A Method and System for Predicting Manufacturing Process Quality Based on Multimodal Data and Adaptive Hybrid Models

By using multimodal data and an adaptive hybrid model, the optimal quality prediction model is selected for different processing characteristics, which solves the problem of low accuracy in quality prediction in existing technologies and achieves high-precision manufacturing process quality prediction.

CN117787761BActive Publication Date: 2026-11-17NORTHWESTERN POLYTECHNICAL UNIV
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202311524728.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-11-16
Publication Date
2026-11-17
Estimated Expiration
2043-11-16

AI Technical Summary

Technical Problem

In existing digital workshop manufacturing processes, single-modal data and single algorithms result in low accuracy of quality prediction, failing to fully describe the manufacturing process status and characteristics at different stages, thus reducing the accuracy and regression fit of the model in certain sub-domain frequency bands.

Method used

By employing multimodal data and adaptive hybrid models, various intelligent models (such as linear regression, random forest, DNN, and XGBoost models) are obtained. The optimal quality prediction model is selected for different processing characteristics, and an adaptive hybrid model library is constructed to perform high-precision quality prediction.

Benefits of technology

It improves the accuracy of quality prediction in the product manufacturing process, enhances the model's generalization ability and training speed, achieves high-precision prediction of different processing features, and adapts to the performance matching of different local processing geometric features.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117787761B_ABST
    Figure CN117787761B_ABST
Patent Text Reader

Abstract

To address the low prediction accuracy of existing quality prediction schemes used in digital workshop manufacturing processes, this invention proposes a manufacturing process quality prediction method and system based on multimodal data and an adaptive hybrid model. First, this invention collects multimodal data from the manufacturing process of key product components, improving data completeness and quality, and providing a better data foundation for high-precision quality prediction. Then, considering that the relationships between different features in a multimodal data context are not linear regressions, a single quality prediction model is often insufficient to meet practical needs. Therefore, this invention constructs an intelligent model composed of multiple quality prediction models and matches the optimal quality prediction model to each processing feature. This achieves adaptive matching of the quality prediction model with the minimum prediction error for different processing features of different parts to be processed, better fitting the actual situation and improving the accuracy of quality prediction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to a manufacturing process quality prediction method, and more particularly to a manufacturing process quality prediction method and system based on multimodal data of key components collected during the manufacturing process and an adaptive hybrid model. Background Technology

[0002] Currently, manufacturing enterprises face two major challenges. On the one hand, with intensifying market competition, manufacturers are forced to produce high-quality products to attract consumers and expand market share. On the other hand, with the emergence of new demands and the iterative updates of products, product structures are becoming increasingly complex, and manufacturing precision is rising. Furthermore, with the significant progress of the new industrial revolution, new technologies are emerging in rapid succession, and technological development is constantly giving rise to new manufacturing models. As a key link in realizing intelligent manufacturing, quality prediction refers to the use of data and statistical methods to predict the quality level of products or services during the production, service provision, or product delivery process. Quality prediction plays a crucial role in intelligent manufacturing, helping manufacturing enterprises to promptly identify potential problems, take corrective measures, optimize manufacturing processes, and ensure that the final product or service meets the required quality requirements during the manufacturing process.

[0003] In recent years, the main methods for quality monitoring and prediction during the manufacturing process include:

[0004] Chinese patent document CN 116431613 A discloses a system-level M-CPS architecture for smart mine production. This architecture includes a perception layer, a transmission layer, a fusion layer, and a control layer. The perception layer, consisting of the fully mechanized mining subsystem, tunneling subsystem, transportation subsystem, drainage subsystem, ventilation subsystem, and electromechanical system, constitutes the physical world. The transmission layer connects the physical world and the information world, enabling them to interconnect. The fusion layer integrates data and functions from different subsystems to form a unified system and constitute the information world. The control layer includes a decision-making command center, a comprehensive control center, a safety production center, a big data processing center, and an energy management center to solve the problems of resource coordination among subsystems in the coal mine system and unified data management in the production process, thereby promoting collaborative work in the coal mine production process.

[0005] Chinese patent document CN 109270899 B discloses a method for controlling the manufacturing process of key components for marine diesel engines based on digital twins. This method includes five steps: (1) creating a key component manufacturing service platform; (2) real-time data acquisition; (3) establishing a digital twin model system; (4) quality prediction; and (5) synchronous execution and control of the manufacturing process. This achieves real-time, full-cycle control over the quality of key components and the stability of processing resources and equipment during the manufacturing process.

[0006] The methods described above have all played a role in promoting quality monitoring and prediction in the manufacturing process of complex products. However, in the industrial IoT manufacturing environment, the aforementioned patent documents and other current solutions for quality prediction in digital workshop manufacturing processes still have the following key problems:

[0007] 1. Use single-modal data

[0008] In actual production, product quality fluctuations are often affected by multiple factors, which are interrelated. To simplify the complex analysis and processing, existing solutions usually choose to use single-modal data features as system input, which simplifies model construction. However, since single-modal data cannot fully describe the state of the manufacturing process, it reduces the accuracy of the model in certain sub-domain frequency bands.

