High-precision detection system and method for harmful gas in equipment pipeline based on multispectral fusion

Abstract

The present application relates to the technical field of gas detection, and discloses a high-precision detection system and method for harmful gas in equipment pipeline based on multispectral fusion, which first collects and pre-processes multispectral data, then uses an online learning mechanism to identify and reconstruct main background and interference spectrum by a dynamic background and interference modal online learning and reconstruction module, and then subtracts the interference spectrum from the pre-processed data to obtain residual spectrum by a target gas signal extraction and enhancement module, and extracts and enhances target gas characteristic spectral components by orthogonal projection based on standard spectral information. Finally, a variety of gas concentration inversion modules calculate the concentration of target harmful gas based on the characteristic and standard spectral information. Dynamic interference is removed by online learning, the reliability of signal extraction is improved by combining orthogonal projection and dynamic calibration spectral library, the resolution precision of mixed gas is improved by a step-by-step strategy, and the long-term high precision and environmental adaptability of the system are ensured by an adaptive optimization mechanism.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of gas detection, in particular to a device pipeline harmful gas high-precision detection system and method based on multi-spectral fusion. BACKGROUND

[0002] In many key fields such as industrial safety monitoring, environmental quality assessment and emergency response, accurate and reliable detection of harmful gases is of great significance. Especially in the device pipeline system involving flammable, explosive or toxic substance transportation, timely detection and accurate quantification of potential gas leakage are the core links to prevent accidents, ensure personnel safety and protect the environment. Multi-spectral analysis technology can capture rich spectral "fingerprint" information of substances, providing an effective technical approach for identification and quantitative analysis of various gas components, and has been widely studied and applied in the field of gas detection. However, in the actual deployment and application of these detection systems based on multi-spectral analysis, especially in the pursuit of high precision and high robustness, the existing technology still faces several challenges, which limit its performance in complex dynamic environments.

[0003] Specifically, some existing multi-spectral gas detection technologies still have insufficient adaptability and processing capacity when dealing with complex and dynamically changing background signals and unknown disturbances in actual working conditions. For example, fluctuations in the concentration of water vapor in the environment, changes in dust particles, or dynamic disturbances of other non-target coexisting gases may have a significant impact on the spectral signal of the target gas. If the system cannot effectively identify and eliminate these time-varying interference components, it is easy to cause deviations in the detection results, affecting the accuracy of the detection. At the same time, most systems rely on pre-established standard gas spectral library for comparison and quantitative analysis. However, these standard spectra are usually measured under ideal laboratory conditions, and when the environmental conditions such as temperature and pressure change, or the optical properties of the instrument itself drift over time, the actual collected target gas spectral characteristics may not match the standard spectra in the library. This mismatch will directly affect the accuracy of signal extraction and the reliability of subsequent concentration inversion, especially for the detection of trace gases, small spectral deviations may also lead to large quantitative errors. In addition, when there are multiple gas mixtures in the environment to be detected, and their absorption spectra overlap, traditional spectral analysis models may not be able to completely separate the component signals, causing cross interference between components and reducing the accuracy of quantitative analysis of specific target gases in the mixture. Finally, many existing systems have a relatively fixed running mode after parameter setting, lacking the ability to adjust and optimize themselves according to real-time monitoring results and environmental changes, which not only may cause the system performance to gradually decline in long-term operation, but also increases the frequency and complexity of manual maintenance and calibration.

[0004] In view of this, the application provides a high-precision device pipeline harmful gas detection system and method based on multispectral fusion. SUMMARY

[0005] In view of the deficiencies of the prior art, the application provides a high-precision device pipeline harmful gas detection system and method based on multispectral fusion, which solves the problem of low precision and poor adaptability of existing gas detection technologies based on multispectral analysis under dynamic interference, spectral mismatch and multi-component cross-influence.

[0006] To achieve the above object, the application is implemented by the following technical solutions: a high-precision device pipeline harmful gas detection system based on multispectral fusion, comprising:

[0007] A multispectral data acquisition and preprocessing module is configured to acquire original multispectral data of a gas environment to be measured and preprocess the original multispectral data to output preprocessed multispectral data.

[0008] A dynamic background and interference mode online learning and reconstruction module is configured to receive the preprocessed multispectral data and identify and reconstruct main interference spectra caused by dynamic background and unknown interference in real time based on the preprocessed multispectral data using an online learning mechanism.

[0009] A target gas signal extraction and enhancement module is configured to receive the preprocessed multispectral data and the main interference spectra, obtain residual spectra by subtracting the main interference spectra from the preprocessed multispectral data, and perform orthogonal projection processing on the residual spectra using preset target gas standard spectral information to extract and enhance target gas characteristic spectral components.

