Production data management system based on artificial intelligence

Abstract

The present invention provides a production data management system based on artificial intelligence, which relates to the field of data processing technology. The system includes a production data acquisition module, a multi-dimensional time series feature extraction module, a dynamic manifold analysis module, a production path optimization module, a vulcanization process anomaly analysis module and a vulcanization process optimization module. The present invention generates a production manifold feature space containing the complex dynamic characteristics of the vulcanization process by performing multi-dimensional time series feature extraction and dynamic manifold analysis on the key data of the vulcanization process acquired by the data acquisition module, and constructs a target vulcanization process path and process interval based on the feature space to provide a reference for subsequent process deviation detection; the deviation vector is used to detect process anomalies, and the inverse mapping model is combined to achieve accurate restoration and tracing of process parameters, ensuring that the production process is always maintained within the optimal path, making the vulcanization process more intelligent and automated.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a production data management system based on artificial intelligence. Background Art

[0002] In the field of automotive parts production, especially in the manufacturing of key parts such as wipers and filters, the vulcanization process of rubber materials has a direct impact on product performance and service life. The vulcanization process is an important part of rubber material processing. By controlling process parameters such as temperature, pressure and time, it ensures that the rubber material meets strict quality standards while achieving ideal hardness, elasticity and wear resistance. However, since the characteristics of the rubber material itself and the process parameters in the vulcanization process are easily affected by various factors such as environmental fluctuations and equipment status, the vulcanization process often exhibits nonlinear, time-varying and complex dynamic characteristics.

[0003] Some vulcanization process control methods rely on preset process parameters and static quality control in data analysis, and lack the ability to conduct in-depth analysis and optimization of multi-dimensional process data. Effectively converting the historical production data of rubber materials into a predictive and guiding process optimization strategy is an important means to solve the process deviation problem in the vulcanization process. Building an intelligent system that can monitor, analyze and optimize the vulcanization process in real time is one of the problems that need to be solved in the current automotive parts manufacturing industry. Summary of the invention

[0004] In order to solve the above technical problems, the present invention proposes a production data management system based on artificial intelligence, which focuses on the optimization of the vulcanization process of rubber materials. By performing dynamic manifold analysis on the multi-dimensional time series data in the vulcanization process, the ideal path interval of the vulcanization process is identified and constructed, and deviation detection and abnormal parameter positioning of the vulcanization process are carried out in a targeted manner. An optimization scheme is generated through inverse mapping, thereby realizing dynamic optimization of process parameters. It can effectively cope with the complex nonlinear characteristics in the vulcanization process, ensure that the production process runs along the optimal path, and improve the stability of product quality and production efficiency.

[0005] To achieve the above objectives, the present invention provides a production data management system based on artificial intelligence, comprising:

[0006] The production data collection module is used to collect the historical production data of rubber materials, including the production process data and production quality data of rubber materials in the vulcanization process;

[0007] The multi-dimensional time series feature extraction module is used to perform data cleaning and sliding window analysis on the historical production data of rubber materials and generate a multi-dimensional production feature matrix corresponding to the historical production data;

[0008] A dynamic manifold analysis module is used to perform dynamic manifold analysis on a multi-dimensional production feature matrix through a time-series dynamic manifold analysis model, including using an improved UMAP algorithm based on time perception to process the multi-dimensional production feature matrix and generate a production manifold feature space corresponding to the multi-dimensional production feature matrix;

[0009] The production path optimization module is used to perform curve fitting on the production manifold feature space to generate a target vulcanization process path, and determine a target vulcanization process interval according to the target vulcanization process path;

[0010] The vulcanization process anomaly analysis module is used to perform process deviation detection on the collected target vulcanization production data based on the target vulcanization process interval, locate the abnormal parameters after the calculated process deviation parameter is greater than the preset deviation threshold, and generate a deviation detection result including the abnormal parameters and the deviation feature vector;

[0011] The vulcanization process optimization module is used to perform inverse mapping processing on the deviated feature vector through the inverse mapping model, generate the target production feature vector corresponding to the deviated feature vector, generate the vulcanization process optimization plan according to the target production feature vector and the abnormal parameters, and optimize the vulcanization process based on the vulcanization process optimization plan.

[0012] Preferably, an improved UMAP algorithm based on time perception is used to process the multidimensional production feature matrix to generate a production manifold feature space corresponding to the multidimensional production feature matrix, including:

[0013] Calculate the distance parameter between any two data points in the multi-dimensional production feature matrix according to the UMAP algorithm, construct a distance matrix corresponding to the multi-dimensional production feature matrix according to multiple distance parameters, and determine the distance neighborhood graph according to the distance matrix;

[0014] Time perception is introduced into the distance neighborhood graph, including determining the time weight between any two data points based on the target time function, optimizing the distance neighborhood graph according to the time weight to generate a spatiotemporal neighborhood graph. The weight between data points in the spatiotemporal neighborhood graph is:

[0015] ;

[0016] In the formula, is the first and The weights between the data points, is an element in the distance matrix, indicating the and The spatial distance between data points, Represents the first and The distance between data points in time steps, , are the spatial scale parameter and the temporal scale parameter respectively;

[0017] Construct the target neighborhood graph corresponding to the spatiotemporal neighborhood graph, and determine the target optimization function based on the spatiotemporal neighborhood graph and the target neighborhood graph :

[0018] ;

[0019] In the formula, is the first and The weights between data points;

[0020] Based on the objective optimization function, the gradient descent algorithm is used to iteratively optimize the positions of multiple data points in the multidimensional production feature matrix to obtain the position of each data point in the multidimensional production feature matrix in the low-dimensional space. According to the positions of multiple data points in the low-dimensional space, the production manifold feature space corresponding to the multidimensional production feature matrix is ​​constructed.