[0009] 2. Using a single algorithm

[0010] Data from different periods have different characteristics, and a single algorithm cannot achieve high-precision fitting in the entire space and frequency domain, and cannot well describe the state at different times. Therefore, the regression fitting degree of a single algorithm for multiple feature combinations of key components is reduced, which leads to lower accuracy in quality prediction. Summary of the Invention

[0011] To address the technical problem of low prediction accuracy in existing manufacturing process quality prediction schemes for digital workshops, this invention proposes a manufacturing process quality prediction method and system based on multimodal data and an adaptive hybrid model.

[0012] The technical solution of the present invention is as follows:

[0013] The manufacturing process quality prediction method based on multimodal data and adaptive hybrid models is characterized by the following steps:

[0014] Step 1: Acquire multimodal data, which includes at least vibration data and noise data in the XYZ directions of the machining coordinate system;

[0015] Step 2: Process the multimodal data to obtain a data frame;

[0016] Step 3: Perform dimensionality reduction on the data frame;

[0017] Step 4: Using the dimensionality-reduced data frame as input features and the parameters used to characterize the processing and manufacturing quality as output labels, construct and train an intelligent model. The intelligent model includes multiple independently constructed and trained quality prediction models, including but not limited to linear regression models, random forest models, DNN models, or XGBoost models.

[0018] Step 5: For each processing feature on each type of part to be processed, obtain the prediction error percentage of each quality prediction model constructed and trained in Step 4. Select the quality prediction model with the smallest prediction error percentage for each processing feature on each type of part to be processed as the optimal quality prediction model for subsequent quality prediction. Based on this, establish an adaptive hybrid model library, which includes the optimal quality prediction model that corresponds one-to-one with each part to be processed and each processing feature.

[0019] Step 6: Based on the parts to be processed and their processing characteristics in the current manufacturing process, select the optimal quality prediction model from the adaptive hybrid model library built in Step 5. Input the multimodal data collected in real time during the manufacturing process into the currently selected optimal quality prediction model. The optimal quality prediction model outputs the quality prediction result, thereby realizing the quality prediction of the product manufacturing process.

[0020] Further, step 1 specifically involves: building a CPS-based manufacturing workshop / system prototype system, using the prototype system to acquire multimodal data corresponding to different processing parameters during the manufacturing process, and storing them in file formats that can retain complete data.

[0021] Furthermore, the multimodal data obtained in step 1 also includes tool and / or workpiece temperature data during processing, displacement data of cutting tool and / or workpiece features, cutting force data, cutting fluid injection pressure data, and / or light sensing data.

[0022] Furthermore, step 2 specifically involves:

[0023] Step 2.1 Locate, segment, and extract data related to the monitored object from the raw multimodal data collected in Step 1. This extracted data is time-domain data.

[0024] Step 2.2 converts the captured data into frequency domain data;

[0025] Step 2.3 Based on the frequency domain data, obtain the frequency domain characteristics, and through the analysis of the frequency domain characteristics, assign the corresponding confidence level to the time domain data of Step 2.1, and perform time-frequency domain data fusion to form a data frame.

[0026] Furthermore, based on the confidence level assigned to the time-domain data, time-domain data of less than or equal to 20% are removed to obtain a data frame with better data quality.

[0027] Furthermore, in step 3, principal component analysis or feature-selective dimensionality reduction is used to reduce the dimensionality of the data frame.

[0028] Furthermore, the intelligent models described in step 4 include linear regression models, random forest models, DNN models, and XGBoost models.

[0029] Furthermore, when constructing and training each quality prediction model in the intelligent model, cross-validation is used to train and validate the constructed quality prediction models.

[0030] This invention also provides a manufacturing process quality prediction system based on multimodal data and an adaptive hybrid model, which is characterized by including a multimodal data acquisition unit, a data processing unit, a model building and training unit, an adaptive hybrid model library, and a quality prediction unit.

[0031] The multimodal data acquisition unit is used to collect and store multimodal data during the processing and manufacturing of parts and / or products;

[0032] The data processing unit is used to extract data related to the part to be processed from the multimodal data and convert it into frequency domain data, then form a data frame by merging the data, and perform data dimensionality reduction processing on the formed data frame;

[0033] The model building and training unit is used to build an intelligent model, which includes multiple quality prediction models, including but not limited to linear regression models, random forest models, DNN models, or XGBoost models.

[0034] The adaptive hybrid model library includes optimal quality prediction models that correspond one-to-one with each part to be processed and each of its processing features. The optimal quality prediction model is selected from the intelligent models. For a certain processing feature, the quality prediction model with the smallest percentage error in its prediction result is selected as the optimal quality prediction model for that processing feature.

[0035] The quality prediction unit is used to select the best matching quality prediction model from the adaptive hybrid model library based on the parts to be processed and the processing features in the current processing process, and output the quality prediction result of the current processing features.

