A method, device, medium and product for methane concentration inversion for a farm scene

Abstract

The application discloses a methane concentration inversion method and device for a farm scene, a medium and a product, relates to the fields of greenhouse gas spectrum detection signal processing and intelligent inversion, and comprises the following steps: synchronously acquiring multi-modal auxiliary data; performing joint noise reduction processing on the measurement beam signal to obtain a methane absorption spectrum signal; performing photoelectric endogenous error adaptive compensation on the methane absorption spectrum signal based on the multi-modal auxiliary data; performing multi-branch environmental interference decoupling processing on the standardized spectrum signal based on a four-branch parallel deep learning interference decoupling network; extracting a multi-dimensional feature vector according to the methane characteristic spectrum and the multi-modal auxiliary data; and determining a methane concentration inversion result by using a pre-trained inversion model. The application can realize accurate and stable inversion of methane concentration in the full scene and full concentration range of a farm.

Description

Technical Field

[0001] This application relates to the field of greenhouse gas spectral detection signal processing and intelligent inversion, and in particular to a method, equipment, medium and product for methane concentration inversion in a livestock farm setting. Background Technology

[0002] Methane, as one of the major greenhouse gases, has a greenhouse effect 28 times greater than that of carbon dioxide per unit mass. Livestock farming is the main source of methane emissions in agriculture, accounting for more than 40% of total agricultural methane emissions. Accurate monitoring of methane concentrations and emission fluxes in livestock farms is of significant practical importance and application value for formulating scientific emission reduction policies, evaluating the effectiveness of emission reduction technologies, and fulfilling greenhouse gas emission reduction responsibilities.

[0003] Currently, methane detection methods based on tunable semiconductor laser absorption spectroscopy (TDLAS) technology have become the mainstream technology for methane concentration detection due to their advantages such as fast response speed, high detection sensitivity, and strong selectivity. However, the unique and complex operating conditions in aquaculture farms lead to many shortcomings in traditional TDLAS detection solutions in this scenario, making it impossible to meet the requirements for high-precision detection. Specific problems are as follows: (1) Significant interference from internal errors in the photoelectric system: The operating characteristics of the laser source and photodetector are easily affected by the environment, which in turn introduces errors. Specifically, the fluctuation of the laser's operating current and the core temperature drift will cause the laser center wavelength to shift. For every 1°C change in temperature, the wavelength of the distributed feedback semiconductor laser (DFB) will drift by 0.01~0.02nm; for every 1mA change in current, the wavelength will drift by about 0.005nm, making it impossible for the laser to accurately align with the characteristic absorption spectrum of methane, resulting in distortion of the absorption signal intensity. The temperature drift of the photodetector will cause an increase in dark current and a rise in noise level, resulting in irregular fluctuations in the baseline, which will directly drown out the weak absorption signal of low-concentration methane and reduce the signal-to-noise ratio. At the same time, the power fluctuation of the laser source will cause the incident light intensity to be unstable, further aggravating the distortion of the spectral signal and affecting the concentration inversion accuracy. Traditional algorithms often employ linear compensation with fixed coefficients, which cannot accurately quantify the nonlinear mapping relationship between photoelectric parameter drift and spectral signal distortion. This makes it difficult to accurately correct intrinsic errors, resulting in a significant increase in inversion errors in scenarios with drastic temperature changes in aquaculture farms.

[0004] (2) Complex and difficult-to-eliminate external interference in the aquaculture environment: The aquaculture environment is characterized by high humidity, high dust, coexistence of multiple gases, and drastic temperature and pressure fluctuations. Various environmental factors will cause multiple interferences to the laser spectral signal, and the interference mechanism is complex. Traditional solutions cannot achieve effective decoupling. 1) High humidity interference: The relative humidity in aquaculture farms is maintained at 50%~90% year-round. Water vapor has a strong absorption band in the characteristic absorption spectrum of methane (around 1653.72nm), which partially overlaps with the methane absorption spectrum, forming spectral overlap interference. This causes traditional single-wavelength detection algorithms to be unable to distinguish the absorption signals of methane and water vapor, resulting in falsely high methane concentration measurements with errors exceeding 10%. The 2025 review of TDLAS water vapor concentration detection, "Advances in enhancing thensitivity of TDLAS for water vapor concentration detection — A review," clearly points out that temperature and pressure compensation and background noise suppression are key technical routes to improve the detection limit. However, traditional water vapor interference elimination schemes are either technically complex or have poor error control effects, making them difficult to adapt to the high humidity conditions of aquaculture farms.

[0005] 2) High dust disturbance: The concentration of organic dust generated by feed, manure, feathers, dander, etc. in the farm can reach 10 mg / m³. 3 ~50mg / m 3 Dust particles scatter laser light, causing uniform attenuation of the laser intensity, generating scattering noise, disrupting the integrity of the spectral signal, and introducing broadband absorption interference, altering the absorption peak morphology. Traditional solutions can only perform simple intensity normalization or use an indirect compensation method of "external parameter input + empirical model fitting," which cannot accurately compensate for the loss caused by dust scattering and does not consider the multi-dimensional interference caused by dust. The latest research proposed in 2025, the "Anti-Particle Interference (API) Optical Model," uses a beam splitting structure to divide the laser into a measurement beam and a reference beam, and eliminates common-mode dust noise in the two beams through differential absorption processing, at a dust concentration of 8 mg / m³. 3 ~10 mg / m 3 Under these conditions, the localization accuracy of methane characteristic peaks reached 100%, while that of traditional direct absorption spectroscopy (DAS) was only 21.5%, reducing the root mean square error (RMSE) by 69.9%–86.9%. This approach, through the synergy of hardware and algorithms, achieved a more thorough elimination of dust interference. Another study from 2025 developed a Residual-MSCNN-LSTM model, utilizing the spectral shape characteristics and temporal correlation of methane absorption peaks to distinguish methane signals from dust interference at a concentration of 2 mg / m³. 3 ~10 mg / m 3 At dust concentrations, RMSE reductions of 47.2% to 85.3% were achieved. In contrast, traditional dust compensation schemes are clearly insufficient in terms of mechanism characterization.

[0006] 3) Crosstalk from Multiple Stray Gases: Farms often contain various interfering gases such as ammonia (NH3), hydrogen sulfide (H2S), and carbon dioxide (CO2). Ammonia has a weak absorption peak near the characteristic absorption wavelength of methane, which can overlap with the methane absorption signal, creating spectral crosstalk. Furthermore, the absorption lines of these stray gases also overlap. Traditional algorithms lack specific strategies to suppress this interference and cannot effectively filter out stray gas absorption components. In practical applications, the core method of TDLAS technology to address crosstalk from multiple stray gases is to optimize laser wavelength selection—selecting spectral lines where the target gas has strong absorption while the interfering gas has weak absorption, thus avoiding interference at its source. If complete avoidance is not possible, WMS harmonic detection and nonlinear least squares fitting are used to distinguish overlapping peaks. Traditional orthogonal projection schemes are less feasible in engineering and may project effective signal components into the stray gas space, resulting in incorrect rejection.