[0021] Preferably, curve fitting is performed on the production manifold feature space to generate a target vulcanization process path, and a target vulcanization process interval is determined according to the target vulcanization process path, including:

[0022] The production manifold feature space is clustered based on the DBSCAN algorithm, and the production manifold feature space is divided into multiple sub-feature spaces. Bezier curve fitting is performed on multiple data points in each sub-feature space to obtain the local Bezier curve of each sub-feature space.

[0023] Analyze multiple data points in each sub-feature space, determine the density parameter and distance parameter of each sub-feature space, determine the target weight of each sub-feature space according to the density parameter and distance parameter, fuse the local Bezier curve of each sub-feature space according to the target weight, and generate the target vulcanization process path corresponding to the production manifold feature space;

[0024] Each sub-feature space is analyzed according to the target vulcanization process path, and the distances between multiple data points in each sub-feature space and the target vulcanization process path on each low-dimensional feature parameter are determined, so as to obtain local vulcanization process intervals containing distances corresponding to multiple low-dimensional feature parameters, and generate a target vulcanization process interval containing multiple local vulcanization process intervals.

[0025] Preferably, the process deviation detection is performed on the collected target vulcanization production data based on the target vulcanization process interval, including:

[0026] Inputting the target vulcanization production data into the time-series dynamic manifold analysis model to generate the target manifold features, respectively analyzing the association between the target process features and multiple sub-feature spaces and determining the target sub-feature space;

[0027] Performing a difference analysis on a target vulcanization process path based on a target sub-feature space to generate a first manifold deviation vector corresponding to the target vulcanization process path, and performing a difference analysis on a target manifold feature based on a target sub-feature space to generate a second manifold deviation vector corresponding to the target manifold feature;

[0028] The process deviation detection is performed on the target vulcanization production data according to the first manifold deviation vector and the second manifold deviation vector, including calculating the distance between the first manifold deviation vector and the second manifold deviation vector and using it as the process deviation parameter of the target vulcanization production data.

[0029] Preferably, the inverse mapping model further includes:

[0030] A sample data set for training an inverse mapping model is constructed based on a multidimensional production feature matrix and a production manifold feature space, a plurality of data points in the production manifold feature space are used as inputs of the inverse mapping model, and data points corresponding to the plurality of data points in the production manifold feature space in the multidimensional production feature matrix are used as training targets, and an inverse mapping model is obtained through training;

[0031] After generating a deviation feature vector according to the first manifold deviation vector and the second manifold deviation vector, the deviation feature vector is input into the inverse mapping model to generate a target production feature vector corresponding to the deviation feature vector.

[0032] Preferably, the target weight of the sub-feature space further includes:

[0033] The ratio of the number of data points in the sub-feature space to the total number of data points in the production manifold feature space is recorded as the density parameter, and the mean of the distances between multiple data points in the sub-feature space and the center point of the sub-feature space is recorded as the distance parameter. Multiple density parameters and distance parameters are normalized respectively, and the sum of the normalized density parameters and distance parameters of each sub-feature space is taken as the target weight of the sub-feature space.

[0034] Preferably, the distances between the target process feature and the center points of the plurality of sub-feature spaces are calculated respectively, and the sub-feature space with the shortest distance to the target process feature is selected and recorded as the target sub-feature space.

[0035] Preferably, an inverse mapping model is constructed based on a GAN network.

[0036] The present invention has the following beneficial effects:

[0037] The present invention comprehensively monitors and optimizes the vulcanization production data of rubber materials, and realizes the whole process management from data collection, feature extraction, process path optimization to anomaly detection and intelligent adjustment. The data acquisition module obtains the key data of the vulcanization process, and generates the production manifold feature space containing the complex dynamic characteristics of the vulcanization process through multi-dimensional time series feature extraction and dynamic manifold analysis. Based on the feature space, the target vulcanization process path and process interval are constructed to provide a reference for subsequent process deviation detection; the vulcanization process anomaly analysis module uses the deviation vector to detect process anomalies on the basis of real-time monitoring of production data, and combines the inverse mapping model to realize the accurate restoration and tracing of process parameters, ensuring that the production process is always maintained within the optimal path. The system effectively identifies and locates process deviations through quantitative analysis of process deviation detection parameters, adjusts production parameters in time, and finally improves the stability of the production process and the consistency of product quality. The multi-dimensional analysis and dynamic optimization capabilities of the system make the vulcanization process more intelligent and automated. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 A schematic diagram of the structure of an artificial intelligence-based production data management system provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0039] In order to make those skilled in the art better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0040] Figure 1 A schematic diagram of the structure of a production data management system based on artificial intelligence provided by an embodiment of the present invention is shown in FIG. Figure 1 As shown, the system includes a production data acquisition module, a multi-dimensional time series feature extraction module, a dynamic manifold analysis module, a production path optimization module, a vulcanization process abnormality analysis module and a vulcanization process optimization module. Specifically:

[0041] The production data collection module is used to collect the historical production data of rubber materials, including the production process data and production quality data of rubber materials in the vulcanization process.

