A mixed gas recognition method based on one-dimensional convolutional neural network
By combining one-dimensional convolutional neural networks (1D-CNN) with data preprocessing and transfer learning, the problems of high computational complexity, strong dependence on labeled data, and insufficient robustness in mixed gas identification are solved, achieving efficient and accurate mixed gas identification, which is suitable for embedded systems and edge computing platforms.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ELECTRIC POWER RES INST OF GUANGXI POWER GRID CO LTD
- Filing Date
- 2026-02-10
- Publication Date
- 2026-06-16
AI Technical Summary
Existing mixed gas identification technologies suffer from problems such as low identification accuracy, poor robustness, high computational resource consumption, strong dependence on labeled data, and insufficient generalization ability in complex environments. They are particularly difficult to deploy and adapt to environmental changes in embedded systems and wearable devices.
A one-dimensional convolutional neural network (1D-CNN) is used for mixed gas identification. Combining data preprocessing and a lightweight structure, features are automatically extracted through multi-layer convolution and pooling operations, supporting multi-class prediction. A transfer learning mechanism is also introduced to reduce the dependence on labeled data.
It enables efficient identification of mixed gases in complex environments, reduces computational load, improves identification accuracy and robustness, adapts to different sensor configurations and environmental changes, and is suitable for embedded systems and edge computing platforms.
Smart Images
Figure CN122224344A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of mixed gas recognition technology, and in particular to a mixed gas recognition method based on a one-dimensional convolutional neural network. Background Technology
[0002] Mixed gas identification technology plays a crucial role in several key application scenarios, including environmental monitoring, industrial production safety control, medical and health diagnosis, food safety, and public safety, and has broad research and application prospects. Traditional mixed gas identification methods mainly rely on the response characteristics of physicochemical sensor arrays. They typically utilize the selective response behavior of multiple sensors to a target gas, combined with pattern recognition algorithms, to achieve qualitative and quantitative analysis of a single gas or gas mixture. However, in practical applications, due to factors such as sensor cross-sensitivity, response drift, environmental noise interference, and fluctuations in the concentration of multi-component gases, these methods suffer from significant limitations when handling complex mixed gas identification tasks, including low identification accuracy, poor robustness, and sensitivity to external environments.
[0003] With the rapid development of artificial intelligence technology, especially deep learning methods, researchers have gradually introduced it into the field of gas identification. Deep neural networks can automatically extract discriminative features from high-dimensional, nonlinear and redundant sensor data by constructing multi-layer nonlinear mapping structures, thereby effectively improving the accuracy and generalization ability of mixed gas identification. Among them, convolutional neural networks (CNNs) perform well in extracting local spatial features and are widely used in image recognition and modeling of one-dimensional time series data; while recurrent neural networks (RNNs) are suitable for capturing dynamic change patterns in sensor signals due to their ability to model time dependencies. Although neural network-based mixed gas identification methods have made significant progress, there are still several problems to be solved in the existing technology. These are mainly reflected in the following aspects: (1) Strong dependence on labeled data: The performance of deep learning models depends to a large extent on sufficient and high-quality labeled data. Especially in gas identification tasks, different gas components, concentration combinations and environmental variables constitute an extremely large sample space. However, in practical applications, due to limited experimental conditions, high annotation costs, and uneven sample distribution, it is often very difficult to obtain a wide range of high-quality labeled datasets, which seriously restricts the sufficiency of model training and the space for performance improvement. (2) High computational resource consumption: The current mainstream deep neural network models are generally complex in structure, containing a large number of parameters and computational layers. Their training and inference processes require high computational power and significant storage resources. Especially in application scenarios with high real-time requirements or limited deployment conditions (such as wearable devices, embedded sensing nodes, etc.), traditional models are often difficult to deploy directly, which limits their engineering promotion and universal application feasibility. (3) Limited robustness and environmental adaptability: Gas sensing signals are often affected by a variety of factors, such as gas concentration fluctuations, humidity, temperature changes, and equipment aging, resulting in noise and nonlinear drift in the sensing data. In this context, traditional neural network models are poorly adaptable to outliers, disturbances, or non-ideal working conditions, exhibiting certain robustness defects, which can easily lead to a decrease in recognition accuracy and affect the stable operation of the system. (4) Insufficient generalization ability: Although deep models may show high recognition performance on specific datasets, their recognition performance is often difficult to maintain stability when sensor configurations are different, application environments change, or new gas components are introduced, demonstrating poor cross-scene transfer ability. This problem of insufficient generalization ability is particularly prominent in multi-source heterogeneous sensing systems and real complex environments, which limits the scalability and practicality of the model.
