A natural gas odorant concentration detection system based on multi-sensor fusion

Abstract

The present invention provides a natural gas odorant concentration detection system based on multi-sensor fusion, which includes: a multi-sensor array, including at least three different types of sensors; a data acquisition module, respectively collecting detection signals of each type of sensor in the multi-sensor array; a signal preprocessing module, preprocessing the collected detection signals of each type of sensor to obtain a preprocessed signal corresponding to each type of sensor; a feature extraction module, extracting corresponding feature vectors from the preprocessed signal of each type of sensor by wavelet transform and principal component analysis methods; a multi-modal fusion module, using a deep neural network algorithm to fuse the feature vectors of all sensors in the multi-sensor array to obtain fusion features; a concentration estimation module, using a plurality of machine learning regression algorithms to perform parallel predictions on the fusion features to determine the odorant concentration. The present invention realizes high-precision, real-time and reliable detection of natural gas odorant concentration.

Description

Technical Field

[0001] The invention relates to the field of natural gas odorant, and in particular to a natural gas odorant concentration detection system based on multi-sensor fusion. Background Art

[0002] Natural gas is widely used as a clean energy source, but it is colorless and odorless. For safety reasons, odorants are usually added. Accurately detecting the concentration of odorants in natural gas is crucial to ensure safe use and meet relevant regulatory requirements.

[0003] Traditional methods for detecting odorant concentration mainly include gas chromatography and single sensor detection. Gas chromatography has the advantages of high precision and high selectivity, and can accurately separate and quantitatively analyze various components in complex mixtures. However, this method has the disadvantages of expensive equipment, large size, professional operation, and inability to achieve real-time online monitoring.

[0004] The single sensor detection method has the advantages of simple operation, fast response and low cost. However, this method is easily affected by environmental factors such as temperature and humidity, and has poor selectivity for interfering gases, resulting in low reliability and accuracy of the detection results. Summary of the invention

[0005] In view of this, an embodiment of the present invention provides a natural gas odorant concentration detection system based on multi-sensor fusion to solve the above technical problems.

[0006] To achieve the above object, the present invention provides a natural gas odorant concentration detection system based on multi-sensor fusion, which comprises:

[0007] a multi-sensor array including at least three different types of sensors;

[0008] A data acquisition module, used to respectively acquire detection signals of each type of sensor in the multi-sensor array;

[0009] A signal preprocessing module is used to filter, amplify and normalize the collected detection signals of each type of sensor to obtain a preprocessed signal corresponding to each type of sensor;

[0010] A feature extraction module is used to extract corresponding feature vectors from the preprocessed signals of each type of sensor by wavelet transform and principal component analysis method;

[0011] A multimodal fusion module, used to fuse the feature vectors of all sensors in the multi-sensor array using a deep neural network algorithm to obtain a fusion feature;

[0012] A concentration estimation module is used to perform parallel prediction on the fused features through multiple machine learning regression algorithms to determine the odorant concentration.

[0013] In some possible implementations, the multi-sensor array includes:

[0014] At least one metal oxide semiconductor sensor for detecting odorant molecules in natural gas and outputting a resistance change signal related to the concentration of the odorant; wherein the odorant molecules react with the surface of the metal oxide semiconductor sensor as a reducing gas;

[0015] at least one quartz crystal microbalance sensor, configured to detect a mass change of the quartz crystal microbalance sensor caused by adsorption of odorant molecules on the surface of the quartz crystal microbalance sensor and output a frequency signal corresponding to the mass change;

[0016] at least one surface acoustic wave sensor for detecting a change in acoustic wave propagation characteristics caused by adsorption of odorant molecules on a surface of the surface acoustic wave sensor and outputting an acoustic wave signal corresponding to the change in acoustic wave propagation characteristics;

[0017] At least one electrochemical sensor is used to detect the current change caused by the redox reaction of the odorant molecules on the electrode surface of the electrochemical sensor and output a current signal corresponding to the current change.

[0018] In some possible implementations, the signal preprocessing module specifically includes:

[0019] An adaptive filter is used to automatically adjust filter parameters according to the spectrum characteristics of the detection signal, and use the adjusted filter parameters to remove noise in the detection signals of different types of sensors to obtain a filtered signal;

[0020] A nonlinear amplification unit, used for nonlinearly amplifying the filtered signal using a piecewise linear function to obtain an amplified signal;

[0021] A dynamic baseline correction unit, used for real-time monitoring and correction of the baseline drift of each sensor, and performing baseline correction on the amplified signal to obtain a corrected signal;

[0022] An adaptive normalization processing unit, used for dynamically adjusting normalization parameters according to historical data of each type of sensor, and performing normalization processing on the corrected signal to obtain a normalized signal;

[0023] The outlier detection and processing unit is used to identify outliers in the normalized signal using a local outlier factor algorithm; and process the identified outliers to obtain the preprocessed signal.

[0024] In some possible implementations, the feature extraction module specifically includes:

[0025] A wavelet decomposition unit is used to perform wavelet decomposition on the preprocessed signal of each type of sensor to obtain multi-scale wavelet coefficients with a number equal to the order of the wavelet basis function;

[0026] A characteristic scale selection unit, configured to select, from the multi-scale wavelet coefficients, a number of scales whose energy proportion exceeds a preset proportion of the total energy as characteristic scales;

[0027] A statistical feature extraction unit, used to extract statistical features of the wavelet coefficients corresponding to the characteristic scale to obtain statistical features including mean, variance, skewness and kurtosis;

[0028] An initial feature vector generating unit, used for sequentially arranging the statistical features of each feature scale in the order of mean, variance, skewness, and kurtosis, and combining them into an initial feature vector;

[0029] A principal component analysis unit, used to perform principal component analysis on the initial feature vector and select a principal component whose cumulative contribution rate reaches a preset threshold;

[0030] The feature vector output unit is used to output the selected principal component as the final feature vector.

[0031] In some possible implementations, the multimodal fusion module specifically includes:

[0032] A spatial feature extraction unit, used for extracting spatial features from feature vectors of each type of sensor using a convolutional neural network to obtain spatial feature representations of each type of sensor;

[0033] A temporal dependency analysis unit is used to capture the temporal dependency between the feature vectors of each type of sensor using a long short-term memory network to obtain a temporal correlation feature;

[0034] A dynamic weighting unit is used to dynamically weight the feature importance of the feature vector of each type of sensor through an attention mechanism to obtain a weighted feature representation;

[0035] The feature comprehensive processing unit is used to comprehensively process the spatial feature representation, the temporal correlation feature and the weighted feature representation through a fully connected layer, and output a fused feature.

[0036] In some possible implementations, the concentration estimation module specifically includes:

[0037] A multi-algorithm prediction unit is used to predict the odorant concentration of the fused features using three target algorithms including a support vector regression algorithm, a random forest algorithm, and a gradient boosting decision tree algorithm, respectively, to obtain three preliminary odorant concentration prediction results;

[0038] A weighted average unit, used to synthesize the three preliminary concentration prediction results by a weighted average method, wherein the weights are dynamically adjusted according to the performances of the three target algorithms on the validation set, respectively, to obtain a comprehensive odorant concentration prediction result;

[0039] A hyperparameter optimization unit, used to automatically adjust the hyperparameters of the three target algorithms by using a Bayesian optimization method to obtain optimized hyperparameters corresponding to the three target algorithms respectively;

[0040] An optimization prediction unit, used for applying the optimized hyperparameters to each of the three target algorithms accordingly, re-predicting the odorant concentration, and obtaining an optimized comprehensive odorant concentration prediction result;

[0041] In some possible implementations, the system further includes:

[0042] Environmental parameter monitoring module, used to monitor the environmental parameters including temperature, humidity and air pressure of the natural gas odorant concentration detection system in real time;

[0043] The compensation algorithm module is used to dynamically compensate the detection signal output by each type of sensor according to the environmental parameters, and the dynamic compensation specifically includes:

[0044] Modeling the sensitivity of each type of sensor to the environmental parameters;

[0045] Calculating compensation coefficients through a multiple linear regression algorithm according to real-time environmental parameters and the sensitivity model;

[0046] The compensation coefficient is applied to the original detection signal to obtain a compensated detection signal.

[0047] In some possible implementations, the system further includes:

[0048] A remote monitoring module, used to view the detection signal and the status of the natural gas odorant concentration detection system in real time through a remote device;

[0049] Artificial intelligence-assisted decision-making module for:

[0050] Acquiring historical data including historical detection signals, the environmental parameters and the operating status of the system, and using a time series analysis method to predict the trend of odorant concentration changes based on the historical data;

[0051] According to the prediction result of the odorant concentration variation trend and the preset system optimization goal, an optimization suggestion for the odorant injection strategy is generated through a reinforcement learning algorithm.

[0052] In some possible implementations, the system further includes:

[0053] The data encryption module is used to perform one or more of the following: encrypt the transmitted detection signal using an encryption algorithm; use a distributed ledger structure to ensure the immutability of data storage; and implement role-based data access control based on the user roles and permission matrix predefined by the system administrator.