[0036] Compared with the prior art, the beneficial effects of the present invention are:

[0037] 1. This invention first collects multimodal data during the manufacturing process of key product components, improving data integrity and quality, and providing a better data foundation for high-precision quality prediction. Then, considering that the relationship between different features in the context of multimodal data is not linear regression, a single quality prediction model is often insufficient to meet actual needs. Therefore, this invention constructs an intelligent model composed of multiple quality prediction models and matches the optimal quality prediction model to each processing feature. This achieves adaptive matching of the quality prediction model with the minimum prediction error for each processing feature of different parts to be processed, better fitting the actual situation and improving the accuracy of quality prediction in the product manufacturing process.

[0038] 2. In the intelligent model constructed in this invention, each quality prediction model is constructed and trained independently, and the prediction performance of the models will not affect each other. The generalization performance and flexibility performance are better, and the training time of a single quality prediction model is short and the required computing power is small. The selection of the optimal model for each processing feature depends on the geometric features of the processed part, the characteristics of the quality prediction model and the hyperparameters. Therefore, the establishment and enrichment of the adaptive hybrid model library is completely unaffected by time, space, intelligent models and computing power, and is easier to implement in engineering.

[0039] 3. This invention establishes an adaptive model library to match the best-performing quality prediction model for different local machining geometric features of the machined parts, thereby achieving high-precision quality prediction globally.

[0040] 4. When processing multimodal data, this invention first extracts time-domain data related to the monitored object, and then introduces confidence scores (or weights) based on frequency domain characteristics into these time-domain data to fuse the time-frequency domain data, forming a data frame for training various quality prediction models in the subsequent intelligent model. This improves the generalization ability of the model, enabling the quality prediction model to still have high prediction accuracy when predicting different processing characteristics.

[0041] 5. Based on the confidence level (or weight) assigned to the time-domain data, the present invention deletes no more than 20% of the time-domain data in the data frame to improve the quality of the training data of each quality prediction model in the subsequent intelligent model, so as to obtain the optimal parameters of the model in a shorter time and improve the training speed of the quality prediction model. Attached Figure Description

[0042] Figure 1 This is a flowchart of the quality prediction method of the present invention.

[0043] Figure 2 This is a schematic diagram of the quality prediction system of the present invention.

[0044] Figure 3This is a wiring diagram of the edge-end actual physical space multimodal data sensing and acquisition device involved in the acquisition of multimodal data in this invention.

[0045] Figure 4 This is the actual processing prototype system based on CPS built by the present invention.

[0046] Figure 5 This is a schematic diagram showing the positions of the three measuring surfaces for the surface roughness value of an open cavity in an embodiment of the quality prediction method of the present invention.

[0047] Figure 6 This is a flowchart illustrating the construction process of the intelligent model in an embodiment of the quality prediction method of the present invention.

[0048] Figure 7 This is a visualization of the random forest model in an embodiment of the quality prediction method of the present invention.

[0049] Figure 8 This is an iterative loss diagram of the DNN model training process in the quality prediction method of this invention.

[0050] Figure 9 This is a visualization of the process in the embodiment of the quality prediction method of the present invention, which involves selecting the quality prediction model with the smallest percentage of prediction error for each processing feature. Detailed Implementation

[0051] The present invention will now be described in detail with reference to the accompanying drawings, so that the technical solutions and advantages of the present invention will be clearer and easier to understand.

[0052] Reference Figure 1 The manufacturing process quality prediction method based on multimodal data and adaptive hybrid model provided by this invention includes the following steps:

[0053] Step 1: Acquire multimodal data during the manufacturing process

[0054] Step 1.1 Set up as follows Figure 4 The actual processing prototype system of the manufacturing workshop / system based on Cyber-Physical System (CPS) is shown.

[0055] To acquire multimodal, multi-domain sensing data of the manufacturing process, this invention first follows... Figure 3 The wiring diagram shown illustrates the connection process, which, after wiring, yields a multimodal data sensing and acquisition device for the actual physical space at the edge. Based on this, a system is built as follows: Figure 4 The CPS-based prototype system shown mainly includes a physical domain - multimodal data physical perception layer, an information domain - multimodal data transmission and information space construction layer, and a system domain - system function application distributed cloud layer.

[0056] The physical domain – multimodal data physical sensing layer – mainly includes customized fixtures added to the existing actual manufacturing workshop / system, as well as various data sensors, data acquisition instruments, data storage software, data transmission cables and interfaces for realizing multimodal data acquisition and transmission during the processing. The customized fixtures are used to install data acquisition, sensing, and interactive sensing devices while clamping and positioning the workpiece without affecting normal processing. For example, one or more of the following can be added to the existing actual manufacturing workshop: a built-in active miniature accelerometer, multiple active vibration sensors (for acquiring vibration signals in the X, Y, and Z directions respectively), a noise sensor, a temperature sensor, a displacement sensor, and a pressure sensor.