[0042] Specifically, the historical production data of rubber materials collected include but are not limited to key parameters such as temperature, pressure, and time during the vulcanization process, such as the collection of relevant production data through temperature sensors installed in the heating area of ​​the vulcanizer, pressure sensors at the pressure chamber position, and other equipment, as well as production quality data such as performance indicators such as rubber hardness, elasticity, and wear resistance, such as the detection of product production quality through quality monitoring equipment such as hardness testers and elastic modulus testers. For the collected historical production data, the sensor data is transmitted to the edge computing device in real time or periodically using the industrial Internet of Things protocol, and uploaded to the database after local processing such as data cleaning and denoising, so as to conduct in-depth analysis of the data later, providing a reliable data foundation for the intelligent analysis and optimization of the vulcanization process.

[0043] The multi-dimensional time series feature extraction module is used to perform data cleaning and sliding window analysis on the historical production data of rubber materials and generate a multi-dimensional production feature matrix corresponding to the historical production data.

[0044] Specifically, after the historical production data of rubber materials is collected through the production data collection module, the data cleaning technology used in the preprocessing process can remove noise data and abnormal data to ensure data quality. Then, the sliding window analysis method is used to divide the time series production data in the vulcanization process into adjacent window segments, thereby retaining the dynamic change characteristics of the production process, constructing the multidimensional feature vector corresponding to each set of data, and generating the corresponding multidimensional production feature matrix, which contains the relevant information between the process and quality parameters in the vulcanization process.

[0045] The dynamic manifold analysis module is used to perform dynamic manifold analysis on the multidimensional production feature matrix generated by the multidimensional time series feature extraction module through the time series dynamic manifold analysis model, including using the improved UMAP algorithm based on time perception to process the multidimensional production feature matrix and generate the production manifold feature space corresponding to the multidimensional production feature matrix.

[0046] Specifically, in the vulcanization process of rubber materials, there are complex nonlinear relationships between process parameters such as temperature, pressure, elastic modulus, and hardness. In order to reveal these nonlinear relationships and extract the implicit features in the vulcanization process, a nonlinear dimensionality reduction technique is used to map the multidimensional time series matrix into a latent space. In this embodiment, the multidimensional production feature matrix is ​​processed based on the time series dynamic manifold analysis model to reveal the potential complex dynamic features in the production data.

[0047] Among them, the time series dynamic manifold analysis model is based on the UMAP algorithm (Uniform Manifold Approximation and Projection). UMAP is a manifold learning technology that can map high-dimensional data to low-dimensional space while retaining the global and local structure of the data. The core idea of ​​manifold learning is that high-dimensional data actually exists on a low-dimensional manifold. The goal of manifold learning is to find this low-dimensional manifold through nonlinear dimensionality reduction methods. As a common manifold learning method, UMAP performs well when processing static data. However, the data of the vulcanization process is a dynamic time series, so time-aware manifold learning is adopted to introduce time dependencies into traditional UMAP to retain the time series characteristics of the vulcanization process. The multi-dimensional production feature matrix is ​​projected into a production manifold feature space through the time series dynamic manifold analysis model to capture nonlinear changes in the production process.

[0048] It is worth noting that the production manifold feature space contains the state information of the vulcanization process at different times of historical production data. Each node represents the state of the vulcanization process at a certain moment, and the distance between different nodes reflects the state change of the vulcanization process at different time steps. The structure and shape of the production manifold feature space reflect the global dynamic evolution law of the vulcanization process. Through time-aware manifold learning, the comprehensive change trends of implicit features in the vulcanization process such as temperature, pressure, hardness, etc. are extracted and expressed as low-dimensional feature representations in the latent space, which facilitates a better understanding of the global changes in the vulcanization process.

[0049] The production path optimization module is used to perform curve fitting on the production manifold feature space generated by the dynamic manifold analysis module to generate a target vulcanization process path, and determine a target vulcanization process interval based on the target vulcanization process path.

[0050] Specifically, the production path optimization module generates the target vulcanization process path through curve fitting technology based on the production manifold feature space generated by the dynamic manifold analysis module, forming an evolution trajectory that describes the ideal vulcanization process. Path fitting in the low-dimensional feature space can more effectively identify the optimal process path in the vulcanization process, avoiding the difficulty of processing complex nonlinear data in the original high-dimensional space. According to the target vulcanization process path, a target vulcanization process interval can be further determined to provide a dynamic reference standard for subsequent deviation detection, thereby ensuring that the production process can fluctuate within a reasonable range and improving the stability of the process.

[0051] It is worth noting that the target vulcanization process path refers to the process parameters in the vulcanization process being adjusted according to the ideal dynamic change mode, so that the properties of the rubber, such as hardness and elastic modulus, reach the best state. Through path fitting, we can model the historical data of the vulcanization process and identify those paths that show excellent quality in the latent space. In the fitting process, we can use the successful vulcanization processes in the historical data, that is, the processes with the best product quality, as a reference, and find the optimal path by fitting these successful data points. Considering that in actual production, the process parameters will not change strictly according to a certain path, but there is a certain range of fluctuations. Therefore, a target vulcanization process interval is defined around the fitted optimal path. This interval reflects the reasonable fluctuation range of the process parameters, and its role is to allow the process parameters to fluctuate within a certain range without significantly affecting the product quality. When the process parameters exceed this interval, it can be considered that the vulcanization process deviates from the ideal path, which may cause product quality problems.

