A TDLAS gas remote sensing system and method for suppressing albedo mutation

Abstract

The present application relates to a kind of TDLAS gas telemetry systems and methods for inhibiting albedo mutation, belong to the technical field of spectral analysis and telemetry.The present application generates original signal sequence by combination;Signal quality factor is output by preset multi-factor weighted decision algorithm;Distortion insensitive feature vector is generated by splicing;First gas concentration fitting value is calculated;Second gas concentration estimate value is mapped out by executing forward propagation operation;Final gas concentration measurement value is output;Online tuning is executed to machine learning regression model.The present application solves the problem that existing TDLAS gas telemetry technology is invalid in the scene of the albedo mutation of reflecting surface, the concentration inversion mode of model fitting is easy to fail, resulting in inaccurate or interrupted measurement results.

Description

Technical Field

[0001] This invention belongs to the technical field of spectral analysis and telemetry, and in particular to a TDLAS gas telemetry system and method for suppressing abrupt changes in albedo. Background Technology

[0002] In scenarios such as industrial safety, environmental monitoring, and pipeline inspection, tunable diode laser absorption spectroscopy (TDLAS) technology is widely used for long-distance, non-contact remote sensing of gas concentration. This technology inverts gas concentration by analyzing the spectral absorption characteristics of the laser after reflection from a distant reflector. However, in actual remote sensing, the reflector is often an uncontrollable object such as the ground or building exterior walls, and its optical characteristics change dynamically. In particular, when the probe spot sweeps across a medium with a large difference in albedo, it can cause abrupt changes in the intensity of the echo signal and the baseline. This optical interference caused by factors other than gas concentration poses a challenge to the reliability of the measurement results.

[0003] To improve the stability and accuracy of TDLAS detection, relevant optimized technical solutions have emerged in the industry. For example, a trace hydrogen concentration sensing system based on TDLAS absorption spectroscopy, with application number 202510528087.8, is proposed. This solution constructs a symmetrical dual-optical-path structure to simultaneously extract the fundamental and second harmonic signals of the target optical path and the reference optical path. After normalization processing, the ratio of the main signal and the compensation ratio are formed. Combined with a piecewise fitting model, the system dynamically compensates for the power drift of the light source and the optical path disturbance, achieving high linearity detection over a wide concentration range.

[0004] While the aforementioned existing technologies can mitigate some of the effects of system drift and optical path disturbances through dual-optical-path compensation and model fitting, their core still relies on a fitting algorithm based on the overall signal amplitude envelope for concentration inversion. When encountering abrupt changes in the albedo of the reflective surface that cause severe distortion of the second harmonic signal morphology, such fitting algorithms are prone to convergence anomalies due to difficulty in finding the optimal solution. This results in instantaneous spikes or null values ​​in the output concentration values, failing to guarantee the continuity and accuracy of measurements under extreme reflection conditions. Consequently, they are ill-suited for open-path telemetry scenarios with complex and variable reflective surface conditions.

[0005] Existing TDLAS gas telemetry technology is prone to failure in scenarios with sudden changes in reflectivity of the reflective surface, and the concentration inversion method that relies on model fitting is prone to failure, resulting in inaccurate or interrupted measurement results. This is the core technical problem that this invention needs to solve. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and propose a TDLAS gas telemetry system and method to suppress abrupt changes in albedo. This solves the problem that the concentration inversion method relying on model fitting is prone to failure in scenarios with abrupt changes in reflective surface albedo in existing TDLAS gas telemetry technology, leading to inaccurate or interrupted measurement results.

[0007] The technical problem solved by this invention is achieved through the following technical solution: A TDLAS gas telemetry system for suppressing abrupt changes in albedo includes a raw signal sequence construction module, a signal quality factor calculation module, a distortion-insensitive feature extraction module, a model-driven inversion module, a data-driven estimation module, a dual-drive causal decision module, and an online model tuning module, wherein the raw signal sequence construction module, signal quality factor calculation module, distortion-insensitive feature extraction module, model-driven inversion module, data-driven estimation module, dual-drive causal decision module, and online model tuning module are connected sequentially. The original signal sequence construction module is used to acquire the echo optical signal from the probe end and perform phase-locked amplification to obtain a continuous second harmonic signal that represents the current scanning cycle, and align it with the preset background modulation waveform template to generate the original signal sequence. The signal quality factor calculation module is used to extract the signal-to-noise ratio, waveform symmetry ratio, baseline drift rate, and waveform distortion of the background modulation waveform template of the continuous second harmonic signal in the original signal sequence, and outputs the signal quality factor through a preset multi-factor weighted decision algorithm. The distortion-insensitive feature extraction module is used to perform point-to-point subtraction mapping on the original signal sequence to generate a differential waveform sequence, extract local waveform segments and locate local peak and valley points and inflection points, calculate the relative amplitude ratio of peaks and valleys and the curvature parameter of inflection points, and splice them to generate a distortion-insensitive feature vector. The model-driven inversion module is used to inject continuous second harmonic signals into a nonlinear evolution model containing linear parameters, and apply nonlinear least squares iterative fitting to calculate the fitted value of the first gas concentration. The data-driven estimation module is used to input distortion-insensitive feature vectors into a pre-trained machine learning regression model and perform forward propagation to map out the estimated value of the second gas concentration. The dual-drive causal decision module performs dual-drive causal decision based on the fitted value of the first gas concentration and the estimated value of the second gas concentration, and outputs the final gas concentration measurement value. The online model tuning module counts the frequency of triggering events that are adopted for the second gas concentration estimate. When the frequency of triggering events exceeds the frequency accumulation threshold within a single time window, the module encapsulates the failed training samples using the fitting residual parameters of the distortion-insensitive feature vector and the fitted value of the first gas concentration, and performs online tuning on the machine learning regression model.

[0008] A telemetry method for a TDLAS gas telemetry system that suppresses abrupt changes in albedo includes the following steps: Step 1: The original signal sequence construction module acquires the echo optical signal from the detector end and performs phase-locked amplification to obtain a continuous second harmonic signal representing the current scanning period. It then aligns the signal with the preset background modulation waveform template and combines them to generate the original signal sequence. Step 2: The signal quality factor calculation module extracts the signal-to-noise ratio, waveform symmetry ratio, baseline drift rate, and waveform distortion of the background modulation waveform template of the continuous second harmonic signals in the original signal sequence, and outputs the signal quality factor through a preset multi-factor weighted decision algorithm. Step 3: The distortion-insensitive feature extraction module performs point-to-point subtraction mapping on the original signal sequence to generate a differential waveform sequence, extracts local waveform segments and locates local peak and valley points and inflection points, calculates the relative amplitude ratio of peaks and valleys and the curvature parameter of inflection points, and splices them to generate a distortion-insensitive feature vector. Step 4: The model-driven inversion module injects the continuous second harmonic signal into the nonlinear evolution model containing linear parameters, applies the nonlinear least squares method for iterative fitting, and calculates the fitted value of the first gas concentration. Step 5: The data-driven estimation module inputs the distortion-insensitive feature vector into the pre-trained machine learning regression model and performs forward propagation to map out the estimated value of the second gas concentration. Step 6: The dual-drive causal decision module introduces a signal quality factor as a path selection verification criterion, combines the fitted value of the first gas concentration with the estimated value of the second gas concentration to perform dual-drive causal decision, and outputs the final gas concentration measurement value. Step 7: The online model tuning module counts the frequency of triggering events that accept the second gas concentration estimate. When the frequency of triggering events exceeds the frequency accumulation threshold within a single time window, the failed training samples are encapsulated using the fitting residual parameter of the distortion-insensitive feature vector and the fitted value of the first gas concentration, and online tuning is performed on the machine learning regression model.

