Single electrostatic forming nanowire semiconductor VOCs identification method based on table deep learning and model fusion

Abstract

The invention belongs to the technical field of VOCs detection, and discloses a single electrostatic forming nanowire semiconductor VOCs identification method based on table deep learning and model fusion, and the method comprises the steps: obtaining a first data set, the first data set comprises a change value of back gate threshold voltage, a change value of back gate sub-threshold swing and a change value of back gate source-drain current before to-be-identified VOCs form a nanowire semiconductor through single static electricity; acquiring a second data set, wherein the second data set comprises a change value of a junction gate threshold voltage, a change value of junction gate sub-threshold swing and a change value of junction gate source-drain current before and after the VOCs to be identified form a nanowire semiconductor through single static electricity; performing Stacking model fusion on the first data set and the second data set to obtain a final prediction result of the to-be-identified VOCs type; when Stacking model fusion is carried out, the base model and the meta-model adopt a table deep learning model. According to the method, table deep learning and model fusion are combined, and the VOCs recognition rate of the nanowire (EFN) gas sensor formed by static electricity is increased.

Description

Technical Field

[0001] The present invention belongs to the technical field of volatile organic compounds (VOCs) detection, and specifically relates to a single electrostatically formed nanowire semiconductor VOCs identification method based on table deep learning and model fusion. Background Art

[0002] Volatile Organic Compounds (VOCs) are a class of organic compounds that are volatile at room temperature and usually come from the combustion of fossil fuels, industrial production, automobile exhaust, household products, etc. VOCs participate in photochemical reactions with nitrogen oxides to form ozone and other harmful compounds, causing air pollution and affecting human health and environmental quality. Long-term exposure to VOCs may cause health problems such as headaches, eye, nose and throat irritation, breathing difficulties, nausea, and central nervous system effects. Some VOCs are also considered carcinogenic. VOCs may enter surface water and groundwater through rainwater erosion and seepage, posing a threat to aquatic life and drinking water quality. VOCs are an important component of photochemical smog in the atmosphere, which has a negative impact on visibility and also affects the growth of crops and vegetation. Therefore, the detection and control of VOCs is crucial to improving air quality and protecting health and the environment. Electrostatically formed nanowire (EFN) gas and VOCs sensor is a rugged, ultra-miniaturized sensor based on conventional CMOS process, with very low power consumption and compatible with VLSI manufacturing. It is a multi-gate field effect transistor (FET) with several degrees of freedom (such as surrounding back gate and junction gate) that enables the extraction of parameters (such as threshold voltage, subthreshold swing, etc.) as input features for machine learning implementations.

[0003] Currently, most VOC detection equipment is bulky and expensive. With the development of the Internet of Things, sensors need to be miniaturized, low-power, and compatible with large-scale integration. The electrostatically formed nanowire (EFN) gas sensor is a rugged, ultra-miniaturized sensor based on traditional CMOS technology, with extremely low power consumption and compatible with very large-scale integration (VLSI) manufacturing.

[0004] The nanowire EFN gas sensor is a multi-gate field effect transistor (FET) (such as the surrounding back gate and junction gate). By adjusting the bias voltage of the field effect tube, the changes in threshold voltage and source-drain on-current are obtained before and after VOCs adsorption, and these changes are used as characteristic parameters to realize the sensor's recognition of target VOCs through machine learning. Previous work by Niharendu et al. has demonstrated that machine learning methods can achieve selectivity even in homologous alcohols. Using the AdaBoost classifier algorithm, the classification accuracy of VOCs is about 93%. Yang et al. proposed a unilateral classification method based on the physical properties of the sensor, and improved the classification accuracy through automatic learning and ensemble learning. .

[0005] The prior art has proven that the use of EFN transverse deformation fingerprint features and longitudinal fingerprint features alone has achieved good classification results. For the FET sensor as a whole, the advantages of the superposition of the two types of features are not fully utilized to make the two types of features complementary and integrated to further improve the classification accuracy. On the other hand, considering that homogeneous data training is used in traditional machine learning and the model needs to be retrained when the data is updated, it is not conducive to the training of non-homogeneous data and incremental learning data, nor is it conducive to the future integration in embodied intelligence. Summary of the invention

[0006] In order to solve the problems existing in the prior art, the purpose of the present invention is to provide a single electrostatically formed nanowire semiconductor VOCs identification method based on table deep learning and model fusion. The present invention combines table deep learning and model fusion to improve the VOCs recognition rate of electrostatically formed nanowire (EFN) gas sensor.