[0004] Therefore, how to build a mixed gas recognition model that is computationally efficient, lightweight, and has low dependence on training samples while ensuring recognition accuracy remains a key technical challenge in this field. Summary of the Invention
[0005] To address the above shortcomings, this invention provides a mixed gas identification method based on a one-dimensional convolutional neural network (1D-CNN). This method fully leverages the structural advantages of 1D-CNN in processing sensor time-series data, eliminating the need for complex data preprocessing or manual feature engineering, and can automatically extract effective features directly from the raw sensor response signal. By constructing a multi-layer convolutional and pooling structure, this method not only enhances the expressive power of gas component features but also possesses good generalization performance and noise resistance. Furthermore, this method supports multi-class prediction of the components of mixed gases and their probability output, contributing to more detailed and interpretable identification results. This approach offers advantages such as lightweight model, efficient training, and accurate identification, providing a promising technical path for the intelligent identification of mixed gases in complex environments.
[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for identifying mixed gases based on a one-dimensional convolutional neural network includes the following steps: S1. Acquire the raw response signals of the sensor array during the acquisition phase, record metadata, and store the data; the metadata includes sampling rate, sensor ID, timestamp, and ambient temperature and humidity; S2. Preprocess the original response signal, including the following steps: (1) Standardize and denoise the original response signal to obtain a multidimensional time series; (2) Extract frequency domain feature information from multidimensional time series based on the time-frequency characteristics of multidimensional time series data; (3) Principal component analysis is used to reduce the dimensionality of high-dimensional features. Principal component features are extracted from the frequency domain feature information in step (2) to obtain the feature sequence. S3. Feature Extraction: The feature sequence processed in step S2 is input into the designed 1D-CNN network to extract deep features; the 1D-CNN network includes an input layer, a convolutional layer, an average pooling layer, a fully connected layer, and an output layer connected in sequence; Dropout operation is added between the convolutional layers; the average pooling layer performs average pooling operation; the classifier is connected to the average pooling layer and the fully connected layer respectively; the network parameters of the 1D-CNN network are optimized using the backpropagation algorithm with the binary logarithmic loss function as the target; S4. The classifier is a multi-label sigmoid output layer, which outputs the probability values of each gas component through the sigmoid activation function.
[0007] Furthermore, in step S1, the acquisition of the raw response signal of the sensor array specifically involves: each gas sample to be tested undergoing a cyclic test according to the acquisition stage; the acquisition stage includes a baseline stage, a target gas exposure stage, and a cleaning and recovery stage; The baseline phase is the period during which the sensor is exposed to clean air or a known reference gas; The target gas exposure phase is the time period during which the sensor is exposed to the target gas to be measured; The cleaning and recovery phase is the period during which the target gas is stopped and clean air or reference gas is introduced again, allowing the sensor to release the adsorbed gas molecules and return to the baseline.
[0008] Furthermore, in step S1, data storage specifically involves: storing the acquired raw signals in real time as a binary file or a structured data file in the form of a time-series matrix; each row of the time-series matrix represents one test sample, and each column represents a sensor feature.
[0009] Furthermore, in step S2, the preprocessing of the original response signal specifically involves normalization and noise suppression relative to the baseline; the normalization relative to the baseline is performed based on the average response of the baseline stage; the noise suppression process employs adaptive moving average filtering.
[0010] Furthermore, in step S2, the frequency domain feature information of the multidimensional time series is extracted using short-time Fourier transform.
[0011] Furthermore, each convolutional layer is followed by a batch normalization and non-linear activation function.
[0012] Furthermore, the loss function uses a bivariate logarithmic loss function to measure the deviation between the predicted probability and the true label for each gas component, and its expression is as follows: ; Where N represents the number of samples; This represents the true label of the i-th sample (or the i-th gas component); This represents the probability value predicted by the model for the i-th sample (or the i-th gas component); the loss function is used to measure the deviation between the predicted probability of each gas component and the true label, and the mean of all loss values is taken as the final loss value.
[0013] Furthermore, in step 3, the network parameters of the 1D-CNN network are optimized using the backpropagation algorithm with the binary logarithmic loss function as the target.
[0014] Furthermore, after step S4, a visualization step is also included, where the 1D-CNN network outputs the probability values of each gas component, and the visualization module is used to display the recognition process and feature map.
[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention aims to propose an efficient mixed gas identification method based on one-dimensional convolutional neural networks (1D-CNN). It provides a solution with good practicality and scalability to address the key problems faced by traditional gas identification models in practical applications, such as high computational complexity, strong dependence on large-scale labeled data, and poor environmental adaptability.
[0016] This method integrates data preprocessing strategies for sensor array signals with a lightweight convolutional neural network structure, significantly reducing model computational load and storage overhead while maintaining recognition accuracy, thus adapting to deployment requirements in resource-constrained scenarios such as embedded systems or edge computing platforms. Furthermore, by constructing a multi-layer feature extraction module, the model can effectively mine temporal and local variation features in gas signals, improving the recognition accuracy and feature representation capability for complex gas mixtures.