[0009] Furthermore, step 1 includes the following steps: Step 1.1: The original signal sequence construction module controls the laser emitting component to output a wavelength-modulated probe laser beam to the target area. After being reflected by the target reflective surface, the echo light signal is collected by the receiving optical system. Step 1.2: After photoelectric conversion and lock-in amplification of the echo optical signal, obtain the continuous second harmonic signal characterizing the current scanning cycle; Step 1.3: Call the system's preset background modulation waveform template, which is generated using the same set of modulation parameters under controlled ideal reflective surface conditions with no gas absorption; Step 1.4: Align the scan period timestamps of the continuous second harmonic signals and the background modulation waveform template in the time domain dimension, and combine them to generate the original signal sequence that enters the subsequent processing link.

[0010] Furthermore, step 2 includes the following steps: Step 2.1: The signal quality factor calculation module performs an overall cross-correlation operation between the continuous second harmonic signal and the background modulation waveform template to calculate the waveform distortion between the two. Step 2.2: Using a preset multi-factor weighted decision algorithm, the signal-to-noise ratio parameter, waveform symmetry ratio, baseline drift rate, and waveform distortion are assigned corresponding weights and summed to output a continuously quantitative signal quality factor. in, For signal quality factor, The weighting coefficients are floating-point numbers preset in the system configuration file. This is the signal-to-noise ratio parameter value. This is the waveform symmetry ratio value. Baseline drift rate, This represents the waveform distortion.

[0011] Furthermore, step 3 includes the following steps: Step 3.1: The distortion-insensitive feature extraction module applies a point-to-point subtraction mapping operation to the aligned continuous second harmonic signals in the current time domain and the background modulation waveform template to generate a differential waveform sequence corresponding to the current scanning period. Step 3.2: Configure the frequency domain anchor point of the target gas absorption line center frequency, and extract local waveform segments of the differential waveform sequence within a preset bandwidth range before and after the frequency domain anchor point; Step 3.3: Traverse and locate the local peak points, local valley points, and inflection points of valley transitions within the local waveform segment. Calculate the peak-valley relative amplitude ratio using the level values ​​of the local peak points and local valley points. in, The ratio of peak to valley relative amplitude. This refers to the level value at the local peak point; This refers to the level value at a local valley point. Calculate the curvature parameters at the inflection point: in, For curvature parameters, The first derivative represents the slope of the waveform at that point; The second derivative represents the rate of change of curvature of the waveform at that point; by concatenating the peak-to-valley relative amplitude ratio with the curvature parameter, a distortion-insensitive feature vector is generated to quantify the topological correlation of the local waveform.

[0012] Furthermore, step 4 includes the following steps: Step 4.1: Load the model-driven inversion module with a nonlinear evolution model containing the target gas absorption linear parameters and laser scanning modulation parameters; Step 4.2: Inject the continuous second harmonic signal into the nonlinear evolution model as the only observation input data, and apply the nonlinear least squares algorithm to perform multi-parameter iterative fitting on the overall amplitude envelope of the observation input data; Step 4.3: Obtain the iterative parameters after the nonlinear least squares algorithm has reached convergence, and calculate the first gas concentration fitting value for the current scanning cycle by substituting them according to the Lambert-Beer law relationship. in, To fit the obtained concentration integral value, The absorption coefficient obtained by fitting; The absorption line strength is determined by the target gas.

[0013] Furthermore, step 5 includes the following steps: Step 5.1: The data-driven estimation module loads the machine learning regression model that has been pre-trained using multiple simulated reflective surface albedo distortion samples, closes the input interface channel, and only allows the input of specified low-dimensional feature parameters; Step 5.2: Input the obtained distortion-insensitive feature vector as the only active activation feature into the mapping layer of the machine learning regression model to perform forward propagation operation; Step 5.3: Receive the estimated scalar data from the output layer of the machine learning regression model and store it in a cache queue as a second gas concentration estimate to mitigate the current waveform distortion. in, This is the estimated value for the second gas concentration. Using distortion-insensitive feature vectors as input data, For machine learning regression models, The hidden layer weight matrix is... This is the output layer weight matrix. This is the hidden layer bias vector; For output layer bias; For output layer activation / linear transformation, This is the activation function for the hidden layer.

[0014] Furthermore, step 6 includes the following steps: Step 6.1: The dual-drive causal decision module responds to the decision result that the value of the signal quality factor is not less than the waveform tolerance threshold, determines that the nonlinear evolution model has physical applicability, extracts the first gas concentration fitting value from the buffer queue and configures it as the final gas concentration measurement value for system output. Step 6.2: In response to the judgment that the value of the signal quality factor is less than the waveform tolerance threshold, the real-time verification logic is triggered to extract the absolute value of the difference between the fitted value of the first gas concentration and the estimated value of the second gas concentration; if the absolute value of the difference is greater than the preset reasonable deviation limit, it is determined that the nonlinear evolution model has failed under the current extreme reflection conditions, and the estimated value of the second gas concentration is adopted as the final gas concentration measurement value for system output.

[0015] Furthermore, step 7 includes the following steps: Step 7.1: The online model tuning module continuously monitors the frequency of trigger events that use the second gas concentration estimate as the final gas concentration measurement value in the background. Step 7.2: When the frequency of the triggering event exceeds the frequency accumulation threshold within a single time window, read the distortion-insensitive feature vector of the current scanning cycle and the fitting residual parameter of the first gas concentration fitting value, and encapsulate them as a heterogeneous failure training sample. Step 7.3: Import the incremental training samples of heterogeneous failures into the data update pool to drive the machine learning regression model to trigger the weight back-to-back optimization iteration based on extreme value deviation, and complete the data flow closed loop that takes into account both the measurement output in harsh environments and the adaptive evolution of the system environment.

[0016] The advantages and positive effects of this invention are: 1. This invention achieves a quantitative assessment of the reliability of each frame of echo signal by constructing an original signal sequence containing gas absorption information along the optical path and analyzing the global waveform index of the sequence to calculate the signal quality factor. Based on the signal quality factor determination results, the system constructs a dual-path data processing architecture that combines model-driven and data-driven approaches. An optimal decision mechanism selects the effective concentration calculation results. When sudden changes in reflectivity of the reflective surface cause severe signal distortion or model-driven path failure, the system can switch to the data-driven path to complete the concentration estimation. This maintains the continuity of gas concentration measurement under harsh reflection conditions and reduces data interruptions or concentration output deviations caused by instantaneous deterioration of signal quality.

[0017] 2. This invention constructs a process for extracting and applying distortion-insensitive feature vectors. First, background interference is removed through differential operations. Then, focusing on the core region of the absorption spectrum, local geometric features such as the peak-to-valley amplitude ratio and inflection point curvature are extracted and a low-dimensional feature vector is constructed. These local features have normalization properties for the overall signal amplitude scaling and baseline shift. Using them as the sole input to a machine learning model ensures that gas concentration estimation decisions are based solely on the spectral morphology itself. This decouples the concentration estimation process from the overall signal amplitude fluctuations caused by reflectivity abrupt changes, improving the stability of the gas concentration inversion results.

[0018] 3. This invention introduces a closed-loop mechanism for online model tuning. In the system's background, it continuously monitors the frequency of failure events in the model-driven path. When the system is determined to be in a persistently severe reflection environment, it automatically captures the distortion-insensitive feature vectors and the accepted data-driven path concentration estimates for that scenario, constructs new training samples, and incrementally tunes the machine learning regression model. This approach enables the concentration estimation capability of the data-driven path to continuously adapt to the specific interference environment currently faced by the system, achieving self-evolution and self-optimization of system performance during continuous operation, allowing the system's concentration estimation capability to adapt to the reflection environment characteristics in practical applications. Attached Figure Description

[0019] Figure 1 This is a structural diagram of the system of the present invention; Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0020] The present invention will be further described in detail below with reference to the accompanying drawings.