[0007] To achieve the above purpose, the technical solution adopted by the present invention is as follows: A method for identifying semiconductor VOCs in a single electrostatically formed nanowire based on table deep learning and model fusion, comprising: Acquire a first data set, wherein the first data set includes a change value of a back gate threshold voltage, a change value of a back gate subthreshold swing, and a change value of a back gate source leakage current before and after the VOCs to be identified are formed into a nanowire semiconductor through a single electrostatic wire; Acquire a second data set, wherein the second data set includes a change value of a junction gate threshold voltage, a change value of a junction gate subthreshold swing, and a change value of a junction gate source leakage current before and after the VOCs to be identified pass through a single electrostatic wire to form a nanowire semiconductor; The first data set and the second data set are fused by the Stacking model to obtain the final prediction result of the type of VOCs to be identified; wherein, when the Stacking model is fused, the base model adopts a table deep learning model, and the meta-model adopts a table deep learning model.

[0008] Preferably, when the first data set and the second data set are fused by a Stacking model: Using a pre-trained first base model to process the first data set, obtaining a first prediction result of the VOCs type to be identified; The second data set is processed using the pre-trained second base model to obtain a second prediction result of the VOCs type to be identified; splicing the first prediction result and the second prediction result to obtain spliced ​​data; The pre-trained meta-model is used to process the spliced ​​data to obtain the final prediction results of the VOCs types to be identified.

[0009] Preferably: the process of obtaining the change value of the back gate threshold voltage, the change value of the back gate subthreshold swing, and the change value of the back gate source leakage current before and after the VOCs to be identified pass through a single electrostatic electrode to form a nanowire semiconductor includes: for different junction gate voltages and drain voltages, the back gate voltage is changed throughout the process, the drain current is measured, and then the change value of the back gate threshold voltage, the change value of the back gate subthreshold swing, and the change value of the back gate source leakage current are calculated before and after the VOCs to be identified pass through a single electrostatic electrode to form a nanowire semiconductor; The process of obtaining the change value of the junction gate threshold voltage, the change value of the junction gate subthreshold swing, and the change value of the junction gate source leakage current before and after the VOCs to be identified pass through a single static electricity to form a nanowire semiconductor includes: changing the junction gate voltage throughout the process under different back gate voltages and drain voltages, measuring the drain current, and then calculating the change value of the junction gate threshold voltage, the change value of the junction gate subthreshold swing, and the change value of the junction gate source leakage current before and after the VOCs to be identified pass through a single static electricity to form a nanowire semiconductor; Wherein, when calculating the change value of the back gate threshold voltage, the change value of the back gate subthreshold swing, the change value of the back gate source leakage current, the change value of the junction gate threshold voltage, the change value of the junction gate subthreshold swing and the change value of the junction gate source leakage current, the voltages of the two junction gates in a single electrostatically formed nanowire semiconductor are made equal to the junction gate voltage.

[0010] Preferably, the VOCs to be identified include at least one of hexanol, butanol, propanol, ethanol, acetone and acetic acid.

[0011] Preferably, the table deep learning model adopts a NODE model, a SAINT model, a TabNet model, a VIME model or an MLP model.

[0012] Preferably, the table deep learning model adopts the TabNet model.

[0013] The present invention also provides a single electrostatically formed nanowire semiconductor VOCs identification system based on table deep learning and model fusion, comprising: A first data set acquisition unit: used to acquire a first data set, wherein the first data set includes a change value of a back gate threshold voltage, a change value of a back gate subthreshold swing, and a change value of a back gate source leakage current before and after the VOCs to be identified pass through a single electrostatic wire to form a nanowire semiconductor; A second data set acquisition unit: used to acquire a second data set, wherein the second data set includes a change value of a junction gate threshold voltage, a change value of a junction gate subthreshold swing, and a change value of a junction gate source leakage current before and after the VOCs to be identified pass through a single electrostatic nanowire semiconductor; Identification module: used to perform Stacking model fusion on the first data set and the second data set to obtain the final prediction result of the type of VOCs to be identified; wherein, when performing Stacking model fusion, the base model adopts a table deep learning model, and the meta-model adopts a table deep learning model.

[0014] The present invention also provides an electronic device, comprising: one or more processors; a storage device having one or more programs stored thereon; When the one or more programs are executed by the one or more processors, the one or more processors implement the single electrostatically formed nanowire semiconductor VOCs identification method based on table deep learning and model fusion as described above in the present invention.

[0015] The present invention also provides a storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the single electrostatically formed nanowire semiconductor VOCs identification method based on table deep learning and model fusion as described above in the present invention is implemented.

[0016] The present invention also provides a computer program product, which includes computer instructions, and the computer instructions instruct a computer to execute the single electrostatically formed nanowire semiconductor VOCs identification method based on table deep learning and model fusion as described above.

