A Vacuum Furnace AI Pressure Control Method Based on Multimodal Data Fusion

Abstract

This invention discloses an AI-based pressure control method for vacuum furnaces based on multimodal data fusion, belonging to the field of vacuum furnace pressure control technology. The method synchronously acquires furnace pressure data, including furnace cavity pressure time-series data, heating zone temperature time-series data, and gas flow time-series data, through the NI CompactRIO platform and IEEE 1588 protocol. After processing with a local outlier factor algorithm, wavelet transform, and Kalman filtering, a furnace pressure feature vector is extracted, including time-domain features, frequency-domain features, and multimodal correlation features. An AI pressure control model is constructed, and the furnace pressure feature vector is input into the AI ​​pressure control model to generate furnace pressure control parameters and convert them into furnace pressure control commands for execution. Simultaneously, safety monitoring is performed during execution; in case of anomalies, a preset safety strategy is executed, and the AI ​​pressure control model is incrementally updated. This invention achieves precise control of vacuum furnace pressure, improving the safety and adaptability of vacuum furnace pressure control.

Description

Technical Field

[0001] This invention relates to the field of vacuum furnace pressure control technology, and specifically to an AI-based pressure control method for vacuum furnaces based on multimodal data fusion. Background Technology

[0002] As a key piece of equipment in the industrial field, the pressure control accuracy of vacuum furnaces directly affects product quality and process stability. Traditional vacuum furnace pressure control often adopts a single-mode control scheme based on PID algorithms, relying solely on pressure sensor data for closed-loop regulation. This approach struggles to cope with nonlinear pressure fluctuations caused by the coupling of multiple factors, including temperature fluctuations in the heating zone and changes in gas flow rate, resulting in problems such as control lag and poor robustness.

[0003] With the development of industrial sensor technology, multi-sensor collaborative monitoring has become a trend. However, existing methods suffer from insufficient multimodal data fusion, reliance on experience in feature engineering, weak adaptability of AI models, and imperfect safety mechanisms. Therefore, achieving high-precision synchronous acquisition and fusion of multimodal data, constructing an adaptively updated AI control model, and establishing a sound safety monitoring mechanism have become key technical challenges in improving the pressure control accuracy of vacuum furnaces. To address this, we propose an AI-based pressure control method for vacuum furnaces based on multimodal data fusion. Summary of the Invention

[0004] The purpose of this invention is to provide an AI-based pressure control method for vacuum furnaces based on multimodal data fusion, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A vacuum furnace AI pressure control method based on multimodal data fusion includes the following steps:

[0007] S1. Real-time acquisition of furnace pressure data of the vacuum furnace;

[0008] S2. Preprocess the furnace pressure data to obtain preprocessed furnace pressure data, and extract features from the preprocessed furnace pressure data to obtain the furnace pressure feature vector;

[0009] S3. Construct an AI pressure control model, input the furnace pressure feature vector into the AI ​​pressure control model, and obtain the furnace pressure control parameters;

[0010] S4. Convert the furnace pressure control parameters into furnace pressure control commands, and execute the furnace pressure control commands accordingly. At the same time, perform safety monitoring on the process of the vacuum furnace executing the furnace pressure control commands, and determine whether to update the AI ​​pressure control model.

[0011] Preferably, the furnace pressure data includes furnace cavity pressure time-series data, heating zone temperature time-series data, and gas flow time-series data. This data is collected in real time by arranging pressure sensors, temperature sensors, and gas flow sensors inside the vacuum furnace cavity, in the heating zone, and in the inlet pipe, respectively. The NI CompactRIO industrial-grade data acquisition platform controls the pressure sensors, temperature sensors, and gas flow sensors to collect data synchronously. Simultaneously, sub-millisecond time synchronization is achieved based on the IEEE 1588 precise time protocol, ensuring the time consistency of the furnace cavity pressure time-series data, heating zone temperature time-series data, and gas flow time-series data.