[0017] Furthermore, this invention introduces a transfer learning mechanism, leveraging prior knowledge from existing models in similar tasks to efficiently adapt to new tasks, thereby reducing reliance on a large number of labeled samples and improving the model's training efficiency and generalization ability under small sample conditions. Combined with multi-source data augmentation and model regularization techniques, this method further enhances robustness and stability against interference factors such as environmental noise, gas concentration fluctuations, and sensor drift.
[0018] In summary, this invention aims to construct a hybrid gas identification system with low computational cost, strong feature representation capability, good generalization and environmental adaptability, providing a feasible and efficient technical path for the field of intelligent sensing and complex gas detection, and has important engineering application value and research significance. Attached Figure Description
[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below.
[0020] Figure 1 This is a schematic diagram illustrating the working principle of a hybrid deep learning model for a hybrid gas identification method based on a one-dimensional convolutional neural network, according to an embodiment of the present invention.
[0021] Figure 2 This is a flowchart illustrating the calculation of the binary logarithmic loss function for a mixed gas identification method based on a one-dimensional convolutional neural network, according to an embodiment of the present invention. Detailed Implementation
[0022] 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, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0023] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.
[0024] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. Furthermore, the technical features involved in the different embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.
[0025] A method for identifying mixed gases based on a one-dimensional convolutional neural network includes the following steps: S1. Acquire the raw response signals of the sensor array during the acquisition phase, record metadata, and store the data.
[0026] The acquisition of the raw response signal of the sensor array is specifically as follows: each gas sample to be tested undergoes a cyclic test according to the acquisition stage; the acquisition stage includes a baseline stage, a target gas exposure stage, and a cleaning and recovery stage; The baseline phase is the period during which the sensor is exposed to clean air or a known reference gas; The target gas exposure phase is the time period during which the sensor is exposed to the target gas to be measured; The cleaning and recovery phase is the period during which the target gas is stopped and clean air or reference gas is introduced again, allowing the sensor to release the adsorbed gas molecules and return to the baseline.
[0027] Metadata includes sampling rate, sensor ID, timestamp, and ambient temperature and humidity; data storage specifically involves storing the collected raw signals in real time as a binary file or structured data file in the form of a time-series matrix; each row of the time-series matrix represents one test sample, and each column represents a sensor feature.
[0028] S2. To improve data quality and model training efficiency, the original sensor array signals are first standardized and denoised. Preprocessing of the original response signals includes the following steps: (1) The original response signal is standardized and denoised to obtain a multidimensional time series; the original response signal is preprocessed by normalization and noise suppression relative to the baseline; the normalization relative to the baseline is based on the average response of the baseline stage; the noise suppression process adopts adaptive moving average filtering.
[0029] (2) Based on the time-frequency characteristics of multidimensional time series data, frequency domain feature information is extracted from the multidimensional time series; short-time Fourier transform is used to extract frequency domain feature information from the multidimensional time series. Short-time Fourier transform (STFT) is used to extract frequency domain feature information to enhance the model's ability to perceive the differences in the spectra of different gases.
[0030] (3) Use principal component analysis to reduce the dimensionality of high-dimensional features, extract principal component features from the frequency domain feature information in step (2), and obtain feature sequences; use principal component analysis (PCA) to reduce the dimensionality of high-dimensional features, retain the main trend of change, reduce redundancy, and improve data expression efficiency.
[0031] S3. Feature Extraction: The feature sequence processed in step S2 is input into the designed 1D-CNN network to extract deep features; the 1D-CNN network includes an input layer, a convolutional layer, an average pooling layer, a fully connected layer, and an output layer connected in sequence; Dropout operation is added between the convolutional layers; the average pooling layer performs average pooling operation; the classifier is connected to the average pooling layer and the fully connected layer respectively; the network parameters of the 1D-CNN network are optimized using the backpropagation algorithm with the binary logarithmic loss function as the target; batch normalization and non-linear activation functions are connected after each convolutional layer.
[0032] The above design has the following advantages: The preprocessed time series is input into a multi-layer one-dimensional convolutional structure to extract local temporal correlation features. Batch normalization and ReLU are applied after each convolutional layer to improve training stability and network expressive power. To prevent overfitting and improve model robustness, Dropout operation is added between convolutional layers to randomly discard some neuron connections and enhance generalization ability. Average pooling is introduced at the end of the feature extraction stage to compress features and remove redundant information, ensuring the representativeness and continuity of the final output features.