[0017] The present invention has the following beneficial effects: The single electrostatically formed nanowire semiconductor VOCs identification method based on table deep learning and model fusion of the present invention takes VOCs with the same responsiveness as the identification object, and takes the physical adsorption of single electrostatically formed nanowire semiconductor VOCs as the basic principle. The physical parameters of the single electrostatically formed nanowire semiconductor response before and after VOCs access are used as identification fingerprints, and then based on the table deep learning model and the Stacking model fusion, the final prediction result obtained by the Stacking model fusion has a higher accuracy, and the selection of the base model has a relatively small impact on the classification accuracy. The principle of this fusion algorithm is applicable to all sensors using EFN structure, including biological, gas, temperature and other sensors based on EFN. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of 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 work.

[0019] Figure 1 This is a bias configuration diagram of an electrostatically formed nanowire semiconductor VOCs sensor in an embodiment of the present invention.

[0020] Figure 2 Schematic diagram of VOCs recognition using deep learning and model fusion of electrostatically formed nanowire sensor tables in an embodiment of the present invention.

[0021] Figure 3 Flow chart of VOCs identification of electrostatically formed nanowire sensor in an embodiment of the present invention.

[0022] FIG. 4( a ) is a diagram comparing the model size and accuracy on the ΔBG dataset in an embodiment of the present invention; FIG. 4( b ) is a diagram comparing the model size and accuracy on the ΔJG dataset in an embodiment of the present invention.

[0023] FIG. 5( a ) is a graph comparing inference time and accuracy on the ΔBG dataset in an embodiment of the present invention; FIG. 5( b ) is a graph comparing model size and accuracy on the ΔJG dataset in an embodiment of the present invention.

[0024] Figure 6 Graph showing the accuracy and inference time of the first-layer prediction results in an embodiment of the present invention.

[0025] FIG7 (a) is a distribution diagram of the accuracy and inference time of the four models of the first layer on the ΔBG dataset in an embodiment of the present invention; FIG7 (b) is a distribution diagram of the accuracy and inference time of the four models of the first layer on the ΔJG dataset in an embodiment of the present invention.

[0026] Figure 8It is the confusion matrix of the prediction results after MLP model fusion in the embodiment of the present invention.

[0027] Fig. 9 It is the ROC curve after the MLP model is integrated in the embodiment of the present invention.

[0028] Fig.10 The model weight after the MLP model fusion in the embodiment of the present invention.

[0029] Fig.11 This is a comparison of the accuracy before and after model fusion in the embodiment of the present invention.

[0030] In the figure, 1-1-first junction gate, 1-2-second junction gate, 2-silicon dioxide film, 3-silicon dioxide back gate. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. The components of the embodiments of the present invention described and shown in the drawings here can usually be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0032] The present invention proposes a gas selectivity research scheme that combines table deep learning and model fusion. The present invention studies the recognition accuracy of table deep learning on two feature data sets, and uses automatic learning to screen and sort the accuracy of table deep learning models. The results under the two biases are fused using the EEN sensor principle, and the fusion results are analyzed. The fusion model has a higher accuracy (99%), and the selection of the base model has little effect on the classification accuracy. The principle of this fusion algorithm is applicable to all sensors using EFN structure, including biological, gas, temperature and other sensors based on EFN.

[0033] The present invention is based on a single electrostatically formed nanowire semiconductor VOCs identification method based on table deep learning and model fusion. Figure 1 The bias configuration shown, the back gate voltage V BG (i.e. the voltage between the back gate and the source) remains unchanged, the junction gate voltage V JG(That is, the voltage between the junction gate and the source. There are two junction gates, marked as the first junction gate 1-1 and the second junction gate 1-2, and the voltages of the first junction gate 1-1 and the second junction gate 1-2 are respectively recorded as V JG1 and V JG2 ) changes in the configuration (corresponding to the ΔJG data set (i.e., the second data set)), at the junction gate voltage V JG Unchanged, back gate voltage V BG The longitudinal changes of FEN in the changing configuration (corresponding to the ΔBG dataset (i.e., the first dataset)) are as follows: Figure 2 As shown. For this purpose, two sets of feature data are measured by two independent biases, and a data set is constructed with these two sets of feature data to identify VOCs through machine learning. The specific introduction of each part is as follows.

[0034] 1. Working bias configuration of VOCs detection device (i.e. single electrostatically formed nanowire semiconductor). The bias configuration of VOCs detection device is as follows: Figure 1 , drain voltage V DS Loaded on the source and drain, back gate voltage V BG Loaded on the back gate, the junction gate voltage V JG Loaded on the node gate, in order to make FEN symmetrical, take the first junction gate voltage V JG1 and the second junction gate voltage V JG2 same.