[0057] The information domain - multimodal data transmission and information space construction layer mainly includes data acquisition instruments / systems and their corresponding data storage software that are synchronously installed at the edge, and data transmission is achieved through edge-end RS-485, RS-232, USB interface, ZigBee, Bluetooth, WIFI, infrared, optical signal or Huawei Star Flash and other methods.

[0058] The system domain - system function application distributed cloud layer mainly includes edge intelligent computing devices and end-edge-distributed cloud collaborative architecture for deep intelligent computing, analysis and processing, which can perform system function applications such as processing status monitoring, manufacturing product tasks and intelligent model construction.

[0059] Step 1.2 Using the actual machining prototype system built in Step 1.1, acquire the multimodal data corresponding to different machining process parameters during the manufacturing process, and store them in file formats that can retain complete data. These multimodal data include at least vibration data and noise data in the XYZ directions of the machining coordinate system. Furthermore, they may also include tool and / or workpiece temperature data, displacement data of cutting tool and / or workpiece features, cutting force data, cutting fluid injection pressure data, and / or light sensing data during machining.

[0060] It is important to note that when saving data, first select a file format based on the amount of data collected. Common data file formats include text data formats such as .csv, .xlsx, .txt, .json, and .xml, and multimedia data formats such as .jpg, .mp3, and .mp4. Then, open the saved data file and observe the data information such as the number of records, the maximum number of records, column names, the meaning of column names, the unit of column names, and the maximum number of columns.

[0061] If the number of rows or columns displayed is less than the maximum number of records or columns, the acquired multimodal data can be saved as a file in a different format that can display more rows or columns. For example, if the acquired multimodal data is initially stored as an .xlsx file, but the actual number of rows displayed is less than the maximum number of records, it indicates that the data is incomplete. In this case, the multimodal data collected by the sensing device should be exported to another supported format, such as .csv. If the data is an Excel file with multiple sheets, tools such as Python can be used to read and merge the multiple sheets into a single .csv file to solve the problem.

[0062] If some rows or columns are missing data, the missing data should be filled in to ensure data integrity. The specific methods for filling in missing data are existing and will not be elaborated further.

[0063] If garbled characters appear in some rows or columns, they should be corrected to ensure data integrity. The specific methods for correcting garbled characters are existing and will not be elaborated further.

[0064] Step 2: Process the multimodal data obtained in Step 1

[0065] Step 2.1 Locate, segment, and extract data related to the monitored object from the raw multimodal data collected in Step 1. This extracted data is time-domain data.

[0066] Step 2.2 converts the captured data into frequency domain data;

[0067] Step 2.3 Based on the frequency domain data of Step 2.2, obtain the frequency domain characteristics. Through the analysis of the frequency domain characteristics of the frequency domain data, assign the corresponding confidence level (or weight) to the time domain data of Step 2.1, and perform time-frequency domain data fusion to form a data frame.

[0068] Furthermore, the data frame formed in step 2.3 can be processed as follows:

[0069] Based on the confidence level (or weight) assigned to the time-domain data, no more than 20% of the time-domain data is removed to obtain a data frame with better data quality.

[0070] Step 3: Perform dimensionality reduction on the data frame obtained after fusion in Step 2.

[0071] Since the data frame obtained in step 2 has a high dimensionality, it needs to be dimensionality reduced before constructing the intelligent model. This invention can use Principal Component Analysis (PCA) or Lasso regression to reduce the dimensionality of the data frame obtained after fusion in step 2.

[0072] Step 4: Build and train the intelligent model

[0073] The intelligent model includes multiple independently constructed and trained quality prediction models, including but not limited to at least two of the following: linear regression model, random forest model, DNN model, and XGBoost model.

[0074] Before training each quality prediction model, it is necessary to first determine the features and labels, and divide the training and test sets; among them,

[0075] The data frame after dimensionality reduction in step 3 is used as the input feature x, and the parameters used to characterize manufacturing quality (such as surface roughness, dimensional accuracy and / or machining deformation error, etc.) are used as the output label y. The method for obtaining the output label y is to detect and obtain the part corresponding to the input feature x between processes and after processing.

[0076] The method for dividing the data into training and test sets is as follows: Add the output label y as a column of data to the data frame after dimensionality reduction in step 3, resulting in a new data frame with the output label y. Then, divide this new data frame into training and test sets according to a set ratio. Generally, 60-80% of the data in the new data frame is allocated to the training set, and the remainder to the test set. Ideally, 80% of the data in the new data frame is allocated to the training set, and the remainder to the test set.

[0077] Step 5: Build an adaptive hybrid model library

[0078] For each machining feature on each type of part to be machined (e.g., inner side straight line, inner fillet, outer straight line, outer fillet, inclined surface, curved surface, profile, disk surface, disk center hole, thin-walled ring, etc.), obtain the prediction result error percentage or prediction probability value of each quality prediction model constructed and trained in step 4. Select the quality prediction model with the smallest prediction result error percentage or prediction probability value for each machining feature on each type of part to be machined as the optimal quality prediction model for subsequent quality prediction, and establish an adaptive hybrid model library based on this. This adaptive hybrid model library includes the optimal quality prediction model that corresponds one-to-one with each part to be machined and each of its machining features.