[0021] A TDLAS gas telemetry system that suppresses abrupt changes in albedo, such as Figure 1 As shown, it includes a raw signal sequence construction module, a signal quality factor calculation module, a distortion-insensitive feature extraction module, a model-driven inversion module, a data-driven estimation module, a dual-drive causal decision module, and an online model tuning module, wherein the raw signal sequence construction module, signal quality factor calculation module, distortion-insensitive feature extraction module, model-driven inversion module, data-driven estimation module, dual-drive causal decision module, and online model tuning module are connected in sequence; The original signal sequence construction module is used to acquire the echo optical signal from the probe end and perform phase-locked amplification to obtain a continuous second harmonic signal that represents the current scanning cycle, and align it with the preset background modulation waveform template to generate the original signal sequence. The signal quality factor calculation module is used to extract the signal-to-noise ratio, waveform symmetry ratio, baseline drift rate, and waveform distortion of the background modulation waveform template of the continuous second harmonic signal in the original signal sequence, and outputs the signal quality factor through a preset multi-factor weighted decision algorithm. The distortion-insensitive feature extraction module is used to perform point-to-point subtraction mapping on the original signal sequence to generate a differential waveform sequence, extract local waveform segments and locate local peak and valley points and inflection points, calculate the relative amplitude ratio of peaks and valleys and the curvature parameter of inflection points, and splice them to generate a distortion-insensitive feature vector. The model-driven inversion module is used to inject continuous second harmonic signals into a nonlinear evolution model containing linear parameters, and apply nonlinear least squares iterative fitting to calculate the fitted value of the first gas concentration. The data-driven estimation module is used to input distortion-insensitive feature vectors into a pre-trained machine learning regression model and perform forward propagation to map out the estimated value of the second gas concentration. The dual-drive causal decision module introduces a signal quality factor as a path selection verification criterion. It performs dual-drive causal decision based on the first gas concentration fitting value and the second gas concentration estimation value, and outputs the final gas concentration measurement value. The online model tuning module counts the frequency of triggering events that are adopted for the second gas concentration estimate. When the frequency of triggering events exceeds the frequency accumulation threshold within a single time window, the module encapsulates the failed training samples using the fitting residual parameters of the distortion-insensitive feature vector and the fitted value of the first gas concentration, and performs online tuning on the machine learning regression model.

[0022] A telemetry method for a TDLAS gas telemetry system that suppresses abrupt changes in albedo, such as Figure 2 As shown, it includes the following steps: Step 1: The original signal sequence construction module acquires the echo optical signal from the probe end and performs phase-locked amplification to obtain a continuous second harmonic signal representing the current scanning period. It then aligns the signal with a preset background modulation waveform template and combines them to generate the original signal sequence.

[0023] Step 1, as the initial stage of data acquisition and preprocessing, is executed by the original signal sequence construction module. The original signal sequence construction module sends a preset composite modulation signal to the laser driving circuit through its digital output port. The composite modulation signal includes a low-frequency sawtooth wave signal for broadband scanning and a high-frequency sine wave signal for wavelength modulation.

[0024] After receiving the drive, the laser emitting component emits a probe laser beam whose wavelength dynamically changes with the composite modulation signal towards the target area. This probe laser beam is reflected by an externally uncertain target reflective surface and converged by the receiving optical system onto the surface of a photodetector. The photodetector converts the received echo light signal into an analog current signal in real time.

[0025] The analog current signal is then fed into a lock-in amplifier. The lock-in amplifier uses the harmonics of the high-frequency sine wave signal as a reference to demodulate the second harmonic component related to the gas absorption characteristics and outputs a continuous second harmonic signal characterizing the current scan period. The scan period refers to the time required for the low-frequency sawtooth wave signal to complete a full scan. Its setting is based on the need to fully cover one or more absorption spectral characteristics of the target gas. The typical setting range is 10 ms to 100 ms, preferably 20 ms. The setting is based on the target gas, such as methane, whose absorption spectral full width at half maximum (FWHM) is about 1 GHz. The laser scan range needs to cover 1.5 times the spectral width. 20 ms can balance the scan integrity and real-time performance. The continuous second harmonic signal is a one-dimensional time-series voltage array. Each element represents the voltage value output by the lock-in amplifier at the sampling time. The overall shape of this array contains both the modulation characteristics of the laser itself and the absorption line shape information of the target gas in the optical path.

[0026] The continuous second harmonic signal is digitized at a fixed sampling rate via a high-speed analog-to-digital converter and stored in a first-in-first-out (FIFO) buffer queue. Simultaneously, the original signal sequence construction module accesses the onboard non-volatile memory, calls and reads the pre-calibrated background modulation waveform template. This template is a reference data sequence acquired, averaged, and then stored under a standardized ideal reflective surface in a controlled experimental environment without target gas absorption, using the exact same modulation parameters as the current online measurement. It should be noted that the background modulation waveform template is a one-dimensional time-series digital array of the same length as the continuous second harmonic signal. The shape of this array only reflects the inherent response curve of the system under conditions without gas absorption and serves as the reference for subsequent differential processing.

[0027] The original signal sequence construction module uses the starting trigger edge of the low-frequency sawtooth wave signal as a synchronization reference to precisely locate the starting point of the continuous second harmonic signal data segment of the current scan cycle in the buffer queue in the time domain dimension, and aligns it with the starting point of the data sequence of the background modulation waveform template in terms of index. After alignment, the original signal sequence construction module encapsulates these two discrete data sequences of the same length into a structured data pair, which together constitute the original signal sequence entering the subsequent signal quality assessment and feature extraction links. This original signal sequence is a data structure containing the continuous second harmonic signal of the current scan cycle and the background modulation waveform template, providing the basic data input for all subsequent processing steps.

[0028] Assuming the current system has a scan period of 20 ms and an analog-to-digital converter (ADC) sampling rate of 50 kS / s, the original signal sequence construction module triggers a new scan at timestamp t0. During the time interval from t0 to t0+20 ms, the ADC acquires 1000 voltage sampling points, forming an array containing 1000 floating-point numbers, i.e., a continuous second harmonic signal, denoted as . For example, its value is [-10.2, -10.5, ..., 120.4, ..., -9.8] mV, where 120.4 mV is the harmonic peak value after being affected by gas absorption.

[0029] The original signal sequence construction module loads an array containing 1000 floating-point numbers, i.e., the background modulation waveform template, from its non-volatile memory, denoted as . For example, its value is [-10.1, -10.3, ..., 150.7, ..., -9.7] mV, where 150.7 mV is the harmonic peak value without gas absorption. The system creates a data object to store the continuous second harmonic signals. and background modulation waveform template This object is stored as two parallel members, with a timestamp t0 appended as metadata. With template The data object, as a raw signal sequence, is passed to the processing task that performs subsequent steps. Step 2: The signal quality factor calculation module extracts the signal-to-noise ratio, waveform symmetry ratio, baseline drift rate, and waveform distortion of the background modulation waveform template of the continuous second harmonic signal in the original signal sequence, and outputs the signal quality factor through a preset multi-factor weighted decision algorithm.