[0035] 2. Construction of the data set. The change in back gate threshold voltage ΔV before and after VOCs pass through EFN th (BG), the change in back gate subthreshold swing ΔSS(BG) and the change in back gate source leakage current ΔI on (BG) is used as the feature quantity, and a data set is constructed and named ΔBG data set (i.e., the first data set). Similarly, the change value of the gate threshold voltage before and after VOCs passes through EFN is ΔV th (JG), the change in junction-gate subthreshold swing ΔSS(JG) and the change in junction-gate-source leakage current ΔI on (JG) is the feature quantity, and a dataset is constructed and named ΔJG dataset (i.e., the second dataset). The two datasets (i.e., ΔBG dataset and ΔJG dataset) are independent but have the same category labels.

[0036] 3. Automatic learning and tabular deep learning screening models. In order to screen out deep learning models that are more suitable for the ΔBG dataset and the ΔJG dataset from the numerous tabular deep learning models, the present invention uses the automatic learning tool provided by TabSurvey to screen popular algorithm models such as NODE, SAINT, TabNet, VIME, MLP, DeepFM, and DeepGBM. Automatic learning screening was performed on the ΔBG dataset and the ΔJG dataset respectively, and the F1 score was used as the sorting indicator to display and compare the accuracy, AUC (Area Under Curve) value, training time, inference time, and model size of different models. Table 2 lists the relevant indicators of the top 5 excellent algorithm models with the F1 score (F1>0.83).

[0037] 4. Based on the two models corresponding to the ΔBG dataset and the ΔJG dataset, the classification accuracy of the VOCs to be identified is further improved by Stacking model fusion. Stacking model fusion is an integrated learning method in machine learning, which improves the generalization ability and prediction accuracy of the model by combining the prediction results of multiple basic models. The basic principle of Stacking model fusion is to use the outputs of multiple basic models as input and train a secondary model for final prediction. The top 5 excellent algorithm models are screened out by automatic learning. In view of the high size and inference time of the NODE model, the present invention eliminates the NODE model during the Stacking model fusion process. When performing Stacking model fusion, the present invention first obtains the prediction results of the first layer through 5-fold cross validation. The accuracy results on the ΔBG dataset and the ΔJG dataset are shown in Table 3. On the ΔBG dataset, the best accuracy is the SAINT model (accuracy is 0.96258±0.00590), and on the ΔJG dataset, the highest accuracy is the VIME model (accuracy is 0.93742±0.01465). The inference time results on the two datasets are shown in Table 4. The SAINT model has the highest inference time on both datasets, with inference times of 0.72349±0.54594 and 0.44631±0.04003 respectively.

[0038] The accuracy of the second-layer model fusion is shown in Table 5. It can be seen from Table 5 that the fusion model with the highest accuracy is the TabNet and SAINT algorithm fusion model, with an accuracy of 99.877%, but TabNet has the shortest reasoning time. Therefore, it is best to select the TabNet model as the base model for fusion.

[0039] 5. Meta-model feature weight distribution after model fusion: Taking MLP model fusion as an example, each feature in the MLP fusion model is composed of different weights of the first-layer 6 VOCs classification prediction results of the two models. Fig.10 It is the first feature after the MLP model is fused. It is composed of 12 features of the first layer with different weights. The weight combination graphs of other features in the second layer are similar. From the distribution of features, it can be seen that the information of the two base models is complementary, which provides important support for the accuracy of the model after fusion.

[0040] Based on the above scheme, the present embodiment provides a single electrostatically formed nanowire semiconductor VOCs identification method based on table deep learning and model fusion, including the following steps: Step 1, Experimental Objects: To have concentration-independent selectivity for the target VOCs, 200 ppm hexanol, 1000 ppm butanol, 1700 ppm propanol, 1800 ppm ethanol, 3000 ppm acetone, and 8800 ppm acetic acid were used in order to obtain the same responsivity for all six molecules (~0.95). Class labels for VOCs: Hexanol, butanol, propanol, ethanol, acetone, and acetic acid are numbered in order from 0 to 5.

[0041] Step 2, feature selection: before and after the VOCs to be identified pass through the silicon dioxide film 2 on the electrostatically formed nanowires, the gate voltage V JG (That is, the voltage between the junction gate and the source. There are two junction gates, marked as the first junction gate and the second junction gate, and the voltages of the first junction gate and the second junction gate are recorded as V JG1 and V JG2 ) and the drain voltage V DS (i.e. drain and source voltages), changing the back gate voltage V BG (i.e., the voltage between the back gate and the source), and measure the drain current I DS (i.e. the current from drain to source), calculate the change in back gate threshold voltage before and after VOCs passes through EFN (ΔV th (BG)), the change of back gate subthreshold swing (ΔSS(BG)) and back gate source leakage current (ΔI on (BG)), this set of parameters is recorded as ΔBG. In another different set of experiments, the VOCs to be identified before and after passing through EFN at different back gate voltages V BG and drain voltage V DS Under the full change of the junction gate voltage V JG , measure the drain current I DS , calculate the change in the gate threshold voltage before and after VOCs passes through EFN (ΔV th(JG)), the change in junction-gate subthreshold swing (ΔSS(JG)), and the change in junction-gate-source leakage current (ΔI on (JG)), this set of parameters is recorded as ΔJG. These extracted parameters are used as fingerprint features of the target analyte to be measured. Back gate voltage V BG , Junction gate voltage V JG and drain voltage V DS For configuration, see Figure 1 To make EFN symmetrical, take V JG =V JG1 =V JG2 The two sets of parameters correspond to the two formation directions of EFN, such as Figure 2 shown.

