A metal casting process quality monitoring system and method for intelligent manufacturing

Abstract

The present invention discloses a metal casting process quality monitoring system and method for intelligent manufacturing, and relates to the technical field of metal casting quality monitoring. The system uses Bayesian methods and multi-model hypothesis testing to monitor the casting process in real time, identify whether there are deviations or anomalies, and make corresponding adjustments to ensure that the quality of the casting always meets the predetermined standards; through four pre-trained models in parallel, the sensor data after standardization is extracted and fused, and the pre-trained model is fine-tuned through transfer learning; the integrated method based on the multi-layer perceptron model performs classification prediction or regression prediction on the fused feature vector, and combines the uncertainty quantification method to evaluate the uncertainty of the prediction result. The system and method of the present invention can not only monitor various physical parameters in the casting process in real time, but also accurately predict and promptly discover problems, reduce waste in production, and improve the overall quality of castings.

Description

Technical Field

[0001] The present invention relates to the technical field of metal casting quality monitoring, and in particular to a metal casting process quality monitoring system and method thereof for intelligent manufacturing. Background Art

[0002] As an important manufacturing process, metal casting plays a vital role in aerospace, automobile manufacturing, mechanical equipment and other industries. However, the inevitable defects in the casting process will seriously affect the quality of the product and bring huge losses to the enterprise. At present, traditional defect detection methods usually rely on manual visual inspection or later destructive testing, which are not only inefficient but also difficult to meet the requirements of large-scale production and high-precision quality control.

[0003] In recent years, deep learning and machine learning techniques have been widely used in the fields of defect detection and process control, especially in image classification and quality inspection. However, due to the scarcity of data and quality fluctuations in the casting process, existing technologies are often unable to effectively deal with small data sets. In addition, traditional defect detection methods usually ignore the uncertainty of model prediction results, which may lead to wrong predictions in high-risk situations.

[0004] Therefore, the present invention proposes a metal casting process quality monitoring system and method, which improves the control accuracy of casting quality and reduces production costs through real-time monitoring and intelligent prediction. Summary of the invention

[0005] The purpose of the present invention is to provide a metal casting process quality monitoring system and method for intelligent manufacturing, which adopts transfer learning, deep learning, uncertainty quantification technology and multi-model integrated learning method to solve the limitations of traditional casting defect detection and quality control methods, thereby improving the quality and production efficiency of casting products.

[0006] The present invention achieves the above-mentioned purpose through the following technical solutions:

[0007] On the one hand, the present invention provides a metal casting process quality monitoring system for intelligent manufacturing, the system comprising:

[0008] Sensor module, used to collect sensor data in the metal casting process in real time, including temperature data, vibration data, photoelectric signal data and machine control signal data;

[0009] The data preprocessing module is used to perform denoising, normalization and standardization on the collected raw sensor data;

[0010] The process monitoring module is used to monitor the casting process in real time based on the collected sensor data using Bayesian methods and multi-model hypothesis testing, identify whether there are deviations or anomalies, and make corresponding adjustments to ensure that the quality of the castings always meets the predetermined standards;

[0011] A feature extraction module, which is used to extract and fuse features of the standardized sensor data through parallel processing of four pre-trained models, and fine-tune the pre-trained models through transfer learning. The four pre-trained models include EfficientNetV2, Swin Transformer, Vision Transformer, and ConvNeXT;

[0012] The integrated learning module is used to perform classification prediction or regression prediction on the fused feature vector based on the integrated method of the multi-layer perceptron model, and to evaluate the uncertainty of the prediction results in combination with the uncertainty quantification method;

[0013] The feedback module is used to display the monitoring results of the process monitoring module and the prediction results of the integrated learning module in a visual manner or to issue an early warning.

[0014] A further improvement of the present invention is that the formula for the standardization process is:

[0015] ;

[0016] In the formula, Indicates The original data points of samples; is the percentile of the data set at the 10% position; is the percentile of the data set at the 90% position; For the The amount of mean adjustment for each sample; Used to indicate definition or assignment, that is, it means that the variable on the left is defined or assigned the value of the expression on the right.