[0007] 4) Temperature and pressure fluctuations: The diurnal temperature range in aquaculture farms can reach -10℃ to 40℃, and atmospheric pressure also fluctuates with weather changes. The absorption coefficient of methane molecules is closely related to temperature and pressure. Traditional algorithms use a fixed absorption coefficient for concentration calculation, ignoring the impact of temperature and pressure fluctuations on the absorption coefficient, leading to model failure and decreased inversion accuracy. Recent research in 2025 has widely adopted more advanced intelligent modeling methods. The dynamic temperature compensation model (CIWOA-BP) based on the improved whale optimization algorithm can handle TDLAS gas detection under dynamic temperature conditions, with an average relative error of only 0.184% at a 35% CO2 concentration. The temperature compensation method based on the whale-sparrow search algorithm-backpropagation neural network (WOA-SSA-BP) has also achieved good results. In comparison, the traditional "bilinear interpolation + pre-built database" method is outdated and lacks accuracy and adaptability.

[0008] (3) Insufficient ability to extract weak signals: The methane concentration varies greatly in different areas of the farm. The methane concentration in the pen is usually 10ppm~100ppm, which is a low concentration range. The corresponding laser absorption signal is very weak and the signal-to-noise ratio is extremely low, usually less than 10:1. Traditional noise reduction algorithms, such as single wavelet threshold noise reduction and mean filtering, are prone to losing the effective absorption characteristics of methane while filtering out noise. They cannot achieve accurate extraction of weak signals, resulting in insufficient sensitivity for low-concentration methane detection and failing to meet the needs of monitoring the concentration of the entire farm. At the same time, the traditional two-stage cascade noise reduction of wavelet + SVD may be cumbersome to operate because the threshold parameters need to be adjusted separately, and the nonlinear coupling effect between various noise sources is not considered.

[0009] (4) Poor adaptability of existing technologies and lack of targeted solutions: Currently, some technological improvements for methane detection in livestock farms focus on hardware structure optimization, such as adding dust removal and dehumidification hardware, optimizing optical path structure, and improving laser drive circuits. These solutions not only increase equipment costs and maintenance difficulty, but also fail to fundamentally solve the interference and error problems at the signal level. Other algorithm solutions are mostly designed for general scenarios and do not fully consider the specific complex operating conditions of livestock farms. They lack targeted interference removal and error compensation strategies, have poor generalization ability, and cannot adapt to the dynamic operating conditions of different emission areas in livestock farms, such as pens, manure pits, and anaerobic fermentation zones. In addition, many existing solutions do not fully utilize the latest research results, have insufficient creative demonstration, and lack sufficient demonstration of some key parameters, such as wavelet basis selection and sampling frequency.

[0010] Therefore, there is an urgent need for a solution that can address issues such as photoelectric errors, environmental interference, and weak signal extraction under complex operating conditions in livestock farms, and achieve high-precision and high-stability inversion of methane concentration to meet the actual needs of methane emission monitoring in livestock farms. Summary of the Invention

[0011] The purpose of this application is to provide a method, equipment, medium, and product for methane concentration inversion in aquaculture farms, which can achieve accurate and stable inversion of methane concentration across all scenarios and concentration ranges in aquaculture farms.

[0012] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a method for methane concentration inversion in aquaculture farm scenarios, including: Simultaneously acquire multimodal auxiliary data; the multimodal auxiliary data includes: the measurement beam signal of the methane laser spectrum, the reference beam signal, as well as photoelectric parameters and environmental parameters; The measured beam signal is subjected to joint noise reduction processing to obtain the methane absorption spectrum signal; the joint noise reduction processing combines an adaptive noise complete set empirical mode decomposition algorithm, an energy ratio criterion, and wavelet packet noise reduction processing. Based on the multimodal auxiliary data, the methane absorption spectrum signal is subjected to photoelectric intrinsic error adaptive compensation to obtain a standardized spectral signal; the photoelectric intrinsic error adaptive compensation includes wavelength drift correction, light intensity normalization and baseline drift correction, and harmonic signal amplitude photoelectric correction. The standardized spectral signal is subjected to multi-branch environmental interference decoupling processing based on a four-branch parallel deep learning interference decoupling network to obtain the methane characteristic spectrum; the four-branch parallel deep learning interference decoupling network includes: water vapor interference decoupling branch, dust interference decoupling branch, crosstalk suppression branch and temperature and pressure coupling correction branch. Based on the methane characteristic spectrum and the multimodal auxiliary data, multidimensional feature vectors are extracted; and a pre-trained inversion model is used to determine the methane concentration inversion result.

[0013] Secondly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the methane concentration inversion method for aquaculture farm scenarios.

[0014] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned methane concentration inversion method for aquaculture farm scenarios.

[0015] Fourthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned methane concentration inversion method for aquaculture farm scenarios.

