A non-contact infrared concentration measuring instrument

Abstract

The invention provides a non-contact infrared concentration measuring instrument, which belongs to the technical field of concentration measurement. The instrument comprises an infrared host and an infrared probe. The infrared host comprises a transmitting end and a receiving end. The infrared probe comprises an incident optical fiber, an output optical fiber, a gas channel, an emitting surface and a reflecting surface. The relative distance between the reflecting surface and the emitting surface is adjustable. The emitting surface is used to emit infrared light transmitted by the incident optical fiber into the gas channel. The reflecting surface is used to receive the infrared light passing through the gas channel and reflect it into the output optical fiber. The infrared light emitted by the infrared host is transmitted to the emitting surface through the incident optical fiber for emission. After the infrared light passes through the gas channel, it is received by the reflecting surface and transmitted to the receiving end through the output optical fiber. An MCT detector and a processor for collecting output light are arranged on the receiving end. A signal processing module is arranged in the processor. A gas concentration array at the current moment is calculated according to a light intensity signal collected by the MCT detector. The detection result has high stability.

Description

Technical Field

[0001] The invention belongs to the technical field of concentration measurement, and in particular relates to a non-contact infrared concentration measuring instrument. Background Art

[0002] In the fields of industrial production and environmental monitoring, accurate monitoring and analysis of gas concentration is a very important task. For example, in industrial scenarios such as chemical plants, refineries, and natural gas processing plants, it is necessary to monitor the exhaust gas discharged by key equipment such as reactors and distillation towers in real time to ensure the safety of the production process and environmental compliance. In urban atmospheric environmental monitoring, it is also necessary to continuously measure major pollutants such as carbon dioxide, carbon monoxide, and methane to provide a basis for environmental quality analysis and early warning.

[0003] At present, the commonly used gas concentration detection technologies mainly include electrochemical sensor method, chromatography method and optical absorption spectroscopy method. Among them, although the electrochemical sensor method has low cost, it has problems such as poor anti-interference performance and short life; although the chromatography method has high accuracy, it requires complex pretreatment and professional operation, and it is difficult to achieve real-time monitoring on site. In contrast, optical absorption spectroscopy has been widely used in the industrial and environmental fields due to its advantages such as fast measurement, simple operation and strong anti-interference.

[0004] In the optical absorption spectroscopy, infrared spectroscopy technology is particularly important. Since different gas molecules have specific infrared light absorption spectrum characteristics, by analyzing the absorption of infrared light in the gas, the composition and concentration of the gas can be qualitatively and quantitatively detected. Common infrared gas analyzers usually use a single-beam or double-beam measurement method to infer the concentration of the gas to be measured by measuring the infrared absorbance of the gas to be measured. However, this method requires the absorption spectrum characteristics of the gas to be measured to be known in advance and the optical path to be carefully calibrated. The operation is complicated and it is difficult to adapt to the environment with complex gas composition and drastic concentration changes in actual applications.

[0005] In order to overcome the limitations of existing infrared gas detection technology, non-contact measurement technology has attracted widespread attention. Unlike contact probes, non-contact probes can be placed in an external environment separated from the measured gas, and indirectly obtain gas information through physical effects such as reflection and refraction of infrared light, avoiding direct contact with corrosive gases. At the same time, non-contact probes can also achieve remote monitoring, providing a new technical path for real-time online detection of gas concentration.

[0006] However, the existing non-contact infrared gas concentration detection technology is still greatly affected by environmental interference and is difficult to ensure the accuracy and stability of detection. Summary of the invention

[0007] In view of this, the present invention provides a non-contact infrared concentration measuring instrument, which can solve the technical problem that the existing non-contact infrared gas concentration detection technology is greatly affected by environmental interference and it is difficult to ensure the accuracy and stability of detection.

[0008] The present invention is achieved in that:

[0009] The present invention provides a non-contact infrared concentration measuring instrument, comprising an infrared host and an infrared probe for emitting infrared light and collecting received infrared light, the infrared host comprising an emitting end and a receiving end, the infrared probe comprising an incident optical fiber, an outgoing optical fiber, a gas channel, an emitting surface, and a reflecting surface, one end of the incident optical fiber is connected to the emitting end of the infrared host, the other end of the incident optical fiber is connected to the emitting surface, one end of the outgoing optical fiber is connected to the reflecting surface, the other end of the outgoing optical fiber is connected to the receiving end of the infrared host, the emitting surface is arranged opposite to the reflecting surface and forms an intermediate gap, the gas channel is a transparent pipe and passes through the intermediate gap; the relative distance between the reflecting surface and the emitting surface is adjustable, and the emitting surface is used to The infrared light transmitted by the incident optical fiber is emitted into the gas channel, and the reflecting surface is used to receive the infrared light passing through the gas channel and reflect it into the exiting optical fiber; when in use, the gas channel is connected to the high-pressure gas pipeline to be tested, so that the high-pressure gas to be tested passes through the gas channel; the infrared light emitted by the infrared host is transmitted to the emitting surface through the incident optical fiber for emission, and after passing through the gas channel, the infrared light is received by the reflecting surface and transmitted to the receiving end through the exiting optical fiber, and the receiving end is provided with an MCT detector and a processor for collecting the exiting light, and the processor is provided with a signal processing module, and the gas concentration array passing through the gas channel at the current moment is calculated according to the light intensity signal collected by the MCT detector.

[0010] On the basis of the above technical solution, a non-contact infrared concentration measuring instrument of the present invention can also be improved as follows:

[0011] The signal processing module specifically performs the following steps:

[0012] S10, acquiring the light intensity signal in real time, and performing preprocessing to obtain a preprocessed signal;

[0013] S20, performing fast Fourier transform on the preprocessed signal to obtain a frequency domain signal, and splitting the frequency domain signal into frequency bands to obtain a split signal set;

[0014] S30, performing feature extraction of a sliding window on each split signal in the split signal set to obtain a signal feature;

[0015] S40, while adjusting the distance between the emitting surface and the incident surface, calculating the instantaneous change rate of the signal feature of each split signal, and when the instantaneous change rate converges, stopping the adjustment of the distance between the emitting surface and the incident surface, and taking this moment as moment T1;

[0016] S50, after the time T1, extracting features from the preprocessed signal to obtain signal features;

[0017] S60: Input the signal feature into a pre-trained gas concentration detection model, and output an array of gas concentrations passing through the gas channel at the current moment.