[0079] Step 6: Quality Prediction in the Processing and Manufacturing Process

[0080] Based on the parts to be processed and their processing characteristics in the current manufacturing process, the optimal quality prediction model is selected from the adaptive hybrid model library constructed in step 5. The multimodal data collected in real time during the manufacturing process is input into the currently selected optimal quality prediction model, which outputs the quality prediction result, thereby realizing the quality prediction of the product manufacturing process.

[0081] Example:

[0082] This embodiment takes the mass prediction of a crucial structural component of an aircraft—the fuselage frame—as an example. Considering the large size and cost-efficiency factors of this critical structural component, a typical open cavity structure on the fuselage frame is selected as the specific research object. The mass prediction includes the following steps:

[0083] Step 1: Acquire multimodal data during the manufacturing process

[0084] Step 1.1 Build a prototype manufacturing workshop / system based on CPS;

[0085] In the physical domain—multimodal data physical perception layer—first, the typical cavity structure of the studied aircraft fuselage frame is determined. Second, a 3D model of the typical cavity structure of the aircraft fuselage frame is established and scaled at a ratio of 0.5. Finally, using JDSoftSurfMill V8.0 X64 Pro software and MasterCAM 2022 software matched with the Beijing Jingdiao CNC machining center, a machining program is developed, and the aircraft fuselage frame is manufactured in the actual machining prototype system of the constructed CPS-based manufacturing workshop / system.

[0086] Step 1.2 After the actual processing prototype system of the CPS-based manufacturing workshop / system is built, the multimodal data corresponding to different processing parameters in the actual manufacturing process can be obtained and stored in real time using this actual processing prototype system.

[0087] Step 2: Process the multimodal data obtained in Step 1

[0088] Step 2.1 Locate, segment, and extract data related to the open cavity from the original multimodal data obtained in Step 1:

[0089]

[0090] in, The X, Y, Z three-dimensional vibration time-domain data of the open cavity were obtained from the actual processing. DS is the raw multimodal data obtained from the industrial noise data at the open cavity.

[0091] Step 2.2 Extract the time-domain data related to the open cavity from Step 2.1. Periodic data is converted to frequency domain data using Fourier series, and non-periodic data is converted to frequency domain data using Fourier transform. Furthermore, considering that the Fast Fourier Transform (FFT) has the advantages of less computation and simpler operation than the traditional Fourier Transform, this embodiment prefers the Fast Fourier Transform (FFT) as the time-frequency domain conversion method.

[0092] Step 2.3 Utilize all the frequency domain data obtained in Step 2.2 The time-domain data related to the open cavity extracted in step 2.1 The data is merged (or combined) to form data frame O1:

[0093]

[0094] in, This indicates that time-domain data and frequency-domain data are connected and fused.

[0095] Step 3: Dimensionality reduction of the data frame O1 obtained in Step 2

[0096] In this embodiment, principal component analysis is selected to reduce the dimensionality of the data frame O1 obtained in step 2.

[0097] The covariance of any two data points x and y in data frame O1 is:

[0098]

[0099] The covariance matrix can be obtained from the covariance, and it is defined as follows:

[0100]

[0101] Singular value decomposition is:

[0102] D T =U∑V T

[0103] Where U is the left singular matrix and V is the right singular matrix.

[0104] Since the covariance matrix is ​​computationally expensive, and the right singular matrix V in singular value decomposition is the principal component of principal component analysis, this invention uses singular value decomposition to solve for the principal components, and uses the solved principal components as the data frame after dimensionality reduction.

[0105] Step 4: Build and train the intelligent model

[0106] Step 4.1 Determine the input feature x and the output label y, where:

[0107] The input feature x is the data frame after dimensionality reduction in step 3;

[0108] The output label 'y' represents the actual measured value after processing and manufacturing. Figure 5 The surface roughness values ​​of the open cavity on the fuselage frame of the aircraft shown are as follows: side surface A, side surface B, and bottom surface C. Side surface A and side surface B were each measured twice, and bottom surface C was measured three times due to its larger area. The maximum surface roughness values ​​of side surface A, side surface B, and bottom surface C were taken as the surface roughness values ​​of their respective surfaces, and the maximum surface roughness value of bottom surface C was taken as the output label y. The specific measured values ​​are shown in Table 1 below.

[0109] Table 1

[0110]

[0111]

[0112] Step 4.2: Divide the training set and the test set;

[0113] The measured output label y is added as a column of data to the data frame after dimensionality reduction in step 3 to obtain a data frame with output label y. 80% of the data in the data frame with output label y is divided into the training set, and the rest is divided into the test set.

[0114] Step 4.3: Build and train the intelligent model;

[0115] In this embodiment, the intelligent models constructed and trained include four quality prediction models: linear regression, random forest, DNN, and XGBoost. The following section will combine these models... Figure 6 The construction and training methods of the four quality prediction models in this embodiment will be described respectively.