[0054] In some possible implementations, the system further includes:

[0055] Self-diagnosis module, used to monitor the working status and performance degradation of each sensor in real time;

[0056] Fault warning module, used for:

[0057] Based on the working status and performance degradation of each sensor provided by the self-diagnosis module, determine whether each sensor is abnormal, wherein the abnormality judgment criteria include: the sensor response time exceeds the preset threshold, the sensitivity decreases by more than 20%, or the baseline drift exceeds 5% of the full scale;

[0058] Send out alarms when abnormal sensors are detected;

[0059] Automatically switch to a backup sensor or adjust the dynamically weighted weight to reduce the impact of the abnormal sensor, which specifically includes: if there is a backup sensor of the same type, automatically switch to the backup sensor; if there is no backup sensor, reduce the dynamically weighted weight of the feature importance of the feature vector of the abnormal sensor in the multimodal fusion module.

[0060] In some possible implementations, the system further includes:

[0061] A gas separation unit for separating water and / or interfering gases from natural gas before concentration detection;

[0062] A microfluidic chip, used for receiving the gas sample processed by the gas separation unit, and controlling the injection amount of the gas sample of each sensor through a programmable micro valve;

[0063] The multi-sensor array is arranged in an array on the microfluidic chip, and a plurality of sensors of each type are provided, and the activation number of each type of sensor can be set through a software interface.

[0064] The natural gas odorant concentration detection system based on multi-sensor fusion of the present invention has the following beneficial technical effects:

[0065] By using at least three different types of sensors to form a multi-sensor array, the system can capture the characteristic information of the odorant from multiple dimensions. The advantages of different types of sensors complement each other, significantly improving the accuracy and reliability of detection, and effectively reducing the errors and instability that may be caused by a single sensor.

[0066] The use of multi-sensor arrays improves the system's resistance to environmental interference. Even if a certain type of sensor is affected by a specific interference factor, other types of sensors can still provide effective information, ensuring the stability and robustness of the system.

[0067] The data acquisition module can collect detection signals from multiple sensors at the same time. Combined with efficient signal preprocessing and feature extraction technology, the system can quickly process large amounts of data, realize real-time detection of odorant concentration, and meet the timeliness requirements in practical applications.

[0068] The signal preprocessing module effectively improves the signal-to-noise ratio and comparability of the original detection signal through filtering, amplification and normalization, laying a good foundation for subsequent feature extraction and fusion.

[0069] The feature extraction module uses wavelet transform and principal component analysis methods to extract the most representative and discriminative features from complex sensor signals, greatly reducing the data dimension and improving the efficiency of subsequent processing.

[0070] The multimodal fusion module adopts a deep neural network algorithm, which can effectively capture the complex nonlinear relationship between different sensor signals, make full use of multimodal information, and improve the expression ability of fusion features.

[0071] The concentration estimation module is based on high-quality fusion features and can estimate the odorant concentration more accurately, with higher precision and stronger generalization ability than traditional methods.

[0072] The modular design of the system makes it adaptable and scalable. It can adjust the sensor type and update the algorithm model according to actual needs to adapt to different application scenarios and detection requirements.

[0073] Although multiple sensors are used, through efficient data processing and fusion algorithms, the system can run on ordinary embedded hardware without the need for expensive dedicated equipment, thus reducing the overall cost.

[0074] The integrated design of the system facilitates installation and maintenance, and the modular structure makes the system easy to upgrade and troubleshoot, improving the maintainability and service life of the system.

[0075] The system can output odorant concentration data in real time, facilitates integration with other systems, supports remote monitoring and data analysis, and is conducive to building an intelligent natural gas monitoring network. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] The accompanying drawings are used to better understand the present invention and do not constitute an improper limitation of the present invention.

[0077] Figure 1 It is a functional block diagram of a natural gas odorant concentration detection system based on multi-sensor fusion according to an embodiment of the present invention;

[0078] Figure 2 is a specific functional block diagram of a multi-sensor array 10 according to an embodiment of the present invention;

[0079] Figure 3 is a specific functional block diagram of the signal preprocessing module 30 according to an embodiment of the present invention;

[0080] Figure 4 is a specific functional block diagram of the feature extraction module 40 according to an embodiment of the present invention;

[0081] Figure 5 is a specific functional block diagram of the multimodal fusion module 50 of an embodiment of the present invention;

[0082] Figure 6 is a specific functional block diagram of the concentration estimation module 60 according to an embodiment of the present invention;

[0083] Figure 7 is a functional block diagram of another natural gas odorant concentration detection system based on multi-sensor fusion according to an embodiment of the present invention;

[0084] Figure 8 This is a functional block diagram of another natural gas odorant concentration detection system based on multi-sensor fusion according to an embodiment of the present invention. DETAILED DESCRIPTION

[0085] The following is a description of exemplary embodiments of the present invention in conjunction with the accompanying drawings, including various details of the embodiments of the present invention to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present invention. Similarly, for clarity and conciseness, the description of well-known functions and structures is omitted in the following description.

[0086] The natural gas odorant concentration detection system based on multi-sensor fusion proposed in the embodiment of the present invention is as follows: Figure 1 As shown, it includes a multi-sensor array 10, a data acquisition module 20, a signal preprocessing module 30, a feature extraction module 40, a multi-modal fusion module 50 and a concentration estimation module 60.

[0087] The multi-sensor array 10 includes at least three different types of sensors. In this embodiment, the multi-sensor array 10 may include but is not limited to any three different types of sensors among a metal oxide semiconductor (MOS) sensor, a quartz crystal microbalance (QCM) sensor, a surface acoustic wave (SAW) sensor, and an electrochemical sensor, or use the above four types of sensors at the same time. This diverse sensor combination can capture the characteristics of the odorant from different angles, thereby improving the comprehensiveness and reliability of the detection.

[0088] The data acquisition module 20 is used to collect the detection signals of each type of sensor in the multi-sensor array 10. The detection signal refers to the electrical signal generated by the sensor in response to the odorant, which can be a voltage, current or frequency change. The data acquisition module 20 adopts multi-channel synchronous sampling technology to ensure the time consistency of different sensor signals, laying the foundation for subsequent multi-modal fusion.

[0089] The signal preprocessing module 30 filters, amplifies and normalizes the detection signals collected from each type of sensor to obtain preprocessed signals corresponding to each type of sensor. The filtering process is used to remove noise from the signal, the amplification process is used to enhance the signal strength, and the normalization process makes the outputs of different sensors comparable. The signal preprocessing module 30 significantly improves the quality of the original signal and provides high-quality input for subsequent processing.

[0090] The feature extraction module 40 extracts the corresponding feature vector from the preprocessed signal of each type of sensor through wavelet transform and principal component analysis. Wavelet transform can analyze the signal in both time and frequency domains, and principal component analysis is used to reduce the dimension and extract the main features. This combined method can effectively capture the key features of the signal, while reducing the data dimension and improving the efficiency of subsequent processing.

[0091] The multimodal fusion module 50 uses a deep neural network algorithm to fuse the feature vectors of all sensors in the multi-sensor array 10 to obtain fusion features. A deep neural network is a machine learning model that can learn complex nonlinear relationships. Through this advanced fusion method, the system can make full use of the complementary information of different sensors to improve the accuracy and robustness of detection.

[0092] The concentration estimation module 60 is used to perform parallel prediction of the fused features through multiple machine learning regression algorithms to determine the odorant concentration. This module converts the highly abstract fused features into specific concentration values ​​and is the final output link of the entire system. By using advanced regression algorithms, such as support vector regression, this module can accurately estimate the odorant concentration.

[0093] The natural gas odorant concentration detection system based on multi-sensor fusion of the embodiment of the present invention achieves multiple technical effects through the synergistic effect of multi-sensor arrays, efficient data acquisition, advanced signal processing and intelligent fusion algorithms. The system significantly improves the accuracy and reliability of odorant concentration detection, enhances resistance to environmental interference, and ensures real-time response performance. The modular design gives the system adaptability and scalability, enabling it to flexibly respond to different detection requirements and environmental conditions. The intelligent fusion of multi-dimensional information not only improves the comprehensiveness of detection, but also enhances the robustness of the system. The efficient data processing process ensures that the system can implement complex algorithms with limited computing resources, thereby reducing overall costs. In addition, the system's real-time data output capability can realize remote monitoring and large-scale deployment, which is conducive to building an intelligent natural gas safety monitoring network.

[0094] like Figure 2 As shown, the multi-sensor array 10 comprises:

[0095] At least one metal oxide semiconductor sensor 110, for detecting odorant molecules in natural gas and outputting a resistance change signal related to the odorant concentration; wherein the odorant molecules react with the surface of the metal oxide semiconductor sensor as a reducing gas;

[0096] at least one quartz crystal microbalance sensor 120, for detecting a mass change of the quartz crystal microbalance sensor caused by odorant molecules adsorbed on the surface of the quartz crystal microbalance sensor and outputting a frequency signal corresponding to the mass change;

[0097] at least one surface acoustic wave sensor 130, for detecting changes in acoustic wave propagation characteristics caused by adsorption of odorant molecules on the surface of the surface acoustic wave sensor and outputting an acoustic wave signal corresponding to the changes in the acoustic wave propagation characteristics;

[0098] At least one electrochemical sensor 140 is used to detect the current change caused by the redox reaction of the odorant molecules on the electrode surface of the electrochemical sensor and output a current signal corresponding to the current change.