[0042] Step 3, algorithm architecture design: Volatile organic compound classification is a multi-classification problem. The features and category labels extracted independently for the deformation of electrostatically formed nanowires under different bias configurations (i.e., lateral deformation and longitudinal deformation) form two independent data sets ΔBG (see Table 1 (a)) and data set ΔJG (see Table 1 (b)): Table 1(a)

[0043] Table 1(b)

[0044] The classification of VOCs gases can be obtained by using table deep learning. Many excellent algorithms have emerged in table deep learning. In order to screen out the table deep learning algorithms suitable for the experimental data of the present invention, the present invention uses an automatic learning framework to sort the table deep learning algorithms, and then performs stacking model fusion on the screened optimal algorithms. Therefore, in the overall algorithm architecture, the present invention adopts a strategy that combines table deep learning, automatic learning and model fusion. The specific flow chart is shown in Figure 3 .

[0045] Step 4, outlier check: Data set check aims to identify and correct identifiable errors in data files, including checking data consistency, handling invalid values, missing values, outliers, etc. Here we focus on the outlier problem, and outliers at the boundary should be retained.

[0046] Step 5, interpolation: In machine learning, due to the time-consuming nature of experiments, the data set is usually expanded to obtain better results. Interpolation is a common method for expanding tabular data; therefore, the present invention uses linear interpolation to perform data expansion here.

[0047] Step 6, model screening and optimization: In order to select deep learning models that are more suitable for the ΔBG and ΔJG datasets from the numerous tabular deep learning models, the NODE model, SAINT model, TabNet model, VIME model, MLP model, DeepFM model, DeepGBM model and other popular algorithm models are screened through automatic learning tools. The common models are compared by F1 score, and the top five models in F1 index are obtained: NODE model, SAINT model, TabNet model, VIME model and MLP model. The comparison results of the tabular deep models of the ΔBG and ΔJG datasets are shown in Table 2: Table 2

[0048] As can be seen from Table 2, for the ΔBG dataset, the NODE model has the highest accuracy, with an accuracy of 0.96324±0.00392. The accuracy of other models from high to low are TabNet model (accuracy of 0.95864±0.00857), SAINT model (accuracy of 0.95788±0.01086), VIME model (accuracy of 0.89056±0.06220) and MLP model (accuracy of 0.86289±0.07215). For the ΔJG dataset, the TabNet model has the highest accuracy, reaching 0.95558±0.00896. The accuracy of other models from high to low are NODE model (accuracy is 0.94869±0.00896), SAINT model (accuracy is 0.94716±0.01705), VIME model (accuracy is 0.94409±0.01345) and MLP model (accuracy is 0.83766±0.01972).

[0049] For the later implementation of VOCs detection system using embedded systems, special attention should be paid to inference time and model size. Too long inference time affects the timeliness of product detection. Too large a model cannot be conveniently deployed in an embedded system.

[0050] Figure 4 (a) and Figure 4 (b) show the relationship between model size and accuracy in two different datasets, ΔBG and ΔJG. As can be seen from Figure 4 (a), the NODE model has the highest accuracy and the smallest standard deviation, but the model size is also the largest. As can be seen from Figure 4 (b), the TabNet model has the highest accuracy in the ΔJG dataset, the NODE model occupies the largest space, and the standard deviations of the TabNet model and the NODE model are the smallest. The MLP model has the lowest accuracy and the largest standard deviation on both datasets (i.e., the ΔBG dataset and the ΔJG dataset), but also the shortest inference time.

[0051] Figure 5 (a) and Figure 5 (b) show the relationship between the model inference time and accuracy under the ΔBG and ΔJG datasets, respectively. The MLP model has the shortest inference time on the two datasets (i.e., the ΔBG dataset and the ΔJG dataset), while the NODE model has the longest inference time. The present invention finds that the inference time is inversely proportional to the model size.

[0052] Combining Figure 4(a), Figure 4(b), Figure 5(a) and Figure 5(b), we can see that the TabNet model has moderate inference time and model size, and its accuracy is relatively high.