[0017] A further improvement of the present invention is that, in the process monitoring module, the method for real-time monitoring of the casting process using the Bayesian method and multi-model hypothesis testing comprises:

[0018] Build a normal process model :The normal process model is established through Gaussian process regression based on the normal data of the historical casting process and the casting quality assessment data. It is used to describe the standard range that the sensor data in the casting process should have under ideal conditions;

[0019] Building anomaly models , , …, : Establish multiple hypothesis testing models for possible abnormal situations, and each abnormal model is used to describe the changes of parameters under certain abnormal situations;

[0020] Calculate the posterior probability: Based on Bayes’ theorem, compare the real-time sensor data with the predictions of each model. (in =0,1,2,…, ), calculate the probability that the current sensor data belongs to the model, the formula is:

[0021] ;

[0022] In the formula, For the given data After that, the model The posterior probability of Indicated in the model The data observed below possibility; For Model The prior probability of For data The overall probability of

[0023] Compare each model The posterior probability :

[0024] If the normal process model If the posterior probability of is high, it means that the process is normal;

[0025] If any abnormal model , , …, If the posterior probability is higher than the preset threshold, the current casting process is considered abnormal, triggering an alarm and requiring the operator to check the equipment or adjust the temperature parameters;

[0026] If the posterior probabilities of multiple anomaly models are simultaneously higher than the threshold, an anomaly alert is issued, requiring comprehensive inspection and intervention.

[0027] A further improvement of the present invention is that the feature extraction module inputs each type of sensor data into four pre-trained models respectively, wherein EfficientNetV2 extracts casting surface quality and temperature fluctuations using a composite scaling method, Swin Transformer extracts temporal variation patterns of temperature and vibration using a sliding window self-attention mechanism, Vision Transformer captures global features of temperature variation, vibration mode, stress and deformation mode, and material flow using a global self-attention mechanism, and ConvNeXT combines a convolutional neural network with a Transformer structure to extract variation trends of machine control signals;

[0028] The feature vectors extracted by the four pre-trained models are concatenated together to form a comprehensive feature vector as the input of the subsequent multi-task learning model.

[0029] A further improvement of the present invention is that the feature extraction module freezes the convolution layer of the pre-trained model by means of transfer learning, and fine-tunes the classification layer or the regression layer. The learning rate adjustment formula during fine-tuning is:

[0030] ;

[0031] In the formula, is the learning rate; is the maximum learning rate, is the minimum learning rate; is the current training step number, is the total number of training steps;

[0032] The Adam optimizer is used to gradually adjust the model weights.

[0033] A further improvement of the present invention is that the ensemble learning module inputs the feature vector extracted and fused by the pre-training model into the multi-layer perceptron model to perform classification or regression tasks;

[0034] In the classification task, the information entropy of the classification prediction is calculated. The information entropy formula is:

[0035] ;

[0036] In the formula, is the entropy of the prediction distribution; The model is for Categories The predicted probability of , C is the total number of categories;

[0037] The model performance is evaluated in combination with uncertainty accuracy. The uncertainty of the model prediction is evaluated by calculating the match between the true category and the predicted category. The calculation formula is:

[0038] ;

[0039] In the formula, is the uncertainty accuracy; is the number of samples for which the prediction is correct and uncertain; is the number of samples for which the prediction is correct and certain; is the number of samples for which the prediction is wrong and uncertain; is the number of samples for which the prediction is wrong and certain;

[0040] In the regression task, the following regression model formula is used to predict the severity of the defect:

[0041] ;

[0042] In the formula, is the predicted defect severity; is the process parameter of the casting; is the distribution characteristics of the material; is the thermophysical property of the material; is the normalization of process parameters; is the thermal diffusivity of the material; and are the mean and standard deviation of the materials, respectively; is the standard deviation of the casting layer.

[0043] A further improvement of the present invention is that the integrated learning module is trained by a multi-layer perceptron, uses the features and labels of the training set for learning, adjusts the network weights to minimize the loss function, and uses the Adam optimizer to dynamically adjust the learning rate during the training process to ensure that the model finds the best decision boundary among all features;

[0044] In the prediction stage, the multi-layer perceptron model processes the input features according to the trained weights and outputs the category or severity of the defect. For regression tasks, the output is a continuous defect severity value, and for classification tasks, the output is the probability distribution of the defect category.

[0045] Through the input of real-time sensor data, the integrated learning module continuously performs defect prediction and quality assessment. If the uncertainty of the model prediction is higher than the threshold, or the quality control results do not meet expectations, the system will recommend adjusting the process parameters to optimize the casting quality.