[0016] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a method, device, medium, and product for methane concentration inversion in aquaculture farms. By simultaneously acquiring the measurement beam, reference beam, photoelectric parameters, and environmental parameters, and employing joint noise reduction processing, it can adaptively decompose the signal into multiple scales, accurately eliminating high-frequency random noise and noise coupling components, while completely preserving the effective absorption signal of methane, thus solving the problem of insufficient sensitivity in detecting low-concentration methane in aquaculture farms. Based on multimodal auxiliary data, the wavelength axis is adaptively corrected, the global light intensity is normalized, and the harmonic amplitude is dynamically corrected, ensuring that the laser is always accurately aligned with the characteristic absorption lines of methane, completely eliminating signal distortion caused by nonlinear drift of the photoelectric system, and guaranteeing the physical authenticity of the spectral data from the source. Based on a four-branch parallel deep learning interference decoupling network, the standardized spectral signal is processed for multi-branch environmental interference decoupling, which can cope with the extreme working conditions of high humidity, high dust, mixed gases, and temperature and pressure fluctuations in aquaculture farms. This application effectively overcomes the technical defects of traditional TDLAS technology in complex working conditions of aquaculture farms, such as easy loss of weak signals, difficulty in correcting photoelectric nonlinear drift, difficulty in decoupling multi-source environmental interference, and poor generalization ability of the inversion model, by combining multi-modal data synchronous acquisition, joint noise reduction, photoelectric intrinsic error adaptive compensation, multi-branch environmental interference decoupling, and intelligent inversion of multi-dimensional features. It achieves accurate and stable inversion of methane concentration in all scenarios and concentration ranges of aquaculture farms. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of a methane concentration inversion method for a livestock farm scenario according to one embodiment of this application; Figure 2 A diagram illustrating the overall technical workflow architecture for methane detection in livestock farms; Figure 3 This is a schematic diagram of the structure of a hybrid model of attention-enhanced convolutional neural network and bidirectional long short-term memory network (attention-enhanced CNN-BiLSTM hybrid inversion model); Figure 4 A schematic diagram of the joint noise reduction algorithm consisting of the adaptive noise complete set empirical mode decomposition algorithm (CEEMDAN), energy ratio criterion (EP), and wavelet packet denoising (WTD). Figure 5 A schematic diagram comparing inversion error and signal-to-noise ratio under different operating conditions; Figure 6 This is a schematic diagram of a multi-branch environmental interference decoupling network structure. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] In one exemplary embodiment, such as Figure 1 and Figure 2 As shown, a method for methane concentration inversion in a livestock farm scenario is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is described using a server as an example, including the following steps S101 to S105. Wherein: S101, synchronously acquire multimodal auxiliary data; the multimodal auxiliary data includes: the measurement beam signal of the methane laser spectrum, the reference beam signal, as well as photoelectric parameters and environmental parameters; As a specific example, the hardware collaborative deployment and multimodal synchronous data acquisition were carried out using standardized commercial devices. The specific configuration is as follows: LDC0300 laser driver module, current noise <1μA, temperature control accuracy ±0.001℃, supporting 1MHz external modulation; DP-GC-H25M-G Heriot-Tropsch gas cell, effective optical path 25m, calcium fluoride window, adaptable to 1~500kPa working pressure; InAsSb-2 / 5um-AG8C photodetector, equipped with a two-stage TEC cooling module, minimum cooling temperature -30℃, 8-level adjustable gain, 400ns fast response; the embedded terminal adopts an FPGA+DSP architecture, supports a sampling rate of 10Msps, 14-bit high-precision resolution, and realizes real-time data transmission through UDP network port. The hardware parameters are fully adapted to the long-term stable operation requirements of the complex working conditions of the farm.

[0022] After equipment deployment, multimodal data synchronous acquisition is initiated. A 1:1 high-precision fiber optic beam splitter equally divides the 1653.72nm laser output from the DFB laser into a measurement beam and a reference beam. The measurement beam passes through a 25m Heriot-Lewis gas cell to achieve methane gas absorption, and a photodetector collects the spectral signal containing absorption characteristics. The reference beam, without gas absorption, is used for optical path error calibration and noise benchmark calibration. Raw laser spectral time-series data in the 1600~1700nm band is continuously acquired at a 1kHz standardized sampling frequency and a 1s sampling period, correcting the inconsistencies between traditional high and low sampling parameters, effectively reducing the computational burden on the embedded terminal, and simultaneously acquiring high-precision photoelectric parameters and environmental parameters in real time. The photoelectric parameters include laser operating current with an accuracy of ±0.01mA, laser core temperature with an accuracy of ±0.01℃, detector temperature with an accuracy of ±0.1℃, and light intensity fluctuation coefficient. The environmental parameters include ambient temperature with an accuracy of ±0.1℃, ambient humidity with an accuracy of ±1%RH, atmospheric pressure with an accuracy of ±0.1kPa, and atmospheric pressure with an accuracy of ±0.01mg / m³. 3 The dust equivalent attenuation coefficient is determined, and all multimodal data are strictly synchronized in time and transmitted in real time to the embedded terminal to build a complete dataset, providing standardized data support for deep learning algorithm operations.

[0023] S102, perform joint noise reduction processing on the measured beam signal to obtain the methane absorption spectrum signal; the joint noise reduction processing combines adaptive noise complete set empirical mode decomposition algorithm, energy ratio criterion and wavelet packet noise reduction processing. like Figure 4As shown, S102 adopts the "CEEMDAN-EP-WTD joint noise reduction algorithm", which is superior to the traditional wavelet + SVD noise reduction. Combined with the attention mechanism, it achieves high-sensitivity extraction of weak signals, while taking into account both noise filtering and effective feature preservation.

[0024] S102 specifically includes: S21, the measurement beam signal is decomposed into multiple scales using an adaptive noise complete set empirical mode decomposition algorithm to obtain multiple intrinsic mode components and a residual component. Specifically, the original spectral time series signal is adaptively decomposed using the adaptive noise complete set empirical mode decomposition algorithm to obtain several intrinsic mode components (IMFs) and a residual component. The high-frequency IMFs contain random high-frequency noise and noise coupling components, while the low-frequency IMFs and the residual component contain the effective methane absorption signal and low-frequency baseline drift. S22, the intrinsic mode components and residual components are screened using the energy ratio criterion to obtain the effective components containing the effective absorption signal of methane, and the pure noise IMF components are eliminated to reduce invalid processing and improve algorithm efficiency. S23, perform wavelet packet denoising on the effective components, and use an adaptive threshold function to perform thresholding on the wavelet packet coefficients of each layer; Specifically, the Symlets wavelet basis is selected, and each component is decomposed into three layers of wavelet packets. An adaptive soft thresholding function is used to threshold the coefficients of each layer of wavelet packets, adaptively adjusting the threshold value to filter out high-frequency random noise and noise coupling components, while retaining weak high-frequency absorption features. ; In the formula, Let be the coefficients of the k-th wavelet packet in the j-th layer. These are the wavelet packet coefficients after noise reduction. The adaptive threshold for the j-th layer is determined adaptively by the variance and energy ratio of the wavelet packet coefficients of each layer, with a clear function definition to avoid symbol overlap; The processed effective components are reconstructed to obtain a pure methane absorption spectrum signal, achieving high-sensitivity extraction of weak signals at low concentrations, improving the signal-to-noise ratio to over 30:1, and effectively identifying methane signals at low concentrations below 10 ppm. From the reconstructed pure spectrum, the following multi-dimensional effective feature vectors are extracted as inputs for subsequent concentration inversion. Combined with the reference beam features acquired by the hardware, the inversion accuracy is improved.

[0025] S24, the processed effective components are reconstructed to obtain the methane absorption spectrum signal.

[0026] S102 employs a CEEMDAN-EP-WTD combined deep learning noise reduction mechanism to perform weak signal layering purification processing on the spectral signal of the measurement beam, achieving accurate filtering of multi-source coupling noise while preserving the original physical characteristics of the spectrum.

