Gas sensor low latency classification method based on sliding window and confidence accumulation

Abstract

This invention provides a low-latency classification method for gas sensors based on sliding window and confidence accumulation, relating to the fields of gas sensing and intelligent diagnostics. Addressing the high detection latency caused by existing temperature-modulated gas sensors relying on complete-cycle signals, this invention proposes a SWSD-CAD framework. This framework employs a small-window sliding detection (SWSD) strategy to perform high-frequency online analysis of real-time data streams. Simultaneously, it arbitrates continuous prediction results through confidence accumulation decision CAD logic, making a final decision only when the prediction result continuously meets a confidence threshold within a preset time, ensuring the reliability of the results. To implement this framework, this invention designs a ConvMamba hybrid network model for efficiently extracting discriminative features within short signal windows. This invention can significantly reduce detection latency while maintaining high classification accuracy, achieving early, rapid, and reliable identification of gas types.

Description

Technical Field

[0001] This invention relates to the field of gas sensing and intelligent diagnostics, and in particular to a low-latency classification method for gas sensors based on sliding windows and confidence accumulation. Background Technology

[0002] Semiconductor gas sensing: Temperature modulation strategies (TMS) can significantly improve the selectivity of a single sensor. This strategy generates unique dynamic response curves, or "gas fingerprints," for different gases by dynamically changing the sensor's operating temperature. However, the traditional TMS analysis paradigm has an inherent drawback: it requires waiting for at least one complete temperature modulation cycle to finish before the acquired signal can be analyzed and a decision made. This "acquire first, analyze later" offline mode results in significant detection delays, greatly limiting its use in applications requiring high response speeds, such as rapid early warning and real-time monitoring.

[0003] Existing improvement methods primarily focus on hardware or signal generation, such as designing special heating pulses or optimizing heating waveforms to accelerate the sensor's physical response process. While these methods have some effect, they essentially remain within the traditional framework of "relying on the complete effective cycle," with the lower limit of their detection latency constrained by the physical response time. They lack fundamental innovation at the data processing and decision-making algorithm levels. Therefore, a novel solution driven by data and algorithmic intelligence is urgently needed to overcome the latency bottleneck of traditional TMS. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a low-latency classification method for gas sensors based on sliding windows and confidence accumulation; overcoming the defects of high detection latency and inability to meet the requirements of rapid response in temperature-modulated gas sensors in existing technologies.

[0005] On one hand, the present invention provides a low-latency classification method for gas sensors based on sliding window and confidence accumulation, comprising the following steps:

[0006] Step 1: Acquire the dynamic response signal data stream of the gas sensor under temperature modulation in real time. Use a fixed-size sliding window with a size of 30% of the modulation period to continuously segment the dynamic response signal data stream at high frequency. At each time step, generate a signal window containing the latest signal segment. ;

[0007] Step 2: Design a ConvMamba hybrid network model; at each time step, input the current signal window into the pre-trained ConvMamba hybrid network model;

[0008] The ConvMamba hybrid network model includes a one-dimensional convolutional layer, a Mamba state space model layer, and a fully connected classification layer. The one-dimensional convolutional layer is used to extract local pattern features of the signal window. The Mamba state space model layer is connected after the one-dimensional convolutional layer and is used to model the long-range temporal dependencies of the local pattern features and output a hidden state vector. The fully connected classification layer is used to map the hidden state vector to the instantaneous probability prediction distribution. A linear projection layer is also included between the one-dimensional convolutional layer and the Mamba state space model layer to project the local pattern features into the dimensional space required by the Mamba state space model layer.

[0009] Step 2.1: First, capture local morphological features and key pattern elements within the signal window using a multi-channel one-dimensional convolutional CNN, as shown in the following formula:

[0010] ;

[0011] in It is the convolution kernel of the i-th channel. yes Perform a convolution operation with a convolution kernel. It is the signal after convolution.