[0012] Preferably, a method for preprocessing furnace pressure data is as follows:

[0013] The local outlier factor algorithm was used to identify and repair outliers in the time series data of furnace pressure, heating zone temperature, and gas flow. The local outlier factor algorithm model was constructed by setting the number of neighborhood samples to 20 and the contamination rate to 0.05. Outliers identified as outliers were repaired by k-nearest neighbor interpolation, and the furnace pressure data with outlier processing was obtained.

[0014] Wavelet transform filtering was applied to the outlier furnace pressure data. The db4 wavelet basis function was used to perform a three-level decomposition of the outlier furnace pressure data. High-frequency noise was eliminated by retaining low-frequency coefficients and setting high-frequency coefficients to zero. The decomposition process was implemented recursively using the Mallat algorithm. After each level of decomposition, the high-frequency components were thresholded to reconstruct the denoised furnace pressure data. Based on this, Kalman filtering was applied to the denoised furnace pressure data to repair missing values. Finally, the pre-processed furnace cavity pressure time series data, heating zone temperature time series data, and gas flow time series data were obtained, which together constitute the pre-processed furnace pressure data.

[0015] The k-nearest neighbor interpolation method is a missing value imputation method based on local similarity;

[0016] The Mallat algorithm is a fast implementation method for wavelet transform.

[0017] Preferably, the method for feature extraction of the preprocessed furnace pressure data is as follows:

[0018] Preprocessed furnace pressure time series data The mean pressure was calculated using the mean formula, crest factor formula, and standard deviation formula, respectively. Pressure peak factor and pressure standard deviation Based on this, the average pressure and pressure standard deviation Pressure skewness is obtained using the skewness formula and the kurtosis formula. and pressure peak ;

[0019] The formula for the mean is: ;

[0020] The formula for the crest factor is: ;

[0021] The formula for the standard deviation is: ;

[0022] The skewness formula is: ;

[0023] The kurtosis formula is: ;

[0024] Preprocessed furnace pressure time series data The data is converted into the frequency domain spectrum of furnace cavity pressure using Fast Fourier Transform. Based on this, the frequency domain spectrum data of the furnace cavity pressure The pressure frequency center is obtained by calculating using the formulas for the main frequency center, total energy, root mean square frequency, and dominant frequency. Total pressure energy , pressure root mean square frequency and pressure-dominant frequency ;

[0025] The formula for calculating the main frequency center is: ;

[0026] The formula for calculating the total energy is as follows: ;

[0027] The formula for calculating the root mean square frequency is as follows:

[0028] The formula for calculating the dominant frequency is: ;

[0029] Pre-processed furnace cavity pressure time series data and heating zone temperature time series data The pressure-temperature correlation coefficient was calculated using the Pearson correlation coefficient formula. Meanwhile, the pressure-temperature hysteresis correlation was calculated using the hysteresis correlation formula. ;

[0030] Pre-processed furnace cavity pressure time series data Gas flow time series data The pressure-flow correlation coefficient was calculated using the Pearson correlation coefficient formula. Meanwhile, the pressure-flow lag correlation was calculated using the lag correlation formula. ;

[0031] The formula for the Pearson correlation coefficient is:

[0032] ;

[0033] in, yes and covariance, and They are and of , Furnace cavity pressure time series data , Time series data of heating zone temperature or gas flow time series data ;

[0034] The formula for the lagged correlation is as follows: ;

[0035] in, and yes and Another form of expression, yes or .

[0036] Preferably, the furnace pressure feature vector includes time-domain features, frequency-domain features, and multimodal correlation features, wherein the time-domain features include the pressure mean. Pressure peak factor Pressure standard deviation Pressure deviation and pressure peak Frequency domain characteristics include pressure frequency domain center Total pressure energy , pressure root mean square frequency and pressure-dominant frequency Multimodal correlation features include pressure-temperature correlation coefficient. Pressure-temperature hysteresis correlation Pressure-flow correlation coefficient Correlation with pressure-flow hysteresis .