[0046] On the other hand, the present invention provides a metal casting process quality monitoring method for intelligent manufacturing, using the metal casting process quality monitoring system for intelligent manufacturing as described above, the method comprising the following steps:

[0047] Step 1: Real-time collection of sensor data during the metal casting process, including temperature data, vibration data, photoelectric signal data, and machine control signal data;

[0048] Step 2: De-noise, normalize and standardize the collected raw sensor data;

[0049] Step 3: Based on the collected sensor data, use Bayesian methods and multi-model hypothesis testing to monitor the casting process in real time, identify whether there are deviations or anomalies, and make corresponding adjustments to ensure that the quality of the castings always meets the predetermined standards;

[0050] Step 4: Extract and fuse features of the standardized sensor data through parallel processing of four pre-trained models, and fine-tune the pre-trained models through transfer learning. The four pre-trained models include EfficientNetV2, Swin Transformer, Vision Transformer, and ConvNeXT.

[0051] Step 5: Based on the integration method of the multi-layer perceptron model, the fused feature vector is classified or regressed, and the uncertainty of the prediction result is evaluated by combining the uncertainty quantification method;

[0052] Step 6: Feedback the process monitoring results or prediction results to the production control system and adjust the production parameters in real time to optimize the product quality during the casting process.

[0053] The present invention also provides a computer device, which stores a computer program, and when the computer program is executed by a processor, it implements the metal casting process quality monitoring system for intelligent manufacturing as described above.

[0054] The beneficial effects of the present invention are: through the combination of deep learning and transfer learning, the system can extract effective features from small data sets, thereby improving the accuracy of casting defect detection; the uncertainty quantification method can provide prediction uncertainty for the model, avoid making wrong predictions under low-confidence conditions, and improve the robustness of the system under complex working conditions; real-time process monitoring and fault detection: through the combination of Bayesian methods and integrated learning, real-time monitoring and defect prediction of the casting process are realized, reducing the risks and failure rates in production; reducing costs and improving production efficiency: through non-destructive real-time monitoring and automatic adjustment of production parameters, the present invention can significantly reduce the scrap rate and maintenance costs in production, while improving production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor. Among them:

[0056] Figure 1 It is a system modular diagram of the present invention;

[0057] Figure 2 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solution and advantages of the embodiment of the present invention clearer, the technical solution of the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings of the embodiment of the present invention. Obviously, the described embodiment is a part of the embodiment of the present invention, not all of the embodiments. Based on the described embodiment of the present invention, all other embodiments obtained by ordinary technicians in this field belong to the scope of protection of the present invention.

[0059] like Figure 1 As shown in FIG. 1 , an embodiment of the present invention is provided, which provides a metal casting process quality monitoring system for intelligent manufacturing, including:

[0060] (1) Sensor module

[0061] It is used to collect sensor data such as temperature, vibration, photoelectric signals and machine control signals in real time during the metal casting process, ensuring that various physical parameters in the casting process can be captured comprehensively and accurately, providing original information for subsequent processing and analysis.

[0062] Temperature data: Real-time monitoring of the temperature changes of the molten pool during the casting process through temperature sensors;

[0063] Vibration data: Use vibration sensors to monitor the vibration status of casting equipment or castings to reflect whether the equipment is operating normally;

[0064] Photoelectric signal data: Through the photodiode sensor, the casting surface temperature, light reflection, etc. are monitored in real time to help determine the molten pool state and metal solidification process;

[0065] Machine control signal data: Collect machine control signals, such as laser power, scanning speed, etc., for comparison with sensor data to determine whether there is a process deviation.

[0066] (2) Data preprocessing module

[0067] Before further analysis, the collected raw sensor data needs to be preprocessed by denoising, normalizing, and standardizing, which helps to eliminate the influence of different dimensions and scales, so that subsequent deep learning and machine learning models can extract features and make predictions more accurately.

[0068] In one embodiment, the formula for the normalization process is:

[0069] ;

[0070] In the formula, Indicates The raw data points of each sample (can be parameters in the casting process such as temperature and vibration);

[0071] It is the percentile of the data set at the 10% position (usually a percentile of the lowest value of the sample, indicating the 10% minimum value in the data set), which is used to extract low-end values ​​from the data set for standardization, and may correspond to a relatively "hard" or low-temperature, low-strength casting area;

[0072] The 90% percentile of the data set (usually a percentile of the maximum value of the sample, representing the 90% maximum value in the data set), used to extract high-end values ​​from the data set for standardization, which may correspond to relatively "soft" or high-temperature, high-strength casting areas;

[0073] For the The mean adjustment for each sample (usually the mean of the data set) After some form of processing, it is used to adjust the data offset). This adjustment is to decentralize the data and remove the data offset, so that all data is relatively concentrated, which helps to remove possible errors.

