A gas concentration identification method based on short-period temperature modulation signal of semiconductor oxide sensor array

Abstract

The application discloses a kind of gas concentration identification methods based on short period temperature modulation signal of semiconductor oxide sensor array.The method is aimed at the dynamic characteristics of sensing signal under temperature modulation condition, and a kind of double-layer feature extraction and timing modeling structure is designed.Firstly, the sensing response signal in each temperature modulation period is extracted in parallel using the path signature mechanism, and multidimensional periodic characteristics are obtained.Secondly, the characteristics of each modulation period are combined in time sequence to construct a cross-period time series representation.Finally, the global correlation features of the time series are extracted using the Transformer model to realize the regression prediction of the mixed gas component concentration.The application improves the accuracy and prediction stability of gas concentration identification through the algorithm structure design of sensing signal features.

Description

Technical Field

[0001] This invention belongs to the field of gas detection technology, specifically a gas concentration identification method based on short-period temperature modulation signals of a semiconductor oxide sensor array. Background Technology

[0002] Metal oxide semiconductor (MOS) gas sensors are well-suited for field gas detection applications due to their low cost, fast response, and high sensitivity. However, MOS sensors inherently suffer from low gas selectivity, especially for gases with similar properties. Similar charge transfer processes occur on the sensing layer, often resulting in overlapping conductivity responses, significantly reducing their effectiveness in complex real-world environments. Temperature modulation techniques, by adjusting the driving voltage across the microheater in the MOS gas sensor, allow the sensor to operate at different temperatures, acquiring dynamic response characteristics under multiple temperature conditions and enriching the dimensions of the response signal. This technique has been extensively studied and is widely considered an effective strategy for improving the selectivity of MOS gas sensors.

[0003] The response signal of a temperature-modulated sensor is influenced by both gas concentration and heating temperature modulation, exhibiting both a long-term response trend that changes with gas concentration and short-term dynamic fluctuations caused by temperature modulation. This multi-timescale characteristic significantly increases feature dimensionality and information complexity, placing higher demands on modeling methods. However, existing methods still treat the temperature-modulated sensor response signal as a single-timescale sequence for modeling. Rapid changes within the modulation period and slow evolution trends across modulation periods are simultaneously encoded in the time-series representation, increasing the complexity of learning features at different timescales and failing to fully utilize the coexisting long and short-timescale characteristics of the temperature-modulated signal. Summary of the Invention

[0004] This invention proposes a gas concentration identification method based on short-cycle temperature modulation signals from a semiconductor oxide sensor array. Path signatures are used to encode the sensor responses within a single temperature modulation cycle, extracting temporal dependencies and dynamic correlation features between sensors. Then, a Transformer is used to model the feature sequences corresponding to different modulation cycles to characterize the correlation and long-term dependence between cycles, thereby improving the accuracy of mixed gas concentration identification.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A gas concentration identification method based on short-period temperature modulation signals from a semiconductor oxide sensor array includes the following steps: Step 1: Receive the response signal from the semiconductor oxide sensor array.X Divided according to temperature modulation period N One modulation period signal, N To constitute X Number of temperature modulation cycles; Step 2: Trajectory Encoding Feature Extraction within a Period: Parallel extraction using third-order path signatures. N The characteristics of each modulation period signal are used to obtain the first, second, and third order path signature features of each modulation period signal. Step 3: For each segment of the modulation period signal, sequentially splice the corresponding first, second, and third-order path signature features to obtain the splicing features of the modulation period signal; Step 4: Arrange the splicing features of all modulation periodic signals according to the original time order of the modulation periodic signals to form a feature sequence. ; Step 5: Cross-period temporal correlation feature extraction: feature sequence The trained Transformer network is input to capture the temporal correlation feature representation Z between different modulation periods. The temporal correlation feature representation Z is input into the fully connected layer to obtain the gas concentration prediction result.

[0006] Further improvements include the response signal of the semiconductor oxide sensor array. X Divided into N A signal with one modulation period is represented as:

[0007] in, T Indicates the total duration of the response signal. d Indicates the number of channels in a semiconductor oxide sensor. X i Then it means the first i The response value of the semiconductor oxide sensor array within each modulation cycle is expressed as:

[0008] in, x j i Indicates the first i The first modulation cycle j The response value of the semiconductor oxide sensor array at any given time. L This indicates the duration of a single temperature modulation cycle.