[0116] a) Construction and training of linear regression models

[0117] First, create a linear regression object.

[0118] Next, the training set obtained in step 4.2 is imported into the linear regression object for fitting training, and tested on the test set until the fit meets the set requirements, thus obtaining the trained linear regression model.

[0119] Furthermore, following the method described in step 4.3 above for constructing a data frame with output label y, a new data frame can be constructed, and the mean squared error (MSE) can be used to evaluate the performance of the linear regression model on the new data frame. If the requirements are not met, the training set and test set can be divided using a new data frame, and training and testing can be performed again until the goodness of fit of the linear regression model meets the requirements.

[0120] Here, in this embodiment, the parameters Coefficients of the linear regression model are obtained through testing and verification. The mean squared error of the linear regression model is 0 on the training set and 0.00234 on the test set, thus obtaining the final linear regression model:

[0121] y=-8.82403e-08+3.79204e-08x1+1.333368e-08x2+...+1.15040e-07x k

[0122] Where -8.82403e-08 is the bias term of the linear regression model, x1,x2,...,x k These are the parameters for the linear regression model, k = 1,000,000, with a total of 1,000,000 fine-tuning parameters.

[0123] b) Construction and training of the random forest model

[0124] First, create a regression object for the random forest model and set the random forest parameters n_estimators (100 in this example), random_state (42 in this example), and max_depth (10 in this example).

[0125] Next, the training set obtained in step 4.2 is imported into the regression object of the random forest model for fitting training, and tested on the validation set until the fit meets the set requirements, thus obtaining the trained random forest model.

[0126] Similarly, a new data frame can be constructed using the method described in step 4 above, which involves constructing a data frame with output label y. The mean squared error (MSE) can then be used to evaluate the performance of the random forest model on the new data frame. If the performance does not meet the requirements, the training and testing sets can be divided using a new data frame, and the training and testing can be repeated until the fit of the random forest model meets the requirements.

[0127] In this embodiment, the Random Forest model is evaluated using MSE, and the model is presented in an intuitive and easy-to-understand visual graphical format, such as... Figure 7 As shown.

[0128] In this random forest model, each box represents a node. The first row of each node contains the decision criteria, the second row contains the squared error, the third row contains the number of samples, and the fourth row contains the predicted value. Taking the first node as an example, 375096-AI 2 (m / s) 2 ) <= -0.04, meaning that for each row X entering this model, only their "375096-AI 2(m / s" are considered at the first node. 2 If the feature value is less than or equal to -0.04 (related to the input training data and the set model hyperparameters), then this row of X is assigned to the next node on the left. The judgment condition of the next node is then checked, and so on, until there are no more nodes to go. The predicted y value is then taken from the last row of the last node.

[0129] c) Construction and training of DNN models

[0130] Set the network structure parameters of the DNN model, import the training set from step 4.1, and iteratively train the model until the loss function value meets the preset requirements. If the number of iterations has reached the maximum value, but the loss function value still cannot meet the preset requirements, then it is necessary to adjust the network structure parameters and / or hyperparameters.

[0131] In DNN model construction, the function used to measure error is called the loss function (LF), such as... Figure 8 The diagram shows the iterative loss during DNN model training. During DNN model training, the DNN model parameters are iterated a finite number of times using an optimization algorithm to minimize the value of the loss function, thereby obtaining a solution to the DNN model; this type of solution is called a numerical solution.

[0132] The DNN model used in this embodiment has a 4-layer structure: the input layer has 128 nodes and uses the ReLU activation function; the first hidden layer has 64 nodes and uses the sigmoid activation function; the second hidden layer has 32 nodes and uses the sigmoid activation function; and the output layer has 1 node and uses the linear activation function. It is worth noting that since this is a regression problem, the number of nodes is set to 1, and the linear activation function is used.

[0133] It should be emphasized that, in the DNN model training process, this embodiment sets the number of iterations of the hyperparameter epochs to 50 and the batch size (the amount of data input for model training in one session) to 10 based on empirical values.

[0134] d) Construction and training of the XGBoost model

[0135] eXtreme Gradient Boosting (XGBoost) is an optimized distributed gradient boosting library, an improved version of GBDT (Gradient Boosting Decision Tree).

[0136] First, create an XGBoost regression object by setting parameters;

[0137] Secondly, the training dataset is imported into the XGBoost regression object for fitting training;

[0138] Finally, XGBoost model predictions were performed on the test set, and the model prediction performance was evaluated using MSE.

[0139] To further optimize the quality prediction model and improve its data utilization, robustness, and generalization ability, this embodiment can also use cross-validation to train and validate the constructed quality prediction model, depending on the specific characteristics of the quality prediction model, during the training of each quality prediction model.