[0030] The main execution unit in this step is the signal quality factor calculation module. After receiving the original signal sequence constructed in the previous step, it starts the global waveform index analysis thread. This analysis thread performs multi-dimensional feature extraction on the continuous second harmonic signals in the original signal sequence to calculate four core indicators: signal-to-noise ratio (SNR), waveform symmetry ratio, baseline drift rate, and waveform distortion. Based on these four indicators, it calculates the signal quality factor (SQF). The analysis thread calls a preset multi-factor weighted decision algorithm to multiply the previously calculated SNR, waveform symmetry ratio, baseline drift rate, and waveform distortion by preset weight coefficients. Then, it sums all the products to generate a single, continuous, and quantitative scalar value, which is used to dynamically characterize the current reflector state and signal reliability. This factor is then output to the subsequent path optimization decision module. for: Among them, the signal-to-noise ratio parameter value in the formula Needs to go through a function After processing, x) is used in the calculation. This function is a logarithmic transformation or threshold normalization function, which is used to map the signal-to-noise ratio parameter value with a large dynamic range to the interval between 0 and 1, so as to balance its influence in the weighted summation. The weighting coefficients are floating-point numbers preset in the system configuration file, with a sum of 1. Their specific values ​​are determined based on statistical analysis of a large amount of experimental data, aiming to reflect the primary and secondary relationships of the influence of different indicators on the final inversion accuracy. In application scenarios where the signal-to-noise ratio is the main limiting factor, they can be set... The weights range from 0.4 to 0.6, based on the fact that the signal-to-noise ratio has the greatest impact on the accuracy of concentration inversion, with the remaining weights allocated in descending order of importance.

[0031] Signal-to-noise ratio parameter value The signal-to-noise ratio (SNR) is a core indicator for evaluating signal clarity. This thread selects a baseline segment that does not contain gas absorption features at the beginning and end of the continuous second harmonic signal. The preferred length of the baseline segment is 50 sampling points, corresponding to 1 ms, with a sampling rate of 50 kS / s. The basis for this setting is that 50 sampling points can accurately calculate the noise level, fewer than 30 points will result in too large a statistical error, and more than 100 points will contain gas absorption features. The standard deviation of these data points is calculated as the noise level, and the maximum amplitude value in the entire signal sequence is located as the signal peak value. The ratio of the two is determined as the SNR parameter value.

[0032] Waveform symmetry ratio This reflects whether there is asymmetric distortion of the absorption spectrum caused by factors such as optical interference. Its ideal value is 1. The thread calculates the waveform integral area in the preset windows on the left and right sides of the signal peak point as the center, and takes the ratio of the smaller area to the larger area as the waveform symmetry ratio.

[0033] Baseline drift rate The stability of the signal baseline is quantified; a smaller value indicates higher signal quality. Therefore, it is used in the weighted summation. The form of the curve ensures that its contribution to the signal quality factor is consistent with other indicators. The thread calculates the average level of the starting baseline segment and the average level of the ending baseline segment, and then normalizes the absolute value of the difference between the two with the signal peak value to obtain the baseline drift rate.

[0034] This thread retrieves the background modulation waveform template from the original signal sequence, takes the continuous second harmonic signal and the background modulation waveform template as input, performs a standardized full-sequence cross-correlation operation, and obtains the maximum correlation coefficient value between 0 and 1. This maximum correlation coefficient value is defined as the waveform distortion degree characterizing the similarity between the two. The waveform distortion By comparing with an ideal template, the overall waveform distortion caused by factors such as abrupt changes in reflectivity was macroscopically evaluated, with an ideal value of 1.

[0035] The signal quality factor calculation module receives a signal containing continuous second harmonics. and background modulation waveform template The original signal sequence. Signal quality factor calculation module analysis. Given an array with a peak value of 120.4 mV and a baseline noise standard deviation of 0.6 mV, the original signal-to-noise ratio is 200.67. This is achieved through a mapping function. (x) is obtained after normalization. It is 0.92; the calculation yields The ratio of the integral areas on both sides of the peak is 0.96, which is the waveform symmetry ratio. It is 0.96; calculated The ratio of the average difference between the starting and ending baseline levels to the peak value is 0.05, which is the baseline drift rate. It is 0.05.

[0036] Module computing and The cross-correlation coefficient is used to obtain the waveform distortion. The value is 0.88. The system retrieves the preset weights, assuming they are respectively... =0.4, =0.2, =0.1, =0.3, the signal quality factor is calculated by weighted summation. =0.4×0.92+0.2×0.96+0.1×(1-0.05)+0.3×0.88=0.919. This signal quality factor of 0.919 will be passed on to subsequent steps as a basis for decision-making.

[0037] Step 3: The distortion-insensitive feature extraction module performs point-to-point subtraction mapping on the original signal sequence to generate a differential waveform sequence, extracts local waveform segments and locates local peak and valley points and inflection points, calculates the relative amplitude ratio of peaks and valleys and the curvature parameter of inflection points, and splices them to generate a distortion-insensitive feature vector.

[0038] This step is performed by the distortion-insensitive feature extraction module, which aims to extract deep topological features from the original signal sequence that are insensitive to changes in the overall signal amplitude. The distortion-insensitive feature extraction module retrieves the time-domain aligned continuous second harmonic signals and the background modulation waveform template from the original signal sequence, and applies an element-wise point-to-point subtraction mapping operation to these two one-dimensional arrays. This operation subtracts the corresponding sampling point value from the background modulation waveform template at each sampling point value of the continuous second harmonic signal, thereby generating a differential waveform sequence. It should be noted that the differential waveform sequence is a representation of the pure gas absorption signal after eliminating the nonlinear modulation effect of the laser.

[0039] The distortion-insensitive feature extraction module loads preset frequency domain anchors from its internal parameter library. It should be understood that the frequency domain anchors are pre-calibrated based on the target gas characteristic absorption frequencies in a standard spectral database such as HITRAN. This module converts these frequency domain anchors into timestamp index positions on the differential waveform sequence, and uses these indexes as centers to extract local waveform segments containing the core features of the absorption lines according to a preset bandwidth range. This preset bandwidth range is set based on the broadening characteristics of the target gas absorption spectral lines under specific temperature and pressure, and is usually 1.5 to 3 times the full width at half maximum (FWHM) of the spectral lines, preferably 2 times the FWHM. The setting is based on the fact that the FWHM of the target gas spectral lines under standard atmospheric pressure is approximately 500 MHz. 2 times the bandwidth can completely contain the peak, valley, and inflection point features of the absorption peaks, 1.5 times is prone to truncation, and 3 times will introduce irrelevant noise, ensuring that the extracted local waveform segments completely contain the peak, valley, and inflection point features required for analysis.

[0040] The distortion-insensitive feature extraction module applies a one-dimensional morphological analysis algorithm to the local waveform segment. First, it performs amplitude normalization, converting the waveform level values ​​into dimensionless relative intensities. Then, it calculates the first and second derivatives to traverse and locate local peaks, local valleys, and inflection points during the upward transition of the valleys within the segment. The distortion-insensitive feature extraction module extracts the level values ​​of the local peaks and valleys, calculates their absolute difference, and compares it with the absolute level value of the valley to obtain the peak-valley relative amplitude ratio. The formula is as follows: Among them, the peak-to-valley relative amplitude ratio It characterizes the relative proportion of absorption depth, and its scaling on the overall signal amplitude has normalization properties. This refers to the level value at the local peak point; This represents the voltage level at a local valley point. The processing module calculates the curvature parameter of the waveform curve at the inflection point. It should be noted that the curvature parameter quantifies the degree of bending of the waveform at the inflection point, reflecting the rate of absorption linear change, and is similarly insensitive to changes in the overall signal amplitude. Satisfy the following formula: In the curvature parameter formula, The first derivative represents the slope of the waveform at that point, i.e., the rate of change of voltage with respect to the sampling point / time. The second derivative represents the rate of change of the curvature of the waveform at that point, i.e., the rate of change of the slope, and characterizes its concavity or convexity.

[0041] The distortion-insensitive feature extraction module concatenates the calculated peak-to-valley relative amplitude ratio and curvature parameters into a two-dimensional vector in a predetermined order, generating and outputting a distortion-insensitive feature vector that can quantify the local waveform topological correlation. It should be understood that the distortion-insensitive feature vector is a low-dimensional feature representation that can stably characterize the morphology of the gas absorption core. Even when the amplitude of the original signal changes drastically due to abrupt changes in reflectivity, the value of this vector can remain relatively stable for use in subsequent data-driven paths.