[0099] In some embodiments, all sensors are installed on the bypass or sampling point of the natural gas main pipeline and are located in a sealed detection room. The detection room is located on the bypass pipeline. This detection room is close to the natural gas main pipeline, but maintains a safe distance. It can be installed in a special detection box or a small station building next to the pipeline. Such an arrangement ensures that the sensor can continuously contact the natural gas sample while controlling the environmental conditions and preventing external interference. Specifically, the natural gas main pipeline is the main pipeline for transporting natural gas, and a bypass pipeline with a smaller diameter is led out from the main pipeline for sampling and detection purposes. The existence of this bypass ensures that the natural gas sample can be continuously delivered to the detection device without affecting the normal transportation of the main pipeline. The detection room can be installed in a windproof and rainproof detection box, so that it can adapt to the outdoor environment, or it can be placed in a small station building to provide better environmental control and safety protection.

[0100] A bypass is drawn from the main pipeline to introduce the natural gas sample into the detection chamber through a flow control valve and a filter. Four sensors (MOS, QCM, SAW and electrochemical sensors) can be arranged in parallel in the detection chamber. Each sensor is connected to the data acquisition module 20 through a dedicated signal line and powered by the power management unit. The data acquisition module 20 is connected to the central processing unit through a data bus, which is in turn connected to the display screen and the communication module.

[0101] The flow control valve is installed before the filter. The function of the flow control valve is to regulate the flow of natural gas entering the detection chamber to ensure a stable supply of natural gas samples. The filter is installed after the flow control valve to remove impurities and particulate matter in the natural gas to prevent these substances from interfering with the normal operation of the sensor or damaging the sensor.

[0102] The central processing unit includes a signal preprocessing module, a feature extraction module, a multimodal fusion module, and a concentration estimation module. It is set up on-site, close to the detection room, so as to process the sensor data in a timely manner. However, in some remote monitoring systems, some processing functions can also be set up in the remote control center.

[0103] Although the four sensors are arranged in parallel in the detection chamber to ensure simultaneous contact with the same gas sample, they are independent of each other. Each sensor may require specific working conditions (temperature, humidity, etc.), so it is best to set up an isolation area or independent small cavity in the detection chamber to install the four sensors separately. This can provide the best working environment for each sensor while avoiding possible interference between them. For example, a partition can be used to divide the detection chamber into four independent but connected small spaces, each space is equipped with a sensor, and small holes or pipes are used to ensure that the gas sample can be evenly distributed to each space. This design not only ensures that the sensors can detect the same gas sample at the same time, but also provides the most suitable working conditions for each sensor, thereby improving the detection accuracy and reliability of the entire system.

[0104] The following is a detailed explanation:

[0105] The multi-sensor array of the embodiment of the present invention includes four different types of sensors, each of which uses different physical or chemical principles to detect odorant molecules in natural gas.

[0106] Metal oxide semiconductor (MOS) sensor is a sensor that uses metal oxide semiconductor materials to be sensitive to gases. Its working principle is: when odorant molecules (such as tetrahydrothiophene THT or tert-butylmercaptomethane TBM) come into contact with the surface of the MOS sensor as a reducing gas, they react with the oxygen ions adsorbed on the surface of the sensor, causing the sensor resistance to change, and this resistance change is related to the concentration of the odorant. In specific implementation, this embodiment can use a MOS sensor with SnO2 as the sensitive material, and the operating temperature is set to 350°C. The resistance change signal output by the sensor is converted into a 0-5V voltage signal through a voltage divider circuit. MOS sensors are highly sensitive to low-concentration odorants, have a fast response speed, and are relatively low in cost. In alternative embodiments, other metal oxide materials such as ZnO, WO3 or In2O3 can be used, or nanostructures can be used to improve sensitivity.

[0107] The quartz crystal microbalance (QCM) sensor is a high-precision sensor that uses the piezoelectric effect to detect mass changes. Its working principle is: when odorant molecules are adsorbed on the surface of the QCM sensor, the mass of the sensor will increase, causing the resonant frequency to decrease. The frequency change is proportional to the adsorbed mass (i.e., the odorant concentration). In specific implementation, this embodiment can use an AT-cut quartz crystal with a base frequency of 10MHz, and a molecular imprinting polymer film is coated on the surface to enhance selectivity. The frequency change is measured by a frequency counter with an accuracy of up to 0.1Hz. The QCM sensor has extremely high mass sensitivity and can detect nanogram-level mass changes, which is suitable for low-concentration odorant detection. In an extended implementation, the following methods can be used: high-frequency QCM, which can use a QCM with a higher base frequency (such as 30MHz or 50MHz) to further improve the mass sensitivity. High-frequency QCM can detect even smaller mass changes, thereby improving the detection capability of low-concentration odorants. Nanobalance technology, which uses ultra-thin QCM manufactured using nanotechnology to achieve femtogram-level (10 -15 g) The mass detection sensitivity is greatly improved, which greatly improves the detection limit of the system. Temperature compensation technology, which introduces temperature sensors and compensation algorithms to reduce the impact of ambient temperature changes on the QCM response and improve the accuracy and stability of the measurement. Multi-harmonic QCM, which uses a QCM that can work in multiple harmonic modes, obtains more information about the properties of the adsorption layer by analyzing the responses of different harmonics, and improves the selectivity of detection. Microfluidic QCM, which integrates QCM into a microfluidic system to achieve precise control and rapid exchange of samples, and improve detection efficiency and repeatability. These extended implementations can further improve the performance of QCM sensors in natural gas odorant detection, including sensitivity, selectivity, stability, and applicability.

[0108] Surface acoustic wave (SAW) sensor is a sensor that detects gas by using the change of propagation characteristics of surface acoustic waves. Its working principle is: the adsorption of odorant molecules on the surface of SAW sensor will change the mass and elastic properties of the sound wave propagation path, resulting in changes in the speed and attenuation of the sound wave. These changes can be detected by changes in phase or amplitude. In specific implementation, this embodiment can use a Love wave type SAW sensor with an operating frequency of 433MHz, and the sensitive layer is a polyisobutylene film. Love wave is a transverse surface wave that propagates in a softer thin layer covering a harder substrate. In the application of natural gas odorant detection, the advantages of Love wave type SAW sensor are: high sensitivity helps to detect low concentrations of odorants; good performance in humid environments can reduce the interference of moisture on detection; the high operating frequency of 433MHz further improves the sensitivity of the sensor; polyisobutylene film as a sensitive layer can effectively adsorb odorant molecules and enhance the detection effect. SAW sensor has high sensitivity, fast response and good temperature stability, and is suitable for real-time monitoring of odorant concentration changes. In alternative embodiments, a Shear Horizontal Surface Acoustic Wave (SH-SAW) or Lamb Wave Device may be used, and different polymer coatings may be selected to optimize the response to specific odorants. In natural gas odorant detection, SH-SAW devices may provide better performance, especially in the presence of moisture interference. By coating the surface of the SH-SAW device with an appropriate polymer layer, its selectivity and sensitivity to specific odorant molecules may be further enhanced. In natural gas odorant detection applications, Lamb wave devices may provide high sensitivity and multi-parameter detection capabilities. By selecting the appropriate operating mode and frequency, combined with a specific polymer coating, Lamb wave devices can achieve accurate detection of different odorant molecules.

[0109] Electrochemical sensor, a sensor that detects gas through electrochemical reaction. Its working principle is: odorant molecules undergo redox reaction on the electrode surface of the electrochemical sensor to generate current. This current is proportional to the concentration of the odorant. In specific implementation, this embodiment can use a three-electrode system, the working electrode is a platinum electrode, the reference electrode is a silver / silver chloride (Ag / AgCl) electrode, and the auxiliary electrode is a carbon electrode. The electrolyte is a 0.1M H2SO4 solution. The current between the working electrode and the auxiliary electrode is measured by a potentiostat. The electrochemical sensor has good selectivity and linear response, and is suitable for precise quantitative analysis. In an alternative implementation, an ion-selective field-effect transistor (ISFET, Ion-Selective Field-EffectTransistor) can be used as a new type of electrochemical sensor to improve the level of integration and miniaturization.

[0110] By combining these four sensors with different principles, the multi-sensor array of the embodiment of the present invention can fully capture the various physical and chemical properties of the odorant. MOS sensors provide fast response, QCM sensors and SAW sensors provide high-sensitivity detection, and electrochemical sensors provide precise quantitative capabilities. This multi-dimensional detection greatly improves the accuracy, reliability and anti-interference ability of the system. At the same time, the advantages of different sensors complement each other, overcoming the limitations of a single sensor, and enabling the system to adapt to various complex actual detection environments.

[0111] like Figure 3 As shown, the signal preprocessing module 30 specifically includes: an adaptive filter 310, which is used to automatically adjust the filtering parameters according to the spectral characteristics of the detection signal, and use the adjusted filtering parameters to remove the noise in the detection signals of different types of sensors to obtain a filtered signal; a nonlinear amplification unit 320, which is used to nonlinearly amplify the filtered signal using a piecewise linear function to obtain an amplified signal; a dynamic baseline correction unit 330, which is used to monitor and correct the baseline drift of each sensor in real time, and perform baseline correction on the amplified signal to obtain a corrected signal; an adaptive normalization processing unit 340, which is used to dynamically adjust the normalization parameters according to the historical data of each type of sensor, and perform normalization processing on the corrected signal to obtain a normalized signal; an outlier detection and processing unit 350, which is used to identify outliers in the normalized signal using a local outlier factor algorithm; and process the identified outliers to obtain a final preprocessed signal.