[0074] Used to indicate definition or assignment, that is, it means that the variable on the left is defined or assigned the value of the expression on the right.

[0075] The standardized method based on material distribution characteristics in this embodiment is highly targeted and adaptable, and can take into account the differences in thermodynamics, vibration response, etc. among different material groups (hard, soft, reference samples). It is particularly suitable for accurately distinguishing and processing the characteristics of different material groups in the metal casting process, and effectively improves the accuracy and reliability of quality monitoring.

[0076] (3) Process monitoring module

[0077] It is used to monitor the casting process in real time based on sensor data using Bayesian methods and multi-model hypothesis testing (MMHT), identify whether there are deviations or anomalies, and make corresponding adjustments to ensure that the quality of castings always meets the predetermined standards.

[0078] The Bayesian method is a statistical method based on probabilistic reasoning, which can infer the probability of uncertain events based on prior knowledge and newly acquired data. Using the Bayesian method for multi-model hypothesis testing can help the system to evaluate in real time whether various parameters in the casting process meet normal expectations and detect any potential anomalies or deviations in a timely manner.

[0079] The specific implementation methods include:

[0080] Build a normal process model : The normal process model is established through Gaussian process regression based on the normal data of the historical casting process and the casting quality assessment data. It is used to describe the standard range that sensor data (such as temperature, vibration, etc.) in the casting process should have under ideal conditions. For example, the range of a certain temperature in the normal casting process should be within a certain interval. If it deviates from this interval, it means there is a problem.

[0081] Building anomaly models , , …, : Establish multiple hypothesis testing models for possible abnormal situations (such as excessive temperature, excessive equipment vibration, etc.). Each abnormal model is used to describe the change of parameters under certain abnormal conditions. For example, if the temperature is too high, establish a model to describe the distribution of abnormal temperature.

[0082] Calculate the posterior probability: Based on Bayes’ theorem, compare the real-time sensor data with the predictions of each model. (in =0,1,2,…, ), calculate the probability that the current sensor data belongs to the model, the formula is:

[0083] ;

[0084] In the formula, For the given data After that, the model The posterior probability of Indicated in the model The data observed below possibility; For Model The prior probability of , usually expressed as the probability of occurrence of the model; For data The overall probability of , as a normalizing constant;

[0085] Compare each model The posterior probability , if the normal process model The posterior probability of is high, indicating that the process is normal; if any abnormal model , , …, If the posterior probability of a model is higher than the preset threshold, the current casting process is considered abnormal, triggering an alarm and requiring the operator to check the equipment or adjust the temperature parameters. If the posterior probability of multiple abnormal models is higher than the threshold at the same time, an abnormal alarm is issued, requiring a comprehensive inspection and intervention.

[0086] As the amount of data in the casting process continues to increase, the system can adaptively adjust the parameters of each model so that the model can more accurately describe the changes in the actual process. By feeding back historical data and real-time data, the Bayesian method can continuously update the prior probability of each model. , so that the model always maintains accuracy

[0087] This method uses Bayesian reasoning to calculate the posterior probability of each hypothetical model, ensuring that the system can respond promptly when deviations or anomalies occur in the casting process, and continuously improves prediction accuracy and system reliability through model optimization and feedback mechanisms.

[0088] The process monitoring module focuses on real-time process monitoring and deviation detection. Its main purpose is to determine whether anomalies occur during the casting process by comparing with the established normal process model, and to identify deviations in the casting process by relying on real-time sensor data and known normal ranges. In addition to the Bayesian method and MMHT, the embodiments of the present invention also provide long-term quality assessment through feature extraction and quality prediction, and adjust production parameters or take corresponding measures based on the predicted quality problems.

[0089] (4) Feature extraction module

[0090] It is used to extract and fuse features of standardized sensor data through parallel processing of four pre-trained models, and fine-tune the pre-trained models through transfer learning. The four pre-trained models include EfficientNetV2, SwinTransformer (Swin-T), Vision Transformer (ViT) and ConvNeXT.