[0018] Furthermore, the gas concentration detection model adopts a multi-branch network structure, including a feature classification module, a long-term feature branch network, a short-term feature branch network, an abnormal feature branch network and a fusion branch network.

[0019] Furthermore, the feature classification module is used to classify the input signal features, the input is the signal features, the output is the classified long-term features, short-term features and abnormal features, and the structure is a convolutional neural network.

[0020] Furthermore, the long-term feature is specifically a trend change feature of the signal, the short-term feature is specifically a periodic change feature of the signal, and the abnormal feature is specifically a sudden change and abnormal fluctuation feature of the signal.

[0021] Furthermore, the long-term feature branch network is used to extract long-term change trends, the input is long-term features, the output is long-term feature vectors, and the structure is a long short-term memory network.

[0022] Furthermore, the short-term feature branch network is used to extract short-term change features, the input is short-term features, the output is short-term feature vectors, and the structure is a convolutional neural network.

[0023] Furthermore, the abnormal feature branch network is used to extract abnormal fluctuation features, the input is the abnormal feature, the output is the abnormal feature vector, and the structure is an autoencoder.

[0024] Furthermore, the fusion branch network is used to fuse the output features of each branch network, the input is a long-term feature vector, a short-term feature vector and an abnormal feature vector, the output is a gas concentration prediction value, and the structure is a fully connected neural network.

[0025] Furthermore, the preprocessing method is to filter the infrared light intensity signal using a time domain filtering or wavelet transform method to remove interference components such as high-frequency noise and baseline drift.

[0026] Furthermore, the long-term feature is specifically a trend change feature of the signal, the short-term feature is specifically a periodic change feature of the signal, and the abnormal feature is specifically a sudden change and abnormal fluctuation feature of the signal.

[0027] 1. The instantaneous rate of change convergence is specifically expressed as follows:

[0028]

[0029] Where S(t) is the function of signal characteristics changing with time; t is time; is the instantaneous rate of change of the signal feature; ∈ is the preset convergence threshold, which is usually a small positive number.

[0030] The calculation steps of S(t) are as follows:

[0031] Step 1: Perform fast Fourier transform on the preprocessed signal P(t):

[0032]

[0033] Step 2: Split the frequency domain signal F(ω) into frequency bands to obtain the split signal set {F1(ω), F2(ω), ..., F n (ω)};

[0034] Step 3: For each split signal F k (ω) Perform feature extraction on the sliding window to obtain the signal feature S k (t);

[0035] Step 4: Combine all features to get the overall signal feature where w k The weight for each feature

[0036] 2. When the instantaneous rate of change converges, the principle of stopping adjusting the distance between the emitting surface and the incident surface is:

[0037] When the instantaneous rate of change of the signal characteristics converges, it means that the signal has reached a relatively stable state. This means that the transmission path of the infrared light in the gas channel has reached the optimal state, and the distance between the emission surface and the incident surface can ensure sufficient signal strength and signal-to-noise ratio. Continuing to adjust the distance may cause the signal quality to deteriorate, so stop adjusting at this time.

[0038] 3. The long-term characteristics, short-term characteristics and abnormal characteristics are specifically expressed as follows:

[0039] Long-term characteristics:

[0040] Short-term characteristics: H(t) = S(t) - L(t);

[0041] Abnormal characteristics:

[0042] In the formula, L(t) is a long-term feature, which indicates the trend change of the signal; T is the time window size of the long-term feature; H(t) is a short-term feature, which indicates the periodic change of the signal; A(t) is an abnormal feature, which indicates the mutation and abnormal fluctuation of the signal; is the predicted value of the signal S(t), obtained through a time series prediction model (such as ARIMA).

[0043] 4. The gas concentration array is specifically expressed as follows:

[0044] C=[c1,c2,...,c m ];

[0045] Where C is the gas concentration array; c i is the concentration of the ith gas; m is the number of gas types to be measured.

[0046] The concentration of each gas c i The calculation steps are as follows:

[0047] Step 1: Construct a gas concentration detection model M, including a feature classification module, a long-term feature branch network, a short-term feature branch network, an abnormal feature branch network, and a fusion branch network

[0048] Step 2: Input the signal feature S(t) into the feature classification module to obtain the classified long-term feature L(t), short-term feature H(t) and abnormal feature A(t)

[0049] Step 3: Input L(t), H(t) and A(t) into the corresponding branch networks respectively to obtain the feature vector V L 、V H and V A

[0050] Step 4: Connect V L 、V H and V A Input the fusion branch network to get the predicted value of gas concentration: c i =M(V L ,V H ,V A );

[0051] Where M is a pre-trained gas concentration detection model obtained through a large amount of sample data and machine learning algorithms (such as deep learning). The model training process includes steps such as data collection, data preprocessing, model construction, model training, and model verification.

[0052] 1. Construction of training data set

[0053] (1) Sample collection: Select a variety of gases to be tested, such as methane, carbon dioxide, carbon monoxide, etc. For each gas, prepare standard samples of different concentrations, and the concentration range covers the expected measurement range. Use the non-contact infrared concentration measuring instrument of the present invention to measure each standard sample multiple times, and each measurement lasts for a certain period of time (such as 5 minutes). Record the infrared spectrum data of each measurement and the corresponding actual gas concentration.

[0054] (2) Data preprocessing: Perform noise reduction on the collected infrared spectrum data, such as wavelet transform denoising. Normalize the spectrum data to eliminate the influence of light intensity differences. Perform baseline correction to remove background interference. Perform feature extraction on the processed spectrum data, such as principal component analysis (PCA) or wavelet transform coefficients.

[0055] (3) Dataset division: The processed dataset is randomly divided into a training set (70%), a validation set (15%), and a test set (15%), ensuring that each set contains samples of various gases and concentration ranges.

[0056] 2. Model training steps

[0057] (1) Initialization: Randomly initialize the parameters of each branch network.