[0112] The signal preprocessing module 30 of the embodiment of the present invention is conducive to improving the quality of the original detection signal and laying a foundation for subsequent feature extraction and fusion. The module includes five main units, which are: adaptive filter, nonlinear amplification unit, dynamic baseline correction unit, adaptive normalization processing unit and outlier detection and processing unit in the order of signal processing.

[0113] The adaptive filter automatically adjusts the filter parameters according to the spectral characteristics of the detection signal to effectively remove noise from different types of sensor signals. In one embodiment, an adaptive filter with a minimum mean square error (LMS) algorithm can be used. The system first performs a fast Fourier transform (FFT) on the input signal to analyze its spectral characteristics. Based on the results of the spectrum analysis, the cutoff frequency and order of the filter are automatically adjusted. For example, for metal oxide semiconductor sensors, the typical cutoff frequency is set in the range of 10-100Hz; while for piezoelectric sensors, a higher cutoff frequency is required, such as 1-10kHz. After the filter parameters are adjusted, the original detection signal is filtered to obtain a filtered signal. This step significantly improves the signal-to-noise ratio of the signal, laying the foundation for subsequent processing.

[0114] The nonlinear amplification unit uses a piecewise linear function to nonlinearly amplify the filtered signal to enhance the detection capability of weak signals while avoiding saturation of strong signals. In one embodiment, a three-segment piecewise linear function can be used: for low-intensity signals (e.g., less than 20% of full scale), a higher gain (e.g., 5 times) is used; for medium-intensity signals (20%-80% of full scale), a medium gain (e.g., 2 times) is used; for high-intensity signals (greater than 80% of full scale), a low gain (e.g., 1.2 times) is used. This nonlinear amplification method can effectively improve the dynamic range of the system, enabling it to process weak and strong odorant signals at the same time.

[0115] The dynamic baseline correction unit monitors and corrects the baseline drift of each sensor in real time to ensure the accuracy of long-term detection. In one embodiment, the working process of the dynamic baseline correction unit includes the following steps:

[0116] S1: Initialization step. This step sets the sliding window size to 60 seconds. Initialize the exponentially weighted moving average (EWMA) parameters α The mutation detection threshold θ is set to 3 times the standard deviation.

[0117] S2: Data receiving step: This step continuously receives signal data from the nonlinear amplification unit.

[0118] S3: Sliding window update step. In this step, the newly received data points are added to the sliding window. If the window is full, the oldest data point is removed.

[0119] S4: Baseline estimation calculation step. This step applies the EWMA algorithm to the data in the sliding window:

[0120] .in, Bt is the baseline estimate at the current time, Xt is the current signal value, is the baseline estimate for the previous moment. is the smoothing factor .

[0121] S5: Mutation detection step. This step calculates the standard deviation σ of the data in the sliding window. , a mutation is considered to have occurred.

[0122] S6: Baseline update decision step. If no mutation is detected, continue to update the baseline estimate. If a mutation is detected, pause the baseline update and keep using the baseline estimate at the last non-mutation moment.

[0123] S7: Baseline correction step. This step subtracts the current baseline estimate from the current signal value:

[0124] .in, Y t is the corrected signal value.

[0125] S8: Output correction signal step. In this step, the corrected signal Y t Passed to the next processing unit (adaptive normalization processing unit).

[0126] S9: Parameter adaptive adjustment (optional). This step periodically (e.g., every hour) evaluates the baseline correction effect. Based on the evaluation results, the EWMA parameters are automatically adjusted. α and the mutation detection threshold θ.

[0127] S10: Abnormal handling step. If the baseline cannot be updated for 30 consecutive minutes (mutations are continuously detected), a warning signal is triggered, indicating that there may be a sensor failure or environmental abnormality.

[0128] S11: Loop through the steps. Return to step S2 and continue processing new data points.

[0129] This dynamic baseline correction method effectively compensates for sensor drift by tracking the long-term trend of the signal in real time, while being able to quickly respond to true concentration changes. It improves the long-term stability and detection accuracy of the system and is particularly suitable for continuous odorant concentration monitoring in complex and changing environments.

[0130] The adaptive normalization processing unit dynamically adjusts the normalization parameters according to the historical data of various types of sensors to make the outputs of different sensors comparable. In one embodiment, for each type of sensor, a circular buffer containing the most recent 1000 data points is maintained. The mean (μ) and standard deviation (σ) of the data in the buffer are calculated regularly (for example, every 10 minutes). The normalization process uses the Z-score method: (x-μ) / σ, where x is the input signal value. This adaptive normalization method ensures the consistency of the outputs of different types of sensors, which is beneficial to subsequent multimodal fusion.

[0131] The outlier detection and processing unit uses the local outlier factor (LOF) algorithm to identify and process outliers to improve the reliability of the preprocessed signal. In a specific implementation, outlier detection uses the LOF algorithm to calculate the local outlier factor for each data point. The LOF threshold is set to 2.0, and points exceeding this threshold are marked as potential outliers. When processing outliers, for the marked outliers, if there are less than 3 consecutive outliers, they are replaced by linear interpolation; if there are more than 3 but less than 10 consecutive outliers, median filtering is used; if there are more than 10 consecutive outliers, the original value is retained and a warning is issued to the system, indicating a sensor failure or a real abnormal event. This outlier detection and processing method can effectively remove burst noise and abnormal points in the signal while retaining possible important information.

[0132] Through the series connection of the above five units, the signal preprocessing module of the embodiment of the present invention can significantly improve the quality of the original detection signal and provide high-quality input data for subsequent feature extraction and fusion. This not only improves the detection accuracy and reliability of the system, but also enhances the system's adaptability to environmental changes and interference.

[0133] like Figure 4 As shown, the feature extraction module 40 specifically includes:

[0134] The wavelet decomposition unit 410 is used to perform wavelet decomposition on the preprocessed signal of each type of sensor to obtain multi-scale wavelet coefficients of the same number as the order of the wavelet basis function;

[0135] A characteristic scale selection unit 420, configured to select, from the multi-scale wavelet coefficients, a number of scales whose energy proportion exceeds a preset proportion of the total energy as characteristic scales;

[0136] A statistical feature extraction unit 430 is used to extract statistical features from the wavelet coefficients corresponding to the characteristic scale to obtain statistical features including mean, variance, skewness and kurtosis;

[0137] An initial feature vector generating unit 440 is used to sequentially arrange the statistical features of each feature scale in the order of mean, variance, skewness, and kurtosis to form an initial feature vector;

[0138] A principal component analysis unit 450 is used to perform principal component analysis on the initial feature vector and select a principal component whose cumulative contribution rate reaches a preset threshold;

[0139] The feature vector output unit 460 is used to output the selected principal component as the final feature vector.

[0140] Each step included in the working process of the feature extraction module 40 is explained in detail below:

[0141] (1) Wavelet decomposition

[0142] In this step, the preprocessed signal of each type of sensor is subjected to wavelet decomposition to obtain multi-scale wavelet coefficients equal in number to the order of the wavelet basis function. Wavelet decomposition is a time-frequency analysis method that can decompose a signal into components in different frequency ranges. This embodiment uses Daubechies wavelet (db4) as the basis function and performs a 5-layer decomposition. During the decomposition process, the signal is divided into approximate coefficients (low-frequency part) and detail coefficients (high-frequency part) at each layer.

[0143] For example, suppose the length of the preprocessed signal of a metal oxide sensor is 1024 points. After 5 layers of wavelet decomposition, the following wavelet coefficients are obtained: the 5th layer approximation coefficient (a5) is 32 points, the 5th layer detail coefficient (d5) is 32 points, the 4th layer detail coefficient (d4) is 64 points, the 3rd layer detail coefficient (d3) is 128 points, the 2nd layer detail coefficient (d2) is 256 points, and the 1st layer detail coefficient (d1) is 512 points.

[0144] (2) Feature scale selection

[0145] This step selects several scales whose energy proportion exceeds a preset proportion of the total energy from the multi-scale wavelet coefficients as characteristic scales. This step calculates the energy of each scale wavelet coefficient and calculates its proportion of the total energy. The scale whose energy proportion exceeds a preset threshold (e.g., 80%) is selected as the characteristic scale. This can retain the most important frequency components in the signal while reducing the computational complexity.

[0146] Assume that the energy proportions of each scale are calculated as follows:

[0147] a5: 45%, d5: 25%, d4: 15%, d3: 8%, d2: 5%, d1: 2%.

[0148] If the cumulative energy percentage threshold is set to 85%, the selection process is as follows:

[0149] Select a5 (45%);

[0150] Select d5 (45% + 25% = 70%).

[0151] Choose d4 (70% + 15% = 85%).

[0152] At this time, the cumulative energy ratio reaches 85%, and the selection stops. Therefore, a5, d5, and d4 are finally selected as the characteristic scales.