[0053] Step 7, model fusion: In combination with the principle of FEN VOCs sensor, the present invention uses a 5-fold cross-validation two-layer stacked model. In order to consider the needs of incremental learning and transfer learning, the base model (i.e., basic model) selects a deep learning model. In order to prevent overfitting, the meta-model (i.e., secondary model) uses logistic regression. In addition, the training of the first layer model (i.e., basic model) is as follows Figure 6 As shown in the figure, cross validation is used. The training process of the fusion model includes: first, the training set is divided into 5 parts, and each basic model is cross-validated by 5 folds, that is, 4 parts of data are used for training the model, and the remaining 1 part is used for prediction. The cycle is performed in the data set, and all 5 parts of data are independently involved in the verification. Then, the cross-validation prediction results of the first-layer model are used as new features, and the prediction results of the two groups of validation data marked with different colors in the first layer are spliced ​​as the training data of the second-layer meta-model, and the meta-model is trained to obtain the fusion model. For the two-layer fusion model, in order to avoid data information leakage, the test data does not participate in the training of any of the two layers. The average values ​​of the 5 groups of test results of the two models in the first layer (i.e., the models that process (i.e., predict) the ΔBG data set and the ΔJG data set respectively) are spliced ​​and used as the test data of the meta-model, and the final prediction results are obtained by data processing through the meta-model. The first-layer prediction results are averaged to reduce the deviation and variance of a single model, and to use the diversity of the model to improve the stability and accuracy of the overall prediction. In this way, the generalization ability of the model can be better improved, and more accurate and reliable prediction results can be obtained.

[0054] The top 5 excellent algorithm models are selected through automatic learning. In view of the high size and inference time of the NODE model, the present invention removes the NODE model during the model fusion process. When performing model fusion, the present invention first obtains the prediction results of the first layer model through 5-fold cross validation. The accuracy results on two data sets (i.e., ΔBG data set and ΔJG data set) are shown in Table 3: Table 3

[0055] As can be seen from Table 3, on the ΔBG dataset, the SAINT model has the best accuracy (accuracy is 0.96258±0.00590), while on the ΔJG dataset, the VIME model has the highest accuracy (accuracy is 0.93742±0.01465). The inference time (in seconds) on the two datasets (i.e., the ΔBG dataset and the ΔJG dataset) is shown in Table 4: Table 4

[0056] It can be seen from Table 4 that the SAINT model has the highest inference time on the two datasets (i.e., the ΔBG dataset and the ΔJG dataset), and the inference time is 0.72349±0.54594 and 0.44631±0.04003 respectively.

[0057] Figure 7 (a) and Figure 7 (b) intuitively show the distribution of accuracy and inference time of the four models in the first layer on two datasets (i.e., ΔBG dataset and ΔJG dataset). The difference in inference time is relatively small relative to the total time of the detection process, and the accuracy is mainly used to measure the effect of the model. From Figure (a) and Figure 7 (b), it can be seen that the SAINT model performs best on the ΔBG dataset, while the VIME model performs best on the ΔJG dataset and the TabNet model.

[0058] The accuracy of the second-layer model fusion is shown in Table 5: Table 5

[0059] As can be seen from Table 5, the fusion model with the highest accuracy is the TabNet and SAINT algorithm fusion model, with an accuracy of 99.877%, but the TabNet model has the shortest inference time. Therefore, it is best to select the TabNet model as the base model for fusion.

[0060] Step 8, model evaluation: For the evaluation of the two-layer model, the present invention directly uses the data that has not participated in any stage of training as the test data to use the confusion matrix and ROC diagram to see the overall prediction effect. The confusion matrix can show the classification of the model in different categories, including the number of true positive examples, false positive examples, true negative examples and false negative examples. For multi-classification problems, the confusion matrix can be presented in the form of a color heat map to more intuitively compare the performance of different models. Figure 8It is the confusion matrix of a set of data in which the base model adopts the MLP model on two data sets (i.e., the ΔBG data set and the ΔJG data set). The ROC curve is an important tool for evaluating model performance and is used to measure the classification ability of the model. The AUC (Area Under Curve) value is the area under the ROC curve and is used to measure the overall performance of the classification model. The AUC value ranges from 0 to 1. The closer the value is to 1, the better the performance of the model. Fig. 9 The ROC curve is nearly a right angle, indicating that the model classification performance is very good.

[0061] Step 9, base model weights: For model fusion, a model weight graph can be drawn to show the weights of each base learner and how these weights affect the final prediction results. Taking MLP model fusion as an example, each feature in the MLP fusion model is composed of different weights of the first layer 6 VOCs classification prediction results of the two models. Fig.10 This is the distribution of different weights of each base model of the first feature after the MLP model fusion. It is composed of different weight combinations of the 12 features in the first layer. The weight combination graphs of other features in the second layer are similar. Fig.10 From the distribution of the medium-weighted bar chart, it can be seen that the two models are highly complementary.

