Multi-gas real-time decoupling system based on AI drive

Abstract

The invention relates to the field of mixed gas decoupling, in particular to a multi-gas real-time decoupling system based on AI driving. According to the system, firstly, wavelet transformation is performed on a gas detection signal collected by a gas sensor and an environment detection signal collected by an environment sensor, a recurrent neural network is combined, a hidden state vector of each time point in a preset historical time period is obtained, and a hidden state vector of each time point in the preset historical time period is obtained by utilizing an attention mechanism; obtaining an attention score of each preset gas at each time point in the preset historical period, and obtaining a comprehensive feature vector of each preset gas according to the attention score of each preset gas at each time point in the preset historical period and the hidden state vector of each time point, and inputting the comprehensive feature vector of each preset gas into a full-connection neural network layer to obtain the real concentration of each preset gas. The problem of cross infection in the gas decoupling process can be avoided, and the decoupling precision is guaranteed when gas types are increased and environmental parameters fluctuate violently.

Description

Technical Field

[0001] This invention relates to the field of gas decoupling, and more specifically to an AI-driven real-time multi-gas decoupling system. Background Technology

[0002] In industrial production, environmental monitoring, and medical diagnosis, it is often necessary to detect multiple gas components simultaneously. The significance of multi-gas decoupling lies in its ability to separate or analyze complex and interfering gas mixtures into independent, accurately identifiable, and quantifiable single components. This allows for clear identification of the concentrations of various gas components, providing reliable data for environmental monitoring, chemical control, and other applications.

[0003] Existing methods typically involve deploying multi-sensor arrays and combining them with mathematical methods such as principal component analysis or least squares to process mixed gas signals, thereby decoupling the concentrations of various gases in the mixed gas. However, since a certain sensor responds to multiple gases, there is cross-interference between the gases, resulting in highly nonlinear responses of traditional sensor arrays or spectroscopic techniques to mixed gases. Existing decoupling algorithms (such as the classic PLS, PCA, and some shallow machine learning models) experience a sharp decline in decoupling accuracy when the number of gas types increases or environmental parameters (temperature and humidity) fluctuate drastically. Summary of the Invention

[0004] To address the technical problem that existing decoupling algorithms experience a sharp decline in decoupling accuracy when the number of gas types increases and environmental parameters fluctuate drastically, this invention aims to provide an AI-driven real-time multi-gas decoupling system. The specific technical solution adopted is as follows: This invention also proposes an AI-driven multi-gas real-time decoupling system, the system comprising: The data acquisition module is used to acquire gas detection signals of mixed gas from different types of gas sensors, and environmental detection signals of the environment from different types of environmental sensors. The data feature analysis module is used to perform wavelet transform processing on the gas detection signals of each gas sensor and the environmental detection signals of each environmental sensor within a preset historical period, and combine it with a recurrent neural network to obtain the hidden state vector of each time point in the preset historical period. The attention analysis module is used to process the hidden state vectors of all time points in a preset historical period using an attention mechanism to obtain the attention score of each preset gas at each time point in the preset historical period; and to obtain the comprehensive feature vector of each preset gas based on the attention score of each preset gas at each time point in the preset historical period and the hidden state vector at each time point. The gas decoupling module is used to input the comprehensive feature vector of each preset gas into a fully connected neural network layer and output the true concentration of each preset gas.

[0005] Furthermore, obtaining the hidden state vector for each point in time within a preset historical period includes: Wavelet transform processing is performed on the gas detection signals of each gas sensor and the environmental detection signals of each environmental sensor within a preset historical time period to obtain the time-frequency diagrams of the gas detection signals of each gas sensor and the environmental detection signals of each environmental sensor. The gas detection signals from each gas sensor and the environmental detection signals from each environmental sensor are input into a recurrent neural network, which then outputs the hidden state vector for each time point in a preset historical period.

[0006] Furthermore, the wavelet basis used in the wavelet transform is the Daubechies wavelet, and the recurrent neural network is a gated recurrent unit network.