[0052] The vulcanization process anomaly analysis module is used to perform process deviation detection on the collected target vulcanization production data based on the target vulcanization process interval determined by the production path optimization module, locate the abnormal parameters after the calculated process deviation parameter is greater than the preset deviation threshold, and generate a deviation detection result including the abnormal parameters and the deviation feature vector.

[0053] Specifically, the vulcanization process anomaly analysis module monitors the collected target vulcanization production data in real time based on the target vulcanization process interval, performs deviation detection and anomaly analysis, and determines whether the vulcanization process deviates from the ideal process trajectory. Once the deviation parameter exceeds the preset deviation threshold, the module will trigger the anomaly detection mechanism, and can further analyze the abnormal conditions of specific parameters such as temperature or pressure during the vulcanization process to determine the abnormal process parameters or quality parameters, quantify the deviation between the target vulcanization production data and the ideal path to generate the corresponding deviation feature vector, and finally generate a detection result containing abnormal parameters and deviation feature vectors, providing real-time feedback for further optimization of the vulcanization process.

[0054] The vulcanization process optimization module is used to process the deviation detection results generated by the vulcanization process anomaly analysis module through an inverse mapping model, including inverse mapping of the deviation feature vector, generating a target production feature vector corresponding to the deviation feature vector, generating a vulcanization process optimization plan according to the target production feature vector and the abnormal parameters, and optimizing the vulcanization process based on the vulcanization process optimization plan.

[0055] Specifically, the vulcanization process optimization module is used to perform intelligent optimization and adjustment for deviations in the vulcanization process. Specifically, the deviation feature vector generated during the deviation detection process is inversely mapped to the original high-dimensional process parameter space through the inverse mapping model to generate the target production feature vector corresponding to the deviation feature vector. Based on the target feature vector and abnormal parameters, the corresponding optimization scheme can be determined for the abnormal parameters in the data, including adjusting the vulcanization process parameters such as temperature, pressure, and time according to the deviation of the parameters, so that the vulcanization process returns to the target path interval, thereby ensuring that the vulcanization process can continue to run along the optimal path, ensuring product quality consistency and stability of production efficiency, and realizing intelligent optimization of the vulcanization process.

[0056] The embodiment of the present invention provides an artificial intelligence-based production data management system, which collects process parameters and production quality data on the production line through a production data acquisition module, specifically including sensors distributed in different parts of the vulcanizer to collect process parameter data, and automatic detection equipment and laboratory sampling detection equipment distributed at the end of the production line to collect production quality data, which are uploaded to the database in the cloud server after preliminary preprocessing. The multidimensional time series feature extraction module, the dynamic manifold analysis module and the production path optimization module can be deployed in the cloud server. The multidimensional time series feature extraction module connected to the production data acquisition module processes the data in the database and transmits the generated multidimensional production feature matrix to the dynamic manifold analysis module. The dynamic manifold analysis module receives the multidimensional production feature matrix from the multidimensional time series feature extraction module and analyzes it to generate a production manifold feature space and transmits it to the production path optimization module. The production path optimization module receives the production manifold feature space generated by the dynamic manifold analysis module and analyzes it, outputs the target vulcanization process path and target process interval, and transmits the target vulcanization process path and process interval to the vulcanization process anomaly analysis module through the service interface. The vulcanization process anomaly analysis module deployed on the edge computing device analyzes the target vulcanization production data collected in real time based on the target vulcanization process path and target process interval, generates a deviation detection report and sends it to the vulcanization process optimization module. The vulcanization process optimization module deployed on the edge computing device and connected to the production control system generates a corresponding vulcanization process optimization plan after processing the deviation detection report, and then sends the vulcanization process optimization plan to the production line equipment to optimize the vulcanization process.

[0057] For example, the sensor detects that the current temperature is 165°C, and the hardness tester detects that the hardness is 75 Shore A. After analyzing the historical production data of rubber materials, the ideal production path is determined to have a temperature of 175°C-185°C and a hardness of 85 Shore A. After determining that the hardness of the vulcanized material is abnormal, it is detected that the vulcanization has deviated. The corresponding optimization plan is generated through inverse mapping, such as increasing the heating power from 75% to 80% to compensate for the temperature difference, and restoring the current production path to the target path, realizing intelligent optimization of the vulcanization process and providing product quality assurance for the vulcanization production process of rubber products.

[0058] Furthermore, in the process of dynamic manifold analysis on the multidimensional production feature matrix through the time-series dynamic manifold analysis model, the dynamic manifold analysis module uses the improved UMAP algorithm based on time perception to process the multidimensional production feature matrix and generate the production manifold feature space corresponding to the multidimensional production feature matrix, which specifically includes:

[0059] According to the UMAP algorithm, the distance parameter between any two data points in the multidimensional production feature matrix is ​​calculated, and the distance matrix corresponding to the multidimensional production feature matrix is ​​constructed according to multiple distance parameters. The distance neighborhood graph is determined according to the distance matrix. It is worth noting that when analyzing the distance between data points, the traditional UMAP algorithm considers that the distance metric varies between different regions, rather than being universal in the entire space. The UMAP algorithm can extract the distance parameters between different data points in the multidimensional production feature matrix. In this way, a distance neighborhood graph containing the spatial distance information between data can be constructed.