[0027] As a specific example, such as Figure 4 As shown, the specific parameter configuration is as follows: 8 CEEMDAN decomposition layers, energy ratio screening threshold of 0.3, and three-layer wavelet packet decomposition using the sym4 wavelet basis. The adaptive soft threshold is set to 1.2 times the variance of the wavelet packet coefficients of each layer. The system first performs multi-scale adaptive decomposition on the original mixed spectral signal, obtaining 8 layers of intrinsic mode components and 1 layer of residual components. The noise component and the effective signal component are distinguished by the energy ratio criterion, and the high-frequency pure noise component is removed, while the effective component containing methane absorption characteristics is retained. Then, the wavelet packet coefficients of each layer are optimized by a deep learning adaptive threshold function to filter out low-frequency baseline drift and multi-source coupling noise. Finally, the effective component signal is reconstructed, and a secondary calibration is performed in combination with the real-time reference beam noise characteristics to completely eliminate inherent hardware noise and random interference. The final output is a high-purity methane spectral signal with a signal-to-noise ratio greater than 30:1, which fully preserves the detailed morphological features of the 1653.72nm characteristic absorption peak.

[0028] S103, based on the multimodal auxiliary data, the photoelectric intrinsic error adaptive compensation is performed on the methane absorption spectrum signal to obtain a standardized spectral signal; the photoelectric intrinsic error adaptive compensation includes wavelength drift correction, light intensity normalization and baseline drift correction, and harmonic signal amplitude photoelectric correction; for the intrinsic errors caused by the operating characteristics of the laser source and photodetector itself, combined with the reference beam data of the beam splitting structure, this application establishes a four-dimensional nonlinear mapping model of "current-temperature-spectral distortion-reference beam" to achieve accurate quantification and adaptive compensation of photoelectric errors; S103 specifically includes: S31, using formula The center wavelength offset of the laser is determined, and the wavelength drift correction of the methane absorption spectrum signal is performed using the center wavelength offset of the laser; wherein, This is the center wavelength offset of the laser. These are the fitting coefficients. The intensity of the reference beam is used to assist in correcting the wavelength shift. This is the operating current. Core temperature; in, A wavelength shift correction function was constructed by fitting extensive experimental data with wavelength stability data of a reference beam; the laser's operating current was acquired in real time. Core temperature and the intensity of the reference beam Substitute into the above function to calculate the wavelength offset. The abscissa (wavelength axis) of the spectrum is adaptively corrected to ensure that the laser is always aligned with the characteristic absorption line of methane (1653.72nm), thus eliminating the absorption signal distortion caused by wavelength drift.

[0029] Regarding incident light intensity caused by laser power fluctuations To address the instability and baseline drift caused by detector temperature drift, a two-step correction strategy is adopted, combined with reference beam data, to correct the problem of unreasonable traditional baseline reference points. S32, performs light intensity normalization and baseline drift correction; Specifically, a background wavelength with no gas absorption is selected. (Below 1570nm, far from all gas absorption peaks), the transmitted light intensity at this wavelength (reference beam) is collected in real time. Using this as a reference intensity, the transmitted light intensity (measurement beam) of all wavelengths is normalized to eliminate the intensity attenuation error caused by laser power fluctuations. ; The "non-absorption baseline fitting" method is employed, which involves fitting the transmitted light intensity in the non-absorption region using a polynomial and extrapolating it to the full scanning range. Simultaneously, the detector temperature is established. Baseline drift The nonlinear fitting model is obtained through polynomial fitting: ; in, For reference beam, The measured spectral signal, The normalized spectral signal, The fitting coefficients were determined through baseline calibration experiments at different detector temperatures. Subtract the baseline drift from the normalized spectral signal. By combining the baseline fitting results, a pure spectral baseline is restored, eliminating errors caused by detector temperature drift.

[0030] S33, using formula Photoelectric correction of harmonic signal amplitude is performed; among which, This represents the ratio of the original second harmonic 2f to the first harmonic 1f. This is the photoelectric correction coefficient, determined by the laser's operating current, core temperature, detector temperature, and the intensity of the reference beam. It is obtained through fitting experimental data and is used to correct harmonic amplitude errors caused by photoelectric parameter drift, further improving signal stability. This is the corrected ratio of the second harmonic 2f to the first harmonic 1f.

[0031] In TDLAS technology, wavelength modulation spectroscopy (WMS) is typically used to extract the first harmonic (1f) and second harmonic (2f) signals, and the influence of light intensity fluctuations is eliminated by the 2f / 1f ratio. However, photoelectric parameter drift can cause distortion of the harmonic signal amplitude. This application adds normalization correction of photoelectric parameters and reference beam data to the 2f / 1f ratio. As a specific implementation, real-time data on the current laser current, laser core temperature, detector temperature, and reference beam intensity are retrieved and input into a trained, converged deep network. The network iteratively optimizes and dynamically solves for all fitting coefficients related to wavelength shift correction, baseline drift correction, intensity normalization, and harmonic amplitude correction. This accurately calculates the spectral wavelength shift and corrects wavelength axis deviation, ensuring the laser spectral lines are precisely aligned with the characteristic absorption band of methane. Simultaneously, the system selects the intensity of the reference beam in the 1570nm non-absorption blank band to perform global intensity normalization, eliminating laser power fluctuation errors. The detector temperature is then monitored. Baseline drift The nonlinear fitting model is used to remove the interference of spectral baseline drift, and finally the dynamic amplitude correction of the 2f / 1f harmonic ratio is completed, which completely solves the signal distortion problem caused by the nonlinear drift of the optoelectronic system and outputs a standardized methane spectral signal without intrinsic error.

[0032] S104, based on a four-branch parallel deep learning interference decoupling network, multi-branch environmental interference decoupling processing is performed on the standardized spectral signal to obtain the methane characteristic spectrum; as follows Figure 6 As shown, the four-branch parallel deep learning interference decoupling network includes: a water vapor interference decoupling branch, a dust interference decoupling branch, a crosstalk suppression branch for mixed gases, and a temperature and pressure coupling correction branch. S104 addresses the specific interference caused by high humidity, high dust, mixed gases, and temperature and pressure fluctuations in farms. This application adopts a hierarchical elimination strategy of "hardware collaboration + algorithm decoupling" and combines the latest research results to achieve accurate separation and compensation of various interferences. S104 specifically includes: S41, in the water vapor interference decoupling branch, based on the real-time collected ambient temperature, air pressure and relative humidity, the water vapor concentration is dynamically estimated, and combined with the pre-constructed water vapor absorption coefficient database, the contribution of the water vapor absorption signal is calculated, and the contribution of the water vapor absorption signal is differentially extracted from the standardized spectral signal. Specifically, water vapor has a strong absorption band at the characteristic absorption wavelength of methane (around 1653.72 nm), which causes spectral overlap and interference. Based on the core ideas of the 2025 TDLAS water vapor detection review, this application adopts an algorithm of "adaptive separation first, then differential spectral stripping", combined with real-time temperature and pressure estimation, to achieve accurate separation of water vapor interference.