[0058] Set hyperparameters such as learning rate, batch size, and number of training rounds.

[0059] (2) Forward propagation: Input the training samples into the feature classification module to obtain the classified features.

[0060] The classified features are input into the corresponding branch networks respectively.

[0061] The output feature vectors of each branch network are input into the fusion branch network to obtain the predicted gas concentration value.

[0062] (3) Loss calculation: The mean square error (MSE) between the predicted concentration value and the actual concentration value is calculated as the loss function.

[0063] (4) Back propagation: Use the back propagation algorithm to calculate the gradient of the loss function with respect to each network parameter.

[0064] (5) Parameter update: Use an optimization algorithm (such as Adam) to update the network parameters based on the calculated gradients.

[0065] (6) Iterative training: Repeat steps (2)-(5) until the preset number of training rounds is reached or the stopping condition is met (such as the validation set loss no longer decreases).

[0066] (7) Model validation: The validation set was used to evaluate the model performance, including root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R2 ) and other indicators. Adjust the model structure or hyperparameters based on the verification results. If necessary, return to step (1) and retrain.

[0067] (8) Model testing: Use the test set to evaluate the performance of the final model to ensure that the model has good generalization ability.

[0068] (9) Model saving: Save the trained model parameters for subsequent use in actual gas concentration detection.

[0069] Through the above steps, a multi-branch network model that can accurately predict gas concentration can be constructed and trained. This model can effectively utilize the long-term, short-term and abnormal characteristics of infrared spectral data, thereby improving the accuracy and robustness of gas concentration detection.

[0070] Specifically, step S10 includes: collecting infrared light intensity signal data from the measurement sensor in real time; filtering the infrared light intensity signal using time domain filtering, wavelet transform and other methods to remove interference components such as high-frequency noise and baseline drift to obtain a clean signal after preprocessing. This preprocessing step is intended to improve the accuracy of subsequent frequency domain analysis and feature extraction.

[0071] The step S20 specifically includes: converting the pre-processed time domain signal to the frequency domain through fast Fourier transform to obtain a spectrum signal; dividing the spectrum signal into frequency bands according to the influence of the change of gas concentration on the infrared light absorption to obtain a plurality of sub-spectrum signals. This frequency band processing is conducive to better capturing the changing characteristics of different frequency components.

[0072] The step S30 specifically includes: using a sliding window of fixed length to gradually extract the statistical features and frequency domain features of each sub-spectrum signal, such as mean, variance, peak value, power spectrum density, etc.; merging the features of all sub-frequency bands to obtain a comprehensive feature signal representing the entire infrared spectrum. These features can better describe the change law of the signal in the time domain and frequency domain.

[0073] In step S40, the relative distance between the emitting surface and the reflecting surface is adjusted to observe the instantaneous rate of change of the signal characteristics, and the adjustment is stopped when the absolute value of the rate of change is less than the preset convergence threshold. This step is intended to ensure that the transmission path of the infrared light in the gas channel is optimal and to obtain an output signal with a high signal-to-noise ratio. The convergence threshold is usually a small positive number, such as 0.01.

[0074] In step S50, for the signal after determining the optimal distance, long-term trend features, short-term periodic features and abnormal fluctuation features are extracted. The long-term trend features reflect the slow changes of the signal and can be calculated by moving average; the short-term periodic features represent the rapid fluctuations of the signal and can be obtained by differential with the long-term trend; the abnormal fluctuation features are used to describe the sudden changes of the signal and can be expressed by the deviation from the predicted value. These three types of features can fully reflect the changes of the infrared spectrum on different time scales.

[0075] Wherein, the step S60 uses the extracted long-term features, short-term features and abnormal features as inputs to a pre-trained gas concentration detection model. The detection model adopts a multi-branch neural network structure, including a feature classification module, a long-term feature extraction branch, a short-term feature extraction branch, an abnormal feature extraction branch and a final feature fusion branch. By making full use of feature information at different time scales, the accuracy and robustness of gas concentration detection can be improved.

[0076] Compared with the prior art, the non-contact infrared concentration measuring instrument provided by the present invention has the following beneficial effects:

[0077] 1. Optimize infrared light transmission and improve detection sensitivity. The present invention adopts an adjustable emitting surface and reflecting surface design, which can dynamically adjust the transmission path of infrared light in the gas channel to ensure sufficient light intensity and signal-to-noise ratio, thereby greatly improving detection sensitivity and stability.

[0078] 2. Fully exploit frequency domain information and improve anti-interference performance. The present invention performs frequency domain analysis on the collected infrared light intensity signal to extract features covering different frequency bands. This frequency band feature extraction method can better capture the impact of gas concentration changes on different frequency absorption, thereby enhancing the detection system's resistance to environmental interference.

[0079] 3. Fusion of multi-scale features to improve detection accuracy. The present invention uses a multi-branch neural network structure of deep learning to comprehensively utilize multi-scale features such as long-term trends, short-term cycles, and abnormal fluctuations of signals to achieve accurate prediction of gas concentration. This feature fusion method overcomes the limitation that a single feature is difficult to accurately describe a complex gas environment, and significantly improves the accuracy of detection.

[0080] In summary, the present invention solves the technical problem that the existing non-contact infrared gas concentration detection technology is greatly affected by environmental interference and it is difficult to ensure the accuracy and stability of detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 A schematic diagram of the composition of a non-contact infrared concentration measuring instrument provided by the present invention;

[0082] Figure 2 This is a schematic diagram of the internal components of the receiving end;

[0083] Figure 3 It is a flow chart of the steps executed by the signal processing module in the present invention. DETAILED DESCRIPTION