[0042] The distortion-insensitive feature extraction module received arrays and The array processing module performs point-to-point subtraction to generate a differential waveform sequence. = - For example, at the absorption peak, the difference is 120.4 mV - 150.7 mV = -30.3 mV. Assume the system's preset frequency domain anchor point corresponds to the 502nd index position in a sequence of 1000 sampling points, and the preset bandwidth corresponds to 50 sampling points before and after it. Therefore, the processing module truncates... A total of 101 data points from index 452 to 552 constitute a local waveform segment.

[0043] The processing module analyzes this data segment and locates a local valley point at index 502, with a voltage level of [value missing]. The voltage is -30.3mV; and a local peak is located at index 480, with a voltage level of -30.3mV. The value is -5.2 mV. The processing module calculates the peak-to-valley relative amplitude ratio. ,get The value is 0.828. Simultaneously, the processing module finds an inflection point at index 515, performs amplitude normalization on the local waveform segment, and calculates the first derivative at that point. The second derivative is 0.5. The value is 0.18. Substituting this into the formula yields the curvature parameter. ,get The result is 0.129. The system concatenates these two values ​​to generate and output a distortion-insensitive feature vector of [0.828, 0.129]. This vector will be sent to the data-driven estimation path for processing.

[0044] Step 4: The model-driven inversion module injects the continuous second harmonic signal into the nonlinear evolution model containing linear parameters, applies the nonlinear least squares method for iterative fitting, and calculates the fitted value of the first gas concentration.

[0045] This step, as one of the dual-drive paths, is the model-driven inversion path, executed in parallel by the model-driven inversion module. The model-driven inversion module loads a pre-built nonlinear evolution model from its firmware library. This nonlinear evolution model is a function describing the physical process of second harmonic signal generation. It encapsulates the absorption Voigt line shape parameters of the target gas and the laser scanning modulation parameters set during system operation. The Voigt line shape parameters include the absorption line center frequency, spectral line intensity, and pressure broadening coefficient, etc. These parameters are obtained from standard spectral databases such as HITRAN and pre-set in the model. The laser scanning modulation parameters are known constants set by the system, such as modulation depth and scan amplitude.

[0046] The processing module extracts the continuous second harmonic signals from the original signal sequence generated in the previous step, treating them as a complete one-dimensional array, and injects them into the nonlinear evolution model as the sole observation input data. The processing module then initiates an iterative fitting program based on the nonlinear least squares algorithm. This algorithm is an optimization technique used to find the optimal solution in a multidimensional parameter space, maximizing the fit between the model output and the actual observation data. In this embodiment, the Levenburg-Marquardt algorithm is used. This algorithm uses the overall amplitude envelope of the observation input data as the fitting target, and gas concentration, baseline offset, and phase difference as variable fitting parameters. By iteratively adjusting these parameters, it drives the nonlinear evolution model to generate the theoretical second harmonic waveform and minimizes the sum of squared residuals between the theoretical waveform and the observation input data, satisfying the following formula: in, This is the integral value of gas concentration, in ppm·m; This is the baseline offset; This is the phase offset. For the first The measured second harmonic signal values ​​at each sampling point are expressed in V. (·) represents the theoretical level value output by the nonlinear evolution model at the corresponding time, also in V; No. The laser frequency at each sampling time.

[0047] When the algorithm determines that the fit has converged, the criteria for convergence are: the relative rate of change of the sum of squared residuals between two consecutive iterations is less than a preset threshold, or the algorithm reaches the maximum number of iterations, at which point the fitting process terminates. The preset threshold can be set to 10. -6The threshold was set based on the premise that the relative change rate of the sum of squared residuals in the fitting is ≤0.0001%, ensuring that the concentration inversion error is ≤1%, which meets the accuracy requirements of industrial remote sensing. The model-driven inversion module extracts the specific iterative parameter of gas concentration from the converged fitting results and converts it into a concentration integral value with physical units according to the Lambert-Beer law relationship embedded in the model. The Lambert-Beer law is a fundamental law of absorption spectroscopy, which establishes the concentration... absorption coefficient The linear relationship between the two values ​​is defined as the fitted value of the first gas concentration for the current scan cycle, and the concentration conversion relationship is shown in the following formula: in, The integral value of the fitted concentration (the first gas concentration fitted value) is expressed in ppm·m and reflects the combined effect of gas concentration and optical path. The absorption coefficient obtained by fitting; The absorption line intensity of the target gas is expressed in ppm·m. -1 It should be noted that the first gas concentration fitting value is the output of this physical model driving path. It represents the most likely concentration value that can be calculated based on physical laws under the current signal conditions, and is stored in the designated data buffer area, waiting to be called in subsequent steps.

[0048] The model-driven inversion module receives continuous second harmonic signals. An array with elements of value [-10.2, ..., 120.4, ..., -9.8] mV is used as the observation input data. The processing module loads a nonlinear evolution model for the target gas, methane. And initialize the iteration parameters, such as the initial gas concentration. 0 = 0 ppm·m. The Levenberg-Marquardt algorithm is initiated; in the first iteration, the nonlinear evolution model... based on 0 = 0 ppm·m produces the theoretical waveform with no absorption; calculate its relationship with... The residual sum of squares. The algorithm adjusts the concentration value based on the gradient information. 1 = 80 ppm·m, recalculated, and found that the sum of squared residuals decreased. This iterative process continued, for example, after 12 iterations, the concentration parameter converged to 12 =131.2 ppm·m, at which point the rate of change of the sum of squared residuals is less than 10. -6The fitting process terminates. The processing module extracts the converged concentration parameter value and, based on the Lambert-Beer law relationship built into the model, formally calculates 131.2 ppm·m as the first gas concentration fitting value. This value of 131.2 ppm·m is written into a specific cache unit for use in subsequent path optimization verification steps.

[0049] Step 5: The data-driven estimation module inputs the distortion-insensitive feature vector into the pre-trained machine learning regression model and performs forward propagation to map out the estimated value of the second gas concentration.

[0050] This step, as an alternate data processing path parallel to the model-driven inversion path, is scheduled and executed by the data-driven estimation module. The data-driven estimation module loads a pre-trained machine learning regression model from its non-volatile memory into memory. This model was developed during the development phase by simulating various abrupt changes in reflectivity of reflective surfaces in a laboratory environment, collecting a large number of distorted waveform samples, and calculating the corresponding distortion-insensitive feature vector for each sample along with the known real gas concentration. This supervised learning training resulted in the training dataset.

[0051] The regression model can employ a feedforward neural network. The software interface of this model is specially designed to form a closed input interface channel. The closed input interface channel is a software-level mandatory constraint that ensures the purity of the data-driven path, making it completely dependent on the extracted robust features and avoiding interference from the original distorted signal. That is, its prediction function can only receive and process low-dimensional feature parameters of specific dimensions and formats.

[0052] The data-driven estimation module takes the distortion-insensitive feature vector generated in the previous step and treats it as a two-dimensional column vector. This vector serves as the sole active activation feature in the current processing cycle and is input into the mapping layer of the machine learning regression model, i.e., the model's input layer. This input vector undergoes a complete forward propagation operation within the model. The forward propagation operation is the standard computation process by which a neural network generates predicted values ​​based on the input data. Its component values ​​are multiplied by the weight matrix of the first layer's neurons and a bias is added. After processing by a non-linear activation function, the propagation is passed layer by layer to the next network layer, until the output layer.