[0012] The signals after multi-channel convolution are combined into a local feature sequence Z(t);

[0013] Step 2.2: Input the local feature sequence Z(t) into the input layer of the ConvMamba hybrid network model; obtain the evolution law of the local feature sequence within the time window, and finally output the hidden state vector. , Let h(t) be the hidden state vector at the last time step in the small window sampling signal. The formula for the hidden state vector h(t) is as follows:

[0014] ;

[0015] ;

[0016] Where A, B, C, and D are learnable parameter matrices. It is the hidden state vector. Rate of change of state This is the output.

[0017] Step 2.3: Hidden State Vector After passing through a fully connected classification layer and a Softmax function, an instantaneous probability prediction distribution is generated, representing the likelihood that the current window belongs to each gas category; the formula for the fully connected classification layer is as follows:

[0018] ;

[0019] The formula for the Softmax function is as follows:

[0020] ;

[0021] in It's weight. It's a bias. This is the output of the ConvMamba hybrid network model, specifically the score for each gas species. It is a score set for gas types. It refers to the types and quantities of gases. It is the probability of the i-th type of gas, where i and j represent the i-th and j-th types of gas, respectively;

[0022] Step 3: Introduce CAD logic: through two key hyperparameters, namely the confidence threshold. and consecutively satisfy the number of steps The high-frequency prediction sequences generated by the SWSD strategy are examined; the specific steps are as follows:

[0023] Step 3.1: Initialize a cumulative counter (counter) to 0 and a candidate category. Empty;

[0024] Step 3.2: Probability distribution p of the i-th type of gas at each time step t t In the middle, obtain the maximum probability value and its corresponding prediction category The confidence output selects the value with the highest probability.

[0025] Step 3.3: Set hyperparameters: confidence threshold and consecutively satisfy the number of steps ;

[0026] Step 3.4: Check if both of the following conditions are met simultaneously: and ;

[0027] If the condition is met, the cumulative counter is incremented by 1; otherwise, the counter is reset to 0, and... Update to the current .

[0028] Step 3.5: If and only if the counter reaches At that time, the final decision will be confirmed. The final identification result is output, and the current time step t is recorded as the decision time point.

[0029] On the other hand, the present invention also provides a low-latency classification system for gas sensors based on sliding window and confidence accumulation, for implementing the aforementioned low-latency classification method for gas sensors based on sliding window and confidence accumulation, including a data processing module, a data inference module, and a decision arbitration module;

[0030] The data processing module acquires the dynamic response signal data stream of the gas sensor in real time and continuously segments it using a sliding window of fixed size, generating a signal window at each time step.

[0031] The model inference module contains a ConvMamba hybrid network model, which receives a signal window at each time step and outputs an instantaneous probability prediction distribution.

[0032] The decision arbitration module performs continuous arbitration on the instant probability prediction distribution. When the prediction confidence of a certain gas category is not lower than the preset confidence threshold within a continuous preset number of time steps, the gas category is output as the final classification result.

[0033] Thirdly, this application proposes a computer-readable storage medium storing executable instructions that, when executed, cause a processor to perform the described low-latency classification method for gas sensors based on sliding windows and confidence accumulation.

[0034] Fourthly, this application proposes a computer program product, including a computer program or instructions that, when executed by a processor, implement the aforementioned low-latency classification method for gas sensors based on sliding windows and confidence accumulation.

[0035] The beneficial effects of adopting the above technical solution are as follows:

[0036] This invention provides a low-latency classification method for gas sensors based on sliding window and confidence accumulation, significantly reducing detection latency: Through the SWSD strategy, the model does not need to wait for the entire cycle to end; once a sufficiently discriminative "golden feature" is captured within the sliding window, the decision-making process can begin. Combined with CAD logic, the decision point can be significantly advanced from the end of the cycle while ensuring reliability. Experiments show that decision latency can be reduced by nearly 40% in transitional response; achieving high-reliability decision-making: The CAD logic effectively avoids misjudgments caused by early signal instability by imposing dual constraints on prediction confidence and temporal continuity, achieving a balance between speed and reliability; flexible and adjustable performance trade-offs: users can adjust the confidence threshold (… ) and the number of consecutive steps ( Two parameters allow for flexible trade-offs between decision accuracy, decision speed, and decision rate based on actual application needs; efficient feature extraction capability: The ConvMamba model, designed specifically for short-time-series signals, combines the local perception capability of CNNs with the long-range modeling advantages of Mamba, enabling it to efficiently extract highly discriminative features from incomplete signal segments. Attached Figure Description