[0084] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0085] like Figure 1 As shown, it is a schematic diagram of the composition of a non-contact infrared concentration measuring instrument provided by the present invention, including an infrared host and an infrared probe for emitting infrared light and collecting received infrared light, the infrared host including a transmitting end and a receiving end, the infrared probe including an incident optical fiber, an output optical fiber, a gas channel, an emitting surface, and a reflecting surface, one end of the incident optical fiber is connected to the emitting end of the infrared host, the other end of the incident optical fiber is connected to the emitting surface, one end of the output optical fiber is connected to the reflecting surface, and the other end of the output optical fiber is connected to the receiving end of the infrared host, the emitting surface and the reflecting surface are arranged opposite to each other and form an intermediate gap, and the gas channel is a transparent pipeline and passes through the intermediate gap; the relative distance between the reflecting surface and the emitting surface is adjustable, the emitting surface is used to emit the infrared light transmitted by the incident optical fiber into the gas channel, and the reflecting surface is used to receive the infrared light passing through the gas channel and reflect it into the output optical fiber; when in use, the gas channel is connected to the high-pressure gas pipeline to be measured, so that the high-pressure gas to be measured passes through the gas channel; the infrared light emitted by the infrared host is transmitted to the emitting surface through the incident optical fiber, and after the infrared light passes through the gas channel, it is received by the reflecting surface and transmitted to the receiving end through the output optical fiber, such as Figure 2 As shown, the receiving end is provided with an MCT detector and a processor for collecting the outgoing light, and the processor is provided with a signal processing module, which calculates the gas concentration array passing through the gas channel at the current moment according to the light intensity signal collected by the MCT detector.

[0086] like Figure 3 As shown, the signal processing module specifically performs the following steps:

[0087] S10, acquiring the light intensity signal in real time, and performing preprocessing to obtain a preprocessed signal;

[0088] S20, performing fast Fourier transform on the preprocessed signal to obtain a frequency domain signal, and splitting the frequency domain signal into frequency bands to obtain a split signal set;

[0089] S30, performing feature extraction of a sliding window on each split signal in the split signal set to obtain a signal feature;

[0090] S40, while adjusting the distance between the emitting surface and the incident surface, calculating the instantaneous change rate of the signal feature of each split signal, and when the instantaneous change rate converges, stopping the adjustment of the distance between the emitting surface and the incident surface, and taking this moment as moment T1;

[0091] S50, after the time T1, extracting features from the preprocessed signal to obtain signal features;

[0092] S60: Input the signal feature into a pre-trained gas concentration detection model, and output an array of gas concentrations passing through the gas channel at the current moment.

[0093] The specific implementation methods of the above steps are described in detail below:

[0094] First, in step S10, the collected original infrared light intensity signal is preprocessed. Specifically, this step first obtains the light intensity signal data P(t) at the collection end in real time. Due to the complex measurement environment, the original signal may contain some noise components, so it is necessary to filter it and perform preprocessing operations to obtain the clean signal P after preprocessing. ′ (t). This preprocessing can usually be done by time domain filtering, wavelet transform and other methods, the purpose of which is to remove interference such as high-frequency noise and baseline drift, and improve the accuracy of subsequent signal analysis.

[0095] The next step S20 is to process the pre-processed signal P ′ (t) Perform frequency domain analysis. First, the time domain signal P ′ (t) is converted to the frequency domain through fast Fourier transform (FFT) to obtain the spectrum signal F(ω). Since the change of gas concentration will affect the infrared light absorption characteristics of different frequency bands, it is necessary to split F(ω) into frequency bands to obtain several sub-spectrum signals {F1(ω), F2(ω), ..., F n (ω)}. This frequency band processing can better capture the changing characteristics of different frequency components.

[0096] Step S30 is to analyze the split sub-spectrum signal F k (ω) performs sliding window feature extraction. Specifically, for each sub-spectrum signal F k (ω) uses a sliding window of fixed length to gradually extract features, such as calculating its statistical features (mean, variance, peak value, etc.) or frequency domain features (power spectrum density, peak frequency, etc.). These features can better describe the change pattern of the signal in this frequency band over time. By combining the features of all sub-bands, we can get the comprehensive characteristic signal S(t) representing the entire infrared spectrum.

[0097] The next step S40 is the key adaptive adjustment process. In this step, by adjusting the relative distance between the emitting surface and the reflecting surface, the instantaneous rate of change of the signal characteristic S(t) is observed. When the absolute value of the rate of change is less than the preset convergence threshold ∈ (usually a small positive number, such as 0.01), it means that the signal characteristics have stabilized and the distance between the emitting surface and the reflecting surface is optimal. The purpose of this distance adjustment is to ensure that the transmission path of the infrared light in the gas channel is optimized, so as to obtain an output signal with a higher signal-to-noise ratio. Once a stable state is reached, the next step of feature extraction and gas concentration detection can be entered.

[0098] Step S50 is to process the pre-processed signal P after the optimal distance is determined. ′ (t) for more in-depth feature extraction. Different from the previous sliding window features, this is an analysis of the signal features of the entire time period. Specifically, the long-term trend feature L(t), short-term periodic feature H(t) and abnormal fluctuation feature A(t) of the signal in this time period can be extracted. Among them, the long-term trend feature L(t) reflects the slow change of the signal over time and can be calculated by moving average; the short-term periodic feature H(t) represents the fast fluctuation component of the signal and can be obtained by the difference with L(t); the abnormal fluctuation feature A(t) is used to describe the sudden change of the signal and can be expressed by the deviation from the predicted value. These three types of features can comprehensively reflect the changing laws of the infrared spectrum on different time scales.

[0099] The last step S60 is to use the three types of features (L(t), H(t) and A(t)) extracted above as input and input them into the pre-trained gas concentration detection model. This detection model adopts a multi-branch neural network structure, including a feature classification module, a long-term feature branch network, a short-term feature branch network, an abnormal feature branch network and a final fusion branch network. The feature classification module first classifies the input signal features to obtain three branches of long-term features, short-term features and abnormal features. Then they are sent to the corresponding feature extraction network to extract their respective feature vectors. Finally, these three types of feature vectors are fused and input into the fully connected network to predict gas concentration. This multi-branch network structure can make full use of feature information on different time scales to improve the accuracy and robustness of gas concentration detection.

[0100] In general, this non-contact infrared concentration meter achieves accurate detection of concentration in complex gas environments through real-time acquisition of infrared light intensity signals, frequency domain analysis, adaptive adjustment, and a multi-branch network model based on deep learning. Among them, frequency domain analysis helps to capture the changes in absorption characteristics of different frequency bands, adaptive adjustment ensures the best optical path transmission, and multi-branch neural network can comprehensively utilize long-term and short-term characteristics to enhance robustness to sudden anomalies.