[0091] EfficientNetV2 is an efficient convolutional neural network architecture. It improves performance by expanding the network width, depth and resolution through compound scaling methods. It is suitable for processing large-scale data and has high computational efficiency. It focuses on extracting local features in images (such as local change patterns such as temperature and vibration). Swin Transformer is a visual model based on the Transformer architecture. It uses a sliding window mechanism and can handle long-range dependencies. It is good at extracting features with a large spatial range and is suitable for complex patterns in vibration and temperature data. ViT is a pure Transformer architecture that focuses on global dependency features at the image level and is suitable for capturing global relationships between long time series data. ConvNeXT is a new convolutional network architecture that combines the advantages of convolution and Transformer, has powerful multi-scale feature extraction capabilities, and is suitable for extracting rich features from complex data. Each model extracts features from the data independently and uses its different architectural advantages to capture various regularities in the data.

[0092] Since these four pre-trained models have been trained on large-scale datasets (such as ImageNet), they can learn universal feature representations. Through transfer learning, the features learned by these pre-trained models on large-scale datasets (such as ImageNet) are used to improve their performance on casting process data (such as temperature, vibration signals, photoelectric signals, etc.).

[0093] The fine-tuning process is mainly performed on the last few layers of each model, especially the classification layer or regression layer. Fine-tuning enables the model to focus on specific patterns in the casting process data by training these layers. During fine-tuning, the mini-batch gradient descent (SGD) or Adam optimization algorithm is used to gradually adjust the model weights so that the pre-trained model adapts to the data characteristics of the casting process, such as temperature, vibration, and photoelectric signals.

[0094] The features output by the four models are concatenated (or weighted averaged) together to form a comprehensive feature vector;

[0095] Concatenation: concatenate the feature vectors of the four models along the feature dimension to form a longer vector;

[0096] Weighted average: assign different weights to the features of each model based on the performance of each model (such as accuracy, stability), and then add up the weighted features;

[0097] Final output: The fused feature vector will be used as the input of the subsequent multi-task learning model for further classification, regression and other tasks.

[0098] In one embodiment, the feature extraction module inputs each type of sensor data into four pre-trained models respectively, where EfficientNetV2 extracts casting surface quality and temperature fluctuation using a compound scaling method, SwinTransformer extracts temporal variation patterns of temperature and vibration using a sliding window self-attention mechanism, VisionTransformer captures global features of temperature variation, vibration mode, stress and deformation mode, and material flow using a global self-attention mechanism, and ConvNeXT combines a convolutional neural network with a Transformer structure to extract variation trends of machine control signals;

[0099] The feature vectors extracted by the four pre-trained models are concatenated together to form a comprehensive feature vector as the input of the subsequent multi-task learning model.

[0100] With the help of transfer learning, freeze the convolutional layer of the pre-trained model and fine-tune the classification layer or regression layer. The learning rate adjustment formula during fine-tuning is:

[0101] ;

[0102] In the formula, is the learning rate; is the maximum learning rate, is the minimum learning rate; is the current training step number, is the total number of training steps;

[0103] The Adam optimizer is used to gradually adjust the model weights.

[0104] (5) Integrated learning module

[0105] An ensemble method based on a multi-layer perceptron (MLP) model is used to perform classification or regression prediction on the extracted features, and the uncertainty quantification (UQ) method is used to evaluate the uncertainty of the prediction results.

[0106] MLP is a feedforward neural network, usually composed of multiple fully connected layers. The MLP model accepts the fused feature vector as input, and outputs the prediction result after multiple layers of nonlinear transformation.

[0107] In a specific embodiment, the implementation method of the integrated learning module includes:

[0108] The feature vector extracted and fused by the pre-trained model is input into the multi-layer perceptron model for classification or regression tasks;

[0109] In the classification task (determining whether a casting has defects and the classification of defect types), in order to better evaluate the model uncertainty, the information entropy of the classification prediction is first calculated. This is an important indicator for measuring the uncertainty of the model output distribution. The information entropy formula is:

[0110] ;

[0111] In the formula, is the entropy of the prediction distribution; The model is for Categories The predicted probability of , C is the total number of categories; the larger this value is, the more uncertain the model's prediction is;

[0112] Information entropy can be used to guide model adjustments and decision-making. For example, in the casting process, if the information entropy is high, it means that the prediction uncertainty is large, and it may be necessary to further optimize the casting parameters or use additional data to improve the prediction accuracy.

[0113] The model performance is evaluated in combination with uncertainty accuracy. The uncertainty of the model prediction is evaluated by calculating the match between the true category and the predicted category. The calculation formula is:

[0114] ;

[0115] In the formula, is the uncertainty accuracy; is the number of samples for which the prediction is correct and uncertain; is the number of samples for which the prediction is correct and certain; is the number of samples for which the prediction is wrong and uncertain; is the number of samples for which the prediction is wrong and certain.