[0037] Preferably, the AI ​​stress control model is jointly constructed by a fusion layer, a decision layer, a temporal feature subnetwork of a CNN structure, a frequency domain feature subnetwork of a fully connected neural network structure, and a multimodal correlation feature subnetwork. The temporal feature subnetwork of the CNN structure includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a fully connected layer, and a first output layer. The frequency domain feature subnetwork of the fully connected neural network structure includes a first hidden layer, a second hidden layer, a third hidden layer, and a second output layer. The multimodal correlation feature subnetwork of the fully connected neural network structure includes a fourth hidden layer, a fifth hidden layer, and a third output layer. The fusion layer includes a feature splicing layer, a batch normalization layer, a fusion network layer, and a fourth output layer. The decision layer includes a first decision layer, a second decision layer, and a fifth output layer.

[0038] Preferably, the process of inputting the furnace pressure feature vector into the AI ​​pressure control model to obtain the furnace pressure control parameters is as follows:

[0039] Temporal features are input into a temporal feature subnetwork. The first convolutional layer of the temporal feature subnetwork uses 16 one-dimensional convolutional kernels of length 3, stride 1 and ReLU activation function to extract short-term dependency patterns in the temporal features. After max pooling dimensionality reduction by the first pooling layer with a window size of 2, the second convolutional layer extracts features again with 32 convolutional kernels of length 3. After pooling by the second pooling layer, the features are mapped to a high-dimensional space through a fully connected layer of 64 neurons. Finally, the first output layer of 32 neurons generates the intermediate representation of temporal features.

[0040] The frequency domain features are input into the frequency domain feature subnetwork. The first hidden layer (64 neurons), the second hidden layer (48 neurons), and the third hidden layer (32 neurons) of the frequency domain feature subnetwork all employ... The LeakyReLU activation function is set to 0.01. The frequency domain features are processed sequentially through the first, second, and third hidden layers, and finally the intermediate expression of the frequency domain features is generated by the second output layer of 16 neurons.

[0041] The multimodal association features are input into the multimodal association feature subnetwork. The 48 neurons in the fourth hidden layer and the 32 neurons in the fifth hidden layer all use the Swish activation function. The multimodal association features are processed through the fourth and fifth hidden layers in sequence, and finally the intermediate vector of multimodal association features is generated by the third output layer with 16 neurons.

[0042] The 32-dimensional temporal feature intermediate representation, the 16-dimensional frequency domain feature intermediate representation, and the 16-dimensional multimodal correlation feature intermediate vector are concatenated into a 64-dimensional feature vector through a feature concatenation layer. After batch normalization by a batch normalization layer, the vector is input into the fusion network layer. The fusion network layer sequentially passes through a 128-neuron GELU activation function layer, a 96-neuron GELU layer, a 64-neuron GELU layer, and a 32-neuron fourth output layer to generate fused features. After the fused features are input into the decision layer, they undergo nonlinear transformation through a 32-neuron ReLU first decision layer and a 16-neuron ReLU second decision layer. Finally, the vector is mapped to the [0,1] interval by a 2-neuron fifth output layer to obtain valve opening control parameters. Pump frequency control parameters The furnace pressure control parameters.

[0043] Preferably, the furnace pressure control parameters are converted into values ​​including the actual furnace opening degree using a control parameter command conversion formula. and actual frequency value Furnace pressure control commands;

[0044] The conversion formula for the control parameter command is:

[0045] ;

[0046] Where X is V or F;

[0047] when When x is V, x is v. and They are and ,and and These are the set maximum valve opening and minimum valve opening, respectively;

[0048] when When x is F, x is f. and They are and ,and and These are the set rated operating frequency and the minimum operating frequency, respectively.

[0049] Preferably, a method for safely monitoring the process of the vacuum furnace executing furnace pressure control commands and determining whether to update the AI ​​pressure control model is provided:

[0050] The pressure and temperature inside the vacuum furnace are acquired in real time by pressure sensors and furnace wall temperature sensors arranged inside the furnace cavity during the execution of furnace pressure control commands. and furnace temperature The upper and lower safety limits of vacuum furnace pressure and temperature are compared with those of vacuum furnace, respectively. The upper and lower safety limits of vacuum furnace pressure and temperature are set with reference to the design parameters and process requirements of vacuum furnace.