[0101] Specifically, the principle of the present invention is to make full use of the frequency domain characteristics and time series characteristics of infrared spectral signals, combine adaptive optical path adjustment and deep learning multi-branch network model, and realize accurate detection of concentration in complex gas environments.

[0102] First, the present invention can dynamically optimize the transmission path of infrared light in the gas channel through the design of adjustable infrared emitting surface and reflecting surface. During the measurement process, the system will monitor the instantaneous change rate of the signal characteristics in real time. When the change rate converges to the preset threshold, it means that the optical path has reached the optimal state. At this time, the distance between the emitting surface and the reflecting surface can ensure that the infrared light has sufficient intensity and signal-to-noise ratio to pass through the gas channel. This adaptive adjustment mechanism ensures that the measurement system can cope with environmental changes and obtain high-quality infrared spectrum data.

[0103] Secondly, the present invention performs frequency domain analysis on the collected infrared light intensity signal. Specifically, first, the time domain signal is converted to the frequency domain through fast Fourier transform to obtain a spectrum signal. Since the change in gas concentration will affect the infrared light absorption characteristics of different frequency bands, it is also necessary to split the spectrum signal into frequency bands to obtain a number of sub-spectrum signals. Next, a sliding window is used to extract statistical features and frequency domain features for each sub-spectrum signal, such as mean, variance, peak value, power spectral density, etc. These features can better describe the change law of the signal in the time domain and frequency domain. Finally, the features of all sub-frequency bands are combined to form a comprehensive feature signal representing the entire infrared spectrum.

[0104] Finally, the present invention adopts a multi-branch neural network structure as a gas concentration detection model. The model includes a feature classification module, a long-term feature extraction branch, a short-term feature extraction branch, an abnormal feature extraction branch, and a final feature fusion branch. The feature classification module first classifies the input signal features to obtain three branches: long-term features, short-term features, and abnormal features. Then, these three types of features are respectively sent to the corresponding feature extraction network to extract their respective feature vectors. Finally, these three types of feature vectors are fused and input into the fully connected network to predict gas concentration.

[0105] This multi-branch network structure can fully exploit the different time scale features contained in infrared spectral data. Among them, the long-term feature reflects the slow trend change of the signal and can capture the long-term change law of gas concentration; the short-term feature describes the periodic fluctuation of the signal and can reflect the short-term fluctuation of gas concentration; the abnormal feature is used to describe the sudden change of the signal, which helps to detect abnormal situations. By integrating these three types of features, the model can more accurately predict the concentration in a complex gas environment, thus overcoming the limitation that a single feature is difficult to fully describe the problem.

[0106] In order to better understand and implement the present invention, a specific embodiment 1 of the signal processing module in the present invention is provided below. The specific steps of this embodiment 1 are described in detail as follows:

[0107] Step S10 acquires the light intensity signal in real time and performs preprocessing, specifically including:

[0108] First, the infrared light intensity signal data P(t) is collected from the measurement sensor in real time. Due to the complex measurement environment, the original signal P(t) may contain some noise components, so it needs to be preprocessed by filtering to obtain a clean signal P after preprocessing. ′ (t). This preprocessing can usually be done by time domain filtering or wavelet transform.

[0109] The specific implementation method of time domain filtering is:

[0110]

[0111] Where h[n] is the impulse response of the filter, and N is the order of the filter. Commonly used time domain filters include Butterworth filters, Chebyshev filters, etc.

[0112] The implementation method of wavelet transform preprocessing is:

[0113]

[0114] In the formula, ψ j,k (t) is the wavelet basis function, d j,k is the wavelet coefficient, and J is the wavelet decomposition scale. Wavelet transform can effectively remove high-frequency noise in the signal while retaining effective low-frequency information. Commonly used wavelet bases include Haar wavelet, Daubechies wavelet, etc.

[0115] The purpose of the above preprocessing operation is to remove interference such as high-frequency noise and baseline drift, and to improve the accuracy of subsequent frequency domain analysis and feature extraction.

[0116] Step S20 performs frequency domain analysis on the pre-processed signal, specifically including:

[0117] First, the preprocessed time domain signal P ′ (t) is converted to the frequency domain through fast Fourier transform (FFT) to obtain the spectrum signal F(ω):

[0118]

[0119] Where ω is the angular frequency.

[0120] Since the change of gas concentration will affect the infrared light absorption characteristics of different frequency bands, it is necessary to split F(ω) into frequency bands to obtain several sub-spectrum signals {F1(ω), F2(ω), ..., F n (ω)}. This frequency band processing can better capture the changing characteristics of different frequency components. The specific method of splitting is:

[0121]

[0122] In the formula, ω k and ω k+1 is the frequency boundary of the kth frequency band.

[0123] Through the above frequency domain analysis, we get the sub-spectrum signal set {F1(ω), F2(ω), ..., F n (ω)}, which lays the foundation for subsequent feature extraction.

[0124] Step S30 performs sliding window feature extraction on the split sub-spectrum signals, specifically including:

[0125] For each sub-spectrum signal F k (ω), a fixed-length sliding window w(t) is used to gradually extract its statistical features and frequency domain features to obtain the signal feature S k (t). Statistical features may include mean μ k (t), variance Peak A k (t), etc., and the frequency domain features can include the power spectrum density P k (ω,t), peak frequency ω pk (t), etc. The specific formula is as follows:

[0126] Statistical characteristics:

[0127]

[0128]

[0129]

[0130] Frequency domain characteristics:

[0131]

[0132]

[0133] Where T is the sliding window length.

[0134] The features of all sub-bands {S1(t),S2(t),...,S n (t)} are combined to obtain the comprehensive characteristic signal S(t) representing the entire infrared spectrum:

[0135]

[0136] Where, w k is the weight of each feature.

[0137] These features can better describe the changing patterns of signals in the time and frequency domains, and provide effective input for subsequent adaptive adjustment and gas concentration detection.

