Vacuum furnace AI pressure control method based on multi-modal data fusion

Abstract

The invention discloses a vacuum furnace AI pressure control method based on multi-modal data fusion, and belongs to the technical field of vacuum furnace pressure control. According to the method, furnace pressure data including furnace chamber pressure time sequence data, heating area temperature time sequence data and gas flow time sequence data are synchronously collected through an NI CompactRIO platform and an IEEE 1588 protocol, and furnace pressure feature vectors including time domain features, frequency domain features and multi-modal correlation features are extracted after local outlier factor algorithm, wavelet transform and Kalman filtering processing. And constructing an AI pressure control model, inputting the furnace pressure feature vector into the AI pressure control model, generating a furnace pressure control parameter, converting the furnace pressure control parameter into a furnace pressure control instruction, executing the furnace pressure control instruction, performing safety monitoring on the execution process, executing a preset safety strategy when the execution process is abnormal, and incrementally updating the AI pressure control model. Accurate control over the pressure of the vacuum furnace is achieved, and the safety and the self-adaptive capacity of pressure control over the vacuum furnace are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of vacuum furnace pressure control, and in particular to a vacuum furnace AI pressure control method based on multimodal data fusion. Background Art

[0002] Vacuum furnaces are critical industrial equipment, and their pressure control accuracy directly impacts product quality and process stability. Traditional vacuum furnace pressure control often uses a single-mode control scheme based on the PID algorithm, relying solely on pressure sensor data for closed-loop regulation. This approach struggles to cope with the nonlinear pressure fluctuations caused by the multi-factor coupling of heating zone temperature fluctuations and gas flow variations, resulting in 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 empirical feature engineering, weak AI model adaptability, and imperfect safety mechanisms. Therefore, achieving high-precision synchronous acquisition and fusion of multimodal data, building an adaptively updated AI control model, and establishing a comprehensive safety monitoring mechanism are key technical challenges in improving the accuracy of vacuum furnace pressure control. To this end, we propose an AI-based pressure control method for vacuum furnaces based on multimodal data fusion. Summary of the Invention

[0004] The purpose of the present invention is to provide a vacuum furnace AI pressure control method based on multimodal data fusion to solve the problems raised in the above background technology.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

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

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

[0008] S2 preprocesses the furnace pressure data to obtain preprocessed furnace pressure data, and extracts features from the preprocessed furnace pressure data to obtain a furnace pressure feature vector;

[0009] S3 builds an AI pressure control model, inputs the furnace pressure feature vector into the AI ​​pressure control model, and obtains the furnace pressure control parameters;

[0010] S4. Convert the furnace pressure control parameters into furnace pressure control instructions, and the vacuum furnace executes the furnace pressure control instructions accordingly. At the same time, the process of the vacuum furnace executing the furnace pressure control instructions is safely monitored, and it is determined whether the AI ​​pressure control model is updated.

[0011] Preferably, the furnace pressure data includes furnace chamber pressure time series data, heating zone temperature time series data and gas flow time series data. The furnace chamber pressure time series data, heating zone temperature time series data and gas flow time series data are collected in real time by arranging pressure sensors, temperature sensors and gas flow sensors inside the furnace chamber, heating zone and air inlet duct of the vacuum furnace respectively. The pressure sensors, temperature sensors and gas flow sensors are controlled by the NI CompactRIO industrial-grade data acquisition platform for synchronous acquisition. At the same time, sub-millisecond time synchronization is achieved based on the IEEE 1588 precision time protocol to ensure the time consistency of the furnace chamber pressure time series data, heating zone temperature time series data and gas flow time series data.

[0012] Preferably, the method for preprocessing the furnace pressure data is:

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

[0014] The furnace pressure data with outlier processing is subjected to wavelet transform filtering, and the db4 wavelet basis function is used to perform a three-layer decomposition on the furnace pressure data with outlier processing. High-frequency noise is eliminated by retaining low-frequency coefficients and setting high-frequency coefficients to zero. The decomposition process is recursively implemented using the Mallat algorithm. After each layer of decomposition, the high-frequency components are threshold-processed to reconstruct the denoised furnace pressure data. Based on this, the Kalman filter is used in turn to repair missing values ​​in the denoised furnace pressure data. Finally, the pre-processed furnace chamber pressure time series data, heating zone temperature time series data, and gas flow time series data are obtained, which together constitute the pre-processed furnace pressure data.