[0153] (3) Statistical feature extraction

[0154] This step extracts statistical features from the wavelet coefficients corresponding to the characteristic scale to obtain statistical features including mean, variance, skewness and kurtosis. Four statistical features are calculated for the wavelet coefficients of each selected characteristic scale: mean, which reflects the overall level of the signal; variance, which indicates the degree of fluctuation of the signal; skewness, which describes the asymmetry of the distribution; kurtosis, which reflects the sharpness of the distribution.

[0155] For example, the statistical characteristics of the a5, d5 and d4 coefficients are calculated as follows:

[0156] a5: mean = 0.5, variance = 0.1, skewness = 0.2, kurtosis = 3.1;

[0157] d5: mean = 0.1, variance = 0.3, skewness = -0.1, kurtosis = 2.8;

[0158] d4: mean = 0.05, variance = 0.2, skewness = 0.1, kurtosis = 3.0.

[0159] (4) Initial eigenvector combination

[0160] This step arranges the statistical features of each feature scale in the order of mean, variance, skewness, and kurtosis to form an initial feature vector. This step arranges the statistical features calculated in the previous step in a fixed order to form a vector. This unified arrangement ensures that the features of different sensors and different time points have a consistent structure.

[0161] For example, based on the calculation results of the previous step, the initial eigenvector is:

[0162] [0.5, 0.1, 0.2, 3.1, 0.1, 0.3, -0.1, 2.8, 0.05, 0.2, 0.1, 3.0];

[0163] The first four values ​​correspond to the statistical characteristics of a5, the middle four values ​​correspond to the statistical characteristics of d5, and the last four values ​​correspond to the statistical characteristics of d4.

[0164] (5) Principal component analysis

[0165] This step performs principal component analysis on the initial eigenvectors and selects the principal components whose cumulative contribution reaches a preset threshold. Principal component analysis (PCA) is a dimensionality reduction technique that can find the main direction of change in the data. By calculating the eigenvalues ​​and eigenvectors, each principal component and its contribution can be obtained. The first few principal components whose cumulative contribution reaches a preset threshold (e.g. 95%) are selected.

[0166] For example, suppose that after performing PCA on the initial eigenvector, the following principal components and their contribution rates are obtained:

[0167] PC1: 55%, PC2: 20%, PC3: 15%, PC4: 5%, PC5: 3%, PC6: 1%.

[0168] If the threshold is set to 95%, the first four principal components (PC1, PC2, PC3, PC4) are selected because their cumulative contribution rate (95%) reaches the preset threshold.

[0169] (6) Final feature vector output

[0170] This step outputs the selected principal components as the final feature vector. This step combines the selected principal components into the final feature vector. This feature vector is the projection of the original features in the principal component space, which retains the main information of the original data while greatly reducing the data dimension.

[0171] For example, based on the four principal components selected in the previous step, the final eigenvector might be as follows: [2.5, -1.2, 0.8, 0.3].

[0172] Each value represents the projection coordinate of the original feature on the corresponding principal component. This 4-dimensional vector contains most of the information of the original 12-dimensional feature.

[0173] Through these six steps, the feature extraction module successfully transforms complex sensor signals into low-dimensional, high-information feature vectors, laying the foundation for subsequent multimodal fusion and concentration estimation. This modified method can better preserve the important features of the signal while maintaining computational efficiency.

[0174] like Figure 5 As shown, the multimodal fusion module 50 specifically includes:

[0175] A spatial feature extraction unit 510 is used to extract spatial features from feature vectors of each type of sensor using a convolutional neural network to obtain spatial feature representations of each type of sensor;

[0176] A timing dependency analysis unit 520 is used to capture the timing dependency relationship between the feature vectors of each type of sensor using a long short-term memory network to obtain a timing correlation feature;

[0177] A dynamic weighting unit 530, used to dynamically weight the feature importance of the feature vector of each type of sensor through an attention mechanism to obtain a weighted feature representation;

[0178] The feature comprehensive processing unit 540 is used to comprehensively process the spatial feature representation, the temporal correlation feature and the weighted feature representation through a fully connected layer, and output a fused feature.

[0179] The working process of the multimodal fusion module 50 specifically includes the following steps:

[0180] (1) Spatial feature extraction

[0181] In this step, a convolutional neural network (CNN) is used to extract spatial features from the feature vectors of each type of sensor to obtain a spatial feature representation of each type of sensor.

[0182] In the specific implementation, a one-dimensional CNN structure is used, which includes two convolutional layers and a maximum pooling layer. The first convolutional layer uses 16 3x1 convolutional kernels, and the second convolutional layer uses 32 3x1 convolutional kernels. Finally, a fixed-dimensional output is obtained through global average pooling.

[0183] For example, suppose there are three types of sensors (MOS, QCM, SAW), and the feature vector dimension of each sensor is 4. The input is: [4x1], [4x1], [4x1]; after CNN processing, it is: [32x1], [32x1], [32x1]; in this way, the features of each sensor are mapped to a 32-dimensional spatial feature representation.

[0184] (2) Capturing Timing Dependencies

[0185] In this step, a long short-term memory network (LSTM) is used to capture the temporal dependencies between all different types of sensor signals and obtain temporal correlation features.

[0186] In the specific implementation, the spatial feature representations of different sensors are connected in time order and input into a bidirectional LSTM network. The LSTM network contains 64 hidden units and outputs the hidden state of the last time step as the temporal correlation feature.

[0187] For example, the input sequence is as follows:

[0188] At time t: [[32.1, 15.3, ..., 8.7], [45.2, 22.1, ..., 10.5], [18.9, 33.4,..., 27.6]];

[0189] Time t-1: [[31.8, 15.5, ..., 8.9], [44.9, 22.3, ..., 10.3], [19.1,33.2, ..., 27.8]];

[0190] At time t-2: [[32.3, 15.1, ..., 8.5], [45.5, 21.9, ..., 10.7], [18.7,33.6, ..., 27.4]];

[0191] At t-3: [[31.9, 15.4, ..., 8.8], [44.7, 22.4, ..., 10.2], [19.2,33.1, ..., 27.9]];

[0192] At t-4: [[32.2, 15.2, ..., 8.6], [45.3, 22.0, ..., 10.6], [18.8,33.5, ..., 27.5]];

[0193] LSTM output is: [128x1];

[0194] This 128-dimensional vector contains the temporal dependency information between different sensors.

[0195] (3) Dynamic feature weighting

[0196] This step dynamically weights the importance of features of all different types of sensors through the attention mechanism to obtain weighted feature representation. In the specific implementation, the self-attention mechanism is used to calculate the query matrix (Q), key matrix (K) and value matrix (V), and then the attention weight is obtained through the softmax function. Finally, the weight is multiplied by the value to obtain the weighted feature.

[0197] The following are some examples:

[0198] Input: [[32x1], [32x1], [32x1]] (spatial features of three sensors);

[0199] Attention calculation:

[0200] Q = K = V = input;

[0201] Attention score = ;

[0202] Weighted feature = attention score * V;

[0203] Output: [32x1] (weighted feature representation).

[0204] (4) Comprehensive feature processing

[0205] This step uses a fully connected layer to comprehensively process the spatial feature representation, temporal correlation feature, and weighted feature representation, and outputs the fused feature. In the specific implementation, the outputs of the first three steps are connected into a large feature vector, which is then processed through a two-layer fully connected network. The first layer uses the ReLU activation function, and the second layer directly outputs.

[0206] As an example, the spatial feature is [32x3] (corresponding to 3 types of sensors), the temporal feature is [128x1], the weighted feature is [32x1], and the connected feature is [256x1]. Then it passes through the fully connected layer 1 (containing 128 neurons) and the output is [128x1]. Then it passes through the fully connected layer 2 (containing 64 neurons) and the final output is [64x1]. The above final output [64x1] vector is the fusion feature, which integrates the spatial features, temporal features, and dynamic weighted features of multiple sensors.

[0207] Through these four steps, the multimodal fusion module can effectively integrate information from different types of sensors, capture the complex relationships between them, and generate a highly informative fusion feature that not only contains the unique information of each sensor, but also reflects the interactions and temporal dependencies between them.

[0208] like Figure 6As shown, the concentration estimation module 60 specifically includes:

[0209] A multi-algorithm prediction unit 510 is used to use three target algorithms including support vector regression, random forest and gradient boosting decision tree to respectively predict the odorant concentration of the fused features to obtain three preliminary odorant concentration prediction results;

[0210] The weighted average unit 620 is used to use the weighted average method to integrate the three preliminary concentration prediction results corresponding to the support vector regression algorithm, the random forest algorithm and the gradient boosting decision tree algorithm, wherein the weights are dynamically adjusted according to the performance of the three target algorithms on the validation set, to obtain a comprehensive odorant concentration prediction result.

[0211] In a further embodiment, the concentration estimation module 60 may further include:

[0212] A hyperparameter optimization unit 630 is used to automatically adjust the hyperparameters of the support vector regression algorithm, the random forest algorithm, and the gradient boosting decision tree algorithm by using a Bayesian optimization method to obtain optimized hyperparameters corresponding to the three target algorithms respectively;

[0213] The optimization prediction unit 640 is used for applying the optimized hyperparameters to each of the three target algorithms, re-predicting the odorant concentration, and obtaining an optimized comprehensive odorant concentration prediction result;

[0214] The confidence interval estimation unit 650 is used to perform confidence interval estimation on the optimized comprehensive concentration prediction result to obtain a final odorant concentration estimation result with a credibility range.