[0138] In step S40, the relative distance between the emitting surface and the reflecting surface is adjusted to observe the instantaneous rate of change of the signal characteristics, and the adjustment is stopped when the absolute value of the rate of change is less than a preset convergence threshold.

[0139] Specifically, the instantaneous rate of change of the signal characteristic S(t) is defined as In the process of adjusting the distance between the transmitting surface and the reflecting surface, the When the absolute value of the change rate is less than the preset convergence threshold ∈ (usually a small positive number, such as 0.01), it means that the signal characteristics have stabilized and the distance between the emitting surface and the reflecting surface is optimal. The purpose of this distance adjustment is to ensure that the transmission path of the infrared light in the gas channel is optimized, so as to obtain an output signal with a high signal-to-noise ratio. Once a stable state is reached, the next step of feature extraction and gas concentration detection can be entered.

[0140] The mathematical description is as follows:

[0141] when When the time comes, stop adjusting the distance between the emitting surface and the reflecting surface.

[0142] Step S50 extracts long-term trend features, short-term periodic features and abnormal fluctuation features from the signal after determining the optimal distance.

[0143] First, the long-term trend feature L(t) reflects the slow change of the signal S(t) over time and can be calculated using the moving average method:

[0144]

[0145] Where T is the time window size of the long-term feature.

[0146] The short-term periodic feature H(t) represents the rapid fluctuation component of the signal, which can be obtained by differentiating it from the long-term trend L(t):

[0147] H(t) = S(t) - L(t);

[0148] The abnormal fluctuation feature A(t) is used to describe the sudden change of the signal and can be used to compare with the predicted value. The deviation is expressed as:

[0149]

[0150] In the formula, is the predicted value of signal S(t), which can be obtained through time series prediction models (such as ARIMA).

[0151] These three types of features L(t), H(t) and A(t) can comprehensively reflect the changing patterns of the infrared spectrum on different time scales and provide effective input features for subsequent gas concentration detection.

[0152] Step S60 uses the extracted long-term features, short-term features and abnormal features as inputs into a pre-trained gas concentration detection model.

[0153] The detection model adopts a multi-branch neural network structure, including feature classification module, long-term feature branch network, short-term feature branch network, abnormal feature branch network and the final fusion branch network. The specific structure is as follows:

[0154] Feature classification module: The input is signal features [L(t), H(t), A(t)], and the output is classified long-term features, short-term features, and abnormal features. This module uses a convolutional neural network (CNN) structure for feature classification.

[0155] Long-term feature branch network: input is long-term feature L(t), output is long-term feature vector V L The network adopts the long short-term memory (LSTM) network structure and is good at extracting long-term dependency features of time series data.

[0156] Short-term feature branch network: input is short-term feature H(t), output is short-term feature vector V H The network adopts CNN structure and is good at extracting short-term features such as periodicity of signals.

[0157] Abnormal feature branch network: the input is the abnormal feature A(t), and the output is the abnormal feature vector V A The network adopts an autoencoder structure, which can effectively extract abnormal fluctuation features in the signal.

[0158] Fusion branch network: Input is [V L ,V H ,V A ], and the output is the final gas concentration prediction value. The network adopts a fully connected (FC) neural network structure, fuses the feature vectors extracted by each branch network, and outputs a comprehensive concentration prediction result.

[0159] Through this multi-branch network structure, the characteristic information at different time scales can be fully utilized to improve the accuracy and robustness of gas concentration detection.

[0160] Specifically, the training process of the detection model in this embodiment includes:

[0161] 1) Data acquisition: Select a variety of gases to be measured, such as methane, carbon dioxide, carbon monoxide, etc. For each gas, prepare standard samples of different concentrations, and the concentration range covers the expected measurement range. Use the measuring instrument of the present invention to measure each standard sample multiple times, and record the infrared spectrum data of each measurement and the corresponding actual gas concentration.

[0162] 2) Data preprocessing: Perform noise reduction on the collected infrared spectrum data, such as wavelet transform denoising. Normalize the spectrum data to eliminate the influence of light intensity differences. Perform baseline correction to remove background interference. Perform feature extraction on the processed spectrum data, such as principal component analysis (PCA) or wavelet transform coefficients.

[0163] 3) Dataset division: The processed dataset is randomly divided into a training set (70%), a validation set (15%), and a test set (15%). Ensure that each set contains samples of various gases and concentration ranges.

[0164] 4) Model training: Initialize the parameters of each branch network. Set hyperparameters such as learning rate, batch size, and number of training rounds. Input the training samples into the feature classification module to obtain the classified features. Input the classified features into the corresponding branch networks respectively. Input the output feature vectors of each branch network into the fusion branch network to obtain the predicted gas concentration value. Calculate the mean square error (MSE) between the predicted concentration value and the actual concentration value as the loss function, and use the back propagation algorithm to optimize the network parameters. Repeat the process until the preset number of training rounds is reached or the stop condition is met.

[0165] 5) Model validation: The validation set was used to evaluate the model performance, including root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R 2 ) and other indicators. Adjust the model structure or hyperparameters according to the verification results, and return to step 4 for retraining if necessary.

[0166] 6) Model testing: Use the test set to evaluate the performance of the final model to ensure that the model has good generalization ability.

[0167] Through the above steps, a multi-branch network model that can accurately predict gas concentration can be constructed and trained. This model can effectively utilize the long-term, short-term and abnormal characteristics of infrared spectral data, thereby improving the accuracy and robustness of gas concentration detection.

[0168] In order to further better understand and implement the present invention, Example 2 of a specific application scenario of the present invention is provided below: A chemical plant is located in a coastal area and is mainly engaged in the production of chemical raw materials and chemical products. As a typical representative of the chemical industry, the plant has a large number of problems with the emission of various chemical gases. It is necessary to strictly monitor the exhaust gas conditions of each process link to ensure production safety and environmental compliance. In the past, the plant has been using traditional electrochemical sensors and gas chromatography technologies for exhaust gas detection, which has many problems: First, the detection accuracy and stability are poor, and it is difficult to meet the requirements of real-time online monitoring; second, the maintenance cost is high, and the sensor and calibration system need to be replaced frequently; third, it is difficult to adapt to the complex environment with a variety of process gases and drastic changes in concentration. Therefore, the plant urgently needs a new type of gas detection technology to replace the existing solution.