[0007] Furthermore, obtaining the attention score for each preset gas at each time point within a preset historical period includes: A multi-head attention mechanism is adopted, with an attention head set for each preset gas. The hidden state vectors of all time points in the preset historical period are input into each attention head, and each attention head outputs the attention score of the corresponding preset gas at each time point in the preset historical period.

[0008] Furthermore, obtaining the comprehensive feature vector for each preset gas includes: Use any preset gas as the target preset gas; By using the attention scores of the target preset gas at each time point within a preset historical period, the hidden state vectors at each time point are weighted and summed to obtain the comprehensive feature vector of the target preset gas.

[0009] Furthermore, the output of the actual concentration of each preset gas includes: In the fully connected neural network layer, an independent neuron is set for each preset gas, and the comprehensive feature vector of each preset gas is input into the corresponding neuron, thereby outputting the concentration value of each preset gas.

[0010] Furthermore, the fully connected neural network layer is trained using a mixed gas sample containing each preset gas.

[0011] Furthermore, the real-time decoupling system is deployed on an FPGA chip.

[0012] Furthermore, the gas sensor includes at least an infrared sensor, an electrochemical sensor, and a catalytic combustion sensor.

[0013] Furthermore, the environmental sensor includes at least a temperature sensor and a humidity sensor.

[0014] The present invention has the following beneficial effects: This invention addresses the issue that existing decoupling algorithms suffer a sharp decline in accuracy when the number of gas types increases or environmental parameters fluctuate drastically, as some sensors respond to multiple gases. Therefore, this invention first acquires gas detection signals from different types of gas sensors for mixed gases. These different types of sensors have varying sensitivities to different gases, providing complementary detection signals for subsequent data processing and reducing cross-interference between different gas detection signals. Simultaneously, different types of environmental sensors are incorporated into the decoupling process to detect environmental conditions, preventing sudden environmental changes from affecting the decoupling of the mixed gas. This gives the system strong environmental adaptability, maintaining accuracy under different climatic conditions and meeting the needs of complex scenarios ranging from industrial sites to medical rooms. Then, this invention performs small-scale processing on the gas detection signals and environmental detection signals. Waveform processing extracts the time-frequency features of the signal, and a recurrent neural network is used to process the temporal features of each detected signal to accurately capture the long-term dependencies of various detected signals. Finally, a hidden state vector is generated for each time point of the preset historical period. The hidden state vector of each time point encodes all historical information up to that time point. At the same time, considering that the importance of gas detection signals from different gas sensors varies at different time points for each preset gas, the attention score is obtained to reflect the degree of attention received by each preset gas at each time point of the preset historical period. The obtained comprehensive feature vector is then input into a fully connected neural network layer to output the true concentration of each preset gas, thereby eliminating the cross-interference problem caused by the increase in gas types and the drastic fluctuation of environmental parameters, and improving the gas decoupling accuracy. Attached Figure Description

[0015] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a block diagram of an AI-driven real-time decoupling system for multiple gases, provided as an embodiment of the present invention.

[0017] Figure 2The flowchart illustrates a method for obtaining the hidden state vector at each point in time during a preset historical period, as provided in an embodiment of the present invention. Detailed Implementation

[0018] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of an AI-driven real-time decoupling system proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0020] The following description, in conjunction with the accompanying drawings, details a specific solution for an AI-driven real-time decoupling system for multiple gases provided by this invention.

[0021] Please see Figure 1 The diagram illustrates a block diagram of an AI-driven real-time decoupling system for multiple gases according to an embodiment of the present invention. The system includes a data acquisition module 101, a data feature analysis module 102, an attention analysis module 103, and a gas decoupling module 104.

[0022] The data acquisition module 101 is used to acquire gas detection signals of mixed gas from different types of gas sensors, and at the same time acquire environmental detection signals of the environment from different types of environmental sensors.