[0009] Further improvements are made, and the specific steps of step two are as follows: 2.1) Treat the short-cycle temperature modulation signal of the semiconductor oxide sensor array as a path X t , pathX t Defined as from interval [ a , b A continuous mapping from ] to, denoted as Where d represents the number of channels in the semiconductor oxide sensor, for any t ∈[ a , b The path is represented as:

[0010] in Indicates the first i Each signal channel at time t Response value; path X The response signal of the semiconductor oxide sensor array is represented within a single temperature modulation cycle. Each channel is acquired under the same modulation conditions, forming a multidimensional time-series trajectory. a Indicates the start time of the temperature modulation period. b Indicates the end time of the temperature modulation cycle; 2.2) First-order path signature This indicates that each signal channel is in the interval [ a , b Cumulative increment within:

[0011] The overall variation amplitude of the modulation periodic signal within a single period is described. Under the temperature-modulated semiconductor oxide sensor response signal, the first-order path signature reflects the cumulative effect of gas-sensor interaction on a macroscopic time scale and reflects the overall sensitivity of different sensors to changes in gas concentration. Indicates the first i Each signal channel in t 1 The response value at any given time; 2.3) The second-order path signature is defined as follows:

[0012] For the first i With the j The directed area formed by each signal channel in the time dimension explicitly characterizes the temporal order and coupling relationship between different channels; within a single temperature modulation cycle, the second-order path signature captures the relative dynamic characteristics between the responses of different semiconductor oxide sensors, including the order of response, the difference in rate of change, and the potential hysteresis relationship. 2.4) Third-order path signatures, by encoding the joint changes across multiple channels and time points, further characterize the complex geometric structure of the path in high-dimensional space, representing the high-order interactions between multiple channels and the composite dynamic behavior across multiple time scales. The expression is:

[0013] For the first i , j , k The ordered triple integral features formed by the three channels in the time dimension characterize the high-order temporal dependencies and nonlinear coupling structure among the three channels.

[0014] Further improvements are made, and the specific steps of step three are as follows: Will , and Sequentially concatenating the signals, the concatenation characteristics of the modulated periodic signals are used as a cross-period time series representation:

[0015] in f t For the first t The splicing characteristics of signals with modulation cycles.

[0016] Further improvements are made, and the specific steps of step four are as follows: Arrange the spliced ​​features of all modulation periods in their original order to obtain the feature sequence. :

[0017] Among them, the feature sequence Each element corresponds to a modulation period, and the feature sequence The structure reflects the sequential relationship between modulation cycles.

[0018] Further improvements are made, and the specific steps of step five are as follows: A sequence modeling method based on self-attention mechanism is used to model the feature sequence. Modeling is performed by adaptively assigning weights to different periodic features at the sequence level to achieve unified modeling of multi-period information: In order to explicitly map periodic order information, the feature sequence Temporal encoding is performed, and the encoded data is then used to generate Query, Key, and Value through linear mapping. Attention weights A are then calculated, and the value vectors are weighted and converged to obtain the temporal correlation feature representation Z. Finally, Z is fed into a fully connected layer to obtain the predicted gas concentration. ;

[0019]

[0020] in, Q Query means to search; K Key, meaning key; V Valure represents the value; softmax represents the normalized exponential function. This represents the dimension of the key vector Key. This represents the model's predicted output. This represents the weight matrix of the fully connected layer. The term represents the bias term of the fully connected layer; T represents the matrix transpose. Represents the input feature matrix. This represents the learnable weight matrix.

[0021] A further improvement is made to the loss function of the Transformer network as follows: Training until The trained Transformer network is obtained when the minimum is reached.

[0022] in, k Indicates gas component index, Indicates the first k Predicted concentration values ​​for each gas. Indicates the first k The true concentration of each gas.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention employs a two-layer feature extraction architecture. By modeling the sensor response within a single temperature modulation cycle as a path, the path signature is used to compactly represent its temporal structure and dynamic relationship between channels, effectively capturing dynamic information within the cycle. The feature representations corresponding to different modulation cycles form a high-level temporal sequence in chronological order, and the Transformer further models the correlation and long-term dependence between cycles, improving accuracy compared to existing methods. Attached Figure Description

[0024] Figure 1 This is a structural diagram of the gas concentration identification method based on short-period temperature modulation signal of semiconductor oxide sensor array according to the present invention.