[0116] Through this formula, the system can effectively measure the performance of the model when facing uncertain data, which is of great significance for defect detection and quality control in the metal casting process. If the model's accuracy is low, it can be supplemented by optimizing the training data set or using more sensor data to improve the ability to handle uncertainty.

[0117] In the regression task (predicting the severity of defects, the hardness, strength and other physical properties of castings), the following regression model formula is used to predict the severity of defects:

[0118] ;

[0119] In the formula, is the predicted defect severity; is the process parameter of the casting; is the distribution characteristics of the material; is the thermophysical property of the material; is the normalization of process parameters; is the thermal diffusivity of the material; and are the mean and standard deviation of the materials, respectively; is the standard deviation of the casting layer.

[0120] The training is performed through a multi-layer perceptron (MLP), using the features and labels of the training set to learn and adjust the weights of the network to minimize the loss function. The training process uses the Adam optimizer to dynamically adjust the learning rate to ensure that the model finds the best decision boundary among all features;

[0121] In the prediction phase, the MLP model processes the input features according to the trained weights and outputs the defect category or severity (depending on the task). For regression tasks, the output is a continuous defect severity value, and for classification tasks, the output is a probability distribution of the defect category.

[0122] During the production process, the integrated learning model continuously predicts defects and evaluates quality through the input of real-time sensor data. If the uncertainty of the model prediction is high or the quality control results do not meet expectations, the system will recommend adjusting process parameters (such as temperature, speed, laser power, etc.) to optimize the casting quality.

[0123] For example, if the predicted defect severity If the pressure is above the set threshold, the system can automatically trigger an alarm, requiring the operator to adjust production parameters or conduct equipment inspections.

[0124] (6) Feedback module

[0125] It is used to display the monitoring results of the process monitoring module and the prediction results of the integrated learning module in a visual way or to issue an early warning.

[0126] like Figure 2 As shown, another embodiment of the present invention provides a metal casting process quality monitoring method for intelligent manufacturing, using the metal casting process quality monitoring system for intelligent manufacturing as described above, including the following steps:

[0127] Step 1: Real-time collection of sensor data during the metal casting process, including temperature data, vibration data, photoelectric signal data, and machine control signal data;

[0128] Step 2: De-noise, normalize and standardize the collected raw sensor data;

[0129] Step 3: Based on the sensor data, use Bayesian methods and multi-model hypothesis testing to monitor the casting process in real time, identify whether there are deviations or anomalies, and make corresponding adjustments to ensure that the quality of the castings always meets the predetermined standards;

[0130] Step 4: Extract and fuse features of the standardized sensor data through parallel processing of four pre-trained models, and fine-tune the pre-trained models through transfer learning. The four pre-trained models include EfficientNetV2, Swin Transformer, Vision Transformer, and ConvNeXT.

[0131] Step 5: Based on the integration method of the multi-layer perceptron model, the fused feature vector is classified or regressed, and the uncertainty of the prediction result is evaluated by combining the uncertainty quantification method;

[0132] Step 6: Feedback the process monitoring results or prediction results to the production control system and adjust the production parameters in real time to optimize the product quality during the casting process.

[0133] An embodiment of the present invention further provides a computer device storing a computer program, wherein the computer program, when executed by a processor, implements the metal casting process quality monitoring system for intelligent manufacturing as described above.