[0023] To achieve precise decoupling of the gas mixture and reduce cross-interference during the decoupling process, this invention first integrates different types of gas sensors into a micro-chamber. This ensures that the gas mixture to be measured flows uniformly and rapidly through all sensing units, guaranteeing data synchronization. The gas detection signals of the gas mixture are then acquired in real time using different types of gas sensors. The selection of different types of gas sensors follows a complementary principle. In one embodiment of this invention, the gas sensors include at least three types: infrared sensors, electrochemical sensors, and catalytic combustion sensors. The infrared sensors can provide information about gas molecules. The spectroscopy of gas sensors exhibits high specificity for gases such as CO2 and CH4. Electrochemical sensors demonstrate high selectivity and sensitivity for specific toxic gases (such as CO, H2S, and NO2). Catalytic combustion sensors are relatively sensitive to combustible gases (such as CH4, H2, and alkanes) and respond rapidly. These three types of gas sensors have different sensitivities to different gases, thus providing complementary detection signals for subsequent data processing, reducing cross-interference between different gas detection signals, and improving the decoupling accuracy of gas concentration. In other embodiments of this invention, other types of gas sensors may be selected or added to acquire gas detection signals, which is not limited here.

[0024] Meanwhile, considering that existing decoupling algorithms are difficult to apply to complex scenarios with varying environmental conditions, this invention also integrates different types of environmental sensors into a micro gas chamber and uses these sensors to synchronously collect environmental detection signals. This avoids the impact of sudden environmental changes on the decoupling of the mixed gas, giving the system strong environmental adaptability and maintaining accuracy under different climatic conditions. This meets the needs of complex scenarios from industrial sites to medical rooms. In one embodiment of this invention, the different types of environmental sensors include at least a temperature sensor and a humidity sensor. The temperature sensor collects temperature signals, and the humidity sensor collects humidity signals. In other embodiments of this invention, other types of environmental sensors can be selected or added to collect environmental detection signals, which is not limited here.

[0025] The data feature analysis module 102 is used to perform wavelet transform processing on the gas detection signals of each gas sensor and the environmental detection signals of each environmental sensor within a preset historical time period, and combine them with a recurrent neural network to obtain the hidden state vector of each time point in the preset historical time period.

[0026] Because existing decoupling algorithms suffer from a sharp decline in decoupling accuracy when the number of gas types increases and environmental parameters fluctuate drastically, this invention first performs wavelet transform processing on the gas detection signals of each gas sensor and the environmental detection signals of each environmental sensor within a preset historical time period. This extracts the time-frequency features of the gas and environmental detection signals within the preset historical time period. A recurrent neural network is then used to process these time-frequency features, thereby accurately capturing the long-term dependencies between various gas and environmental detection signals. Finally, a hidden state vector is generated for each time point within the preset historical time period. The hidden state vector at each time point encodes all historical information up to that time point. Subsequently, the concentration of the mixed gas can be accurately decoupled based on the hidden state vector at each time point. The preset historical time period refers to a period of time prior to the current time point. The length of the preset historical time period typically ranges from 1 to 5 seconds. In one embodiment of this invention, the length of the preset historical time period is set to 2 seconds, meaning that wavelet transform processing is performed on the gas and environmental detection signals within the most recent two seconds. The specific value of the preset historical time period length can also be set by the implementer according to the specific implementation scenario and is not limited here.

[0027] Preferably, in one embodiment of the present invention, the method for obtaining the hidden state vector at each time point of a preset historical period specifically includes: Please see Figure 2 The flowchart illustrates a method for obtaining the hidden state vector at each point in time during a preset historical period, according to an embodiment of the present invention.

[0028] Step S1: Perform wavelet transform processing on the gas detection signals of each gas sensor and the environmental detection signals of each environmental sensor within the preset historical time period to obtain the time-frequency diagrams of the gas detection signals of each gas sensor and the environmental detection signals of each environmental sensor.