[0051] like Pressure exceeding the safe upper limit of the vacuum furnace , Temperature exceeds the safe upper limit of the vacuum furnace , Less than the safe lower limit of vacuum furnace temperature or Less than the lower limit of the safe pressure of the vacuum furnace If the preset vacuum furnace safety control strategy is executed, the preset vacuum furnace safety control strategy is either an emergency shutdown procedure or a safety interlock protection. The emergency shutdown procedure is to send a shutdown command, close the air inlet valve, stop the vacuum pump, cut off the heating power supply, and start the furnace cooling fan. The safety interlock protection is to activate the equipment interlock mechanism, prohibit the sending of furnace pressure control commands, until the abnormality is manually confirmed and reset. At the same time, the AI ​​pressure control model is retrained and updated using an incremental learning method.

[0052] Compared with the prior art, the present invention has the following beneficial effects:

[0053] This invention synchronously acquires multimodal time-series data through the NI CompactRIO platform and the IEEE 1588 protocol. Data preprocessing is performed using local outlier factor algorithms and wavelet transforms to construct a furnace pressure model incorporating time-domain, frequency-domain, and multimodal correlation features. An AI pressure control model integrating CNN and fully connected neural networks is used to generate furnace pressure control parameters. Incremental learning and dual safety mechanisms enhance adaptability and system reliability. This effectively solves the problems of insufficient multimodal fusion, reliance on experience in feature engineering, weak model adaptability, and imperfect safety mechanisms in traditional vacuum furnace pressure control schemes. It improves the pressure control accuracy and response speed of the vacuum furnace and reduces the incidence of equipment safety accidents. Attached Figure Description

[0054] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is a flowchart of the method steps of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0057] Examples, such as Figure 1 As shown, an AI-based pressure control method for a vacuum furnace based on multimodal data fusion includes the following steps:

[0058] S001. Real-time acquisition of furnace pressure data of the vacuum furnace;

[0059] S002. Preprocess the furnace pressure data to obtain preprocessed furnace pressure data, and extract features from the preprocessed furnace pressure data to obtain the furnace pressure feature vector;

[0060] S003. Construct an AI pressure control model, input the furnace pressure feature vector into the AI ​​pressure control model, and obtain the furnace pressure control parameters;

[0061] S004. Convert the furnace pressure control parameters into furnace pressure control commands, and execute the furnace pressure control commands accordingly. At the same time, perform safety monitoring on the process of the vacuum furnace executing the furnace pressure control commands, and determine whether to update the AI ​​pressure control model.

[0062] Furthermore, the working principle of the present invention will be illustrated below through embodiments:

[0063] A semiconductor manufacturing company uses a chemical vapor deposition vacuum furnace for depositing silicon dioxide thin films on silicon wafers. The process requires the furnace chamber pressure to be stable within the range of 10 to 50 Pa, the temperature to be controlled within the range of 600 to 800 ℃, and the gas flow rate to be controlled within the range of 50 to 500 sccm.

[0064] By using the NI CompactRIO-9068 platform with a pressure sensor, a K-type thermocouple, and a gas flow sensor, furnace pressure data, including furnace cavity pressure time-series data, heating zone temperature time-series data, and gas flow time-series data, is synchronously collected at a sampling frequency of 100Hz. At the same time, sensor time synchronization is achieved based on the IEEE 1588 protocol, covering the entire process of heating, isothermal and cooling of thin film deposition.

[0065] The furnace pressure data was preprocessed by identifying outliers using a local outlier algorithm and repairing them with k-nearest neighbor interpolation. The preprocessed data was then subjected to 3-level Mallat decomposition using the db4 wavelet basis function for noise reduction, retaining low-frequency coefficients and setting high-frequency coefficients to zero. Kalman filtering was then used to repair missing values, resulting in the furnace pressure feature vector. Taking 1000 data points from the furnace pressure data at the isothermal stage (600℃ furnace temperature and 30Pa furnace pressure) as an example, the mean pressure was extracted. The pressure peak factor is 30.2 Pa. 1.25, pressure standard deviation 0.3 Pa, pressure deviation For 0.12 and pressure kurtosis The time-domain characteristic is 2.8, including the pressure frequency domain center. 2.5Hz, total pressure energy It is 18.7 , pressure root mean square frequency 3.2Hz and pressure-dominant frequency The frequency domain characteristics are 2.3 Hz, including the pressure-temperature correlation coefficient. The correlation coefficient is 0.82, indicating a pressure-temperature hysteresis relationship. The correlation coefficient between pressure and flow rate is 0.75. The correlation coefficient is -0.68 with pressure-flow hysteresis. The multimodal correlation feature is 0.32, and based on this, a furnace pressure feature vector containing 13 dimensions is constructed.