[0033] The loss function uses a bivariate logarithmic loss function to measure the deviation between the predicted probability and the true label for each gas component. Its expression is as follows: ; Where N represents the number of samples; This represents the true label of the i-th sample (or the i-th gas component); This represents the probability value predicted by the model for the i-th sample (or the i-th gas component); the loss function is used to measure the deviation between the predicted probability of each gas component and the true label, and the average of all loss values is taken as the final loss value, thereby ensuring that the model can simultaneously optimize the classification performance of each gas component. Figure 2 The diagram shows the calculation flowchart for the binary logarithmic loss function.
[0034] S4. The classifier is a multi-label sigmoid output layer, which outputs the probability values of each gas component through the sigmoid activation function.
[0035] S5. Visualization Step: The model ultimately outputs the prediction results and probability scores of each gas component. The visualization module can be used to display the recognition process and feature maps.
[0036] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A method for identifying mixed gases based on a one-dimensional convolutional neural network, characterized in that, Includes the following steps: S1. Acquire the raw response signals of the sensor array during the acquisition phase, record metadata, and store the data; the metadata includes sampling rate, sensor ID, timestamp, and ambient temperature and humidity; S2. Preprocess the original response signal, including the following steps: (1) Standardize and denoise the original response signal to obtain a multidimensional time series; (2) Extract frequency domain feature information from multidimensional time series based on the time-frequency characteristics of multidimensional time series data; (3) Principal component analysis is used to reduce the dimensionality of high-dimensional features. Principal component features are extracted from the frequency domain feature information in step (2) to obtain the feature sequence. S3. Feature Extraction: The feature sequence processed in step S2 is input into the designed 1D-CNN network to extract deep features; the 1D-CNN network includes an input layer, a convolutional layer, an average pooling layer, a fully connected layer, and an output layer connected in sequence; Dropout operation is added between the convolutional layers; the average pooling layer performs average pooling operation; the classifier is connected to the average pooling layer and the fully connected layer respectively; the network parameters of the 1D-CNN network are optimized using the backpropagation algorithm with the binary logarithmic loss function as the target; S4. The classifier is a multi-label sigmoid output layer, which outputs the probability values of each gas component through the sigmoid activation function.
2. The method for identifying mixed gases based on a one-dimensional convolutional neural network according to claim 1, characterized in that: In step S1, the acquisition of the raw response signal of the sensor array specifically involves: each gas sample to be tested undergoing a cyclic test according to the acquisition stage; the acquisition stage includes a baseline stage, a target gas exposure stage, and a cleaning and recovery stage; The baseline phase is the period during which the sensor is exposed to clean air or a known reference gas; The target gas exposure phase is the time period during which the sensor is exposed to the target gas to be measured; The cleaning and recovery phase is the period during which the target gas is stopped and clean air or reference gas is introduced again, allowing the sensor to release the adsorbed gas molecules and return to the baseline.
3. The method for identifying mixed gases based on a one-dimensional convolutional neural network according to claim 1, characterized in that: In step S1, data storage specifically involves storing the collected raw signals in real time as a binary file or a structured data file in the form of a time-series matrix; each row of the time-series matrix represents one test sample, and each column represents a sensor feature.
4. The method for identifying mixed gases based on a one-dimensional convolutional neural network according to claim 1, characterized in that: In step S2, the preprocessing of the original corresponding signal specifically involves normalization and noise suppression relative to the baseline. The normalization relative to the baseline is performed by normalizing the average response based on the baseline phase; Noise suppression is achieved using adaptive moving average filtering.
5. The method for identifying mixed gases based on a one-dimensional convolutional neural network according to claim 1, characterized in that: In step S2, short-time Fourier transform is used to extract frequency domain feature information from the multidimensional time series.
6. The method for identifying mixed gases based on a one-dimensional convolutional neural network according to claim 1, characterized in that: Each convolutional layer is followed by a batch normalization and non-linear activation function.
7. The method for identifying mixed gases based on a one-dimensional convolutional neural network according to claim 1, characterized in that: The loss function uses a bivariate logarithmic loss function to measure the deviation between the predicted probability and the true label for each gas component. Its expression is as follows: ； Where N represents the number of samples; This represents the true label of the i-th sample or the i-th gas component; This represents the probability value predicted by the model for the i-th sample or the i-th gas component; the loss function is used to measure the deviation between the predicted probability of each gas component and the true label, and the mean of all loss values is taken as the final loss value.
8. The method for identifying mixed gases based on a one-dimensional convolutional neural network according to claim 1, characterized in that: In step 3, the network parameters of the 1D-CNN network are optimized using the backpropagation algorithm with the binary logarithmic loss function as the objective.
9. The method for identifying mixed gases based on a one-dimensional convolutional neural network according to claim 1, characterized in that: Step S4 is followed by a visualization step, where the 1D-CNN network outputs the probability values of each gas component, and the visualization module displays the recognition process and feature map.