[0134] In summary, the present invention collects temperature, vibration, photoelectric signal and machine control signal data in the metal casting process in real time through the sensor module. The system can comprehensively and accurately obtain various key physical parameters in the metal casting process. These data provide rich original information for subsequent data processing, feature extraction and prediction, which is helpful to achieve efficient quality monitoring; the data preprocessing module ensures the consistency and comparability of different sensor data, provides high-quality input data for subsequent deep learning and machine learning models, and improves the accuracy and stability of feature extraction and prediction; the process monitoring module uses Bayesian methods and multi-model hypothesis testing to monitor the casting process in real time, and can accurately identify deviations or anomalies in the casting process based on sensor data. By comparing with the normal process model, the system can promptly discover potential problems and make corresponding adjustments to ensure that the quality of the casting always meets the predetermined standards, effectively avoiding the production of unqualified products; the feature extraction module uses four advanced pre-trained models (EfficientNetV2, Swin Transformer, Vision Transformer and ConvNeXT) extract features from standardized sensor data. These models have different advantages and can extract local and global features from the data, which helps to capture complex patterns in the metal casting process. The pre-trained model is fine-tuned through transfer learning to better adapt to the casting process data, improving the accuracy and reliability of feature extraction; the integrated learning module uses a multi-layer perceptron (MLP) model to perform classification prediction or regression prediction on the fused features. In the classification task, the system can accurately identify whether the casting has defects; in the regression task, the system can accurately predict the severity of the defects. The integrated learning method effectively combines the advantages of multiple models and improves the accuracy and stability of the prediction. By combining uncertainty quantification methods, the system can evaluate the uncertainty of the prediction results. This technology can measure the uncertainty of the model on the input data, especially in high-risk and low-confidence situations. The system can provide uncertainty measurements to help decision makers make more reliable adjustments and early warnings. For example, when predicting defects, if the uncertainty is high, the system can prompt the operator to check further, thereby reducing the risk of misjudgment.

[0135] The system and method of the present invention can not only monitor various physical parameters in the casting process in real time, but also optimize the production process according to the prediction results and adjust relevant parameters (such as temperature, laser power, machine control signal, etc.). Through integrated learning and uncertainty quantification, the system can accurately predict and detect problems in time, reduce waste in production, improve the overall quality of castings, and ensure that the final product meets the expected quality standards.

[0136] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of various changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A metal casting process quality monitoring system for intelligent manufacturing, characterized in that: The system comprises: Sensor module, used to collect sensor data in the metal casting process in real time, including temperature data, vibration data, photoelectric signal data and machine control signal data; The data preprocessing module is used to perform denoising, normalization and standardization on the collected raw sensor data; The process monitoring module is used to monitor the casting process in real time based on sensor data using Bayesian methods and multi-model hypothesis testing to identify deviations or anomalies and make corresponding adjustments to ensure that the quality of castings always meets the predetermined standards; A feature extraction module, which is used to extract and fuse features of the standardized sensor data through parallel processing of four pre-trained models, and fine-tune the pre-trained models through transfer learning. The four pre-trained models include EfficientNetV2, Swin Transformer, Vision Transformer, and ConvNeXT; The feature extraction module inputs each type of sensor data into four pre-trained models respectively, among which EfficientNetV2 extracts casting surface quality and temperature fluctuation using a compound scaling method, Swin Transformer extracts the temporal variation pattern of temperature and vibration through a sliding window self-attention mechanism, Vision Transformer captures the global features of temperature variation, vibration mode, stress and deformation mode, and material flow through a global self-attention mechanism, and ConvNeXT combines a convolutional neural network with a Transformer structure to extract the variation trend of machine control signals; The feature vectors extracted by the four pre-trained models are concatenated together to form a comprehensive feature vector as the input of the ensemble learning module; The integrated learning module is used to perform classification prediction or regression prediction on the fused feature vector based on the integrated method of the multi-layer perceptron model, and to evaluate the uncertainty of the prediction results in combination with the uncertainty quantification method; The feedback module is used to display the monitoring results of the process monitoring module and the prediction results of the integrated learning module in a visual manner or to issue an early warning.

2. The metal casting process quality monitoring system for intelligent manufacturing according to claim 1 is characterized in that: The formula for the standardization process is: ； In the formula, Indicates The original data points of samples; is the percentile of the data set at the 10% position; is the percentile of the data set at the 90% position; For the The amount of mean adjustment for each sample; Used to indicate definition or assignment, that is, it means that the variable on the left is defined or assigned the value of the expression on the right.

3. The metal casting process quality monitoring system for intelligent manufacturing according to claim 1 is characterized in that: In the process monitoring module, the method of using the Bayesian method and multi-model hypothesis testing to monitor the casting process in real time includes: Build a normal process model :The normal process model is established through Gaussian process regression based on the normal data of the historical casting process and the casting quality assessment data. It is used to describe the standard range that the sensor data in the casting process should have under ideal conditions; Building anomaly models , , …, : Establish multiple hypothesis testing models for possible abnormal situations, and each abnormal model is used to describe the changes of parameters under certain abnormal situations; Calculate the posterior probability: Based on Bayes’ theorem, compare the real-time sensor data with the predictions of each model. (in =0,1,2,…, ), calculate the probability that the current sensor data belongs to the model, the formula is: ； In the formula, For the given data After that, the model The posterior probability of Indicated in the model The data observed below possibility; For Model The prior probability of For data The overall probability of Compare each model The posterior probability : If the normal process model If the posterior probability of is high, it means that the process is normal; If any abnormal model , , …, If the posterior probability is higher than the preset threshold, the current casting process is considered abnormal, triggering an alarm and requiring the operator to check the equipment or adjust the temperature parameters; If the posterior probabilities of multiple anomaly models are simultaneously higher than the threshold, an anomaly alert is issued, requiring comprehensive inspection and intervention.