[0066] The furnace pressure feature vector is input into an AI pressure control model composed of a CNN time-domain feature subnetwork, a frequency-domain feature subnetwork, a multimodal correlation feature subnetwork, a fusion layer, and a decision layer. The time-domain feature subnetwork passes through 16... Thirty-two one-dimensional convolutional kernels are used to extract short-term dependency patterns. The frequency domain sub-network uses the LeakyReLU activation function to process frequency domain features, and the multimodal sub-network uses the Swish activation function to capture parameter correlations. The 32-dimensional time-domain intermediate representation, the 16-dimensional frequency-domain intermediate representation, and the 16-dimensional multimodal intermediate vector are concatenated into 64-dimensional features. After processing by the fusion network layer activated by GELU, the decision layer outputs the valve opening control parameter v and the pump frequency control parameter f. The AI ​​pressure control model is trained using the Adam optimizer with a learning rate of 0.001 and the loss function is MSE combined with L2 regularization.

[0067] Formula conversion via control parameter commands The furnace pressure control parameters are converted into furnace pressure control commands, where the valve opening V ranges from 10% to 100%, and the pump operating frequency F ranges from 100 to 500 Hz. The furnace pressure and temperature are monitored in real time during the execution of the furnace pressure control commands. Actual measurements show that the furnace pressure... If the pressure exceeds the safety limit of 50Pa by 55Pa, an emergency shutdown procedure is triggered within 30ms, which involves closing the valve and stopping the vacuum pump. At the same time, an incremental learning method is used to incorporate the abnormal data into the training. After the AI ​​pressure control model is trained and updated, the prediction error for similar working conditions is reduced.

[0068] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them; modifications to the technical solutions described in the foregoing embodiments, or equivalent substitutions of some of the technical features, do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A vacuum furnace AI pressure control method based on multimodal data fusion, characterized in that, Includes the following steps: S1. Real-time acquisition of furnace pressure data of the vacuum furnace; S2. Preprocess the furnace pressure data to obtain preprocessed furnace pressure data, and extract features from the preprocessed furnace pressure data to obtain the furnace pressure feature vector; S3. Construct an AI pressure control model that includes a fusion layer, a decision layer, a temporal feature subnetwork with a CNN structure, a frequency feature subnetwork with a fully connected neural network structure, and a multimodal correlation feature subnetwork with a fully connected neural network structure. Input the furnace pressure feature vector into the AI ​​pressure control model to process and obtain the furnace pressure control parameters. The method for inputting the furnace pressure feature vector into the AI ​​pressure control model for processing is as follows: The time-domain features are input into the time-domain feature sub-network to obtain the intermediate representation of the time-domain features, the frequency-domain features are input into the time-domain feature sub-network to obtain the intermediate expression of the frequency-domain features, and the multimodal correlation features are input into the multimodal correlation feature sub-network to obtain the intermediate vector of the multimodal correlation features. Based on this, the intermediate representation of the time-domain features, the intermediate expression of the frequency-domain features, and the intermediate vector of the multimodal correlation features are input into the fusion layer to obtain the fused features, and the fused features are input into the decision layer to obtain the furnace pressure control parameters. S4. Convert the furnace pressure control parameters into furnace pressure control commands, and execute the furnace pressure control commands accordingly. At the same time, perform safety monitoring on the process of the vacuum furnace executing the furnace pressure control commands, and determine whether to update the AI ​​pressure control model.