[0140] Cross-validation involves three steps:

[0141] Step 1: Divide the training set obtained in Step 4.2 into k mutually exclusive subsets, each subset maintaining a consistent data distribution. That is:

[0142] D = {A1, A2, ..., A} k}, k>1, k∈N +

[0143] Make:

[0144] A1 = A2 = ... = A k ,A1∪A2∪...∪A k =D,A1∩A2∩...∩A k =φ

[0145] Where A1, A2, ..., A k Let φ be the k subsets, φ be the empty set, and D be the training set.

[0146] Step 2: Each time, randomly select the union of k-1 subsets to form the training set T. r The remaining subset is the verification set T.

[0147] T r =∪A j ,j≠i

[0148] T = A i i = 1, 2, ..., k

[0149] Step 3: Take the mean of the model parameter values ​​of the quality prediction model determined by k joint cross-training and validation as the optimal parameter value M of the trained quality prediction model, as shown in the following formula:

[0150]

[0151] In the above steps, when k = 10, it is 10-fold cross-validation.

[0152] Finally, the cross-trained quality prediction model is tested using the test set from step 4.2.

[0153] Furthermore, in this embodiment, the mean squared error (MSE) is used for evaluation when testing the quality prediction model after fitting and training. In other embodiments, the root mean square error (RMSE) or the mean absolute error (MAE) can also be used for evaluation.

[0154] Step 5: Build an adaptive hybrid model library

[0155] For each machining feature of the open cavity on the aircraft fuselage frame, the percentage of prediction error of each quality prediction model constructed and trained in step 4 is obtained. The quality prediction model with the smallest percentage of prediction error is selected as the optimal quality prediction model for each machining feature, and an adaptive hybrid model library is established.

[0156] In this embodiment, the process of selecting the quality prediction model with the smallest prediction error percentage for each processing feature is visualized as follows: Figure 9 As shown in the figure. The horizontal axis represents the features to be processed, and the vertical axis represents the percentage error of the prediction result. In the figure, Model 1 is Linear Regression, Model 2 is Random Forest, Model 3 is DNN, and Model 4 is XGBoost. Finally, the optimal quality prediction model adaptively selected for each feature to be processed is Model 3, Model 3, Model 2, Model 2, Model 2, Model 1, Model 2, Model 4, Model 4, Model 4, which is shown by the black line in the figure.

[0157] In other embodiments, the MSE prediction probability values ​​of each quality prediction model constructed and trained in step 4 can be obtained separately. For each processing feature, the quality prediction model with the lowest MSE prediction probability value can be selected as its optimal quality prediction model, and an adaptive hybrid model library can be established. The specific implementation method is as follows:

[0158] Construct the probability model matrix:

[0159]

[0160] In the formula:

[0161] p mn This represents the MSE prediction probability value of m quality prediction models simultaneously for n processing features;

[0162] M is an m×n matrix, where each element takes only values ​​from 0 and 1;

[0163] For any quality prediction model p ij ,1≤i≤m,1≤j≤n, find min(p i1 ,p i2 ,…,p in If 1≤i≤m, set the remaining predicted probability values ​​in the column containing the minimum probability value to 0, and set the corresponding element in the matrix M to 1; and so on, the matrix M can be obtained by following this method.

[0164] y n Select labels for the optimal model, [y1, y2, ..., y n ] T To select the optimal model sequence for different processing characteristics.

[0165] Based on the above probability model matrix [y1, y2, ..., y n ] T This allows for the adaptive selection of the optimal quality prediction model based on processing characteristics during subsequent manufacturing process quality prediction, enabling high-precision quality prediction.

[0166] Step 6: Quality Prediction in the Processing and Manufacturing Process

[0167] Based on the machining characteristics of the open cavity structure on the current fuselage frame, the optimal quality prediction model matching the current machining characteristics is selected from the adaptive hybrid model library constructed in step 5. The multimodal data collected in real time during the machining process is input into the currently selected optimal quality prediction model. The optimal quality prediction model outputs the quality prediction result (the surface roughness value in this embodiment), thereby realizing the quality prediction of the machining process of the open cavity structure on the fuselage frame.

[0168] Reference Figure 2 In addition to the manufacturing process quality prediction method described above, this invention also provides a manufacturing process quality prediction system, including a multimodal data acquisition unit, a data processing unit, a model building and training unit, an adaptive hybrid model library, and a quality prediction unit.

[0169] The multimodal data acquisition unit uses a CPS-based actual machining prototype system to collect and store multimodal data during the machining and manufacturing process.

[0170] The data processing unit is used to extract data related to the parts to be processed from the raw multimodal data acquired by the multimodal data acquisition unit, convert it into frequency domain data, form a data frame by data merging, and perform data dimensionality reduction processing on the formed data frame.

[0171] The model building and training unit is used to build an intelligent model, which includes multiple quality prediction models, including but not limited to at least two of the following: linear regression model, random forest model, DNN model, and XGBoost model.

[0172] The adaptive hybrid model library includes optimal quality prediction models that correspond one-to-one with each part to be processed and each of its processing features. The optimal quality prediction model is selected from the intelligent models trained by the model building and training unit. For a certain processing feature, the quality prediction model with the smallest percentage error in its prediction result is selected as the optimal quality prediction model for that processing feature.