[0062] Step 10, fusion results: The present invention compares the accuracy of the model using the ΔBG dataset alone, the ΔJG dataset alone, and the fusion of the two datasets using the same algorithm. Fig.11 It can be seen that the classification accuracy of the fused model of the present invention is improved compared with the model under a single data set, and the MLP model with low classification accuracy under a single data set is also greatly improved. In terms of accuracy alone, from the current sample data classification results, the accuracy results of the SAINT model and the TabNet model are the same, both of which are the highest, with an accuracy of 0.99877±0.00123. However, considering the training time and inference time of the model, the TabNet model as the base model after Stacking fusion is the best model.

[0063] It can be seen from the above scheme that the present invention uses VOCs with the same responsiveness as the identification object, and uses electrostatic formation of nanowire semiconductor physical adsorption as the basic principle. The physical parameters of semiconductor response before and after VOCs access are used as identification fingerprints. Through feature calculation, data set construction, data set inspection, data set balance and expansion, algorithm framework design, automatic learning framework selection, table deep learning algorithm screening, model fusion, model evaluation, and the feature distribution of the two data sets in the Stacking model fusion is demonstrated, indicating that the two data sets have complementary information, which provides a theoretical analysis for model fusion to improve classification accuracy. Automatic screening and model fusion of table deep learning models are realized and the best classification accuracy is obtained. It can be seen that the above scheme of the present invention has the following advantages, First: The deep learning-based model can use incremental learning to accumulate more data to improve the accuracy of gas recognition and the stability of the sensor, and is also conducive to future integration in the embodied intelligent system.

[0064] Second: Apply tabular deep learning to two independent datasets of EFN sensors.

[0065] Third: We compared the table deep learning algorithm of manifold with the table deep learning algorithm of automatic learning on independent data sets. Automatic learning can select the best model, greatly improving efficiency and reducing human involvement.

[0066] Fourth: Using the FET-type EFN sensor structure and principle properties, the two independent models were fused and the results were analyzed.

[0067] Fifth: This method has wide applicability and can be extended to all independent model fusion tasks based on FET-type EFN sensors.

[0068] The present invention also provides a single electrostatically formed nanowire semiconductor VOCs identification system based on table deep learning and model fusion for implementing any of the above embodiments, including: A first data set acquisition unit: used to acquire a first data set, wherein the first data set includes a change value of a back gate threshold voltage, a change value of a back gate subthreshold swing, and a change value of a back gate source leakage current before and after the VOCs to be identified pass through a single electrostatic wire to form a nanowire semiconductor; A second data set acquisition unit: used to acquire a second data set, wherein the second data set includes a change value of a junction gate threshold voltage, a change value of a junction gate subthreshold swing, and a change value of a junction gate source leakage current before and after the VOCs to be identified pass through a single electrostatic nanowire semiconductor; Identification module: used to perform Stacking model fusion on the first data set and the second data set to obtain the final prediction result of the type of VOCs to be identified; wherein, when performing Stacking model fusion, the base model adopts a table deep learning model, and the meta-model adopts a table deep learning model.

[0069] The embodiments of the present invention also provide corresponding electronic devices and computer-readable storage media for implementing the solutions provided by the embodiments of the present invention.

[0070] Among them, the device includes a memory and a processor, the memory is used to store instructions or codes, and the processor is used to execute the instructions or codes so that the device executes the single electrostatically formed nanowire semiconductor VOCs identification method based on table deep learning and model fusion described in any embodiment of the present application.

[0071] The storage medium stores a computer program, wherein the computer program, when executed by the processor, implements the single electrostatically formed nanowire semiconductor VOCs identification method based on table deep learning and model fusion as described in any embodiment of the present application.

[0072] Another object of an embodiment of the present invention is to provide a computer program product, which includes computer instructions, and the computer instructions instruct the computer to execute the single electrostatically formed nanowire semiconductor VOCs identification method based on table deep learning and model fusion.

[0073] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0074] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0075] The present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, readable storage media, optical storage, etc.) containing computer-usable program codes.

[0076] Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0077] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for identifying semiconductor VOCs in single electrostatically formed nanowires based on table deep learning and model fusion, characterized in that: include: Acquire a first data set, wherein the first data set includes a change value of a back gate threshold voltage, a change value of a back gate subthreshold swing, and a change value of a back gate source leakage current before and after the VOCs to be identified are formed into a nanowire semiconductor through a single electrostatic wire; Acquire a second data set, wherein the second data set includes a change value of a junction gate threshold voltage, a change value of a junction gate subthreshold swing, and a change value of a junction gate source leakage current before and after the VOCs to be identified pass through a single electrostatic wire to form a nanowire semiconductor; The first data set and the second data set are fused by the Stacking model to obtain the final prediction result of the type of VOCs to be identified; wherein, when the Stacking model is fused, the base model adopts a table deep learning model, and the meta-model adopts a table deep learning model.