[0060] On this basis, time perception is introduced into the distance neighborhood graph, including the introduction of a target time function, determining the time weight between any two data points based on the target time function, and optimizing the distance neighborhood graph according to the time weight to generate a spatiotemporal neighborhood graph, measuring the joint influence of spatial distance and time step, so as to enhance the feature retention capability in the time dimension. This ensures that the distance matrix takes into account both spatial proximity and temporal similarity when constructing the neighborhood graph. This optimization retains the evolution information of data points between different time steps, so that the time series characteristics of data points can be captured in the reduced-dimensional space.

[0061] Among them, the weights between data points in the spatiotemporal neighborhood graph are:

[0062] ;

[0063] In the formula, is the first and The weights between the data points, is an element in the distance matrix, indicating the and The spatial distance between data points, Represents the first and The distance between data points in time steps, , are the spatial scale parameter and the temporal scale parameter respectively.

[0064] After constructing the spatiotemporal neighborhood graph based on time and space features through the above process, we further construct the target neighborhood graph corresponding to the spatiotemporal neighborhood graph. Each edge in the target neighborhood graph represents the similarity between two data points in the multidimensional production feature matrix in the low-dimensional space. We use the target optimization function to optimize the edge weights of the target neighborhood graph to obtain a more reasonable low-dimensional space representation. Among them, the target optimization function determined according to the spatiotemporal neighborhood graph and the target neighborhood graph is as follows:

[0065] ;

[0066] In the formula, is the first and The objective optimization function measures the difference between the source neighborhood graph, i.e., the spatiotemporal neighborhood graph, and the target neighborhood graph by Laplace smoothing and cross entropy metric, thereby optimizing the distribution of data points in low-dimensional space. In this way, by minimizing the objective optimization function , it is possible to effectively reduce the dimension of multiple data points in the production feature matrix while maintaining the spatial and temporal relationship of the data.

[0067] In order to achieve this dimensionality reduction process, based on the target optimization function, the gradient descent method is used to iteratively optimize the positions of multiple data points in the multidimensional production feature matrix. During each iterative update, the position of the data point in the low-dimensional space is gradually adjusted according to the current gradient, so that its position in the low-dimensional space can better reflect the spatiotemporal neighborhood relationship in the high-dimensional space. After obtaining the optimal representation of each data point in the multidimensional production feature matrix in the low-dimensional space in this way, the production manifold feature space with time dynamics is generated through the low-dimensional representation. This feature space provides a more intuitive perspective and accurate expression ability for analyzing and processing spatiotemporal data. This time-aware UMAP improvement method enables the spatiotemporal features to more accurately retain the dynamic change characteristics after dimensionality reduction, providing a more solid foundation for subsequent data analysis and prediction.

[0068] Furthermore, the production path optimization module performs curve fitting on the production manifold feature space to generate a target vulcanization process path, and determines a target vulcanization process interval according to the target vulcanization process path, specifically including:

[0069] The production manifold feature space is clustered based on the DBSCAN algorithm, and the production manifold feature space is divided into multiple sub-feature spaces. Bezier curve fitting is performed on multiple data points in each sub-feature space to obtain the local Bezier curve of each sub-feature space.

[0070] Specifically, the DBSCAN algorithm is used to cluster multiple data points in the production manifold feature space. The DBSCAN algorithm is a density-based clustering method that can find relatively high-density data areas in the production manifold feature space and effectively identify low-density noise data. For ease of processing, each data point selects a time series corresponding to one of the sets of data, that is, the performance of each set of historical data in the low-dimensional space is taken as an individual, and the data corresponding to each data point contains the expression of one of the data in the low-dimensional space. The clustering operation can divide the original production manifold feature space into several data subsets with similar characteristics, called sub-feature spaces. Each sub-feature space contains multiple closely related data points.

[0071] After completing the segmentation of the production manifold feature space, the data points in each sub-feature space are fitted with Bezier curves. The Bezier curve is a smooth curve that is used to accurately describe the data distribution trend of each sub-feature space. In this way, the complex data distribution of each sub-feature space is expressed by a smooth Bezier curve, ensuring the accuracy of subsequent curve fusion. Each local Bezier curve presents a specific direction and shape in space, reflecting the feature change trend in the sub-feature space.

[0072] Analyze multiple data points in each sub-feature space, determine the density parameter and distance parameter of each sub-feature space, determine the target weight of each sub-feature space according to the density parameter and distance parameter, fuse the local Bezier curves of each sub-feature space according to the target weight, and generate the target vulcanization process path corresponding to the production manifold feature space.

[0073] Specifically, after obtaining the local Bezier curves of each sub-feature space, we further analyze the data points in each sub-feature space to extract density parameters and distance parameters. The density parameter indicates the degree of aggregation of data points in each sub-feature space, while the distance parameter reflects the relative position of each data point in the space. These two parameters together constitute the target weight of the sub-feature space, which reflects the importance of each sub-feature space in generating the target vulcanization process path. The purpose of calculating these parameters is to adjust the weight of each local Bezier curve according to the characteristics of the sub-feature space during the curve fusion process, so as to more accurately generate the global target vulcanization process path. By assigning different weights to different sub-feature spaces, the fused Bezier curve can reflect the overall trend while retaining the local characteristics of each sub-feature space. The target vulcanization process path finally generated is a smooth curve, which represents the ideal parameter change trend of the vulcanization process and can provide a specific reference for process optimization.