[0029] Among them, the method of extracting the time-frequency graph of the signal using wavelet transform has better time-frequency localization capability than the traditional short-time Fourier transform. It can accurately capture the transient and non-stationary signal characteristics generated during gas adsorption or desorption, and can perfectly capture the dynamic characteristics generated when the gas interacts with the sensor, such as response speed, peak value, recovery slope, etc., so that it can observe detailed information that cannot be seen by traditional methods. Wavelet transform is a well-known technique in the art and will not be described in detail here. In one embodiment of the present invention, Daubechies wavelet is selected as the wavelet basis used for wavelet transform because Daubechies wavelet has tight support, can effectively analyze non-stationary signals such as gas detection signals and environmental detection signals, is suitable for capturing transient features such as rising and falling edges of gas and environmental sensor responses, and Daubechies wavelet has good approximate symmetry, small phase distortion, and can effectively smooth high-frequency noise in the signal, improving the accuracy of gas mixture decoupling.

[0030] Step S2: Input the time-frequency graphs of the gas detection signals from each gas sensor and the environmental detection signals from each environmental sensor into the recurrent neural network, and output the hidden state vector of each time point in the preset historical period from the recurrent neural network.

[0031] In one embodiment of the present invention, the selected recurrent neural network is a gated recurrent unit network (GRU). Compared with recurrent neural networks such as long short-term memory networks (LSTM), the gated recurrent unit network has a simpler structure, fewer parameters, and is generally more efficient in computation, thereby ensuring the timeliness of gas mixture decoupling.

[0032] The attention analysis module 103 is used to process the hidden state vectors of all time points in a preset historical period using an attention mechanism to obtain the attention score of each preset gas at each time point in the preset historical period; and to obtain the comprehensive feature vector of each preset gas based on the attention score of each preset gas at each time point in the preset historical period and the hidden state vector at each time point.

[0033] Meanwhile, considering that the importance of the gas detection signals from different gas sensors varies at different time points for each preset gas, this embodiment of the invention utilizes an attention mechanism to process the hidden state vectors of all time points within a preset historical period, obtaining the attention score for each preset gas at each time point within the preset historical period. The attention score reflects the degree of attention received by each preset gas at each time point within the preset historical period. For example, when the system needs to decouple carbon monoxide gas, the attention mechanism learns to assign extremely high weights to the time points near the response peak of the electrochemical sensor, because the signal of carbon monoxide gas is most significant at this time. When decoupling methane gas, it will shift its attention to the joint response mode of the infrared sensor and the catalytic combustion sensor, and may ignore the signal of the electrochemical sensor, thereby reducing the problem of cross-interference during the decoupling of multiple gases. The preset gas should include all gases that may appear in the actual decoupling scenario. For example, in an industrial production scenario, gases that may be generated include hydrogen, carbon monoxide, carbon dioxide, and methane, so the preset gas needs to include all gases that may be generated in this scenario.

[0034] Preferably, in one embodiment of the present invention, the method for obtaining the attention score of each preset gas at each time point in a preset historical period specifically includes: A multi-head attention mechanism is adopted, in which an attention head is set for each preset gas. The hidden state vectors of all time points in the preset historical period are input into each attention head respectively. Each attention head outputs the attention score of the corresponding preset gas at each time point in the preset historical period. The multi-head attention mechanism can provide optimal feature extraction for each gas, thereby improving the subsequent gas decoupling accuracy. The multi-head attention mechanism is a well-known technique in the art and will not be described in detail here.

[0035] It should be noted that the range of attention scores is [range missing]. Furthermore, the sum of the attention scores of the corresponding preset gas output by each attention head at all time points in the preset historical period is a value of 1.

[0036] Attention score is used to represent the degree of attention received by the hidden state vector of each preset gas at each time point in a preset historical period. The higher the attention score, the more important the hidden state vector of the preset gas is at that time point. Therefore, this embodiment of the invention needs to further obtain the comprehensive feature vector of each preset gas based on the attention score of each preset gas at each time point in the preset historical period and the hidden state vector at each time point. The comprehensive feature vector is a vector form for each preset gas used for subsequent concentration decoupling tasks. At the same time, the comprehensive feature vector also amplifies the key features at different time points and suppresses the non-key features at different time points, thereby reducing the interference of other gases and improving the decoupling accuracy of subsequent gases.