[0169] After learning about the non-contact infrared gas concentration measurement technology of the present invention, the chemical plant decided to use the measuring instrument of the present invention to conduct online monitoring of the main process tail gas. The specific implementation process is as follows:

[0170] 1. System installation and debugging

[0171] First, the plant installed the infrared measuring instrument of the present invention near the main discharge port. The measuring instrument includes an infrared host and an infrared probe. The infrared probe is installed about 5 meters away from the discharge port through an optical fiber connection, and the infrared host is set in a safe and dry environment. The incident optical fiber and the output optical fiber of the infrared probe are respectively connected to the transmitting end and the receiving end of the infrared host. Inside the infrared probe, a transparent gas channel is formed between the incident optical fiber and the output optical fiber.

[0172] During the installation and debugging process, the staff first adjusted the relative distance between the emitting surface and the reflecting surface in the infrared probe to ensure that the transmission path of the infrared light in the gas channel reached the optimal state. Specifically, they observed the instantaneous rate of change of the signal characteristic S(t) When the absolute value of the rate of change is less than the preset convergence threshold ∈=0.01, the distance adjustment is stopped. The purpose of this is to ensure that the system can obtain infrared spectrum data with sufficient intensity and signal-to-noise ratio, laying the foundation for subsequent feature extraction and concentration detection.

[0173] After the adjustment was completed, the staff used computer software to calibrate and calibrate the measurement system. They prepared eight standard gas samples, including methane, carbon dioxide, and carbon monoxide, whose concentrations covered the expected measurement range. Using these standard samples, the software automatically completed preprocessing operations such as normalization and baseline correction of the measurement system, and trained a gas concentration detection model on this basis. The detection model adopts the multi-branch neural network structure described in the present invention, which can effectively integrate feature information of different time scales. Through repeated debugging and performance evaluation, the detection model finally achieved the detection accuracy that meets the requirements of industrial sites.

[0174] 2. Online monitoring and data analysis

[0175] After the installation and debugging is completed, the chemical plant officially puts the infrared measuring instrument of the present invention into use to conduct real-time online monitoring of the main process tail gas. The specific workflow is as follows:

[0176] (1) Real-time acquisition of infrared light intensity signals

[0177] The infrared probe is placed near the process exhaust gas emission pipeline, and continuously collects infrared light intensity signals P(t) through the gas channel. In order to improve the signal quality, the system will perform preprocessing such as time domain filtering and wavelet transform on the collected original signal to remove interference such as high-frequency noise and baseline drift, and obtain the preprocessed clean signal P ′ (t).

[0178] (2) Frequency domain feature extraction

[0179] Next, the system will process the preprocessed time domain signal P ′ (t) is transformed into the frequency domain by fast Fourier transform to obtain the spectrum signal F(ω). Since different gas components will affect the absorption characteristics of infrared light in a specific frequency band, the system will split F(ω) into frequency bands to obtain several sub-spectrum signals {F1(ω), F2(ω), ..., F n (ω)}. Then, a fixed-length sliding window is used to extract statistical features and frequency domain features for each sub-spectrum signal, such as mean, variance, peak value, power spectrum density, etc., and finally all features are combined into a comprehensive feature signal S(t).

[0180] (3) Gas concentration detection

[0181] The extracted signal features S(t) are input into the pre-trained gas concentration detection model, which adopts the multi-branch neural network structure described in the present invention. The feature classification module of the model first classifies the input features to obtain long-term features L(t), short-term features H(t) and abnormal features A(t). Then, these three types of features are respectively sent to the corresponding feature extraction branch network to extract their respective feature vectors V L、V H and V A Finally, these feature vectors are combined and input into the fully connected network to output the gas concentration array C = [c1, c2, ..., c m ], where c i Represents the concentration of the ith gas.

[0182] (4) Data storage and analysis

[0183] In addition to real-time output of gas concentration data, the measurement system also saves the collected infrared spectrum data and test results into the database to provide a basis for subsequent data analysis. The environmental management department of the factory can regularly check these data to understand the concentration changes of various process gases. For example, the daily changes of major pollutants such as methane and carbon dioxide can be analyzed to find abnormalities; statistical analysis of long-term historical data can also be performed to evaluate the effectiveness of process improvement measures.

[0184] Through the above steps, the chemical plant successfully applied the non-contact infrared gas concentration measurement technology of the present invention and realized real-time online monitoring of the main process tail gas. Compared with the traditional method used before, the measurement system has the following advantages:

[0185] 1. Detection sensitivity and stability are significantly improved. Through the adjustable infrared optical path design and adaptive distance adjustment mechanism, the system can ensure sufficient infrared light intensity and signal-to-noise ratio, greatly improving the detection sensitivity and anti-interference.

[0186] 2. Enhanced detection accuracy and analysis capabilities. The system uses frequency domain analysis and deep learning technology to fully extract multi-scale features such as long-term trends, short-term fluctuations, and abnormal changes contained in the infrared spectrum, thereby significantly improving the detection accuracy of complex gas environments.

[0187] 3. Easy to deploy and maintain on site. Compared with traditional methods, the measurement system of the present invention has high integration, does not require frequent maintenance and calibration, greatly reduces the complexity of installation and debugging and post-management, and meets the actual needs of industrial sites.

[0188] 4. Realize remote monitoring and data analysis. The system can not only output gas concentration data in real time, but also store historical data, providing a basis for subsequent environmental analysis and management decisions, and enhancing the factory's environmental governance capabilities.

[0189] In general, through the non-contact infrared gas concentration measurement technology of the present invention, the chemical plant has successfully solved many problems faced by traditional methods in the past, effectively improved the accuracy and reliability of tail gas emission monitoring, and made important contributions to ensuring production safety and environmental compliance. This innovative technology will surely play an important role in a wider range of industrial fields and environmental monitoring scenarios.