[0053] The data-driven estimation module receives estimation scalar data from the model's output layer. It should be understood that this estimation scalar data is the model's numerical prediction of the gas concentration represented by the input features. This value, after unit conversion, is formally defined and stored in a dedicated buffer queue as a second gas concentration estimate, independent of the overall signal amplitude under the current waveform distortion state. The formula is as follows: in, This is the second gas concentration estimate, the final concentration result output by the data-driven path, typically in ppm·m; Using distortion-insensitive feature vectors as input data, i.e., the feature vectors calculated in the previous step containing... and A two-dimensional vector of curvature, in units of sample. -1 Machine learning regression model It represents the mapping relationship of the entire neural network model, that is, the nonlinear function from the feature space to the concentration space; This is the hidden layer weight matrix, which serves as the connection weights between the input layer and the hidden layer of the neural network. These are the parameters learned by the model. This is the output layer weight matrix, which serves as the connection weights between the hidden layer and the output layer. This is the hidden layer bias vector; For output layer bias; For output layer activation / linear transformation, The hidden layer activation function, such as the ReLU function, is used to introduce non-linear characteristics. Therefore, its internal weight matrix... , and bias vector , These are parameters learned during training, and the numerical values ​​and combined effects of these parameters imply a dimensional transformation from the input feature space to the output concentration space.

[0054] The second gas concentration estimate is different from the concentration value given by the data-driven model based on the fitting results of the physical model in the model-driven inversion path. Its core advantage is that it is not sensitive to the drastic fluctuations in the amplitude of the original signal.

[0055] The data-driven estimation module obtains distortion-insensitive feature vectors. The vector is [0.828, 0.129]. The signal processing module loads a pre-trained neural network model with two input nodes, a hidden layer containing four neurons, and an output layer containing one neuron. The system retrieves the model's weights and bias parameters from memory. Then, the vector [0.828, 0.129] is input to the model's input layer. In the hidden layer, the output is calculated using the ReLU activation function; for example, the output of the first hidden layer neuron is max(0, w). 11 0.828+w 12 (0.129+b1). After calculation, the activation vector output by the hidden layer is [1.25, 0, 0.98, 0.45].

[0056] The hidden layer output vector is fed into the output layer, where it is weighted, summed, and then the output layer bias is added. No activation function is used in the output layer for linear regression. The calculation process is: h1w'1 + h2w'2 + h3w'3 + h4w'4 + ... ' out =1.25·80.1+0·(-20.5)+0.98·35.2+0.42(-15.8)-2.0=100.125+0+34.496-7.11-2.0=125.511. The processing module receives this estimated scalar data 125.511 and parses it into 125.5 ppm·m, as the second gas concentration estimate. This value of 125.5 ppm·m is then pushed into the FIFO buffer queue, awaiting the decision call in the subsequent path optimization verification step.

[0057] Step 6: The dual-drive causal decision module introduces a signal quality factor as a path selection verification criterion, combines the fitted value of the first gas concentration with the estimated value of the second gas concentration to perform dual-drive causal decision, and outputs the final gas concentration measurement value.

[0058] This step is the final decision and output stage of the entire dual-drive inversion system. It is executed by the path selection verification logic in the dual-drive causal decision module. This logic first reads the signal quality factor generated by the previous steps from the shared data area, and then retrieves the waveform tolerance threshold built into the system from the system configuration memory. The waveform tolerance threshold is a dimensionless floating-point number between 0 and 1 preset in the system firmware.

[0059] The waveform tolerance threshold is set based on a large amount of offline test data. The critical signal quality factor value was determined through statistical analysis. When it is lower than this tolerance threshold, the concentration error generated by the model-driven path will likely exceed the allowable accuracy range of the system. The waveform tolerance threshold is usually set between 0.75 and 0.9, with 0.85 being the preferred value. The setting is based on ROC curve analysis. 0.85 is the critical value for inversion error ≤ 5%. A value lower than 0.75 will cause the misjudgment rate of the model-driven path to rise to more than 10%, while a value higher than 0.9 will cause the data-driven path to over-trigger.

[0060] The system compares the signal quality factor with the waveform tolerance threshold. If the signal quality factor is not less than the waveform tolerance threshold, the system determines that the currently acquired continuous second harmonic signal is reliable and that the nonlinear evolution model based on the signal is physically applicable. Based on this determination, the logic extracts the first gas concentration fitting value from the data buffer of the model-driven inversion path, configures it as the final gas concentration measurement value, and sends it to the system display or recording unit for external output via the data bus. Conversely, if the signal quality factor is less than the waveform tolerance threshold, the system triggers a real-time verification logic. This verification logic extracts the first gas concentration fitting value and the second gas concentration estimate from the buffer queues of the model-driven inversion path and the data-driven inversion path, respectively, and calculates the absolute value of the difference between the two concentration values.

[0061] The logic compares the absolute value of this difference with a preset reasonable deviation limit, which is a numerical value with concentration units, such as ppm·m. This limit represents the maximum possible difference between the results of the two inversion paths under normal perturbation. The limit is typically set based on the system's overall uncertainty index or a fixed percentage of the target concentration range, such as 5% to 15%. If the absolute value of this difference exceeds the preset reasonable deviation limit, the system ultimately determines that the nonlinear evolution model has failed under the extreme reflection conditions of severe signal distortion, and its fitting results are unreliable. Therefore, the system instead accepts the second gas concentration estimate obtained from the data-driven path and configures it as the final gas concentration measurement for this measurement cycle for system output, thus ensuring measurement continuity and robustness under adverse signal conditions.

[0062] When the actual deviation exceeds the preset reasonable deviation limit, it indicates that at least one path has produced a significant error. Combined with the premise of poor signal quality, this error is attributed to the model-driven path. Therefore, the system prioritizes the second gas concentration estimate of the data-driven path. This judgment logic is set based on the robustness advantage of the data-driven path to signal distortion.

[0063] The system has a built-in waveform tolerance threshold of 0.85 and a reasonable deviation limit of 20.0 ppm·m. In a measurement scenario with good reflector conditions, the signal quality factor calculated in the preceding steps is 0.919. The dual-drive causal decision module compares 0.919 with 0.85. Since 0.919 is not less than 0.85, the system determines that the signal quality is reliable. Therefore, the processing module extracts the first gas concentration fitting value of 131.2 ppm·m obtained from the model-driven inversion path from the buffer and outputs it as the final gas concentration measurement value.

[0064] In another subsequent measurement scenario, the drone flew over a puddle, causing a sudden change in reflectivity, and the signal quality factor calculated in the previous steps dropped to 0.72. Since 0.72 is less than 0.85, the system triggered real-time verification logic. This logic read the first gas concentration fitted value of 158.6 ppm·m from the cache, which was severely deviated from the true value due to waveform distortion, and read the second gas concentration estimate of 129.4 ppm·m from the data-driven inversion path. The absolute value of the difference between the two was calculated as |158.6-129.4|=29.2ppm·m. Since 29.2 ppm·m is greater than the preset reasonable deviation limit of 20.0 ppm·m, the system determined that the nonlinear evolution model had failed. The system accepted and output the second gas concentration estimate of 129.4 ppm·m as the final gas concentration measurement value under the current adverse conditions.

[0065] Step 7: The online model tuning module counts the frequency of triggering events that accept the second gas concentration estimate. When the frequency of triggering events exceeds the frequency accumulation threshold within a single time window, the failed training samples are encapsulated using the fitting residual parameter of the distortion-insensitive feature vector and the fitted value of the first gas concentration, and online tuning is performed on the machine learning regression model.

[0066] This step, serving as the closed-loop control logic for achieving system adaptive evolution, is continuously executed by the background monitoring service within the online model tuning module. During system operation, this service continuously acquires the event flag set by the aforementioned path optimization verification step. This flag is set each time the second gas concentration estimate is adopted as the final gas concentration measurement.