[0215] Specifically, the working process of the concentration estimation module 60 includes the following steps:

[0216] (1) Preliminary concentration prediction

[0217] In this step, three algorithms, namely support vector regression (SVR), random forest (RF) and gradient boosted decision tree (GBDT), were used to predict the odorant concentration based on the fused features, and three preliminary concentration prediction results were obtained.

[0218] In the specific implementation, SVR: uses the RBF kernel function, the penalty parameter C=1.0, and the ε (epsilon)=0.1 in the insensitive loss function; RF: uses 100 decision trees with a maximum depth of 10; GBDT: uses 100 weak learners with a learning rate of 0.1.

[0219] Assume that the fused feature is a 64-dimensional vector: [0.5, 0.3, ..., 0.8]; the prediction results are as follows: the SVR model predicted value is 25.3 mg / m³, the RF model predicted value is 27.1 mg / m³, and the GBDT model predicted value is 26.2 mg / m³.

[0220] (2) Weighted average comprehensive

[0221] In this step, the weighted average method is used to combine the preliminary concentration prediction results of the three algorithms, where the weights are dynamically adjusted according to the performance of each algorithm on the validation set to obtain a comprehensive concentration prediction result.

[0222] In the specific implementation, the most recent 1000 samples are used as the sliding validation set, and the root mean square error (RMSE) of each algorithm on the validation set is calculated. The weight is proportional to the inverse of the RMSE.

[0223] The following are some examples:

[0224] Assume RMSE on the validation set: SVR: 1.2; RF: 0.9; GBDT: 1.0.

[0225] The weights are calculated as follows:

[0226] w_SVR = 1 / 1.2 / (1 / 1.2 + 1 / 0.9 + 1 / 1.0) ≈ 0.29;

[0227] w_RF = 1 / 0.9 / (1 / 1.2 + 1 / 0.9 + 1 / 1.0) ≈ 0.39;

[0228] w_GBDT = 1 / 1.0 / (1 / 1.2 + 1 / 0.9 + 1 / 1.0) ≈ 0.32.

[0229] The comprehensive forecast results are as follows:

[0230] 25.3 * 0.29 + 27.1 * 0.39 + 26.2 * 0.32 ≈ 26.3 mg / m³.

[0231] (3) Hyperparameter Optimization

[0232] This step automatically adjusts the hyperparameters of the three algorithms through the Bayesian optimization method. In the specific implementation, Gaussian process regression is used as the proxy model of Bayesian optimization, and expected improvement (EI) is used as the acquisition function. The following hyperparameters are optimized for each algorithm:

[0233] For SVR: penalty parameter C (range is 0.1-10), ε in the insensitive loss function (range is 0.01-1);

[0234] For RF: number of decision trees (range 50-200), maximum depth of trees (range 5-20);

[0235] For GBDT: number of weak learners (range 50-200), learning rate (range 0.01-0.5);

[0236] In one example, after 50 rounds of optimization, the optimal hyperparameters are as follows:

[0237] For SVR: penalty parameter C = 2.3, ε value = 0.05;

[0238] For RF: number of decision trees = 150, maximum depth of tree = 12;

[0239] For GBDT: number of weak learners = 180, learning rate = 0.08.

[0240] (4) Optimized concentration prediction

[0241] In this step, the optimized hyperparameters are applied to each algorithm, and the odorant concentration prediction is re-performed to obtain the optimized comprehensive concentration prediction result. In the specific implementation, the three models are re-trained using the optimized hyperparameters, and then the same fusion feature is predicted.

[0242] In an example, using the fusion features in step 1, the optimized prediction results are as follows:

[0243] The predicted SVR value was 25.8 mg / m³;

[0244] The predicted value of RF is 26.5 mg / m³;

[0245] The GBDT predicted value is 26.1 mg / m³.

[0246] Perform a weighted average using the weights from step 2:

[0247] 25.8 * 0.29 + 26.5 * 0.39 + 26.1 * 0.32 ≈ 26.2 mg / m³.

[0248] (5) Confidence interval estimation

[0249] This step estimates the confidence interval of the optimized comprehensive concentration prediction result to obtain the final odorant concentration estimation result with a credibility range. In the specific implementation, the bootstrap method is used to repeatedly sample 1000 times from the historical data, and the model is retrained and predicted each time using the sampled data set. The 95% confidence interval is calculated based on these 1000 prediction results.

[0250] In one example, a confidence interval is estimated for the predicted result 26.2 mg / m³ in step 4. After bootstrap sampling and calculation, the 95% confidence interval is [25.5, 26.9] mg / m³.

[0251] The final output is as follows: The estimated odorant concentration is 26.2 mg / m³ and the 95% confidence interval is [25.5,26.9] mg / m³.

[0252] Through these five steps, the concentration estimation module can integrate the advantages of multiple machine learning algorithms, adaptively optimize model parameters, and provide reliable confidence intervals, thereby obtaining high-precision and high-confidence odorant concentration estimation results.

[0253] like Figure 7 As shown, the system also includes: an environmental parameter monitoring module 22, which is used to monitor the environmental parameters of the environment in which the natural gas odorant concentration detection system is located, including temperature, humidity and air pressure; a compensation algorithm module 24, which is used to dynamically compensate the detection signal output by each type of sensor according to the environmental parameters, and the dynamic compensation specifically includes: establishing a sensitivity model of each type of sensor to the environmental parameters; calculating the compensation coefficient through a multivariate linear regression algorithm based on the real-time environmental parameters and the sensitivity model; and applying the compensation coefficient to the original detection signal to obtain a compensated detection signal.

[0254] The natural gas odorant concentration detection system of the embodiment of the present invention further includes an environmental parameter monitoring module and a compensation algorithm module, which are used to improve the detection accuracy of the system under different environmental conditions. The specific working process is as follows:

[0255] (1) Environmental parameter monitoring

[0256] The environmental parameter monitoring module monitors the temperature, humidity and air pressure of the system's environment in real time. In the specific implementation, the temperature sensor uses a PT100 platinum resistor with an accuracy of ±0.1°C and a range of -40°C to 85°C; the humidity sensor uses a capacitive humidity sensor with an accuracy of ±2%RH and a range of 0-100%RH; the pressure sensor uses a MEMS pressure sensor with an accuracy of ±0.1hPa and a range of 300-1100hPa; the sampling frequency is to collect environmental parameters every 10 seconds. For example, the environmental parameters collected at a certain moment are as follows: the temperature is 25.3℃; the humidity is 65.7%RH; the air pressure is 1013.2hPa.

[0257] (2) Establishment of sensitivity model

[0258] Establish a sensitivity model to environmental parameters for each type of sensor. In the specific implementation, collect a large amount of historical data, including sensor output under different environmental conditions and the corresponding actual odorant concentration. Use the multivariate linear regression method to establish the sensitivity model. Model form: S = a * T + b * H + c * P + d; where S is the sensor sensitivity, T is temperature, H is humidity, P is air pressure, and a, b, c, d are regression coefficients.

[0259] Assume that the sensitivity model established for the MOS sensor is:

[0260] S_MOS = -0.02 * T + 0.01 * H - 0.005 * P + 1.5.

[0261] (3) Compensation coefficient calculation

[0262] According to the real-time environmental parameters and the sensitivity model, the compensation coefficient is calculated by the multivariate linear regression algorithm. In the specific implementation, the real-time environmental parameters are substituted into the sensitivity model, the sensor sensitivity under the current conditions is calculated, and then the compensation coefficient is calculated.

[0263] The compensation coefficient calculation formula is: K = S_standard / S_current; where K is the compensation coefficient, S_standard is the sensitivity under standard conditions, and S_current is the sensitivity under current conditions.

[0264] Assume that the sensitivity of the MOS sensor under standard conditions (25°C, 50%RH, 1013.25hPa) is 1.0.

[0265] The current environmental parameters are 25.3°C, 65.7%RH, 1013.2hPa;

[0266] Substitute the following sensitivity model:

[0267] S_current = -0.02 * 25.3 + 0.01 * 65.7 - 0.005 * 1013.2 + 1.5 =0.9835;

[0268] Calculate the compensation coefficient and you get: K = 1.0 / 0.9835 = 1.0168.

[0269] (4) Signal compensation

[0270] The compensation coefficient is applied to the original detection signal to obtain the compensated detection signal. In specific implementation, the original detection signal is multiplied by the compensation coefficient.

[0271] Compensation formula: S_compensated = S_original * K; where S_compensated is the compensated signal, S_original is the original signal, and K is the compensation coefficient.

[0272] Assume the original detection signal of the MOS sensor is 2.5V. Apply compensation: S_compensated = 2.5V *1.0168 = 2.542V.

[0273] (5) Compensation result output

[0274] The compensated detection signal is passed to the subsequent signal processing module. In the specific implementation, the compensated signal is transmitted to the signal preprocessing module through the data interface. For example, the compensated MOS sensor signal 2.542V is transmitted to the signal preprocessing module for further processing.