[0037] Figure 1 This is a flowchart of the low-latency classification method for gas sensors in an embodiment of the present invention;

[0038] Figure 2 This is a comparison of the gas sensor response curve in the embodiments of the present invention and SWSD-CAD with traditional detection methods;

[0039] Figure 3 This is the training framework for the ConvMamba hybrid network model proposed in this embodiment of the invention;

[0040] Figure 4 This is the CAD decision logic framework in the embodiments of the present invention. Detailed Implementation

[0041] The specific implementation methods of this application will be further described in detail below with reference to the accompanying drawings and embodiments.

[0042] Example 1:

[0043] On the one hand, this invention provides a low-latency classification method for gas sensors based on sliding windows and confidence accumulation, such as... Figure 1 As shown, it includes the following steps:

[0044] Step 1: Real-time acquisition of the dynamic response signal data stream of the gas sensor under temperature modulation. Unlike acquiring a complete cycle, this invention uses a fixed-size sliding window with a size of 30% of the modulation period to continuously segment the dynamic response signal data stream at high frequency; at each time step, a signal window containing the latest signal segment is generated. ;

[0045] Step 2: To efficiently process short signal windows, this invention designs a ConvMamba hybrid network model; at each time step, the current signal window is input into the pre-trained ConvMamba hybrid network model;

[0046] The ConvMamba hybrid network model includes a one-dimensional convolutional layer, a Mamba state space model layer, and a fully connected classification layer. The one-dimensional convolutional layer is used to extract local pattern features of the signal window. The Mamba state space model layer is connected after the one-dimensional convolutional layer and is used to model the long-range temporal dependencies of the local pattern features and output a hidden state vector. The fully connected classification layer is used to map the hidden state vector to the instantaneous probability prediction distribution. A linear projection layer is also included between the one-dimensional convolutional layer and the Mamba state space model layer to project the local pattern features into the dimensional space required by the Mamba state space model layer.

[0047] Step 2.1: First, capture local morphological features and key pattern elements within the signal window using a multi-channel one-dimensional convolutional CNN, as shown in the following formula:

[0048] ;

[0049] in It is the convolution kernel of the i-th channel. yes Perform a convolution operation with a convolution kernel. It is the signal after convolution.

[0050] The signals after multi-channel convolution are combined into a local feature sequence Z(t);

[0051] Step 2.2: Input the local feature sequence Z(t) into the input layer of the ConvMamba hybrid network model; Mamba utilizes its powerful long-range dependency modeling capabilities to obtain the evolution law of the local feature sequence within the time window, and finally outputs a hidden state vector that summarizes the window information. , Let h(t) be the hidden state vector at the last time step in the small window sampling signal. The formula for the hidden state vector h(t) is as follows:

[0052] ;

[0053] ;

[0054] Where A, B, C, and D are learnable parameter matrices. It is the hidden state vector. Rate of change of state This is the output.