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

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

[0017] Preferably, the method for extracting features from the pre-processed furnace pressure data is as follows:

[0018] The pre-processed furnace chamber pressure time series data The pressure mean is calculated using the mean formula, crest factor formula, and standard deviation formula. , pressure crest factor and pressure standard deviation , based on the pressure mean and pressure standard deviation , the pressure skewness is obtained by the skewness formula and the kurtosis formula and pressure kurtosis ;

[0019] The mean formula is: ;

[0020] The crest factor formula is: ;

[0021] The standard deviation formula is: ;

[0022] The skewness formula is: ;

[0023] The kurtosis formula is: ;

[0024] The pre-processed furnace chamber pressure time series data Converted into furnace chamber pressure frequency domain spectrum data through fast Fourier transform Based on the frequency domain spectrum data of furnace chamber pressure , the pressure frequency center is calculated by the main frequency center calculation formula, total energy calculation formula, root mean square frequency calculation formula and dominant frequency calculation formula , total pressure energy , pressure RMS frequency and pressure dominant frequency ;

[0025] The main frequency center calculation formula is: ;

[0026] The total energy calculation formula is: ;

[0027] The root mean square frequency calculation formula is:

[0028] The dominant frequency calculation formula is: ;

[0029] Pre-processed furnace chamber pressure time series data And heating zone temperature time series data The pressure-temperature correlation coefficient is calculated using the Pearson correlation coefficient formula At the same time, the pressure-temperature hysteresis correlation is calculated by the hysteresis correlation formula ;

[0030] Pre-processed furnace chamber pressure time series data and gas flow time series data The pressure-flow correlation coefficient is calculated using the Pearson correlation coefficient formula At the same time, the pressure flow hysteresis correlation is calculated by the hysteresis correlation formula ;

[0031] The Pearson correlation coefficient formula is:

[0032] ;

[0033] in, yes and The covariance of and They are and of , is the furnace chamber pressure time series data , is the heating zone temperature time series data Or gas flow time series data ;

[0034] The lag correlation formula is: ;

[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 crest factor , pressure standard deviation , pressure deflection and pressure kurtosis , the frequency domain characteristics include the pressure frequency domain center , total pressure energy , pressure RMS frequency and pressure dominant frequency , multimodal correlation features include pressure-temperature correlation coefficient , pressure-temperature hysteresis correlation , pressure-flow correlation coefficient and pressure-flow hysteresis correlation .

[0037] Preferably, the AI ​​pressure control model is jointly constructed by a fusion layer, a decision layer, a time domain feature subnetwork of a CNN structure, and a frequency domain feature subnetwork and a multimodal association feature subnetwork of a fully connected neural network structure, wherein the time domain 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 association 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, and the decision layer includes a first decision layer, a second decision layer and a fifth output layer.

[0038] Preferably, the furnace pressure characteristic vector is input into the AI ​​pressure control model to obtain the furnace pressure control parameters:

[0039] The time domain features are input into the time domain feature sub-network. The first convolutional layer of the time domain feature sub-network uses 16 one-dimensional convolution kernels of length 3, stride 1, and ReLU activation function to extract short-term dependency patterns in the time domain features. After the first pooling layer with a window size of 2 performs maximum pooling dimensionality reduction, the second convolutional layer extracts features with 32 convolution kernels of length 3. After pooling in 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 an intermediate representation of the time domain features.

[0040] The frequency domain features are input into the frequency domain feature subnet. The first hidden layer of 64 neurons, the second hidden layer of 48 neurons and the third hidden layer of 32 neurons of the frequency domain feature subnet are all The LeakyReLU activation function is 0.01, which processes the frequency domain features through the first hidden layer, the second hidden layer, and the third hidden layer in sequence, and finally generates the intermediate representation of the frequency domain features by the second output layer of 16 neurons;

[0041] The multimodal correlation features are input into the multimodal correlation feature sub-network. The 48 neurons in the fourth hidden layer and the 32 neurons in the fifth hidden layer both use the Swish activation function. The multimodal correlation features are processed by the fourth hidden layer and the fifth hidden layer in turn, and finally the multimodal correlation feature intermediate vector is generated by the third output layer of 16 neurons.