[0275] Through the above steps, the environmental parameter monitoring module and the compensation algorithm module can effectively reduce the impact of environmental factors on the sensor output and improve the detection accuracy of the system under different environmental conditions. This dynamic compensation method takes into account the combined effects of temperature, humidity and air pressure, allowing the system to adapt to various complex practical application environments and ensure the reliability and consistency of the detection results.

[0276] In a further embodiment, the system also includes: a remote monitoring module, which is used to view the status of the detection signal and the natural gas odorant concentration detection system in real time through a remote device; a data encryption module, which is used to perform one or more of the following: encrypting the transmitted detection signal using an encryption algorithm; using a distributed ledger structure to ensure the non-tamperability of data storage; implementing role-based data access control according to the user roles and permission matrix predefined by the system administrator; an artificial intelligence-assisted decision-making module, which is used to: obtain historical data including historical detection signals, environmental parameters and system operating status, and use a time series analysis method to predict the odorant concentration change trend based on the historical data; based on the prediction results of the odorant concentration change trend and the preset system optimization goals, generate optimization suggestions for the odorant injection strategy through a reinforcement learning algorithm.

[0277] Specifically, the remote monitoring module allows real-time viewing of detection signals and system status through remote devices.

[0278] In the specific implementation, the WebSocket protocol is used to achieve real-time data transmission, the React.js front-end framework is used to develop an interactive web interface, and the Echarts chart library is integrated to achieve data visualization.

[0279] The features of the remote monitoring module are: displaying the detection signals of each sensor and the calculated odorant concentration in a graphical form. Displaying the working status of each module, environmental parameters and system alarm information. Allowing authorized users to adjust system parameters such as sampling frequency, alarm threshold, etc.

[0280] For example, a remote user can log in to the system through a mobile device and view the odorant concentration change curve for the past 24 hours. The current ambient temperature is 25.3°C, the system is running normally, and there is no alarm information.

[0281] Specifically, the data encryption module is responsible for the secure transmission, storage and access control of data. In the specific implementation, data encryption uses the AES-256 algorithm to encrypt the transmitted detection signal and adopts the TLS 1.3 protocol to ensure communication security.

[0282] In the specific implementation, in order to ensure that the data cannot be tampered with, the blockchain technology based on Hyperledger Fabric is used to build a distributed ledger; each data record contains a timestamp and hash value to ensure data integrity.

[0283] In the specific implementation, regarding access control, an access control mechanism based on RBAC (Role-Based Access Control) is implemented; the system administrator predefines user roles (such as operator, administrator, auditor) and permission matrix.

[0284] For example, the operator role can only view real-time data and basic system status; the administrator role can adjust system parameters and view historical data; and the auditor role can access system logs and operation records.

[0285] The AI-assisted decision-making module uses historical data to predict concentration trends and generate optimization suggestions. The specific implementation process includes:

[0286] In the data acquisition and preprocessing steps, historical detection signals, environmental parameters, and system operation status data for the past three months were collected; the sliding window method was used to handle missing values, and the Z-score method was used for data standardization.

[0287] In the prediction of concentration change trend, the LSTM (Long Short-Term Memory) neural network model is used for time series analysis; the model input is the data of the past 72 hours, and the output is the concentration forecast for the next 24 hours.

[0288] In injection strategy optimization, perform the following steps:

[0289] defining optimization objectives, which include maintaining odorant concentration within a target range while minimizing odorant usage;

[0290] Use the Deep Q-Learning algorithm to build a reinforcement learning model;

[0291] Define a state space that contains the environmental information perceived by the model at each moment. In this embodiment, the state space includes the current odorant concentration, predicted future concentration, environmental parameters, etc. These variables together describe the current state of the system and provide a basis for the AI ​​model to make decisions. By obtaining this state information, the model can better assess the current situation and predict future changes, thereby making more reasonable injection adjustment decisions.

[0292] Define an action space that defines the actions that the system can perform in each state. In the odorant injection strategy, the action space includes three possible actions: increase the injection rate, decrease the injection rate, or maintain the current injection rate. These actions allow the system to flexibly adjust the injection rate based on the current and predicted concentration conditions to ensure that the odorant concentration remains within the target range while improving efficiency and saving resources.

[0293] Set the reward function, which is the optimization goal of the reinforcement learning model and is used to evaluate the effect of each action. In the odorant control system, the reward function is based on two key factors: concentration control accuracy and odorant use efficiency. If the system can accurately control the concentration within the target range and use the odorant efficiently, the model will get a higher reward value.

[0294] For example, after analyzing the data from the past three months, the system predicts that the odorant concentration will decrease in the next 24 hours. Based on this prediction and the current inventory situation, the AI ​​module recommends increasing the odorant injection rate by 5% over the next eight hours to maintain the target concentration range.

[0295] By integrating these modules, the natural gas odorant concentration detection system of the embodiment of the present invention can not only accurately detect and control the odorant concentration, but also has remote monitoring, data security and intelligent decision-making capabilities. This greatly improves the availability, safety and efficiency of the system, enabling it to better adapt to the needs of modern natural gas management.

[0296] like Figure 8 As shown, the system also includes: a self-diagnosis module 12, which is used to monitor the working status and performance degradation of each sensor in real time; a fault warning module 14, which is used to determine whether each sensor is abnormal based on the working status and performance degradation of each sensor provided by the self-diagnosis module, wherein the abnormality judgment criteria include: the sensor response time exceeds a preset threshold, the sensitivity decreases by more than 20%, or the baseline drift exceeds 5% of the full scale; an alarm is issued when a sensor abnormality is detected; automatically switching to a backup sensor or adjusting the dynamically weighted weight to reduce the impact of the abnormal sensor, which specifically includes: if there is a backup sensor of the same type, automatically switching to the backup sensor; if there is no backup sensor, then reducing the dynamically weighted weight of the feature importance of the feature vector of the abnormal sensor in the multimodal fusion module.

[0297] By integrating the self-diagnosis module 12 and the fault warning module 14, the natural gas odorant concentration detection system of the embodiment of the present invention can actively identify and respond to sensor performance problems, greatly improving the reliability and stability of the system. This adaptive mechanism enables the system to adjust in time when sensor performance changes, ensuring the accuracy and consistency of the detection results, while also reducing maintenance costs and system downtime.

[0298] In a further embodiment, the system further includes: a gas separation unit, used to separate moisture and / or interfering gas in natural gas before concentration detection; a microfluidic chip, used to receive gas samples processed by the gas separation unit, and control the injection volume of the gas samples of each sensor through a programmable micro valve; the multi-sensor array is arranged in an array on the microfluidic chip, and each type of sensor is provided with multiple sensors, and the activation number of each type of sensor can be set through the software interface.

[0299] The gas separation unit of the embodiment of the present invention is used to separate moisture and / or other interfering gases in natural gas before concentration detection. The gas separation unit adopts multi-stage separation technology, including condensation dehumidification, molecular sieve adsorption and membrane separation. First, a thermoelectric cooler is used to cool the gas sample to below the dew point, and the condensed moisture is discharged through a steam trap. Then, the gas passes through an adsorption column filled with 5A molecular sieves to selectively adsorb residual moisture and other polar molecules (such as CO2, H2S, etc.). Finally, the gas passes through a polyether sulfone (PES) hollow fiber membrane and is further separated by using the difference in the rate at which different gases pass through the membrane. For example, the original natural gas sample contains 2000ppm of moisture and 500ppm of CO2. After being treated by the gas separation unit, the moisture content can be reduced to below 50ppm and the CO2 content can be reduced to below 100ppm, significantly reducing the interference of these components with odorant detection.

[0300] The microfluidic chip of the embodiment of the present invention is used to receive the gas sample after the gas separation unit is processed, and the gas sample injection amount of each sensor is controlled by a programmable micro valve. The chip adopts polydimethylsiloxane (PDMS) material, and the microchannel and valve structure are made using soft lithography technology. The chip integrates a piezoelectrically driven micro valve system, and the valve opening can be accurately controlled by voltage, ranging from 0-100%. The gas distribution network adopts a tree-like branching structure to ensure that the gas sample is evenly distributed to each sensor unit. Each sensor unit is equipped with an independent injection channel and valve, and the valve operation is controlled by an Arduino microcontroller, and the control command is received by communicating with the main control computer through a serial port. For example, in a specific implementation, 12 sensor units are integrated on the microfluidic chip, including 4 MOS sensors, 4 QCM sensors and 4 SAW sensors, and the injection amount of each sensor unit can be independently controlled, ranging from 0-10mL / min.

[0301] The multi-sensor array of the embodiment of the present invention is arranged in an array on the microfluidic chip, and each type of sensor is provided with multiple, and the activation number of each type of sensor can be set through the software interface. The sensors are arranged in a 3x4 matrix, each row represents a type of sensor, and the sensor spacing is 5mm to ensure thermal and electrical isolation. The system has developed a Python-based graphical user interface (Graphical User Interface, GUI), through which any sensor can be selected to be activated or deactivated. The system supports hot plugging, and the sensor configuration can be changed at runtime. The configuration change will automatically trigger system recalibration. For example, in one operation, the initial configuration activated 4 MOS sensors, 3 QCM sensors, and 3 SAW sensors. During operation, the operator deactivated a MOS sensor with degraded performance through the software interface, and activated an additional SAW sensor at the same time. The system automatically adjusted the data fusion algorithm to adapt to the new configuration. This flexible design enables the system to optimize the sensor configuration according to actual needs, improve detection accuracy and system reliability.