[0037] Preferably, in one embodiment of the present invention, the method for obtaining the comprehensive feature vector of each preset gas specifically includes: Take any preset gas as the target preset gas, and then use the attention scores of the target preset gas at each time point within a preset historical period to perform a weighted summation of the hidden state vectors at each time point to obtain the comprehensive feature vector of the target preset gas.

[0038] As an example, in one embodiment of the present invention, the expression for the comprehensive feature vector of the target preset gas can be specifically as follows: in, This represents the comprehensive feature vector of the target preset gas; This indicates the target preset gas within a preset historical time period. Attention scores at each time point; Indicates the first [number]th ... The hidden state vector at each time point; This indicates the number of time points included in the preset historical period.

[0039] The same method described above can be used to obtain the comprehensive feature vector of each preset gas.

[0040] The gas decoupling module 104 is used to input the comprehensive feature vector of each preset gas into the fully connected neural network layer and output the true concentration of each preset gas.

[0041] The comprehensive feature vector of each preset gas is a concentrated vector form focused on the current concentration decoupling task. Therefore, the comprehensive feature vector of each preset gas can be input into a fully connected neural network layer and the true concentration of each preset gas can be output, thereby realizing the decoupling process of the mixed gas concentration.

[0042] Preferably, in one embodiment of the present invention, the method for obtaining the true concentration of each preset gas specifically includes: First, in the fully connected neural network layer, an independent neuron is set up for each preset gas, so that the concentration estimation of each gas depends only on the comprehensive feature vector corresponding to that gas, rather than the comprehensive feature vector of all gases. This can avoid unnecessary interference and make the model easier to interpret.

[0043] Then, the comprehensive feature vector of each preset gas is input into the corresponding neuron, thereby outputting the concentration value of each preset gas.

[0044] It should be noted that the fully connected neural network layer used for decoupling the concentration of mixed gases is trained using mixed gas samples containing each preset gas. The specific process is as follows: First, a large number of training samples are constructed. The training samples include gas detection signals of mixed gases containing each preset gas collected by different types of gas sensors in the data acquisition module 101, as well as environmental detection signals collected by different types of environmental sensors. The labels are the true concentration values ​​of each preset gas. During the training process, the true concentration values ​​of each preset gas can be measured using high-precision standard instruments (such as gas chromatographs). Then, after a series of calculations by the data acquisition module, data feature analysis module, attention analysis module, and gas decoupling module, a predicted concentration value is finally obtained in the fully connected output layer. The loss is calculated using the preset concentration value and the true concentration value. The loss function can be the mean squared error loss function, thereby realizing the training of the fully connected neural network layer.

[0045] Meanwhile, in order to improve the timeliness of the decoupling system, this embodiment of the invention deploys the real-time decoupling system on an FPGA chip, thereby achieving hardware acceleration of the algorithm and reducing the processing latency from hundreds of milliseconds in traditional solutions to less than 10 milliseconds, meeting the real-time monitoring requirements.

[0046] In summary, this invention first acquires gas detection signals of mixed gases from different types of gas sensors. The sensors used can achieve signal complementarity. Compared with traditional deep learning solutions, the number of sensors required by the system is reduced by 75%, and the computational resource requirements are reduced by 60%, making it suitable for resource-constrained portable devices. At the same time, it acquires environmental detection signals of the environment from different types of environmental sensors, avoiding the impact of excessive environmental changes on the gas decoupling accuracy. Then, it extracts features between gas detection signals and environmental detection signals through wavelet transform and recurrent neural networks. It assigns different weights to the hidden state vectors at each time point within a preset historical period through an attention mechanism to identify key features. Then, it achieves decoupling of the corresponding preset gas concentration through the acquired comprehensive feature vector. In the test environment, the decoupling system has an accuracy of over 98.33% in identifying 20 common industrial gases, which is more than 25% higher than traditional methods. Furthermore, through FPGA hardware acceleration, the decoupling system can complete multi-gas decoupling within 10 milliseconds, meeting the real-time monitoring needs of high-risk scenarios such as mines and chemical plants. It can also be used in various complex mixed gas environments such as oil refineries, chemical industrial parks, and medical breath diagnostics.