[0042] The 32-dimensional time domain feature intermediate representation, the 16-dimensional frequency domain feature intermediate expression and the 16-dimensional multimodal correlation feature intermediate vector are spliced ​​into a 64-dimensional feature vector through the feature splicing layer, and then input into the fusion network layer after batch normalization by the batch normalization layer. The fusion network layer generates fusion features by sequentially passing through the GELU activation function layer of 128 neurons, the GELU layer of 96 neurons, the GELU layer of 64 neurons and the fourth output layer of 32 neurons. After the fusion features are input into the decision layer, they are nonlinearly transformed by the first ReLU decision layer of 32 neurons and the second ReLU decision layer of 16 neurons, and then mapped to the [0,1] interval by the fifth output layer of 2 neurons to obtain the valve opening control parameters. and pump frequency control parameters The furnace pressure control parameters.

[0043] Preferably, the furnace pressure control parameter is converted into the actual opening value by the control parameter instruction conversion formula and the actual frequency value Furnace pressure control instructions;

[0044] The control parameter instruction conversion formula is:

[0045] ;

[0046] Wherein, X is V or F;

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

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

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

[0050] The pressure sensor and furnace wall temperature sensor arranged inside the furnace cavity are used to obtain the pressure and temperature of the vacuum furnace in real time during the process of the vacuum furnace executing the furnace pressure control instruction. and furnace temperature Compare with the upper and lower safety limits of vacuum furnace pressure and the upper and lower safety limits of vacuum furnace temperature respectively. The upper and lower safety limits of vacuum furnace pressure and the upper and lower safety limits of vacuum furnace temperature are set with reference to vacuum furnace design indicators and process requirements;

[0051] like Greater than the safety upper limit of vacuum furnace pressure 、 Greater than the safety upper limit of vacuum furnace temperature 、 Less than the safe lower limit of vacuum furnace temperature or Less than the safety lower limit of vacuum furnace pressure , the preset vacuum furnace safety control strategy is executed. The preset vacuum furnace safety control strategy is the emergency shutdown process or safety interlock protection. The emergency shutdown process is to send a shutdown command, close the air intake valve, stop the vacuum pump, cut off the heating power supply and start the furnace cavity cooling fan. The safety interlock protection is to activate the equipment interlock mechanism and prohibit the furnace pressure control command from being sent until manual confirmation is made that the abnormality has been eliminated and reset. At the same time, the AI ​​pressure control model is retrained and updated using incremental learning.

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

[0053] The present invention synchronously collects multimodal time series data through the NI CompactRIO platform and the IEEE 1588 protocol, adopts the local outlier factor algorithm and wavelet transform for data preprocessing, constructs a furnace pressure containing time domain, frequency domain and multimodal correlation features, and uses an AI pressure control model that integrates CNN and fully connected neural networks to realize the generation of furnace pressure control parameters. The adaptive capability and system reliability are improved through incremental learning and dual safety mechanisms. 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, improves the pressure control accuracy and response speed of vacuum furnaces, and reduces the incidence of equipment safety accidents. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the technical solutions and advantages of the embodiments of the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. 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 paying any creative work.

[0055] Figure 1 The figure is a flow chart of the method steps of the present invention. DETAILED DESCRIPTION

[0056] To make the objectives, technical solutions, and advantages of the present invention more apparent, the technical solutions of the present invention will be described in detail below. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other implementations obtained by those of ordinary skill in the art without inventive effort are within the scope of protection of the present invention.

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

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

[0059] S002 preprocesses the furnace pressure data to obtain preprocessed furnace pressure data, and extracts features from the preprocessed furnace pressure data to obtain a furnace pressure feature vector;

[0060] S003 builds an AI pressure control model, inputs the furnace pressure characteristic vector into the AI ​​pressure control model, and obtains the furnace pressure control parameters;

[0061] S004. Convert the furnace pressure control parameters into furnace pressure control instructions, and the vacuum furnace executes the furnace pressure control instructions accordingly. At the same time, the process of the vacuum furnace executing the furnace pressure control instructions is safety monitored, and it is determined whether the AI ​​pressure control model is updated.

