individual data points can cause significant changes in the model's training gradient, leading to decreased prediction accuracy. If the length is too large, it increases the training time for a single batch of data, making it difficult to promptly deploy the latest trained model to edge nodes when their prediction accuracy declines. The variable-length portion depends on the prediction accuracy of the edge node models. Increasing the variable-length window size allows for more historical data to be included in the cloud-based model training, ensuring model stability. Simultaneously, higher current model accuracy allows for more training time. Conversely, decreasing the variable-length sliding window size discards some historical data, allowing the latest data to be used to update the model, improving its applicability in the current production environment, and reducing training time to facilitate timely deployment of the latest quality prediction model to edge nodes.
[0050] The length of the variable-length sliding window depends on the RMSE predicted by the edge node model, and the variable-length part L c for:
[0051]
[0052] The model prediction accuracy is the root mean square error R, which is determined based on factors such as cloud computing power. cmax For the variable length part L c The maximum value of L, where R1 and R2 are thresholds; k represents a coefficient, and when the root mean square error is greater than R1, L... c When L is fixed at its maximum value, and the root mean square error is less than R², c When k is 0, and k < 0, the window length is inversely proportional to the root mean square error between R1 and R2.
[0053] In a preferred embodiment, the quality prediction model employs a long short-term memory recurrent neural network.
[0054] Quality prediction should consider the temporal characteristics of industrial product production data, perceive the trend of product quality changes, and since there are many factors affecting product quality, it is impossible to describe them all quantitatively or qualitatively. Moreover, most production parameters are set manually, and subtle changes during the production process cannot be perceived. There is a certain upper limit to establishing a mapping relationship between production parameters and quality parameters. The LSTM algorithm is more suitable for modeling temporal characteristics. At the same time, using the quality output parameters of the previous time step as the input of the prediction model at the current time step improves the upper limit of the quality prediction model.
[0055] In a preferred embodiment, the cloud node unit determines whether to push the quality prediction model down to the edge node unit for replacement based on the received model prediction accuracy, specifically including:
[0056] If the model's prediction accuracy is within a preset threshold range where there is a slight decrease (e.g., greater than R), b Less than R a When the quality prediction model parameters are updated, the latest quality prediction model will be deployed to the edge node unit for replacement.
[0057] If the model's prediction accuracy is within a preset threshold range for the severity of the decrease (e.g., less than R), b When the time comes, the currently updated quality prediction model will be immediately deployed to the edge node unit for replacement.
[0058] In this implementation, when the prediction accuracy of the edge node unit's quality prediction model decreases to a certain extent, the model trained using the latest data is deployed after the training of the model currently being trained on the cloud node unit is completed. When the accuracy of the quality prediction model is very low, it indicates that the quality prediction model is no longer suitable for the current production environment and needs to be updated immediately. Therefore, the quality prediction model stored in the cloud is immediately deployed to the edge nodes. This method reduces the number of model deployments, reduces network traffic, and allows the model's prediction accuracy to remain within a relatively ideal range for a longer period, thereby improving the model's prediction performance.
[0059] The visualization terminal in this embodiment is a visualization interface. Mobile and PC visualization software or web pages can be developed simultaneously. The visualization terminal can communicate with the field acquisition node unit, edge node unit, and cloud node unit respectively, so that operators can view the status of the quality prediction system. The functions implemented are: the prediction results of the edge node unit are sent to the visualization terminal to generate the quality change curve. The visualization terminal displays the quality prediction curve on the monitoring interface in real time, and can also display the actual quality value for comparison.
[0060] The cloud node unit is also equipped with a historical database to store the inputs and outputs of the quality prediction model. The visualization terminal in this embodiment can also perform the following functions: data tracing of the historical database and data visualization of historical data in the cloud through various statistical charts.
[0061] The visualization terminal in this embodiment can also perform the following functions: manage each node unit, add or delete edge nodes, and set the connection between edge nodes and cloud nodes; in addition, the visualization terminal can also manage the quality prediction model, set system parameters, and manage user permissions.
[0062] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.
Claims
1. An online quality prediction system, characterized by, The system includes; The data acquisition node unit is used to collect production parameter data and corresponding quality parameter true value data from the production site and send them to the edge node unit. The edge node unit is used to preprocess the received production parameter data to obtain the input data for the quality prediction model, which is then input into the quality prediction model, and the quality prediction model outputs the corresponding predicted values of the quality parameters. It is also used to send the input data of the quality prediction model and the corresponding true values of the quality parameters to the cloud node unit; and to perform error calculation between the true values of the quality parameters corresponding to the input data of the quality prediction model and the predicted values of the quality parameters to obtain the model prediction accuracy, which is then sent to the cloud node unit. The cloud node unit is used to cache the input data of the received quality prediction model and the corresponding true data of the quality parameters, update the parameters of the quality prediction model using the cached data, and determine whether to push the quality prediction model down to the edge node unit for replacement based on the received model prediction accuracy. The cloud node unit updates the parameters of the quality prediction model by selecting cached data using a sliding window based on the received model prediction accuracy. The sliding window is composed of a variable length portion and a fixed length portion both portions composition: The size of the variable length depends on the received model prediction accuracy, if the model prediction accuracy is in a high threshold range, then the variable length part is increased , if the model prediction accuracy is in a low threshold range, then the variable length part is decreased , discarding part of the historical data.
2. The online quality prediction system according to claim 1, characterized in that, It includes multiple edge node units and acquisition node units. Each edge node unit and one acquisition node unit form a group for predicting the quality parameters of a production site. The cloud node unit is used to cache the input data of the quality prediction model of all groups and the corresponding true data of the quality parameters. Based on the prediction accuracy of each group's model, the cached data is used to update the parameters of the quality prediction model of each group. After the parameters are updated, the quality prediction model is sent down to the corresponding edge node unit for replacement.
3. The online quality prediction system according to claim 1, characterized in that, Variable length portion for: The model's prediction accuracy is the root mean square error. , For variable length part The maximum value, , It is a threshold; Represents the coefficient. .
4. The online quality prediction system according to claim 1, characterized in that, The quality prediction model employs a long short-term memory recurrent neural network.
5. The online quality prediction system according to claim 1, characterized in that, The cloud node unit determines whether to push the quality prediction model down to the edge node unit for replacement based on the received model prediction accuracy, specifically including: If the model prediction accuracy is within a preset threshold range where the accuracy decreases slightly, the latest quality prediction model will be deployed to the edge node unit for replacement after the quality prediction model parameters are updated. If the model's prediction accuracy falls within a preset threshold range indicating a decrease in severity, the currently updated quality prediction model will be immediately deployed to the edge node unit for replacement.
6. The online quality prediction system according to claim 2, characterized in that, The system also includes a visualization terminal, which is used to display the quality prediction results in real time.
7. The online quality prediction system according to claim 6, characterized in that, The cloud node unit is also equipped with a historical database for storing the inputs and outputs of the quality prediction model; The visualization terminal is also used to trace historical data by accessing the historical database.
8. The online quality prediction system according to claim 6, characterized in that, The visualization terminal is also used to manage each node unit and the quality prediction model.
9. The online quality prediction system according to claim 1, characterized in that, The preprocessing includes handling missing values and normalization.