2. According to claim 1, a method for identifying semiconductor VOCs of a single electrostatically formed nanowire based on table deep learning and model fusion, characterized in that: When the first and second data sets are fused using the Stacking model: Using a pre-trained first base model to process the first data set, obtaining a first prediction result of the VOCs type to be identified; The second data set is processed using the pre-trained second base model to obtain a second prediction result of the VOCs type to be identified; splicing the first prediction result and the second prediction result to obtain spliced ​​data; The pre-trained meta-model is used to process the spliced ​​data to obtain the final prediction results of the VOCs types to be identified.

3. According to claim 1, a method for identifying semiconductor VOCs of a single electrostatically formed nanowire based on table deep learning and model fusion, characterized in that: The process of obtaining the change value of the back gate threshold voltage, the change value of the back gate subthreshold swing, and the change value of the back gate source leakage current before and after the VOCs to be identified pass through a single static electrode to form a nanowire semiconductor includes: for different junction gate voltages and drain voltages, the back gate voltage is changed throughout the process, the drain current is measured, and then the change value of the back gate threshold voltage, the change value of the back gate subthreshold swing, and the change value of the back gate source leakage current are calculated before and after the VOCs to be identified pass through a single static electrode to form a nanowire semiconductor; The process of obtaining the change value of the junction gate threshold voltage, the change value of the junction gate subthreshold swing, and the change value of the junction gate source leakage current before and after the VOCs to be identified pass through a single static electricity to form a nanowire semiconductor includes: changing the junction gate voltage throughout the process under different back gate voltages and drain voltages, measuring the drain current, and then calculating the change value of the junction gate threshold voltage, the change value of the junction gate subthreshold swing, and the change value of the junction gate source leakage current before and after the VOCs to be identified pass through a single static electricity to form a nanowire semiconductor; Wherein, when calculating the change value of the back gate threshold voltage, the change value of the back gate subthreshold swing, the change value of the back gate source leakage current, the change value of the junction gate threshold voltage, the change value of the junction gate subthreshold swing and the change value of the junction gate source leakage current, the voltages of the two junction gates in a single electrostatically formed nanowire semiconductor are made equal to the junction gate voltage.

4. According to claim 1, a method for identifying semiconductor VOCs of a single electrostatically formed nanowire based on table deep learning and model fusion, characterized in that: The VOCs to be identified include at least one of hexanol, butanol, propanol, ethanol, acetone and acetic acid.

5. A method for identifying semiconductor VOCs in a single electrostatically formed nanowire based on table deep learning and model fusion according to any one of claims 1 to 4, characterized in that: The table deep learning model adopts a NODE model, a SAINT model, a TabNet model, a VIME model or an MLP model.

6. According to claim 3, a method for identifying semiconductor VOCs of a single electrostatically formed nanowire based on table deep learning and model fusion, characterized in that: The table deep learning model adopts the TabNet model.

7. A single electrostatically formed nanowire semiconductor VOCs identification system based on table deep learning and model fusion, characterized in that: include: A first data set acquisition unit: used to acquire a first data set, wherein the first data set includes a change value of a back gate threshold voltage, a change value of a back gate subthreshold swing, and a change value of a back gate source leakage current before and after the VOCs to be identified pass through a single electrostatic wire to form a nanowire semiconductor; A second data set acquisition unit: used to acquire a second data set, wherein the second data set includes a change value of a junction gate threshold voltage, a change value of a junction gate subthreshold swing, and a change value of a junction gate source leakage current before and after the VOCs to be identified pass through a single electrostatic nanowire semiconductor; Identification module: used to perform Stacking model fusion on the first data set and the second data set to obtain the final prediction result of the type of VOCs to be identified; wherein, when performing Stacking model fusion, the base model adopts a table deep learning model, and the meta-model adopts a table deep learning model.

8. An electronic device, characterized in that: include: one or more processors; a storage device having one or more programs stored thereon; When the one or more programs are executed by the one or more processors, the one or more processors implement the single electrostatically formed nanowire semiconductor VOCs identification method based on table deep learning and model fusion as described in any one of claims 1 to 6.

9. A storage medium, characterized in that: A computer program is stored thereon, wherein when the computer program is executed by a processor, the single electrostatically formed nanowire semiconductor VOCs identification method based on table deep learning and model fusion as described in any one of claims 1 to 6 is implemented.

10. A computer program product, comprising computer instructions, characterized in that: The computer instructions instruct the computer to execute the single electrostatically formed nanowire semiconductor VOCs identification method based on table deep learning and model fusion as described in any one of claims 1-6.

Images