[0025] Figure 2 This is a schematic diagram of the third-order path signature feature extraction of the gas concentration identification method based on short-period temperature modulation signal of semiconductor oxide sensor array according to the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 those skilled in the art without creative effort are within the protection scope of the present invention.

[0027] The gas concentration identification method based on short-period temperature modulation signals of semiconductor oxide sensor arrays includes the following steps: Step 1: Divide the sensor array signal into 40 segments according to the temperature modulation period, where 40 is the number of temperature modulation periods that constitute the signal. Step 2: Use third-order path signature to extract features from 40 segments of sensor array signals in parallel; Step 3: For each segment of the sensor array signal, sequentially splice its first, second, and third-order path signature features; Step 4: Arrange the features of all modulation cycles in their original time order to form a cross-cycle time series representation; Step 5: Model the feature sequences from Step 4 using a Transformer network to capture the temporal correlation features between different modulation periods; Step 6: Input the features output from Step 5 into the fully connected layer to obtain the gas concentration prediction results.

[0028] The gas concentration identification method based on short-period temperature modulation signals from a semiconductor oxide sensor array employs five-fold cross-validation during training, with a training set, validation set, and test set ratio of 6:2:2. Bayesian optimization is used to search for the optimal hyperparameters, and the hyperparameter search space is as follows: Table 1. Search Space for Hyperparameters in Bayesian Optimization

[0029] The following is a comparison of the gas concentration identification method of the short-period temperature modulation signal of the semiconductor oxide sensor array proposed in this invention with the mixed gas concentration identification data of existing 1DCNN, LSTM, GRU, ResNet18, and TCN models.

[0030] The data comparison results are shown in Table 2: Table 2 Comparison of recognition method accuracy

[0031] As shown in Table 2, among the various comparison methods, compared with models such as 1D-CNN, LSTM, GRU, ResNet18, and TCN, the gas concentration identification method based on short-period temperature modulation signals of a semiconductor oxide sensor array in this invention exhibits a lower overall error level in the RMSE and MAE indices for both methane and acetylene, indicating its stronger modeling capability in the sensor response to gas concentration identification tasks. Especially in the methane gas identification scenario, the proposed method reduces the prediction error by at least 44.77% compared to the aforementioned comparative models, demonstrating a significant accuracy advantage. These results demonstrate that the gas concentration identification method based on short-period temperature modulation signals of a semiconductor oxide sensor array in this invention achieves competitive prediction performance in mixed gas concentration identification tasks.

[0032] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A gas concentration identification method based on short-period temperature modulation signals from a semiconductor oxide sensor array, characterized in that, Includes the following steps: Step 1: Receive the response signal from the semiconductor oxide sensor array. X Divided according to temperature modulation period N One modulation period signal, N To constitute X Number of temperature modulation cycles; Step 2: Trajectory Encoding Feature Extraction within a Period: Parallel extraction using third-order path signatures. N The characteristics of each modulation period signal are used to obtain the first, second, and third order path signature features of each modulation period signal. Step 3: For each segment of the modulation period signal, sequentially splice the corresponding first, second, and third-order path signature features to obtain the splicing features of the modulation period signal; Step 4: Arrange the splicing features of all modulation periodic signals according to the original time order of the modulation periodic signals to form a feature sequence. ; Step 5: Cross-period temporal correlation feature extraction: feature sequence The trained Transformer network is input to capture the temporal correlation feature representation Z between different modulation periods. The temporal correlation feature representation Z is input into the fully connected layer to obtain the gas concentration prediction result.

2. The gas concentration identification method based on short-period temperature modulation signal of semiconductor oxide sensor array as described in claim 1, characterized in that, The response signal of the semiconductor oxide sensor array X Divided into N A signal with one modulation period is represented as: ； in, T Indicates the total duration of the response signal. d Indicates the number of channels in a semiconductor oxide sensor. X i Then it means the first i The response value of the semiconductor oxide sensor array within each modulation cycle is expressed as: ； in, x j i Indicates the first i The first modulation cycle j The response value of the semiconductor oxide sensor array at any given time. L This indicates the duration of a single temperature modulation cycle.