[0055] Step 2.3: Hidden State Vector After passing through a fully connected classification layer and a Softmax function, an instantaneous probability prediction distribution is generated, representing the likelihood that the current window belongs to each gas category; the formula for the fully connected classification layer is as follows:

[0056] ;

[0057] The formula for the Softmax function is as follows:

[0058] ;

[0059] in It's weight. These are biases, and both are learnable parameter matrices. This is the output of the ConvMamba hybrid network model, specifically the score for each gas species. It is a score set for gas types. It refers to the types and quantities of gases. It is the probability of the i-th type of gas, where i and j represent the i-th and j-th types of gas, respectively;

[0060] Step 3: Introduce CAD logic as a "time decision arbitrator"; such as Figure 4 As shown, the CAD logic uses two key hyperparameters, namely the confidence threshold. and consecutively satisfy the number of steps The high-frequency prediction sequences generated by the SWSD strategy are examined; the specific steps are as follows:

[0061] Step 3.1: Initialize a cumulative counter (counter) to 0 and a candidate category. Empty;

[0062] Step 3.2: Probability distribution p of the i-th type of gas at each time step t t In the middle, obtain the maximum probability value and its corresponding prediction category The confidence output selects the value with the highest probability.

[0063] Step 3.3: Set hyperparameters: confidence threshold and consecutively satisfy the number of steps ;

[0064] Step 3.4: Check if both of the following conditions are met simultaneously: and ;

[0065] If the condition is met, the cumulative counter is incremented by 1; otherwise, the counter is reset to 0, and... Update to the current .

[0066] Step 3.5: If and only if the counter reaches At that time, the final decision will be confirmed. The final identification result is output, and the current time step t is recorded as the decision time point.

[0067] On the other hand, the present invention also provides a low-latency classification system for gas sensors based on sliding window and confidence accumulation, for implementing the aforementioned low-latency classification method for gas sensors based on sliding window and confidence accumulation, including a data processing module, a data inference module, and a decision arbitration module;

[0068] The data processing module acquires the dynamic response signal data stream of the gas sensor in real time and continuously segments it using a sliding window of fixed size, generating a signal window at each time step.

[0069] The model inference module contains a ConvMamba hybrid network model, which receives a signal window at each time step and outputs an instantaneous probability prediction distribution.

[0070] The decision arbitration module performs continuous arbitration on the instant probability prediction distribution. When the prediction confidence of a certain gas category is not lower than the preset confidence threshold within a continuous preset number of time steps, the gas category is output as the final classification result.

[0071] Example 2:

[0072] This embodiment uses a semiconductor gas sensor as an example to classify five gases: hydrogen, methane, carbon monoxide, ethanol, and butanone, and elaborates on the low-delay gas classification method proposed in this invention.

[0073] Step 1: A periodically varying heating voltage (temperature modulation) is applied to the sensor using a programmable DC power supply to collect the sensor's detection signal. Specific voltage parameters are shown in Table 1. Comparison of sensor response signal state, traditional detection methods, and the SWSD-CAD method is also presented. Figure 2 As shown.

[0074] Table 1 Temperature modulation waveform parameters:

[0075]

[0076] Step 2: The core of this invention lies in the SWSD-CAD (Sliding) framework, which outputs the probability of each concentration based on the gas signal, such as... Figure 3 As shown. The specific steps are as follows:

[0077] In this embodiment, the data was collected over 400 time steps starting from the moment the gas appeared, with a window length L of 30. Two convolutional layers were used for feature extraction, with a kernel size of 5 and a total of 32 channels.

[0078] Step 2.1: At any time step t, obtain a signal window of fixed length L from the real-time data stream. .

[0079] Step 2.2: The signal window first passes through two cascaded one-dimensional convolutional layers (Conv1d). For example, the first layer maps a 1-channel input to a 32-channel input, and the second layer performs further processing.

[0080] In this embodiment, Mamba has 128 dimensions.

[0081] Step 2.3: Feature map output by the convolution module Through a linear projection layer, its dimensions are mapped from 32 to the dimensions required by the Mamba model. (For example, 128 dimensions). The formula is as follows:

[0082] .

[0083] Step 2.4: Projected Sequence It was sent to the Mamba backbone network.

[0084] Step 2.5: After processing the entire window sequence, the Mamba network extracts the hidden state vector of the last time step. This vector is considered a final summary of the entire window's information. After layer normalization, Dropout, and a fully connected classification layer, the final logits vector is calculated. Then, the probability distribution of various gases is converted using the Softmax function. .