[0074] Among them, for the calculation of density parameters and distance parameters, the ratio of the number of data points in the sub-feature space to the total number of data points in the production manifold feature space is recorded as the density parameter, and the average of the distances between multiple data points in the sub-feature space and the center point of the sub-feature space is recorded as the distance parameter. Multiple density parameters and distance parameters are normalized respectively, and the sum of the normalized density parameters and distance parameters of each sub-feature space is used as its own target weight.

[0075] Each sub-feature space is analyzed according to the target vulcanization process path, and the distances between multiple data points in each sub-feature space and the target vulcanization process path on each low-dimensional feature parameter are determined, so as to obtain local vulcanization process intervals containing distances corresponding to multiple low-dimensional feature parameters, and generate a target vulcanization process interval containing multiple local vulcanization process intervals.

[0076] Specifically, after obtaining the target vulcanization process path, each sub-feature space is further analyzed to determine the distance between each data point and the target vulcanization process path on each low-dimensional feature parameter. Among them, we calculate the difference between multiple data points in each sub-feature space on the low-dimensional feature parameters. Specifically, each data point contains multiple low-dimensional feature parameters, which are used to characterize the expression of process parameters such as temperature, pressure, and hardness in the low-dimensional space. The same sub-feature space contains multiple values ​​corresponding to each low-dimensional feature parameter. After calculating the distance between each value and the target vulcanization process path, that is, the difference, and taking the average, the local vulcanization process interval of each sub-feature space can be constructed, which contains the difference between multiple low-dimensional feature parameters and the target vulcanization process path.

[0077] After obtaining multiple local vulcanization process intervals, these intervals are summarized to form a target vulcanization process interval containing multiple local vulcanization process intervals. The target vulcanization process interval integrates the characteristic performance of each sub-feature space in the production manifold feature space, and provides multi-dimensional constraints and operation references for the vulcanization process. In the actual process, by referring to the parameter variation range within the target vulcanization process interval, an optimization reference can be provided to ensure that the vulcanization process proceeds along the optimal path.

[0078] Furthermore, under the guidance of the target vulcanization process interval, in order to ensure that the vulcanization production process operates within a reasonable process range, the vulcanization process anomaly analysis module performs process deviation detection on the collected target vulcanization production data based on the target vulcanization process interval, specifically including:

[0079] The target vulcanization production data is input into the time series dynamic manifold analysis model to generate the target manifold feature, and the association between the target process feature and multiple sub-feature spaces is analyzed respectively to determine the target sub-feature space. That is, the dynamic manifold analysis module is used to perform dynamic manifold analysis on the multi-dimensional production feature matrix in the time series dynamic manifold analysis model, and the target vulcanization production data is subjected to dimensionality reduction processing so that it is mapped in a low-dimensional space, and then the association between the feature and multiple sub-feature spaces is analyzed. To determine the best matching area of ​​the target manifold feature in the production manifold feature space, the similarity measure between the target manifold feature and each sub-feature space can be calculated, such as the Euclidean distance, cosine similarity, etc. For example, the Euclidean distance between the target manifold feature and the center point of each sub-feature space is calculated, and the sub-feature space with the closest distance is selected as the target sub-feature space.

[0080] Finally, the area most closely associated with the target manifold features is recorded as the target sub-feature space, which represents the typical behavior pattern of the target vulcanization production data in the low-dimensional space.

[0081] Based on the target sub-feature space, a difference analysis is performed on the target vulcanization process path to generate a first manifold deviation vector corresponding to the target vulcanization process path; based on the target sub-feature space, a difference analysis is performed on the target manifold feature to generate a second manifold deviation vector corresponding to the target manifold feature.

[0082] Specifically, after determining the target sub-feature space, the deviation of the target vulcanization process path in the target sub-feature space is calculated, and a vector representing the degree of difference is generated and recorded as the first manifold deviation vector. Specifically, the path corresponding to the center point of the target sub-feature space can be selected to represent the target sub-feature space, and the difference between it and the target vulcanization process path in each low-dimensional feature parameter is calculated. According to the local vulcanization process interval corresponding to the target sub-feature space, it is judged whether the difference in the low-dimensional feature parameters deviates from the range of the local vulcanization process interval, and the deviation value of each low-dimensional feature parameter with respect to the local vulcanization process interval is recorded to construct the first manifold deviation vector.

[0083] At the same time, the target manifold features themselves are also analyzed for differences. Within the framework of the target sub-feature space, the deviation between the target manifold features and the typical manifold distribution in the sub-feature space is analyzed in the same way to obtain the second manifold deviation vector.