[0062] Furthermore, the working principle of the present invention is described below by way of examples:

[0063] A semiconductor manufacturing company's chemical vapor deposition vacuum furnace is used to deposit silicon dioxide films on the surface of silicon wafers. The process requires that the furnace chamber pressure be stable in the range of 10 to 50 Pa, the temperature be controlled at 600 to 800°C, and the gas flow rate be controlled at 50 to 500 sccm.

[0064] By using the NI CompactRIO-9068 platform with pressure sensors, K-type thermocouples, and gas flow sensors, furnace pressure data, including furnace chamber pressure time series data, heating zone temperature time series data, and gas flow time series data, is synchronized at a 100 Hz sampling frequency. Sensor time synchronization is also achieved based on the IEEE 1588 protocol, covering the entire heating, constant temperature, and cooling process of the thin film deposition process.

[0065] The furnace pressure data is identified by the local outlier factor algorithm and the outliers are repaired by the k-nearest neighbor interpolation method to obtain the preprocessed furnace pressure data. The preprocessed furnace pressure data is subjected to 3-layer Mallat decomposition and denoising using the db4 wavelet basis function. After retaining the low-frequency coefficients and setting the high-frequency coefficients to zero, the Kalman filter is used to repair the missing values ​​and obtain the furnace pressure feature vector. Taking 1000 data points in the furnace pressure data of the constant temperature stage, i.e., the furnace temperature is 600℃ and the furnace pressure is 30Pa, as an example, the pressure mean is extracted. 30.2Pa, pressure crest factor 1.25, pressure standard deviation 0.3Pa, pressure deflection The pressure kurtosis is 0.12 and The time domain characteristics of 2.8, including the pressure frequency domain center 2.5Hz, total pressure energy 18.7 , pressure RMS frequency is 3.2Hz and the pressure dominant frequency The frequency domain characteristics of 2.3Hz, including the pressure-temperature correlation coefficient The pressure-temperature hysteresis correlation is 0.82. The pressure-flow correlation coefficient is 0.75. The correlation between pressure and flow lag is -0.68 The multimodal correlation feature is 0.32, based on which the furnace pressure feature vector containing 13-dimensional features is constructed.

[0066] The furnace pressure feature vector is input into the AI ​​pressure control model composed of CNN time domain feature sub-network, frequency domain feature sub-network, multimodal correlation feature sub-network, fusion layer and decision layer. The time domain feature sub-network is connected through 16 32 one-dimensional convolution kernels 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 captures parameter associations through the Swish activation function. The 32-dimensional time domain intermediate representation, 16-dimensional frequency domain intermediate expression, and 16-dimensional multimodal intermediate vector are spliced ​​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 training uses the Adam optimizer with a learning rate of 0.001 and a loss function of MSE combined with L2 regularization.

[0067] Transformation formula by controlling parameter instructions , the furnace pressure control parameters are converted into furnace pressure control instructions, where the valve opening V range is 10% to 100%, the pump operating frequency F range is 100 to 500Hz, and the furnace pressure and temperature are monitored in real time during the execution of the furnace pressure control instruction. It is found that the furnace pressure For 55 Pa over the safety upper limit of 50 Pa, the emergency shutdown process is triggered within 30 ms, i.e. closing the valve and stopping the vacuum pump, while adopting incremental learning to include this abnormal data into training, and after updating the AI pressure control model, the prediction error for similar working conditions is reduced.