[0033] Absorption spectra of water vapor in the wavelength range of 1600–1700 nm were collected experimentally under different humidity, temperature, and pressure conditions to construct a water vapor absorption coefficient database. ,in, The relative humidity of the environment; based on the synchronously collected ambient temperature. air pressure relative humidity An adaptive estimation algorithm is used to calculate water vapor concentration in real time. This avoids errors associated with traditional indirect calculations; it retrieves the corresponding water vapor absorption coefficient from the database. Using the Lambert-Beer law, the contribution of water vapor absorption signal at the characteristic wavelength of methane is calculated. : ; Standardized spectral signal Subtract the contribution of water vapor absorption By combining the water vapor interference calibration data of the reference beam, a pure methane absorption signal was obtained. : ; In the formula, The contribution of water vapor absorption in the reference beam is used for auxiliary correction. This algorithm can effectively eliminate water vapor overlap interference, reducing the concentration error caused by humidity to less than 1%, and is suitable for high humidity conditions in aquaculture farms.

[0034] S42, in the dust interference decoupling branch, based on the real-time dust concentration and spectral line morphology characteristics, the dynamic light intensity attenuation coefficient is determined by substituting it into the pre-constructed dust attenuation model, and combined with the differential processing of the measured beam signal and the reference beam signal, the light intensity attenuation error caused by dust scattering and broadband absorption is compensated. Specifically, drawing on the core ideas of the 2025 "Avoidance of Particle Interference (API) Optical Model" and the Residual-MSCNN-LSTM model, a "hardware spectral splitting + algorithm compensation" approach is adopted, combined with the spectral morphology characteristics of dust interference, to achieve accurate compensation for scattering loss. The measurement beam and reference beam are acquired using a beam splitter, and the intensity data of the two beams are collected simultaneously. The common-mode dust noise characteristics of the two beams are utilized to assist in interference removal. The experiment simulates different dust concentrations (0~100mg / m³) in a livestock farm. 3 The laser intensity attenuation under the condition of ( ) is measured, and the background wavelength is collected. Using light intensity attenuation data (without gas absorption) and feature extraction methods based on the Residual-MSCNN-LSTM model, dust concentration is established. Light intensity attenuation coefficient A three-dimensional nonlinear fitting model of spectral line morphology characteristics (constructed dust attenuation model): ; In the formula, The fitting coefficients are determined through calibration using experimental data. The spectral morphology feature function of dust interference is used to improve the model fitting accuracy; dust concentration in the farm is collected simultaneously. Based on the spectral line morphology characteristics, the light intensity attenuation coefficient is calculated using the above model. By combining the intensity difference between the reference beam and the measurement beam, the scattering loss calculation results are corrected; finally, the methane absorption signal is... Divide by light intensity attenuation coefficient By combining the differential processing of the two beams, the compensated methane absorption signal is obtained. Eliminate the light intensity attenuation error caused by dust scattering and broadband absorption: ; In the formula, This is used to measure the intensity of the light beam at the background wavelength, and is used for auxiliary compensation.

[0035] S43, construct a stray gas interference feature library in the stray gas crosstalk suppression branch and perform orthogonalization processing, and normalize the spectral signal to the orthogonal complement space of the stray gas interference feature library; Specifically, addressing the dynamic changes in the concentration of stray gases and spectral overlap in livestock farms, an algorithm combining "wavelength selection optimization + orthogonal projection improvement" is employed. This, combined with the wavelength calibration function of a hardware laser, achieves precise suppression of stray gas crosstalk, correcting the shortcomings of traditional orthogonal projection. Through the wavelength calibration function of the hardware laser, the characteristic wavelength (1653.72nm) of strong methane absorption and weak stray gas absorption is selected to avoid stray gas crosstalk at its source. Simultaneously, the absorption spectra of common stray gases in livestock farms (ammonia, hydrogen sulfide, and carbon dioxide) in the 1600–1700nm wavelength range are collected, and the characteristic absorption vectors of each stray gas are extracted to construct a stray gas interference feature library. ; Considering the coupling effect between impurity gases, the impurity gas interference feature library is analyzed. Orthogonalization is performed to eliminate the influence of spectral line overlap between impurities, and then the corrected mixed spectral absorption signal vector is... Projected onto the stray gas interference feature library The orthogonal complement space is used to obtain the signal vector after removing impurity gas interference. : ; In the formula This is the transpose matrix of the impurity gas interference feature library. Its inverse matrix; the signal vector after orthogonal projection The signal is restored to a spectral signal, and then verified and corrected by WMS harmonic detection and nonlinear least squares fitting. This yields the pure methane absorption spectrum after removing crosstalk from impurities, effectively suppressing concentration measurement errors caused by impurities.

[0036] S44, in the temperature and pressure coupling correction branch, based on the real-time collected ambient temperature and air pressure, an intelligent optimization interpolation algorithm is used to dynamically determine the methane molecule absorption coefficient under the current operating conditions from the pre-constructed three-dimensional absorption coefficient database, and substitute it into Lambert-Beer law to correct the spectral absorption parameters; the intelligent optimization interpolation algorithm is an interpolation algorithm optimized based on the improved whale optimization algorithm.

[0037] Specifically, drawing on the advanced intelligent temperature compensation model ideas of 2025, a dynamic correction model for the absorption coefficient of "temperature-pressure coupling + intelligent interpolation" is constructed to replace the traditional and outdated bilinear interpolation method.

[0038] The absorption coefficient of methane at a characteristic wavelength (1653.72 nm) was collected experimentally under different temperatures (-20℃~50℃) and different pressures (80~120 kPa) to construct a three-dimensional database. Supplementing samples under different temperature and pressure coupling conditions improves database accuracy; Simultaneous collection of ambient temperature data from the breeding farm and air pressure An improved interpolation algorithm based on the CIFOOA (Compared Whale Optimization Algorithm) is used to accurately solve the methane absorption coefficient under the current temperature and pressure conditions from the database. Compared to traditional bilinear interpolation, the accuracy is improved by more than 30%; the dynamically solved... By substituting Lambert-Beer's law into the traditional fixed absorption coefficient and combining it with real-time temperature and pressure change trends, the absorption coefficient change is predicted, thereby correcting the error caused by temperature and pressure fluctuations and ensuring stable inversion accuracy under different temperature and pressure conditions.