3. The gas concentration identification method based on short-period temperature modulation signal of semiconductor oxide sensor array as described in claim 1, characterized in that, The specific steps of step two are as follows: 2.1) Treat the short-cycle temperature modulation signal of the semiconductor oxide sensor array as a path X t , path X t Defined as from interval [ a , b A continuous mapping from ] to, denoted as Where d represents the number of channels in the semiconductor oxide sensor, for any t ∈[ a , b The path is represented as: ； in Indicates the first i Each signal channel at time t The response value; path X The response signal of the semiconductor oxide sensor array is represented within a single temperature modulation cycle. Each channel is acquired under the same modulation conditions, forming a multidimensional time-series trajectory. a Indicates the start time of the temperature modulation period. b Indicates the end time of the temperature modulation cycle; 2.2) First-order path signature This indicates that each signal channel is in the interval [ a , b Cumulative increment within: ； The overall variation amplitude of the modulation periodic signal within a single period is described. Under the temperature-modulated semiconductor oxide sensor response signal, the first-order path signature reflects the cumulative effect of gas-sensor interaction on a macroscopic time scale and reflects the overall sensitivity of different sensors to changes in gas concentration. Indicates the first i Each signal channel in t 1 The response value at any given time; 2.3) The second-order path signature is defined as follows: ； For the first i With the j The directed area formed by each signal channel in the time dimension explicitly characterizes the temporal order and coupling relationship between different channels; within a single temperature modulation cycle, the second-order path signature captures the relative dynamic characteristics between the responses of different semiconductor oxide sensors, including the order of response, the difference in rate of change, and the potential hysteresis relationship. 2.4) Third-order path signatures, by encoding the joint changes across multiple channels and time points, further characterize the complex geometric structure of the path in high-dimensional space, representing the high-order interactions between multiple channels and the composite dynamic behavior across multiple time scales. The expression is: ； For the first i , j , k The ordered triple integral features formed by the three channels in the time dimension characterize the high-order temporal dependencies and nonlinear coupling structure among the three channels.

4. The gas concentration identification method based on short-period temperature modulation signal of semiconductor oxide sensor array as described in claim 3, characterized in that, The specific steps of step three are as follows: Will , and Sequentially concatenating the signals, the concatenation characteristics of the modulated periodic signals are used as a cross-period time series representation: ； in f t For the first t The splicing characteristics of signals with modulation cycles.

5. The gas concentration identification method based on short-period temperature modulation signal of semiconductor oxide sensor array as described in claim 1, characterized in that, The specific steps of step four are as follows: Arrange the spliced ​​features of all modulation periods in their original order to obtain the feature sequence. : ； Among them, the feature sequence Each element corresponds to a modulation period, and the feature sequence The structure reflects the sequential relationship between modulation cycles.

6. The gas concentration identification method based on short-period temperature modulation signal of semiconductor oxide sensor array as described in claim 1, characterized in that, The specific steps of step five are as follows: A sequence modeling method based on self-attention mechanism is used to model the feature sequence. Modeling is performed by adaptively assigning weights to different periodic features at the sequence level to achieve unified modeling of multi-period information: In order to explicitly map periodic order information, the feature sequence Temporal encoding is performed, and after encoding, Query, Key and Value are generated by linear mapping. Then, attention weight A is calculated, and the value vectors are weighted and converged to obtain the temporal relevance feature representation Z. Finally, Z is fed into the fully connected layer to obtain the predicted gas concentration value. ; ； ； in, Q Query means to search; K Key, meaning key; V Valure represents the value; softmax represents the normalized exponential function. This represents the dimension of the key vector Key. This represents the model's predicted output. This represents the weight matrix of the fully connected layer. The term represents the bias term of the fully connected layer; T represents the matrix transpose. Represents the input feature matrix. This represents the learnable weight matrix.

7. The gas concentration identification method based on short-period temperature modulation signal of semiconductor oxide sensor array as described in claim 1, characterized in that, The loss function of the Transformer network is Training until The trained Transformer network is obtained when the minimum is reached. ； in, k Indicates gas component index, Indicates the first k Predicted concentration values ​​for each gas. Indicates the first k The true concentration of each gas.

Images