[0085] Step 3: For the online diagnostic method in this invention, the process in Step 2 is performed at every time step. At each time step t, the system captures the latest signal window. This is then fed into the ConvMamba model to generate an instantaneous probability prediction. This strategy ensures that the data pattern during inference is completely consistent with that during training, thus guaranteeing model performance. To filter out unstable predictions of transitional signals, a CAD (Autonomous Data Acquisition) decision logic is introduced. The specific steps are as follows:

[0086] Step 3.1: Initialize a cumulative counter (counter) to 0 and a candidate category. Empty;

[0087] Step 3.2: At each time step t, calculate the probability distribution from the model output. In the middle, obtain the maximum probability value (Confidence level) and its corresponding prediction category ;

[0088] Step 3.3: Set hyperparameters: confidence threshold and consecutively satisfy the number of steps .

[0089] Step 3.4: Check if both conditions are met simultaneously: and .

[0090] Step 3.5: If the condition is met, increment the counter by 1. If not, reset the counter to 0 and... Update to the current .

[0091] Step 3.6: If and only if the counter reaches At that time, the system finally confirms the decision and will The final identification result is output, and the current time step t is recorded as the decision time point.

[0092] Step 4: A lightweight Gas Initial Sensing (GOP) algorithm was designed as a "wake-up" trigger. It determines whether there is a significant signal change by calculating the Euclidean distance between the current signal and the stored clean air baseline signal in real time. The SWSD-CAD framework is only officially activated when this distance consistently exceeds a dynamically set threshold, thus achieving an optimal balance between efficiency and power consumption.

[0093] The evaluation metrics in this embodiment include decision rate, decision accuracy, and average decision delay. The experimental results are shown in Table 2. By adjusting... and The parameters can be flexibly balanced among performance metrics. For example, when pursuing high accuracy (>95%) parameter settings, the SWSD-CAD (Sliding) framework proposed in this invention can make effective decisions.

[0094] Table 2 Gas category prediction results using the SWSD-CAD (Sliding) framework:

[0095]

[0096] Example 3:

[0097] This embodiment proposes a computer-readable storage medium that stores executable instructions. When these instructions are executed, if they are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0098] The computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the low-latency classification method for gas sensors based on sliding window and confidence accumulation described in various embodiments of this application.

[0099] The aforementioned storage media include: flash memory, hard disk, multimedia card, card-type memory (e.g., SD (Secure Digital Memory Card) or DX (Memory Data Register, MDR) memory, random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, server, APP (Application) application store, and other media capable of storing program verification codes, on which computer programs are stored. When the computer program is executed by the processor, it can implement the various steps of the above-mentioned low-latency classification method for gas sensors based on sliding window and confidence accumulation.

[0100] Example 4:

[0101] This embodiment proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the aforementioned low-latency classification method for gas sensors based on sliding windows and confidence accumulation.

[0102] Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a computer program product.

[0103] The various embodiments in this application are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0104] The scope of protection of this application is not limited to the embodiments described above. Obviously, those skilled in the art can make various modifications and variations to this disclosure without departing from the scope and spirit of this disclosure. If such modifications and variations fall within the scope of the methods disclosed herein and their equivalents, then the intent of this disclosure also includes such modifications and variations.