[0039] As a specific implementation, the standardized spectral signal enters a four-branch parallel deep learning interference decoupling network, which, based on the inherent physical mechanisms of various interferences, hierarchically eliminates the four core environmental interferences of the farm. The water vapor interference decoupling branch retrieves a high-density water vapor absorption coefficient database covering -20℃ to 50℃, 80 to 120 kPa, and 0% to 100% RH. It dynamically estimates water vapor concentration using real-time temperature, humidity, and pressure parameters, and accurately removes overlapping absorption interference through a deep learning differential stripping algorithm. The dust interference decoupling branch, based on real-time dust concentration and spectral line morphology characteristics, uses a deep learning-fitted dust attenuation model to solve for dynamic loss coefficients, and compensates for signal errors caused by dust scattering and broadband absorption using a dual-beam differential mechanism. The stray gas crosstalk decoupling branch relies on a dedicated stray gas feature library for aquaculture scenarios, using a deeply optimized orthogonal projection algorithm to strip away spectral crosstalk components of ammonia, hydrogen sulfide, and carbon dioxide. The temperature and pressure coupling correction branch uses the CIWOA deep learning interpolation algorithm to dynamically update the methane molecule absorption coefficient under current operating conditions, replacing traditional fixed parameters to correct spectral absorption parameters. Multiple branches operate in parallel, and results are fused to ultimately obtain a pure methane characteristic spectrum completely free of environmental interference.

[0040] S105, based on the methane characteristic spectrum and the multimodal auxiliary data, extract multi-dimensional feature vectors; and use a pre-trained inversion model to determine the methane concentration inversion result.

[0041] Multidimensional feature vectors include: spectral domain features, optoelectronic domain features, environmental domain features, and hardware-related features.

[0042] Specifically, spectral domain features include the position of methane characteristic absorption peaks, peak height, peak area, 2f / 1f normalized harmonic ratio, full width at half maximum (FWHM) of absorption lines, and spectral morphology; optoelectronic domain features include laser operating current, laser temperature, detector temperature, and intensity fluctuation coefficient; environmental domain features include ambient temperature, ambient humidity, atmospheric pressure, and dust equivalent attenuation coefficient; and hardware coordination features include the intensity ratio of the reference beam to the measurement beam and the noise characteristics of the reference beam.

[0043] As a specific implementation, multi-dimensional effective feature extraction and standardization preprocessing are performed on multi-dimensional feature vectors. Sixteen categories of original features are extracted in batches. Redundant and invalid features with variance less than 0.01 are removed using a variance screening criterion, ultimately retaining thirteen core effective features, including six spectral domain features, four photoelectric domain features, two environmental domain features, and one hardware-related feature. The system unifies the dimensions of all core features through a standardization algorithm, eliminating model fitting interference caused by differences in parameter dimensions, and constructing a standardized feature dataset adapted to the deep learning inversion model, providing high-quality feature support for intelligent concentration inversion.

[0044] The intelligent inversion model is a hybrid model of attention-enhanced convolutional neural network and bidirectional long short-term memory network. It adopts a hybrid optimization algorithm of whale optimization algorithm and sparrow search algorithm to perform global optimization of initial weights and thresholds, balances the model's global search and local fine optimization capabilities, accurately optimizes the network's initial weights and thresholds, and uses the weighted sum of mean square error and mean absolute error as the loss function to effectively improve the model's convergence speed and inversion accuracy.

[0045] This application abandons the traditional linear method of solving concentration using a single harmonic ratio and the DE-BP neural network, and adopts a more advanced "CNN-BiLSTM hybrid model". Combining an attention mechanism, it constructs an intelligent inversion model that integrates four domains: spectral features, photoelectric parameters, environmental parameters, and hardware-coordinated features, achieving high-precision nonlinear inversion of methane concentration. The specific design of the CNN-BiLSTM hybrid model is as follows: The input layer has 13 neurons, corresponding to 13 input features, namely: methane second harmonic peak value, absorption peak area, 2f / 1f normalized harmonic ratio, absorption spectral line half width at half maximum, laser operating current, laser temperature, detector temperature, ambient temperature, ambient humidity, atmospheric pressure, dust equivalent attenuation coefficient, intensity ratio of reference beam to measurement beam, and noise characteristics of reference beam. The feature preprocessing layer employs an attention mechanism to automatically assign weights to each feature, focusing on methane absorption peak features and hardware-coordinated features to reduce interference from invalid features and improve model generalization. The hidden layer adopts a CNN-BiLSTM hybrid structure. The CNN layer is used to extract spatial features of spectral features and hardware co-located features, while the BiLSTM layer is used to extract temporal correlation features, avoiding the gradient vanishing problem and ensuring that the model is lightweight and easy to deploy in embedded systems. The CNN layer contains 2 convolutional layers and 1 pooling layer, and the BiLSTM layer contains 1 hidden layer (20 neurons). The output layer consists of one neuron, corresponding to the methane concentration value (unit: ppm or %VOL). The activation function is the Sigmoid function, which normalizes the output concentration value. The actual concentration is then obtained through inverse normalization.

[0046] like Figure 3As shown, the core parameter configuration is as follows: 13-dimensional neurons in the input layer, convolutional layer 1 with 3×1 kernels and 16 output channels, convolutional layer 2 with 5×1 kernels and 32 output channels, combined with a 2×1 max pooling layer, 20 neurons in the BiLSTM hidden layer, and single-dimensional concentration neurons in the output layer; the WOA-SSA algorithm has a population size of 30, a crossover probability of 0.7, a mutation probability of 0.1, a maximum number of iterations of 100, and a loss function MSE to MAE weight ratio of 0.6:0.4. After the standardized feature dataset is input into the model, the feature weights are adaptively allocated through a multi-head attention mechanism. The CNN layer extracts subtle spectral spatial features, the BiLSTM layer captures the dynamic features of the working conditions over time, and multi-dimensional feature information is fused through multi-layer nonlinear mapping to output a normalized concentration prediction value. The initial methane concentration inversion result is obtained through inverse normalization. The system compares the data with multi-scene calibration library data in real time, and fine-tunes the weights of the model output layer and the attention mechanism through gradient descent to correct the working condition adaptation bias and obtain high-precision concentration inversion values.

[0047] Following S105 are: The methane concentration inversion results are smoothed and optimized using a time-series sliding filter algorithm, and outlier identification and linear interpolation are performed using preset criteria to obtain the final methane concentration inversion results.

[0048] The time-series sliding filter algorithm is used to optimize data stability. Its filtering formula is as follows: ; In the formula, For the final output concentration, The model predicts the concentration at time ti. The value is the sliding window size, ranging from 5 to 10, which effectively suppresses concentration data jumps under dynamic operating conditions. A lightweight iterative calibration mechanism is also configured, eliminating the need for full model retraining; only the parameters of the shallow network and output layer are fine-tuned. Combined with regular calibration and sample library updates using standard gases, this ensures long-term stability in model detection accuracy and generalization ability under operating conditions.