2. The vacuum furnace AI pressure control method based on multimodal data fusion according to claim 1, characterized in that, The furnace pressure data includes furnace cavity pressure time-series data, heating zone temperature time-series data, and gas flow time-series data. The furnace cavity pressure time-series data, heating zone temperature time-series data, and gas flow time-series data are synchronously acquired by physical sensors controlled by the NI CompactRIO platform, and time synchronization is achieved based on the IEEE 1588 protocol. The physical sensors include pressure sensors, temperature sensors, and gas flow sensors.

3. The vacuum furnace AI pressure control method based on multimodal data fusion according to claim 2, characterized in that, The method for preprocessing furnace pressure data: Outliers in the furnace pressure data are identified and repaired using a local outlier factor algorithm to obtain outlier-processed furnace pressure data. Based on this, the outlier-processed furnace pressure data is processed sequentially through wavelet transform and Kalman filtering to obtain preprocessed furnace cavity pressure time series data, heating zone temperature time series data, and gas flow time series data, which together constitute the preprocessed furnace pressure data.

4. The vacuum furnace AI pressure control method based on multimodal data fusion according to claim 3, characterized in that, The method for feature extraction from preprocessed furnace pressure data: Preprocessed furnace pressure time series data The mean pressure was calculated using the mean formula, crest factor formula, and standard deviation formula, respectively. Pressure peak factor and pressure standard deviation Based on this, the average pressure and pressure standard deviation The pressure skewness is obtained. and pressure peak ; Preprocessed furnace pressure time series data The data is converted into the frequency domain spectrum of furnace cavity pressure using Fast Fourier Transform. Based on this, the frequency domain spectrum data of furnace pressure The pressure frequency domain center was obtained. Total pressure energy , pressure root mean square frequency and pressure-dominant frequency ; Pre-processed furnace cavity pressure time series data and heating zone temperature time series data The pressure-temperature correlation coefficient was calculated using the Pearson correlation coefficient formula. Meanwhile, the pressure-temperature hysteresis correlation was calculated using the hysteresis correlation formula. ; Pre-processed furnace cavity pressure time series data Gas flow time series data The pressure-flow correlation coefficient was calculated using the Pearson correlation coefficient formula. Meanwhile, the pressure-flow lag correlation was calculated using the lag correlation formula. .

5. The vacuum furnace AI pressure control method based on multimodal data fusion according to claim 4, characterized in that, The furnace pressure feature vector includes time-domain features, frequency-domain features, and multimodal correlation features, wherein the time-domain features include the pressure mean. Pressure peak factor Pressure standard deviation Pressure deviation and pressure peak Frequency domain characteristics include pressure frequency domain center Total pressure energy , pressure root mean square frequency and pressure-dominant frequency Multimodal correlation features include pressure-temperature correlation coefficient. Pressure-temperature hysteresis correlation Pressure-flow correlation coefficient Correlation with pressure-flow hysteresis .

6. The vacuum furnace AI pressure control method based on multimodal data fusion according to claim 1, characterized in that, The furnace pressure control parameters include valve opening control parameters. Pump frequency control parameters .

7. The vacuum furnace AI pressure control method based on multimodal data fusion according to claim 1, characterized in that, The furnace pressure control parameters are converted into values ​​including the actual furnace opening degree using the control parameter command conversion formula. and actual frequency value Furnace pressure control commands.

8. The vacuum furnace AI pressure control method based on multimodal data fusion according to claim 7, characterized in that, The method for safely monitoring the process of executing furnace pressure control commands in a vacuum furnace and determining whether to update the AI ​​pressure control model: The vacuum furnace's internal pressure and temperature are acquired in real time during the execution of furnace pressure control commands. and furnace temperature Compare with the upper and lower safety limits of vacuum furnace pressure and temperature, respectively; like Pressure exceeding the safe upper limit of the vacuum furnace , Temperature exceeds the safe upper limit of the vacuum furnace , Less than the safe lower limit of vacuum furnace temperature or Less than the lower limit of the safe pressure of the vacuum furnace If the preset vacuum furnace safety control strategy is executed, the AI ​​pressure control model will be retrained and updated using incremental learning.

Images