[0067] The background monitoring service maintains an event counter and a timer. Whenever an event flag is detected, the event counter is incremented. The timer is reset cyclically with a preset single time window, and the event counter is cleared to zero each time it is reset. The single time window is the time period used to count the frequency of events. Its length is set according to the application scenario of the system and the sensitivity requirements for responding to environmental changes. For example, in vehicle-mounted rapid inspection applications, it can be set to 30s to 120s, preferably 60s. The setting is based on the fact that the typical speed of vehicle / drone inspection is 30 km / h, and 60s can cover a 250m path within 60s, which is consistent with the environmental change cycle of the scenario.

[0068] If the event counter value reaches and exceeds the preset frequency accumulation threshold before the end of any single time window, the online optimization process is triggered. It should be noted that the frequency accumulation threshold is the maximum number of model-driven path failures that the system can tolerate within a single time window. Once this accumulation threshold is exceeded, it indicates that the system is continuously operating in an environment with poor signal quality, and it is necessary to optimize the data-driven model. This accumulation threshold is set based on experience, for example, from 10 to 50 times, with 10 times being preferred. The basis for setting it is that 10 failure events indicate a continuously harsh environment, such as continuously passing through a waterlogged area. In this case, optimization can enable the model to adapt quickly. Exceeding 50 times will lead to optimization lag.

[0069] The process first reads the distortion-insensitive feature vector that triggered the current scan from the corresponding data buffer of the current scan cycle. Then, it reads the fitting residual parameter associated with the currently failed first gas concentration fitting value from the log of the nonlinear least squares algorithm. The fitting residual parameter is the final value of the objective function (sum of squared residuals) after the nonlinear least squares algorithm converges in the model-driven inversion path. The magnitude of this value reflects the degree of mismatch between the physical model and the actual observation data, serving as a quantitative indicator of the severity of model failure. Simultaneously, the process also reads the currently accepted second gas concentration estimate. This process uses the distortion-insensitive feature vector as input features and the second gas concentration estimate as the target output ground truth, encapsulating them together into a novel heterogeneous failure training sample.

[0070] Heterogeneous failure training samples are structured data pairs that contain input features that lead to the failure of the physical model, as well as target output values ​​that are more reliable and accepted by the system's decision logic under these conditions, providing high-quality supervision information for the incremental learning of machine learning models.

[0071] This process incrementally imports pre-packaged heterogeneous failed training samples into a dedicated data update pool. The data update pool is a first-in-first-out data buffer used to collect newly generated training samples to support small-batch model weight updates. When the number of samples accumulated in the data update pool reaches the preset batch size, the system drives the machine learning regression model as a low-priority background task, using the newly added samples in the pool to trigger a weight back-end optimization iteration based on extreme value deviation, thereby completing the entire data flow closed loop that balances reliable measurement output under harsh environments and adaptive evolution of the system environment.

[0072] Among them, the weight back-to-back optimization iteration based on extreme value deviation is a targeted optimization for the prediction bias of the machine learning regression model under distorted feature input. Through incremental learning in small batches, the model can quickly adapt to the current harsh measurement environment, improve the concentration estimation accuracy in the same / similar environment, and the low-priority execution mode in the background will not affect the real-time gas concentration detection and output of the system.

[0073] Assume the system's single time window is set to 60 seconds, and the frequency accumulation threshold is 10 times. At the start of the inspection task, the timer of the background monitoring service starts, and the event counter is reset to zero. In the next 45 seconds, due to the path repeatedly passing through vegetation cover and damp ground, the decision logic of the path optimization verification step adopted the second gas concentration estimate a total of 11 times. When the 11th event occurs, the event counter value is 11, exceeding the frequency accumulation threshold of 10. Therefore, the system triggers the online optimization process. This process immediately reads the data associated with the 11th event, including the distortion-insensitive feature vector [0.828, 0.129] generated by the aforementioned steps, and the fitting residual parameter associated with the failed first gas concentration fitting value of 158.6 ppm·m read from the fitting log of the model-driven inversion path; its value is a large residual sum of squares of 450.0.

[0074] The process reads the accepted second gas concentration estimate of 129.4 ppm·m. The system encapsulates the input features [0.828, 0.129] and the target output of 129.4 ppm·m into a new heterogeneous failure training sample. This sample is added to the data update pool. Assuming the batch size of the data update pool is 8, and there are already 7 previously accumulated samples in the pool, the addition of the new sample brings the number of samples in the pool to 8. The system then starts a background training task, using these 8 samples to perform backpropagation and gradient descent updates on the network weights of the machine learning regression model. This allows the model to produce an output closer to 129.4 ppm·m for distorted feature inputs such as [0.828, 0.129], thus completing the online adaptive tuning of the model.

[0075] It should be emphasized that the embodiments described in this invention are illustrative rather than limiting. Therefore, this invention includes, but is not limited to, the embodiments described in the specific implementation. Any other implementations derived by those skilled in the art based on the technical solutions of this invention are also within the scope of protection of this invention.

Claims

1. A TDLAS gas telemetry system for suppressing abrupt changes in albedo, characterized in that: It includes a raw signal sequence construction module, a signal quality factor calculation module, a distortion-insensitive feature extraction module, a model-driven inversion module, a data-driven estimation module, a dual-drive causal decision module, and an online model tuning module, wherein the raw signal sequence construction module, signal quality factor calculation module, distortion-insensitive feature extraction module, model-driven inversion module, data-driven estimation module, dual-drive causal decision module, and online model tuning module are connected in sequence; The original signal sequence construction module is used to acquire the echo optical signal from the probe end and perform phase-locked amplification to obtain a continuous second harmonic signal that represents the current scanning cycle, and align it with the preset background modulation waveform template to generate the original signal sequence. The signal quality factor calculation module is used to extract the signal-to-noise ratio, waveform symmetry ratio, baseline drift rate, and waveform distortion of the background modulation waveform template of the continuous second harmonic signal in the original signal sequence, and outputs the signal quality factor through a preset multi-factor weighted decision algorithm. The distortion-insensitive feature extraction module is used to perform point-to-point subtraction mapping on the original signal sequence to generate a differential waveform sequence, extract local waveform segments and locate local peak and valley points and inflection points, calculate the relative amplitude ratio of peaks and valleys and the curvature parameter of inflection points, and splice them to generate a distortion-insensitive feature vector. The model-driven inversion module is used to inject continuous second harmonic signals into a nonlinear evolution model containing linear parameters, and apply nonlinear least squares iterative fitting to calculate the fitted value of the first gas concentration. The data-driven estimation module is used to input distortion-insensitive feature vectors into a pre-trained machine learning regression model and perform forward propagation to map out the estimated value of the second gas concentration. The dual-drive causal decision module performs dual-drive causal decision based on the fitted value of the first gas concentration and the estimated value of the second gas concentration, and outputs the final gas concentration measurement value. The online model tuning module counts the frequency of triggering events that are adopted for the second gas concentration estimate. When the frequency of triggering events exceeds the frequency accumulation threshold within a single time window, the module encapsulates the failed training samples using the fitting residual parameters of the distortion-insensitive feature vector and the fitted value of the first gas concentration, and performs online tuning on the machine learning regression model.