[0049] As a specific implementation, to avoid instantaneous concentration jumps caused by dynamic scenarios such as livestock activity, manure disturbance, and airflow disturbance in farms, the system employs a time-series sliding filter algorithm with a window size of N=8 to smooth and optimize the initial inversion concentration, ensuring the stability of the data time series. Simultaneously, the 3σ criterion is used to identify outliers, determining that concentration data exceeding three times the standard deviation of the mean are outliers. Outliers are then filled in by linear interpolation of the preceding and following valid data, ensuring the continuity and completeness of the monitoring data. Finally, the system outputs optimized and accurate methane concentration data, synchronously storing the corresponding photoelectric conditions, environmental conditions, and hardware coordination parameters, enabling data traceability and replayability, and completing a single full-process detection output.

[0050] This application incorporates a lightweight, full-domain dynamic calibration and autonomous model maintenance mechanism to ensure the long-term accuracy and stability of equipment testing. The system monitors model inversion errors in real time. When the error exceeds the 1.5% accuracy threshold for five consecutive measurements, or when seasonal changes or significant shifts in operating conditions occur, a lightweight iterative update is automatically triggered. This involves only fine-tuning the parameters of the hidden and output layers, eliminating the need for full retraining. This results in high iteration efficiency and low hardware resource consumption. After iteration, accuracy verification and Allan variance stability analysis are performed using a dedicated validation sample set to ensure that the inversion error remains stable at ≤1.5% after iteration. Simultaneously, the system conducts precise manual calibration every 180 days using a 10-100 ppm gradient standard gas (uncertainty ±0.01 ppm), updating the calibration library sample data for multiple operating conditions, and continuously optimizing the model's generalization ability, achieving long-term maintenance-free, high-precision, and stable operation of the equipment.

[0051] To fully verify the practical engineering application effect of this application, a 30-day all-weather comparative experiment was conducted using three typical operating conditions: livestock and poultry farm pens, manure storage ponds, and anaerobic fermentation zones. These three areas covered the entire range of methane concentrations and interference conditions within the farms. A classic TDLAS scheme, employing traditional wavelet + SVD noise reduction, DE-BP inversion, and bilinear interpolation temperature and pressure correction, was used as a control group. Accuracy verification was completed using gradient standard gases. Figure 5 As shown, the average relative errors of this application's scheme in the pen, manure storage tank, and anaerobic fermentation zone are 0.82%, 0.95%, and 1.13%, respectively, with a stable overall error of ≤1.5%FS, far superior to the traditional schemes' 3.76%, 4.21%, and 4.85%. In the scenario of detecting weak signals at low concentrations of 10-30 ppm, the average error of this application is only 0.78%, with a detection sensitivity of up to 0.1 ppm and a detection limit as low as 34.8 ppb at a 100s integration time. The ability to extract weak signals and the accuracy of low-concentration detection are significantly superior to traditional schemes. Under extremely harsh conditions of high humidity and high dust, the average error of this application is still controlled within 1.3%, and the signal-to-noise ratio is stably maintained above 30:1, with significantly better anti-interference capabilities than traditional schemes. At the engineering deployment level, the lightweight algorithm of this application takes only 85ms for a single processing, which meets the real-time requirement of ≤100ms. The CPU utilization rate of the embedded terminal is stable at 45%~55% and the memory utilization is ≤30%. It can run stably and without failure for a long time. The accuracy can be further improved after dynamic iteration of the model. There is no accuracy decay in long-term operation, which is fully adapted to the high-precision monitoring needs of the whole scene and all weather of the breeding farm.

[0052] This application, based on the Lambert-Beer absorption law of tunable laser absorption spectroscopy and combined with the beam characteristics of a simple spectroscopic structure, constructs a theoretical model of the original absorption spectrum of methane in aquaculture farms, providing a theoretical foundation for subsequent algorithm design. The core expression of the Lambert-Beer law is: ; In the formula, Laser incident light intensity (unit: W / m²) 2 The value is calibrated in real time by the reference beam of the beam splitter. Ideally, it is a constant value, but in reality, it changes due to fluctuations in laser power and light intensity attenuation. Laser transmitted light intensity (unit: W / m) 2 The intensity of the laser beam after absorption by methane gas (measured beam) includes the methane absorption signal, various noise and interference signals; The methane gas concentration (unit: ppm or %VOL) is the target parameter to be retrieved. The equivalent optical path length (in meters) between the laser and methane gas is calculated without the need for hardware measurement. It is indirectly corrected by combining light intensity attenuation compensation at the algorithm level with reference beam data. For methane molecules at wavelength ,temperature air pressure Absorption coefficient under the given conditions (unit: m) -1 ·ppm -1 Its value changes dynamically with temperature and air pressure, and it is a key parameter for concentration inversion.

[0053] Under actual operating conditions in a livestock farm, the measured spectral signal is not an ideal methane absorption spectrum, but rather a mixed and distorted signal superimposed with various noises and interferences. Furthermore, there is nonlinear coupling between the noise sources. Its actual expression is: ; In the formula, This refers to the internal noise of the optoelectronic system, including laser current fluctuations, light intensity noise introduced by temperature drift, detector dark current noise, baseline drift noise, etc. This refers to external noise sources affecting the farm environment, including noise from water vapor overlap absorption, dust scattering, and crosstalk from other gases. This refers to random interference noise, including random noise caused by electromagnetic interference, environmental vibration, etc. This refers to noise coupling interference, which is the additional interference generated by the interaction between various noise sources. Traditional solutions have not effectively addressed this issue.

[0054] The core objective of this application is to achieve, through the collaboration of algorithms and simple hardware, [the following is a separate, unrelated sentence:] From Various noises, interferences, and coupling components are gradually removed to extract the pure methane absorption signal. This allows for the inversion of the precise methane concentration. .

[0055] Compared to traditional TDLAS detection technology and the latest deep learning research solutions, this application has outstanding advantages such as high sensitivity in weak signal extraction, high accuracy in decoupling multi-source coupling interference, small concentration inversion error under all operating conditions, strong long-term operational stability, and embeddable lightweight real-time deployment. Through the synergistic cooperation of algorithms and hardware, it achieves accurate methane concentration inversion in all scenarios of livestock farms. It can fully adapt to the complex operating conditions of livestock farms, including pens, manure storage ponds, anaerobic fermentation zones, and extreme high temperature, high humidity, and dust. Compared with existing technologies, its overall performance is significantly improved. It can solve problems such as low inversion accuracy, poor robustness, and insufficient sensitivity caused by high humidity, high dust, multiple mixed gases, temperature and pressure fluctuations, and photoelectric system disturbances in livestock farms. This application has extremely high scientific research value and engineering application value, and can be widely used in scenarios such as methane emission monitoring and environmental governance assessment in livestock and poultry farms.