[0173] The quality prediction unit is used to select the optimal quality prediction model from the adaptive hybrid model library based on the parts to be processed and the processing features in the current processing process, and output the quality prediction result of the current processing features.

Claims

1. A manufacturing process quality prediction method based on multimodal data and an adaptive hybrid model, characterized in that, Includes the following steps: Step 1: Acquire multimodal data, which includes at least vibration data and noise data in the XYZ directions of the machining coordinate system; Step 2: Process the multimodal data to obtain a data frame; Step 2.1 Locate, segment, and extract data related to the monitored object from the raw multimodal data collected in Step 1. This extracted data is time-domain data. Step 2.2 converts the captured data into frequency domain data; Step 2.3 Based on the frequency domain data, obtain the frequency domain characteristics, and through the analysis of the frequency domain characteristics, assign the corresponding confidence level to the time domain data of Step 2.1, and perform time-frequency domain data fusion to form a data frame; Step 3: Perform dimensionality reduction on the data frame; Step 4: Using the dimensionality-reduced data frame as input features and the parameters used to characterize the processing and manufacturing quality as output labels, construct and train an intelligent model. The intelligent model includes multiple independently constructed and trained quality prediction models, including but not limited to linear regression models, random forest models, DNN models, or XGBoost models. Step 5: For each processing feature on each type of part to be processed, obtain the prediction error percentage of each quality prediction model constructed and trained in Step 4. Select the quality prediction model with the smallest prediction error percentage for each processing feature on each type of part to be processed as the optimal quality prediction model for subsequent quality prediction. Based on this, establish an adaptive hybrid model library, which includes the optimal quality prediction model that corresponds one-to-one with each part to be processed and each processing feature. Step 6: Based on the parts to be processed and their processing characteristics in the current manufacturing process, select the optimal quality prediction model from the adaptive hybrid model library built in Step 5. Input the multimodal data collected in real time during the manufacturing process into the currently selected optimal quality prediction model. The optimal quality prediction model outputs the quality prediction result, thereby realizing the quality prediction of the product manufacturing process.

2. The method of claim 1, wherein: Step 1 specifically involves: building a CPS-based manufacturing workshop / system prototype system, using the prototype system to acquire multimodal data corresponding to different processing parameters during the manufacturing process, and storing them in file formats that can retain complete data.

3. The method of claim 1, wherein: The multimodal data obtained in step 1 also includes tool and / or workpiece temperature data during processing, displacement data of cutting tool and / or workpiece features, cutting force data, cutting fluid injection pressure data, and / or light sensing data.

4. The manufacturing process quality prediction method based on multi-modal data and adaptive mixture model according to any one of claims 1-3, characterized in that: Based on the confidence level assigned to the time-domain data, remove less than or equal to 20% of the time-domain data to obtain a data frame with better data quality.

5. The method of claim 4, wherein: In step 3, principal component analysis or feature-selective dimensionality reduction is used to reduce the dimensionality of the data frame.

6. The method of claim 5, wherein: The intelligent models mentioned in step 4 include linear regression models, random forest models, DNN models, and XGBoost models.

7. The manufacturing process quality prediction method based on multimodal data and an adaptive hybrid model according to claim 6, characterized in that: When constructing and training each quality prediction model in the intelligent model, cross-validation is used to train and validate the constructed quality prediction models.

8. A manufacturing process quality prediction system based on multimodal data and an adaptive hybrid model, characterized in that: It includes a multimodal data acquisition unit, a data processing unit, a model building and training unit, an adaptive hybrid model library, and a quality prediction unit; The multimodal data acquisition unit is used to collect and store multimodal data during the processing and manufacturing of parts and / or products; The data processing unit is used to extract data related to the part to be processed from the multimodal data and convert it into frequency domain data, then form a data frame by merging the data, and perform data dimensionality reduction processing on the formed data frame; The model building and training unit is used to build an intelligent model, which includes multiple quality prediction models, including but not limited to linear regression models, random forest models, DNN models, or XGBoost models. The adaptive hybrid model library includes optimal quality prediction models that correspond one-to-one with each part to be processed and each of its processing features. The optimal quality prediction model is selected from the intelligent models. For a certain processing feature, the quality prediction model with the smallest percentage error in its prediction result is selected as the optimal quality prediction model for that processing feature. The quality prediction unit is used to select the best matching quality prediction model from the adaptive hybrid model library based on the parts to be processed and the processing features in the current processing process, and output the quality prediction result of the current processing features.

Citation Information

Patent Citations

  • A method for process control of heavy marine diesel engine components based on digital twins

    CN109270899B

  • System-level M-CPS architecture for intelligent mine production

    CN116431613A

  • Motor train unit bogie business data cleaning method based on artificial intelligence

    CN114647640A

  • Embedded intelligent manufacturing resource configuration method based on edge computing and terminal system

    CN116128254A