[0084] Finally, the process deviation detection of the target vulcanization production data is performed according to the first manifold deviation vector and the second manifold deviation vector, specifically including calculating the distance between the first manifold deviation vector and the second manifold deviation vector and using it as the process deviation parameter of the target vulcanization production data. This parameter analyzes the difference between the collected data and the optimal path, and takes into account the local pattern of the data, and comprehensively considers the degree of difference between the target vulcanization production data and the target process path under complex actual conditions in the production process, thereby providing a quantitative reference for process control. Single analysis of a certain parameter may easily lead to misjudgment due to data noise. This method of calculating the process deviation parameter based on the overall difference of the comprehensive data takes into account the overall situation of the data. In the actual production process, the obvious abnormality of a process parameter is easy to detect, but the impact of the minor abnormalities between multiple parameters is difficult to identify. In this way, the abnormalities can be mined in advance.

[0085] By analyzing the process deviation parameters, we can identify the deviations in the vulcanization production process and take timely adjustment measures. If the process deviation parameter is greater than the preset deviation threshold, it means that the collected data is different from the ideal situation as a whole, which means that there may be potential deviation risks in the current production process. In this case, the parameters that may be abnormal can be further located to determine multiple abnormal parameters. For example, based on the preset threshold, the deviation difference of each low-dimensional feature parameter is quantified by the first manifold deviation vector and the second manifold deviation vector, and multiple low-dimensional feature parameters greater than the threshold are marked as abnormal parameters.

[0086] Furthermore, the inverse mapping model used by the vulcanization process optimization module to perform inverse mapping processing on the deviated feature vector is constructed in the following manner:

[0087] In the mapping relationship between the multidimensional production feature matrix and the production manifold feature space, each data point in the production manifold feature space has a unique corresponding data point in the multidimensional production feature matrix. In order to train the inverse mapping model, a sample data set for training the inverse mapping model is constructed based on the multidimensional production feature matrix and the production manifold feature space, which ensures that when the inverse mapping model receives any data point in the manifold feature space, it can predict its corresponding position in the multidimensional production feature matrix.

[0088] After the sample data set is constructed, it is used to train the inverse mapping model. In this embodiment, the inverse mapping model is constructed based on the generative adversarial network (GAN). In this application scenario, the generator is used to generate target data points in the multidimensional production feature matrix, and the discriminator is used to distinguish the difference between the generated feature vector and the real feature vector. Multiple data points in the production manifold feature space are used as the input of the inverse mapping model, and the data points corresponding to the multiple data points in the production manifold feature space in the multidimensional production feature matrix are used as training targets. Through this adversarial training mechanism, the generator can gradually learn the mapping relationship between the production manifold feature space and the multidimensional production feature matrix, and finally form a high-precision inverse mapping model.

[0089] After generating the deviation feature vector according to the first manifold deviation vector and the second manifold deviation vector, the deviation feature vector is input into the inverse mapping model, and the model maps the deviation feature vector to the corresponding feature in the multidimensional production feature matrix according to the learned mapping relationship, that is, generates the target production feature vector corresponding to the deviation feature vector. It represents the specific position and features of the deviation state in the multidimensional production feature space, and helps locate the specific manifestation of process deviation in the multidimensional space of production features.

[0090] The production data management system provided by the present invention comprehensively monitors and optimizes the vulcanization production data of rubber materials through multi-module collaboration, and realizes the full process management from data acquisition, feature extraction, process path optimization to anomaly detection and intelligent adjustment. The data acquisition module obtains the key data of the vulcanization process, generates the production manifold feature space containing the complex dynamic characteristics of the vulcanization process through multi-dimensional time series feature extraction and dynamic manifold analysis, and constructs the target vulcanization process path and process interval based on the feature space, providing a reference for subsequent process deviation detection; the vulcanization process anomaly analysis module uses the deviation vector to detect process anomalies on the basis of real-time monitoring of production data, and realizes the accurate restoration and tracing of process parameters in combination with the inverse mapping model, ensuring that the production process is always maintained within the optimal path. The system effectively identifies and locates process deviations through quantitative analysis of process deviation detection parameters, adjusts production parameters in time, and finally improves the stability of the production process and the consistency of product quality. The multi-dimensional analysis and dynamic optimization capabilities of the system make the vulcanization process more intelligent and automated, providing strong support for quality control and process improvement in the automotive parts production industry such as wipers and filters.

[0091] The above is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. Parts not described in detail in this specification belong to the prior art known to those skilled in the art.