[0068] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; modifying the technical solutions described in the foregoing embodiments, or equivalently replacing some of the technical features, does not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A vacuum furnace AI pressure control method based on multimodal data fusion, characterized in that: The following steps are involved: S1. Real-time acquisition of vacuum furnace pressure data; S2 preprocesses the furnace pressure data to obtain preprocessed furnace pressure data, and extracts features from the preprocessed furnace pressure data to obtain a furnace pressure feature vector; S3. Construct an AI pressure control model that includes a fusion layer, a decision layer, a time domain feature subnetwork with a CNN structure, a frequency domain 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 obtain the furnace pressure control parameters. The method for processing the furnace pressure characteristic vector input into the AI ​​pressure control model is: Input the time domain features into the time domain feature sub-network to obtain the time domain feature intermediate representation, input the frequency domain features into the time domain feature sub-network to obtain the frequency domain feature intermediate expression, input the multimodal correlation features into the multimodal correlation feature sub-network to obtain the multimodal correlation feature intermediate vector, based on which the time domain feature intermediate representation, frequency domain feature intermediate expression and multimodal correlation feature intermediate vector are input into the fusion layer to obtain the fusion feature, and the fusion feature is input into the decision layer to obtain the furnace pressure control parameter; S4. Convert the furnace pressure control parameters into furnace pressure control instructions, and the vacuum furnace executes the furnace pressure control instructions accordingly. At the same time, the process of the vacuum furnace executing the furnace pressure control instructions is safely monitored, and it is determined whether the AI ​​pressure control model is updated.

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 chamber pressure time series data, heating zone temperature time series data, and gas flow time series data, wherein the furnace chamber pressure time series data, heating zone temperature time series data, and gas flow time series data are synchronously collected by controlling physical sensors via the NI CompactRIO platform and are used to achieve time synchronization based on the IEEE 1588 protocol; The physical sensors include a pressure sensor, a temperature sensor and a gas flow sensor.

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: The furnace pressure data is identified and repaired based on the local outlier factor algorithm to obtain the outlier-processed furnace pressure data. The outlier-processed furnace pressure data is then processed by wavelet transform and Kalman filtering in turn to obtain the preprocessed furnace chamber 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 is characterized in that: The method for extracting features from the pre-processed furnace pressure data: The pre-processed furnace chamber pressure time series data The pressure mean is calculated using the mean formula, crest factor formula, and standard deviation formula. , pressure crest factor and pressure standard deviation , based on the pressure mean and pressure standard deviation , and the pressure deflection is obtained and pressure kurtosis ; The pre-processed furnace chamber pressure time series data Converted into furnace chamber pressure frequency domain spectrum data through fast Fourier transform Based on the furnace chamber pressure frequency domain spectrum data , get the pressure frequency domain center , total pressure energy , pressure RMS frequency and pressure dominant frequency ; Pre-processed furnace chamber pressure time series data And heating zone temperature time series data The pressure-temperature correlation coefficient is calculated using the Pearson correlation coefficient formula At the same time, the pressure-temperature hysteresis correlation is calculated by the hysteresis correlation formula ; Pre-processed furnace chamber pressure time series data and gas flow time series data The pressure-flow correlation coefficient is calculated using the Pearson correlation coefficient formula At the same time, the pressure flow hysteresis correlation is calculated by the hysteresis correlation formula .

5. The vacuum furnace AI pressure control method based on multimodal data fusion according to claim 4 is 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 crest factor , pressure standard deviation , pressure deflection and pressure kurtosis , the frequency domain characteristics include the pressure frequency domain center , total pressure energy , pressure RMS frequency and pressure dominant frequency , multimodal correlation features include pressure-temperature correlation coefficient , pressure-temperature hysteresis correlation , pressure-flow correlation coefficient and pressure-flow hysteresis correlation .

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 and 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 parameter is converted into the actual opening value by the control parameter instruction conversion formula and the actual frequency value Furnace pressure control instructions.

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 instructions on a vacuum furnace and determining whether to update the AI ​​pressure control model: Real-time acquisition of the pressure and temperature inside the vacuum furnace during the execution of the furnace pressure control instruction. and furnace temperature Compare with the upper and lower safety limits of vacuum furnace pressure and vacuum furnace temperature respectively; like Greater than the safety upper limit of vacuum furnace pressure 、 Greater than the safety upper limit of vacuum furnace temperature 、 Less than the safe lower limit of vacuum furnace temperature or Less than the safety lower limit of vacuum furnace pressure , the preset vacuum furnace safety control strategy is executed, and the AI ​​pressure control model is retrained and updated using incremental learning.

Images