[0056] In an exemplary embodiment, a computer device is provided, which may be a server or a terminal. The computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is connected to the system bus via the I / O interfaces. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The I / O interfaces of the computer device are used for exchanging information between the processor and external devices. The communication interface of the computer device is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a methane concentration inversion method for a livestock farm scenario.

[0057] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0058] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0059] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0060] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0061] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0062] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchain. The processors involved in the embodiments provided in this application may be, but are not limited to, general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc.

[0063] In this application, all actions to acquire signals, information, or data are carried out in compliance with the relevant data protection laws and policies of the country where the location is situated, and with the authorization granted by the owner of the relevant device.

[0064] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0065] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for retrieving methane concentration in a livestock farm setting, characterized in that, include: Simultaneously acquire multimodal auxiliary data; The multimodal auxiliary data includes: the measurement beam signal of the methane laser spectrum, the reference beam signal, as well as photoelectric parameters and environmental parameters; The measured beam signal is subjected to joint noise reduction processing to obtain the methane absorption spectrum signal; the joint noise reduction processing combines an adaptive noise complete set empirical mode decomposition algorithm, an energy ratio criterion, and wavelet packet noise reduction processing. Based on the multimodal auxiliary data, the methane absorption spectrum signal is subjected to photoelectric intrinsic error adaptive compensation to obtain a standardized spectral signal; the photoelectric intrinsic error adaptive compensation includes wavelength drift correction, light intensity normalization and baseline drift correction, and harmonic signal amplitude photoelectric correction. The standardized spectral signal is subjected to multi-branch environmental interference decoupling processing based on a four-branch parallel deep learning interference decoupling network to obtain the methane characteristic spectrum; the four-branch parallel deep learning interference decoupling network includes: water vapor interference decoupling branch, dust interference decoupling branch, crosstalk suppression branch and temperature and pressure coupling correction branch. Based on the methane characteristic spectrum and the multimodal auxiliary data, multidimensional feature vectors are extracted; and a pre-trained inversion model is used to determine the methane concentration inversion result.

2. The methane concentration inversion method for aquaculture farm scenarios according to claim 1, characterized in that, The joint noise reduction processing of the measured beam signal to obtain the methane absorption spectrum signal specifically includes: The measurement beam signal is decomposed into multiple scales using an adaptive noise complete set empirical mode decomposition algorithm to obtain multiple intrinsic mode components and one residual component. The intrinsic mode components and residual components are screened using the energy ratio criterion to obtain the effective components containing the effective absorption signal of methane; The effective components are subjected to wavelet packet denoising, and the wavelet packet coefficients of each layer are thresholded using an adaptive threshold function. The processed effective components are reconstructed to obtain the methane absorption spectrum signal.

3. The methane concentration inversion method for aquaculture farm scenarios according to claim 1, characterized in that, The step of adaptively compensating for photoelectric intrinsic errors in the methane absorption spectrum signal based on the multimodal auxiliary data to obtain a standardized spectral signal specifically includes: Using formula The center wavelength offset of the laser is determined, and the wavelength drift correction of the methane absorption spectrum signal is performed using the center wavelength offset of the laser; wherein, This is the center wavelength offset of the laser. These are the fitting coefficients. The light intensity of the reference beam, This is the operating current. Core temperature; Using formula and Light intensity normalization and baseline drift correction are performed; among them, For reference beam, The measured spectral signal, The normalized spectral signal, This represents the baseline drift. The fitting coefficients were determined through baseline calibration experiments at different detector temperatures. For the detector temperature; Using formula Photoelectric correction of harmonic signal amplitude is performed; among which, This represents the ratio of the original second harmonic 2f to the first harmonic 1f. This is the photoelectric correction factor, determined by the laser's operating current, core temperature, detector temperature, and the intensity of the reference beam. It is used to correct harmonic amplitude errors caused by photoelectric parameter drift. This is the corrected ratio of the second harmonic 2f to the first harmonic 1f.

4. The methane concentration inversion method for aquaculture farm scenarios according to claim 1, characterized in that, The four-branch parallel deep learning interference decoupling network performs multi-branch environmental interference decoupling processing on the standardized spectral signal to obtain the methane characteristic spectrum, specifically including: In the water vapor interference decoupling branch, the water vapor concentration is dynamically estimated based on the real-time collected ambient temperature, air pressure and relative humidity, and the contribution of water vapor absorption signal is calculated by combining the pre-constructed water vapor absorption coefficient database. The contribution of water vapor absorption signal is then differentially extracted from the standardized spectral signal. In the dust interference decoupling branch, based on the real-time dust concentration and spectral line morphology characteristics, the dynamic light intensity attenuation coefficient is determined by substituting it into the pre-constructed dust attenuation model. Combined with the differential processing of the measured beam signal and the reference beam signal, the light intensity attenuation error caused by dust scattering and broadband absorption is compensated. In the crosstalk suppression branch, a crosstalk feature library is constructed and orthogonalized, and the normalized spectral signal is mapped to the orthogonal complement space of the crosstalk feature library. In the temperature-pressure coupling correction branch, based on the real-time collected ambient temperature and air pressure, an intelligent optimization interpolation algorithm is used to dynamically determine the methane molecule absorption coefficient under the current operating conditions from a pre-constructed three-dimensional absorption coefficient database, and substitute it into Lambert-Beer law to correct the spectral absorption parameters; the intelligent optimization interpolation algorithm is an interpolation algorithm optimized based on the improved whale optimization algorithm.

5. The methane concentration inversion method for aquaculture farm scenarios according to claim 1, characterized in that, Multidimensional feature vectors include: spectral domain features, optoelectronic domain features, environmental domain features, and hardware-related features.

6. The methane concentration inversion method for aquaculture farm scenarios according to claim 1, characterized in that, The intelligent inversion model is a hybrid model of attention-enhanced convolutional neural network and bidirectional long short-term memory network; and a hybrid optimization algorithm of whale optimization algorithm and sparrow search algorithm is used to globally optimize the initial weights and thresholds.

7. The methane concentration inversion method for aquaculture farm scenarios according to claim 1, characterized in that, The multi-dimensional feature vector is extracted based on the methane characteristic spectrum and the multimodal auxiliary data; The pre-trained inversion model is used to determine the methane concentration inversion results, followed by: The methane concentration inversion results are smoothed and optimized using a time-series sliding filter algorithm, and outlier identification and linear interpolation are performed using preset criteria to obtain the final methane concentration inversion results.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the methane concentration inversion method for a farm scenario according to any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the methane concentration inversion method for a farm scenario as described in any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the methane concentration inversion method for a farm scenario as described in any one of claims 1-7.

Images