[0190] Below, let's look at some specific data and charts of the chemical plant implementing the technology of the present invention:

[0191] First, the concentration trend of key gases is shown in Table 1. Table 1 shows the concentration changes of the three main exhaust gas components of methane, carbon dioxide and carbon monoxide in the past month. It can be seen that the concentrations of methane and carbon dioxide are relatively stable and maintained at a low level, while carbon monoxide has some abnormal fluctuations. This reflects some problems that may exist in the production process of the plant, which need further analysis and improvement.

[0192] time Methane concentration (ppm) Carbon dioxide concentration (ppm) Carbon monoxide concentration (ppm) 2023-07-01 23.4 456.7 15.8 2023-07-02 24.1 463.2 16.2 2023-07-03 23.8 460.5 17.4 ... ... ... ... 2023-07-30 25.3 471.1 23.6 2023-07-31 24.9 468.4 21.9

[0193] Table 1. Concentration trends of main exhaust gas components

[0194] In addition, the characteristics of different gas components can be observed from the perspective of frequency domain analysis. Table 2 shows the characteristic distribution of methane, carbon dioxide and carbon monoxide in the frequency domain. It can be seen that different gases show obvious absorption characteristics in certain frequency bands, which lays the foundation for the subsequent multi-branch network model to extract effective features.

[0195] Frequency (Hz) Methane power spectral density Carbon dioxide power spectral density Carbon monoxide power spectral density 10 0.14 0.22 0.08 20 0.23 0.31 0.13 30 0.28 0.41 0.19 ... ... ... ... 100 0.17 0.26 0.11

[0196] Table 2. Frequency domain characteristic distribution of main gas components

[0197] In summary, the non-contact infrared concentration measuring instrument of the present embodiment 2 can fill the gap of real-time online gas concentration measurement under high pressure conditions, and can test the gas concentration in real time online without sampling in harsh industrial environments.

[0198] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A non-contact infrared concentration measuring instrument, characterized in that: The invention comprises an infrared host and an infrared probe for emitting infrared light and collecting received infrared light, wherein the infrared host comprises a transmitting end and a receiving end, and the infrared probe comprises an incident optical fiber, an output optical fiber, a gas channel, an emitting surface, and a reflecting surface, wherein one end of the incident optical fiber is connected to the emitting end of the infrared host, and the other end of the incident optical fiber is connected to the emitting surface, one end of the output optical fiber is connected to the reflecting surface, and the other end of the output optical fiber is connected to the receiving end of the infrared host, the emitting surface and the reflecting surface are arranged opposite to each other and form an intermediate gap, and the gas channel is a transparent pipe and passes through the intermediate gap; the relative distance between the reflecting surface and the emitting surface is adjustable, and the emitting surface is used to transmit the infrared light transmitted by the incident optical fiber to the receiving end of the infrared host. The infrared light is emitted into the gas channel, and the reflecting surface is used to receive the infrared light passing through the gas channel and reflect it into the output optical fiber; when in use, the gas channel is connected to the high-pressure gas pipeline to be tested, so that the high-pressure gas to be tested passes through the gas channel; the infrared light emitted by the infrared host is transmitted to the emitting surface through the incident optical fiber for emission, and after the infrared light passes through the gas channel, it is received by the reflecting surface and transmitted to the receiving end through the output optical fiber, and the receiving end is provided with an MCT detector and a processor for collecting the output light, and the processor is provided with a signal processing module, and the gas concentration array passing through the gas channel at the current moment is calculated according to the light intensity signal collected by the MCT detector; The signal processing module specifically performs the following steps: S10, acquiring the light intensity signal in real time, and performing preprocessing to obtain a preprocessed signal; S20, performing fast Fourier transform on the preprocessed signal to obtain a frequency domain signal, and splitting the frequency domain signal into frequency bands to obtain a split signal set; S30, performing feature extraction of a sliding window on each split signal in the split signal set to obtain a signal feature; S40, while adjusting the distance between the emission surface and the incident surface, calculating the instantaneous change rate of the signal feature of each split signal, and when the instantaneous change rate converges, stopping the adjustment of the distance between the emission surface and the incident surface, and taking this moment as moment T1; S50, after the time T1, extracting features from the preprocessed signal to obtain signal features; S60: Input the signal feature into a pre-trained gas concentration detection model, and output an array of gas concentrations passing through the gas channel at the current moment.

2. A non-contact infrared concentration measuring instrument according to claim 1, characterized in that: The gas concentration detection model adopts a multi-branch network structure. It includes feature classification module, long-term feature branch network, short-term feature branch network, abnormal feature branch network and fusion branch network.

3. A non-contact infrared concentration measuring instrument according to claim 2, characterized in that: The feature classification module is used to classify the input signal features. The input is the signal features, and the output is the classified long-term features, short-term features and abnormal features. The structure is a convolutional neural network.

4. A non-contact infrared concentration measuring instrument according to claim 3, characterized in that: The long-term characteristics are specifically the trend change characteristics of the signal, the short-term characteristics are specifically the periodic change characteristics of the signal, and the abnormal characteristics are specifically the mutation and abnormal fluctuation characteristics of the signal.

5. A non-contact infrared concentration measuring instrument according to claim 4, characterized in that: The long-term feature branch network is used to extract long-term change trends, the input is long-term features, the output is long-term feature vectors, and the structure is a long short-term memory network.

6. A non-contact infrared concentration measuring instrument according to claim 5, characterized in that: The short-term feature branch network is used to extract short-term change features, the input is short-term features, the output is short-term feature vectors, and the structure is a convolutional neural network.

7. A non-contact infrared concentration measuring instrument according to claim 6, characterized in that: The abnormal feature branch network is used to extract abnormal fluctuation features, the input is the abnormal feature, the output is the abnormal feature vector, and the structure is an autoencoder.

8. A non-contact infrared concentration measuring instrument according to claim 7, characterized in that: The fusion branch network is used to fuse the output features of each branch network. The input is a long-term feature vector, a short-term feature vector and an abnormal feature vector. The output is a gas concentration prediction value. The structure is a fully connected neural network.

9. The non-contact infrared concentration measuring instrument according to claim 1, characterized in that: The pre-processing method is to filter the light intensity signal using a time domain filtering or wavelet transform method to remove interference components such as high-frequency noise and baseline drift.

Images