Claims

1. A production data management system based on artificial intelligence, characterized in that: include: The production data collection module is used to collect the historical production data of rubber materials, including the production process data and production quality data of rubber materials in the vulcanization process; The multi-dimensional time series feature extraction module is used to perform data cleaning and sliding window analysis on the historical production data of rubber materials and generate a multi-dimensional production feature matrix corresponding to the historical production data; A dynamic manifold analysis module is used to perform dynamic manifold analysis on a multi-dimensional production feature matrix through a time-series dynamic manifold analysis model, including using an improved UMAP algorithm based on time perception to process the multi-dimensional production feature matrix and generate a production manifold feature space corresponding to the multi-dimensional production feature matrix; The production path optimization module is used to perform curve fitting on the production manifold feature space to generate a target vulcanization process path, and determine a target vulcanization process interval according to the target vulcanization process path; The vulcanization process anomaly analysis module is used to perform process deviation detection on the collected target vulcanization production data based on the target vulcanization process interval, locate the abnormal parameters after the calculated process deviation parameter is greater than the preset deviation threshold, and generate a deviation detection result including the abnormal parameters and the deviation feature vector; The vulcanization process optimization module is used to perform inverse mapping processing on the deviated feature vector through the inverse mapping model, generate the target production feature vector corresponding to the deviated feature vector, generate the vulcanization process optimization plan according to the target production feature vector and the abnormal parameters, and optimize the vulcanization process based on the vulcanization process optimization plan; The improved UMAP algorithm based on time perception is used to process the multi-dimensional production feature matrix to generate the production manifold feature space corresponding to the multi-dimensional production feature matrix, including: Calculate the distance parameter between any two data points in the multi-dimensional production feature matrix according to the UMAP algorithm, construct a distance matrix corresponding to the multi-dimensional production feature matrix according to multiple distance parameters, and determine the distance neighborhood graph according to the distance matrix; Time perception is introduced into the distance neighborhood graph, including determining the time weight between any two data points based on the target time function, optimizing the distance neighborhood graph according to the time weight to generate a spatiotemporal neighborhood graph. The weight between data points in the spatiotemporal neighborhood graph is: ； In the formula, is the first and The weights between the data points, is an element in the distance matrix, indicating the and The spatial distance between data points, Represents the first and The distance between data points in time steps, , are the spatial scale parameter and the temporal scale parameter respectively; Construct the target neighborhood graph corresponding to the spatiotemporal neighborhood graph, and determine the target optimization function based on the spatiotemporal neighborhood graph and the target neighborhood graph : ； In the formula, is the first and The weights between data points; Based on the objective optimization function, the gradient descent algorithm is used to iteratively optimize the positions of multiple data points in the multidimensional production feature matrix to obtain the position of each data point in the multidimensional production feature matrix in the low-dimensional space. According to the positions of multiple data points in the low-dimensional space, the production manifold feature space corresponding to the multidimensional production feature matrix is ​​constructed.

2. The production data management system based on artificial intelligence according to claim 1 is characterized in that: Perform curve fitting on the production manifold feature space to generate the target vulcanization process path, and determine the target vulcanization process range according to the target vulcanization process path, including: The production manifold feature space is clustered based on the DBSCAN algorithm, and the production manifold feature space is divided into multiple sub-feature spaces. Bezier curve fitting is performed on multiple data points in each sub-feature space to obtain the local Bezier curve of each sub-feature space. Analyze multiple data points in each sub-feature space, determine the density parameter and distance parameter of each sub-feature space, determine the target weight of each sub-feature space according to the density parameter and distance parameter, fuse the local Bezier curve of each sub-feature space according to the target weight, and generate the target vulcanization process path corresponding to the production manifold feature space; Each sub-feature space is analyzed according to the target vulcanization process path, and the distances between multiple data points in each sub-feature space and the target vulcanization process path on each low-dimensional feature parameter are determined, so as to obtain local vulcanization process intervals containing distances corresponding to multiple low-dimensional feature parameters, and generate a target vulcanization process interval containing multiple local vulcanization process intervals.

3. The production data management system based on artificial intelligence according to claim 2 is characterized in that: Based on the target vulcanization process range, the collected target vulcanization production data is subjected to process deviation detection, including: Inputting the target vulcanization production data into the time-series dynamic manifold analysis model to generate the target manifold features, respectively analyzing the association between the target process features and multiple sub-feature spaces and determining the target sub-feature space; Performing a difference analysis on a target vulcanization process path based on a target sub-feature space to generate a first manifold deviation vector corresponding to the target vulcanization process path, and performing a difference analysis on a target manifold feature based on a target sub-feature space to generate a second manifold deviation vector corresponding to the target manifold feature; The process deviation detection is performed on the target vulcanization production data according to the first manifold deviation vector and the second manifold deviation vector, including calculating the distance between the first manifold deviation vector and the second manifold deviation vector and using it as the process deviation parameter of the target vulcanization production data.

4. The production data management system based on artificial intelligence according to claim 3 is characterized in that: For the inverse mapping model, it also includes: A sample data set for training an inverse mapping model is constructed based on a multidimensional production feature matrix and a production manifold feature space, a plurality of data points in the production manifold feature space are used as inputs of the inverse mapping model, and data points corresponding to the plurality of data points in the production manifold feature space in the multidimensional production feature matrix are used as training targets, and an inverse mapping model is obtained through training; After generating a deviation feature vector according to the first manifold deviation vector and the second manifold deviation vector, the deviation feature vector is input into the inverse mapping model to generate a target production feature vector corresponding to the deviation feature vector.

5. The production data management system based on artificial intelligence according to claim 4 is characterized in that: For the target weight of the sub-feature space, it also includes: The ratio of the number of data points in the sub-feature space to the total number of data points in the production manifold feature space is recorded as the density parameter, and the mean of the distances between multiple data points in the sub-feature space and the center point of the sub-feature space is recorded as the distance parameter. Multiple density parameters and distance parameters are normalized respectively, and the sum of the normalized density parameters and distance parameters of each sub-feature space is taken as the target weight of the sub-feature space.

6. The production data management system based on artificial intelligence according to claim 3 is characterized in that: The distances between the target process feature and the center points of multiple sub-feature spaces are calculated respectively, and the sub-feature space with the closest distance to the target process feature is selected as the target sub-feature space.

7. The production data management system based on artificial intelligence according to claim 4 is characterized in that: Construct an inverse mapping model based on the GAN network.

Images