Claims

1. A low-latency classification method for gas sensors based on sliding window and confidence accumulation, characterized in that, Includes the following steps: Step 1: Real-time acquisition of the dynamic response signal data stream of the gas sensor under temperature modulation; Step 2: Design a ConvMamba hybrid network model; at each time step, input the current signal window into the pre-trained ConvMamba hybrid network model; The ConvMamba hybrid network model includes a one-dimensional convolutional layer, a Mamba state space model layer, and a fully connected classification layer. The one-dimensional convolutional layer is used to extract local pattern features of the signal window. The Mamba state space model layer is connected after the one-dimensional convolutional layer and is used to model the long-range temporal dependencies of the local pattern features and output a hidden state vector. The fully connected classification layer is used to map the hidden state vector to an instantaneous probability prediction distribution. A linear projection layer is also included between the one-dimensional convolutional layer and the Mamba state space model layer to project the local pattern features into the dimensional space required by the Mamba state space model layer. Step 3: Introduce CAD logic: through two key hyperparameters, namely the confidence threshold. and consecutively satisfy the number of steps Examine high-frequency prediction sequences; Step 3.1: Initialize a cumulative counter (counter) to 0 and a candidate category. Empty; Step 3.2: Probability distribution p of the i-th type of gas at each time step t t In the middle, obtain the maximum probability value and its corresponding prediction category The confidence output selects the value with the highest probability. Step 3.3: Set hyperparameters: confidence threshold and consecutively satisfy the number of steps ; Step 3.4: Check if both of the following conditions are met simultaneously: and ; If the condition is met, the cumulative counter is incremented by 1; otherwise, the counter is reset to 0, and... Update to the current ; Step 3.5: If and only if the counter reaches At that time, the final decision will be confirmed. The final identification result is output, and the current time step t is recorded as the decision time point.

2. The low-latency classification method for gas sensors based on sliding window and confidence accumulation according to claim 1, characterized in that, Step 1 specifically employs a fixed-size sliding window with a size of 30% of the modulation period to perform high-frequency continuous segmentation of the dynamic response signal data stream; at each time step, a signal window containing the latest signal segment is generated. .

3. The low-latency classification method for gas sensors based on sliding window and confidence accumulation according to claim 2, characterized in that, Step 2 includes the following steps: Step 2.1: First, capture local morphological features and key pattern elements within the signal window using a multi-channel one-dimensional convolutional CNN, as shown in the following formula: ； in It is the convolution kernel of the i-th channel. yes Perform a convolution operation with a convolution kernel. It is the signal after convolution; The signals after multi-channel convolution are combined into a local feature sequence Z(t); Step 2.2: Input the local feature sequence Z(t) into the input layer of the ConvMamba hybrid network model; obtain the evolution law of the local feature sequence within the time window, and finally output the hidden state vector. , Let h(t) be the hidden state vector at the last time step in the small window sampling signal. The formula for the hidden state vector h(t) is as follows: ； ； Where A, B, C, and D are learnable parameter matrices. It is the hidden state vector. Rate of change of state For output; Step 2.3: Hidden State Vector After passing through a fully connected classification layer and a Softmax function, an instantaneous probability prediction distribution is generated, representing the likelihood that the current window belongs to each gas category; the formula for the fully connected classification layer is as follows: ； The formula for the Softmax function is as follows: ； in It's weight. It's a bias. This is the output of the ConvMamba hybrid network model, specifically the score for each gas species. It is a score set for gas types. It refers to the types and quantities of gases. It is the probability of the i-th type of gas, where i and j represent the i-th and j-th types of gas, respectively.

4. The low-latency classification method for gas sensors based on sliding window and confidence accumulation as described in claim 1 is implemented through a low-latency classification system for gas sensors based on sliding window and confidence accumulation, characterized in that... It includes a data processing module, a data reasoning module, and a decision arbitration module; The data processing module acquires the dynamic response signal data stream of the gas sensor in real time and continuously segments it using a sliding window of fixed size, generating a signal window at each time step. The model inference module contains a ConvMamba hybrid network model, which receives a signal window at each time step and outputs an instantaneous probability prediction distribution. The decision arbitration module performs continuous arbitration on the instant probability prediction distribution. When the prediction confidence of a certain gas category is not lower than the preset confidence threshold within a continuous preset number of time steps, the gas category is output as the final classification result.

5. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed, cause the processor to perform the low-latency classification method for gas sensors based on sliding window and confidence accumulation as described in any one of claims 1-4.

6. A computer program product, characterized in that, Includes a computer program or instructions that, when executed by a processor, implement the low-latency classification method for gas sensors based on sliding window and confidence accumulation as described in any one of claims 1-4.

Images