4. The metal casting process quality monitoring system for intelligent manufacturing according to claim 1 is characterized in that: The feature extraction module freezes the convolution layer of the pre-trained model with the help of transfer learning, and fine-tunes the classification layer or regression layer. The learning rate adjustment formula during fine-tuning is: ； In the formula, is the learning rate; is the maximum learning rate, is the minimum learning rate; is the current training step number, is the total number of training steps; The Adam optimizer is used to gradually adjust the model weights.

5. The metal casting process quality monitoring system for intelligent manufacturing according to claim 1 is characterized in that: The integrated learning module inputs the feature vector extracted and fused by the pre-training model into the multi-layer perceptron model to perform classification or regression tasks; In the classification task, the information entropy of the classification prediction is calculated. The information entropy formula is: ； In the formula, is the entropy of the prediction distribution; The model is for Categories The predicted probability of , C is the total number of categories; The model performance is evaluated in combination with uncertainty accuracy. The uncertainty of the model prediction is evaluated by calculating the match between the true category and the predicted category. The calculation formula is: ； In the formula, is the uncertainty accuracy; is the number of samples for which the prediction is correct and uncertain; is the number of samples for which the prediction is correct and certain; is the number of samples for which the prediction is wrong and uncertain; is the number of samples for which the prediction is wrong and certain; In the regression task, the following regression model formula is used to predict the severity of the defect: ； In the formula, is the predicted defect severity; is the process parameter of the casting; is the distribution characteristics of the material; is the thermophysical property of the material; is the normalization of process parameters; is the thermal diffusivity of the material; and are the mean and standard deviation of the materials, respectively; is the standard deviation of the casting layer.

6. The metal casting process quality monitoring system for intelligent manufacturing according to claim 5 is characterized in that: The integrated learning module is trained by a multi-layer perceptron, using the features and labels of the training set for learning, and adjusting the network weights to minimize the loss function. The training process uses the Adam optimizer to dynamically adjust the learning rate to ensure that the model finds the best decision boundary among all features. In the prediction stage, the multi-layer perceptron model processes the input features according to the trained weights and outputs the category or severity of the defect. For regression tasks, the output is a continuous defect severity value, and for classification tasks, the output is the probability distribution of the defect category. Through the input of real-time sensor data, the integrated learning module continuously performs defect prediction and quality assessment. If the uncertainty of the model prediction is higher than the threshold, or the quality control results do not meet expectations, the system will recommend adjusting the process parameters to optimize the casting quality.

7. A metal casting process quality monitoring method for intelligent manufacturing, characterized in that: The metal casting process quality monitoring system for intelligent manufacturing according to any one of claims 1 to 6 is applied, and the method comprises the following steps: Step 1: Real-time collection of sensor data during the metal casting process, including temperature data, vibration data, photoelectric signal data, and machine control signal data; Step 2: De-noise, normalize and standardize the collected raw sensor data; Step 3: Based on the sensor data, use Bayesian methods and multi-model hypothesis testing to monitor the casting process in real time, identify whether there are deviations or anomalies, and make corresponding adjustments to ensure that the quality of the castings always meets the predetermined standards; Step 4: Extract and fuse features of the standardized sensor data through parallel processing of four pre-trained models, and fine-tune the pre-trained models through transfer learning. The four pre-trained models include EfficientNetV2, Swin Transformer, Vision Transformer, and ConvNeXT. Step 5: Based on the integration method of the multi-layer perceptron model, the fused feature vector is classified or regressed, and the uncertainty of the prediction result is evaluated by combining the uncertainty quantification method; Step 6: Feedback the process monitoring results or prediction results to the production control system and adjust the production parameters in real time to optimize the product quality during the casting process.

8. A computer device storing a computer program, characterized in that: When the computer program is executed by a processor, the metal casting process quality monitoring system for intelligent manufacturing as claimed in any one of claims 1 to 6 is implemented.

Images