[0302] By integrating the gas separation unit and the microfluidic chip and realizing flexible sensor array configuration, the natural gas odorant concentration detection system of the present invention achieves multiple technical effects. First, the gas separation unit effectively removes interfering gases and significantly reduces the influence of cross-sensitivity, thereby greatly improving the accuracy and reliability of detection. Secondly, the programmable microvalves in the microfluidic chip realize precise control of gas sample distribution, so that the system can flexibly adapt to different detection requirements and environmental conditions. Furthermore, the dynamic configuration function of the sensor realized through the software interface allows specific sensors to be activated or deactivated according to actual needs, which not only optimizes resource utilization and extends the overall service life of the system, but also effectively reduces energy consumption. Finally, the system supports hot plugging and dynamic configuration, which greatly improves maintainability and facilitates daily maintenance and system upgrades. The combined effect of these technical effects enables the system to maintain high accuracy and reliability in a complex and changeable natural gas component environment, providing strong technical support for the accurate detection and long-term stable operation of natural gas odorant concentration.

[0303] The technicians in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0304] The units involved in the embodiments of the present invention described in the embodiments of the present invention may be implemented by software or by hardware. The units described may also be arranged in a processor. The above specific implementations do not constitute limitations on the scope of protection of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may occur depending on design requirements and other factors. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention shall be included in the scope of protection of the present invention.

Claims

1. A natural gas odorant concentration detection system based on multi-sensor fusion, characterized in that: include: a multi-sensor array including at least three different types of sensors; A data acquisition module, used to respectively acquire detection signals of each type of sensor in the multi-sensor array; A signal preprocessing module is used to filter, amplify and normalize the collected detection signals of each type of sensor to obtain a preprocessed signal corresponding to each type of sensor; A feature extraction module is used to extract corresponding feature vectors from the preprocessed signals of each type of sensor by wavelet transform and principal component analysis method; A multimodal fusion module, used to fuse the feature vectors of all sensors in the multi-sensor array using a deep neural network algorithm to obtain a fusion feature; a concentration estimation module for performing parallel prediction of the fused features through multiple machine learning regression algorithms to determine the odorant concentration; Wherein, the feature extraction module specifically includes: A wavelet decomposition unit is used to perform wavelet decomposition on the preprocessed signal of each type of sensor to obtain multi-scale wavelet coefficients with a number equal to the order of the wavelet basis function; A characteristic scale selection unit, configured to select, from the multi-scale wavelet coefficients, a number of scales whose energy proportion exceeds a preset proportion of the total energy as characteristic scales; A statistical feature extraction unit, used to extract statistical features of the wavelet coefficients corresponding to the characteristic scale to obtain statistical features including mean, variance, skewness and kurtosis; An initial feature vector generating unit, used for sequentially arranging the statistical features of each feature scale in the order of mean, variance, skewness, and kurtosis, and combining them into an initial feature vector; A principal component analysis unit, used to perform principal component analysis on the initial feature vector and select a principal component whose cumulative contribution rate reaches a preset threshold; A feature vector output unit, used for outputting the selected principal component as a final feature vector; Wherein, the multimodal fusion module specifically includes: A spatial feature extraction unit, used for extracting spatial features from feature vectors of each type of sensor using a convolutional neural network to obtain spatial feature representations of each type of sensor; A temporal dependency analysis unit is used to capture the temporal dependency between the feature vectors of each type of sensor using a long short-term memory network to obtain a temporal correlation feature; A dynamic weighting unit is used to dynamically weight the feature importance of the feature vector of each type of sensor through an attention mechanism to obtain a weighted feature representation; A feature comprehensive processing unit, used for comprehensively processing the spatial feature representation, the temporal correlation feature and the weighted feature representation through a fully connected layer, and outputting a fused feature; Wherein, the concentration estimation module specifically includes: A multi-algorithm prediction unit is used to predict the odorant concentration of the fused features using three target algorithms including a support vector regression algorithm, a random forest algorithm and a gradient boosting decision tree algorithm, respectively, to obtain three preliminary odorant concentration prediction results; A weighted averaging unit, for synthesizing the three preliminary odorant concentration prediction results by a weighted averaging method, wherein the weights are dynamically adjusted according to the root mean square errors of the three target algorithms on the validation set, respectively, to obtain a comprehensive odorant concentration prediction result; A hyperparameter optimization unit, used to automatically adjust the hyperparameters of the three target algorithms by using a Bayesian optimization method to obtain optimized hyperparameters corresponding to the three target algorithms respectively; An optimization prediction unit, used for applying the optimized hyperparameters to each of the three target algorithms accordingly, re-predicting the odorant concentration, and obtaining an optimized comprehensive odorant concentration prediction result; The confidence interval estimation unit is used to perform confidence interval estimation on the optimized comprehensive odorant concentration prediction result by using a self-service sampling method to obtain a final odorant concentration estimation result with a credibility range.

2. The system according to claim 1, characterized in that The multi-sensor array comprises: At least one metal oxide semiconductor sensor for detecting odorant molecules in natural gas and outputting a resistance change signal related to the concentration of the odorant; wherein the odorant molecules react with the surface of the metal oxide semiconductor sensor as a reducing gas; at least one quartz crystal microbalance sensor, configured to detect a mass change of the quartz crystal microbalance sensor caused by adsorption of odorant molecules on the surface of the quartz crystal microbalance sensor and output a frequency signal corresponding to the mass change; at least one surface acoustic wave sensor for detecting a change in acoustic wave propagation characteristics caused by adsorption of odorant molecules on a surface of the surface acoustic wave sensor and outputting an acoustic wave signal corresponding to the change in acoustic wave propagation characteristics; At least one electrochemical sensor is used to detect the current change caused by the redox reaction of the odorant molecules on the electrode surface of the electrochemical sensor and output a current signal corresponding to the current change.

3. The system according to claim 1, characterized in that The signal preprocessing module specifically includes: An adaptive filter is used to automatically adjust filter parameters according to the spectrum characteristics of the detection signal, and use the adjusted filter parameters to remove noise in the detection signals of different types of sensors to obtain a filtered signal; A nonlinear amplification unit, used for nonlinearly amplifying the filtered signal using a piecewise linear function to obtain an amplified signal; A dynamic baseline correction unit, used for real-time monitoring and correction of the baseline drift of each sensor, and performing baseline correction on the amplified signal to obtain a corrected signal; An adaptive normalization processing unit, used for dynamically adjusting normalization parameters according to historical data of each type of sensor, and performing normalization processing on the corrected signal to obtain a normalized signal; The outlier detection and processing unit is used to identify outliers in the normalized signal using a local outlier factor algorithm; and process the identified outliers to obtain the preprocessed signal.

4. The system according to claim 1, characterized in that Also includes: Environmental parameter monitoring module, used to monitor the environmental parameters including temperature, humidity and air pressure of the natural gas odorant concentration detection system in real time; The compensation algorithm module is used to dynamically compensate the detection signal output by each type of sensor according to the environmental parameters, and the dynamic compensation specifically includes: Modeling the sensitivity of each type of sensor to the environmental parameters; Calculating compensation coefficients through a multiple linear regression algorithm according to real-time environmental parameters and the sensitivity model; The compensation coefficient is applied to the original detection signal to obtain a compensated detection signal.

5. The system according to claim 4, characterized in that Also includes: A remote monitoring module, used to view the detection signal and the operating status of the natural gas odorant concentration detection system in real time through a remote device; Artificial intelligence-assisted decision-making module for: Acquiring historical data including historical detection signals, the environmental parameters and the operating status of the system, and using a time series analysis method to predict the trend of odorant concentration changes based on the historical data; According to the prediction result of the odorant concentration variation trend and the preset system optimization goal, an optimization suggestion for the odorant injection strategy is generated through a reinforcement learning algorithm.

6. The system according to claim 1, characterized in that Also includes: Self-diagnosis module, used to monitor the working status and performance degradation of each sensor in real time; Fault warning module, used for: Based on the working status and performance degradation of each sensor provided by the self-diagnosis module, determine whether each sensor is abnormal, wherein the abnormality judgment criteria include: the sensor response time exceeds the preset threshold, the sensitivity decreases by more than 20%, or the baseline drift exceeds 5% of the full scale; Send out alarms when abnormal sensors are detected; Automatically switch to a backup sensor or adjust the dynamically weighted weight to reduce the impact of the abnormal sensor, which specifically includes: if there is a backup sensor of the same type, automatically switch to the backup sensor; if there is no backup sensor, reduce the dynamically weighted weight of the feature importance of the feature vector of the abnormal sensor in the multimodal fusion module.

7. The system according to claim 1, characterized in that Also includes: A gas separation unit for separating water and / or interfering gases from natural gas before concentration detection; A microfluidic chip, used for receiving the gas sample processed by the gas separation unit, and controlling the injection amount of the gas sample of each sensor through a programmable micro valve; The multi-sensor array is arranged in an array on the microfluidic chip, and a plurality of sensors of each type are provided, and the activation number of each type of sensor can be set through a software interface.

Images