[0047] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0048] The various embodiments in this specification 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.

Claims

1. An AI-driven multi-gas real-time decoupling system, characterized in that, The system includes: The data acquisition module is used to acquire gas detection signals of mixed gas from different types of gas sensors, and environmental detection signals of the environment from different types of environmental sensors. The data feature analysis module is used to perform wavelet transform processing on the gas detection signals of each gas sensor and the environmental detection signals of each environmental sensor within a preset historical time period, and combine it with a recurrent neural network to obtain the hidden state vector of each time point in the preset historical time period. The attention analysis module is used to process the hidden state vectors of all time points in a preset historical period using an attention mechanism to obtain the attention score of each preset gas at each time point in the preset historical period; and to obtain the comprehensive feature vector of each preset gas based on the attention score of each preset gas at each time point in the preset historical period and the hidden state vector at each time point. The gas decoupling module is used to input the comprehensive feature vector of each preset gas into a fully connected neural network layer and output the true concentration of each preset gas.

2. The AI-driven multi-gas real-time decoupling system according to claim 1, characterized in that, The process of obtaining the hidden state vector for each point in time during a preset historical period includes: Wavelet transform processing is performed on the gas detection signals of each gas sensor and the environmental detection signals of each environmental sensor within a preset historical time period to obtain the time-frequency diagrams of the gas detection signals of each gas sensor and the environmental detection signals of each environmental sensor. The gas detection signals from each gas sensor and the environmental detection signals from each environmental sensor are input into a recurrent neural network, which then outputs the hidden state vector for each time point in a preset historical period.

3. The AI-driven multi-gas real-time decoupling system according to claim 2, characterized in that, The wavelet basis used in the wavelet transform is the Daubechies wavelet, and the recurrent neural network is a gated recurrent unit network.

4. The AI-driven multi-gas real-time decoupling system according to claim 1, characterized in that, The method of obtaining the attention score for each preset gas at each time point in a preset historical period includes: A multi-head attention mechanism is adopted, with an attention head set for each preset gas. The hidden state vectors of all time points in the preset historical period are input into each attention head, and each attention head outputs the attention score of the corresponding preset gas at each time point in the preset historical period.

5. The AI-driven multi-gas real-time decoupling system according to claim 1, characterized in that, The process of obtaining the comprehensive feature vector for each preset gas includes: Use any preset gas as the target preset gas; By using the attention scores of the target preset gas at each time point within a preset historical period, the hidden state vectors at each time point are weighted and summed to obtain the comprehensive feature vector of the target preset gas.

6. The AI-driven multi-gas real-time decoupling system according to claim 1, characterized in that, The actual concentration of each preset gas output includes: In the fully connected neural network layer, an independent neuron is set for each preset gas, and the comprehensive feature vector of each preset gas is input into the corresponding neuron, thereby outputting the concentration value of each preset gas.

7. The AI-driven multi-gas real-time decoupling system according to claim 1, characterized in that, The fully connected neural network layer is trained using a mixed gas sample containing each preset gas.

8. The AI-driven multi-gas real-time decoupling system according to claim 1, characterized in that, The real-time decoupling system is deployed on an FPGA chip.

9. The AI-driven multi-gas real-time decoupling system according to claim 1, characterized in that, The gas sensor includes at least an infrared sensor, an electrochemical sensor, and a catalytic combustion sensor.

10. The AI-driven multi-gas real-time decoupling system according to claim 1, characterized in that, The environmental sensors include at least a temperature sensor and a humidity sensor.

Images