2. A telemetry method for a TDLAS gas telemetry system as described in claim 1, characterized in that: Includes the following steps: Step 1: The original signal sequence construction module acquires the echo optical signal from the detector end and performs phase-locked amplification to obtain a continuous second harmonic signal representing the current scanning period. It then aligns the signal with the preset background modulation waveform template and combines them to generate the original signal sequence. Step 2: The signal quality factor calculation module extracts the signal-to-noise ratio, waveform symmetry ratio, baseline drift rate, and waveform distortion of the background modulation waveform template of the continuous second harmonic signals in the original signal sequence, and outputs the signal quality factor through a preset multi-factor weighted decision algorithm. Step 3: The distortion-insensitive feature extraction module performs point-to-point subtraction mapping on the original signal sequence to generate a differential waveform sequence, extracts local waveform segments and locates local peak and valley points and inflection points, calculates the relative amplitude ratio of peaks and valleys and the curvature parameter of inflection points, and splices them to generate a distortion-insensitive feature vector. Step 4: The model-driven inversion module injects the continuous second harmonic signal into the nonlinear evolution model containing linear parameters, applies the nonlinear least squares method for iterative fitting, and calculates the fitted value of the first gas concentration. Step 5: The data-driven estimation module inputs the distortion-insensitive feature vector into the pre-trained machine learning regression model and performs forward propagation to map out the estimated value of the second gas concentration. Step 6: The dual-drive causal decision module introduces a signal quality factor as a path selection verification criterion, combines the fitted value of the first gas concentration with the estimated value of the second gas concentration to perform dual-drive causal decision, and outputs the final gas concentration measurement value. Step 7: The online model tuning module counts the frequency of triggering events that accept the second gas concentration estimate. When the frequency of triggering events exceeds the frequency accumulation threshold within a single time window, the failed training samples are encapsulated using the fitting residual parameter of the distortion-insensitive feature vector and the fitted value of the first gas concentration, and online tuning is performed on the machine learning regression model.

3. The telemetry method for a TDLAS gas telemetry system that suppresses abrupt changes in albedo according to claim 2, characterized in that: Step 1 includes the following steps: Step 1.1: The original signal sequence construction module controls the laser emitting component to output a wavelength-modulated probe laser beam to the target area. After being reflected by the target reflective surface, the echo light signal is collected by the receiving optical system. Step 1.2: After photoelectric conversion and lock-in amplification of the echo optical signal, obtain the continuous second harmonic signal characterizing the current scanning cycle; Step 1.3: Call the system's preset background modulation waveform template, which is generated using the same set of modulation parameters under controlled ideal reflective surface conditions with no gas absorption; Step 1.4: Align the scan period timestamps of the continuous second harmonic signals and the background modulation waveform template in the time domain dimension, and combine them to generate the original signal sequence that enters the subsequent processing link.

4. The telemetry method for a TDLAS gas telemetry system that suppresses abrupt changes in albedo according to claim 2, characterized in that: Step 2 includes the following steps: Step 2.1: The signal quality factor calculation module performs an overall cross-correlation operation between the continuous second harmonic signal and the background modulation waveform template to calculate the waveform distortion between the two. Step 2.2: Using a preset multi-factor weighted decision algorithm, the signal-to-noise ratio parameter, waveform symmetry ratio, baseline drift rate, and waveform distortion are assigned corresponding weights and summed to output a continuously quantitative signal quality factor. ； in, For signal quality factor, The weighting coefficients are floating-point numbers preset in the system configuration file. This is the signal-to-noise ratio parameter value. This is the waveform symmetry ratio value. Baseline drift rate, This represents the waveform distortion.

5. The telemetry method for a TDLAS gas telemetry system that suppresses abrupt changes in albedo according to claim 2, characterized in that: Step 3 includes the following steps: Step 3.1: The distortion-insensitive feature extraction module applies a point-to-point subtraction mapping operation to the aligned continuous second harmonic signals in the current time domain and the background modulation waveform template to generate a differential waveform sequence corresponding to the current scanning period. Step 3.2: Configure the frequency domain anchor point of the target gas absorption line center frequency, and extract local waveform segments of the differential waveform sequence within a preset bandwidth range before and after the frequency domain anchor point; Step 3.3: Traverse and locate the local peak points, local valley points, and inflection points of valley transitions within the local waveform segment. Calculate the peak-valley relative amplitude ratio using the level values ​​of the local peak points and local valley points. ； in, The ratio of peak to valley relative amplitude. This refers to the level value at the local peak point; This refers to the level value at a local valley point. Calculate the curvature parameters at the inflection point: ； in, For curvature parameters, The first derivative represents the slope of the waveform at that point; The second derivative represents the rate of change of curvature of the waveform at that point; by concatenating the peak-to-valley relative amplitude ratio with the curvature parameter, a distortion-insensitive feature vector is generated to quantify the topological correlation of the local waveform.

6. The telemetry method for a TDLAS gas telemetry system that suppresses abrupt changes in albedo according to claim 2, characterized in that: Step 4 includes the following steps: Step 4.1: Load the model-driven inversion module with a nonlinear evolution model containing the target gas absorption linear parameters and laser scanning modulation parameters; Step 4.2: Inject the continuous second harmonic signal into the nonlinear evolution model as the only observation input data, and apply the nonlinear least squares algorithm to perform multi-parameter iterative fitting on the overall amplitude envelope of the observation input data; Step 4.3: Obtain the iterative parameters after the nonlinear least squares algorithm has reached convergence, and calculate the first gas concentration fitting value for the current scanning cycle by substituting them according to the Lambert-Beer law relationship. ； in, To fit the obtained concentration integral value, The absorption coefficient obtained by fitting; The absorption line strength is determined by the target gas.

7. The telemetry method for a TDLAS gas telemetry system that suppresses abrupt changes in albedo according to claim 2, characterized in that: Step 5 includes the following steps: Step 5.1: The data-driven estimation module loads the machine learning regression model that has been pre-trained using multiple simulated reflective surface albedo distortion samples, closes the input interface channel, and only allows the input of specified low-dimensional feature parameters; Step 5.2: Input the obtained distortion-insensitive feature vector as the only active activation feature into the mapping layer of the machine learning regression model to perform forward propagation operation; Step 5.3: Receive the estimated scalar data from the output layer of the machine learning regression model and store it in a cache queue as a second gas concentration estimate to mitigate the current waveform distortion. ； in, This is an estimate of the second gas concentration. Using distortion-insensitive feature vectors as input data, For machine learning regression models, The hidden layer weight matrix is... This is the output layer weight matrix. This is the hidden layer bias vector; For output layer bias; For output layer activation / linear transformation, This is the activation function for the hidden layer.

8. The telemetry method for a TDLAS gas telemetry system that suppresses abrupt changes in albedo according to claim 2, characterized in that: Step 6 includes the following steps: Step 6.1: The dual-drive causal decision module responds to the decision result that the value of the signal quality factor is not less than the waveform tolerance threshold, determines that the nonlinear evolution model has physical applicability, extracts the first gas concentration fitting value from the buffer queue and configures it as the final gas concentration measurement value for system output. Step 6.2: In response to the judgment that the value of the signal quality factor is less than the waveform tolerance threshold, the real-time verification logic is triggered to extract the absolute value of the difference between the fitted value of the first gas concentration and the estimated value of the second gas concentration; if the absolute value of the difference is greater than the preset reasonable deviation limit, it is determined that the nonlinear evolution model has failed under the current extreme reflection conditions, and the estimated value of the second gas concentration is adopted as the final gas concentration measurement value for system output.

9. The telemetry method for a TDLAS gas telemetry system that suppresses abrupt changes in albedo according to claim 2, characterized in that: Step 7 includes the following steps: Step 7.1: The online model tuning module continuously monitors the frequency of trigger events that use the second gas concentration estimate as the final gas concentration measurement value in the background. Step 7.2: When the frequency of the triggering event exceeds the frequency accumulation threshold within a single time window, read the distortion-insensitive feature vector of the current scanning cycle and the fitting residual parameter of the first gas concentration fitting value, and encapsulate them as a heterogeneous failure training sample. Step 7.3: Import the incremental training samples of heterogeneous failures into the data update pool to drive the machine learning regression model to trigger the weight back-to-back optimization iteration based on extreme value deviation, and complete the data flow closed loop that takes into account both the measurement output in harsh environments and the adaptive evolution of the system environment.

Images