[0010] A multispectral gas concentration inversion module is configured to receive the target gas characteristic spectral components and the preset target gas standard spectral information, and calculate the concentration of one or more target harmful gases in the gas environment to be measured based on the target gas characteristic spectral components and the preset target gas standard spectral information.

[0011] Preferably, the dynamic background and interference mode online learning and reconstruction module specifically comprises:

[0012] An online algorithm execution unit is configured to execute an online learning algorithm on the preprocessed multispectral data to identify and extract dynamic evolution rules or main change components in the data.

[0013] A main dynamic characteristic identification and selection unit is configured to identify and select spectral change modes representing main dynamic background and unknown interference according to the characteristics of the dynamic evolution rules or main change components output by the online algorithm execution unit.

[0014] a dynamic background and interference spectrum reconstruction unit configured to reconstruct a main interference spectrum based on the spectral variation mode and its corresponding real-time amplitude determined by the main dynamic characteristic identification and selection unit.

[0015] Preferably, the online algorithm execution unit adopts at least one of an online dynamic mode decomposition algorithm or a robust principal component analysis algorithm based on a sliding window.

[0016] Preferably, the target gas signal extraction and enhancement module comprises:

[0017] a residual spectrum calculation unit configured to subtract the reconstructed main interference spectrum from the pre-processed multi-spectrum data to obtain a residual spectrum;

[0018] a target gas standard spectrum subspace construction unit configured to construct a target gas feature subspace based on a pre-set target gas standard spectrum database;

[0019] an orthogonal projection and signal enhancement unit configured to orthogonally project the residual spectrum to the target gas feature subspace to extract and enhance a target gas feature spectrum component.

[0020] Preferably, the orthogonal projection and signal enhancement unit is further configured to calculate a non-target residual component orthogonal to the target gas feature subspace.

[0021] Preferably, the multi-gas concentration inversion module adopts at least one of a classic least squares method, a ridge regression or a partial least squares regression algorithm to perform concentration inversion on the target gas feature spectrum component.

[0022] Preferably, the pre-processing operation in the multi-spectrum data acquisition and pre-processing module includes at least one of dark spectrum deduction, light source reference correction or spectrum normalization, preliminary baseline correction and noise filtering.

[0023] Preferably, the system further comprises:

[0024] a system feedback and adaptive optimization module configured to generate a feedback signal based on the statistical characteristics of the non-target residual component or the stability and reasonableness of the concentration of one or more target harmful gases in the gas environment to be measured, and dynamically adjust the mode selection strategy of the dynamic background and interference mode online learning and reconstruction module, or the calibration parameters of the target gas standard spectrum library in the target gas signal extraction and enhancement module, or the parameters of the concentration inversion model in the multi-gas concentration inversion module based on the feedback signal.

[0025] The application also provides a device pipeline harmful gas high-precision detection method based on multispectral fusion, which is used for the device pipeline harmful gas high-precision detection system based on multispectral fusion and comprises the following steps:

[0026] Obtaining original multispectral data of a gas environment to be detected;

[0027] Pretreating the original multispectral data to obtain pretreated multispectral data;

[0028] Based on the pretreated multispectral data, using an online learning mechanism to identify and reconstruct a main interference spectrum caused by a dynamic background and unknown interference in real time;

[0029] Subtracting the reconstructed main interference spectrum from the pretreated multispectral data to obtain a residual spectrum;

[0030] Further, using preset target gas standard spectrum information to perform orthogonal projection processing on the residual spectrum to extract and enhance a target gas characteristic spectrum component;

[0031] Finally, based on the enhanced target gas characteristic spectrum component and the target gas standard spectrum information, inversely calculating to obtain the concentration of one or more target harmful gases in the gas environment to be detected.

[0032] Preferably, the one or more target harmful gases include at least one of a hydrocarbon gas, a sulfur-containing compound gas, a nitrogen oxide gas, carbon monoxide, chlorine and ammonia.

[0033] The application provides a device pipeline harmful gas high-precision detection system and method based on multispectral fusion. The application has the following beneficial effects:

[0034] 1. The application can identify and subtract a dynamically changing background and interference spectrum from original data in real time by using an online learning technology. This method makes the effective signal of a target gas more prominent, thereby improving the accuracy of detection. Compared with the signal misjudgment or omission that may occur when the prior art processes a changing background, the application solves the problem of insufficient precision in such a scene.

[0035] 2. The application applies an orthogonal projection method in the target gas signal extraction link, and designs a dynamic calibration mechanism for a target gas standard spectrum library used, so that even if the actual gas spectrum changes slightly due to environmental factors or the instrument state drifts, the system can more accurately capture the characteristics of the target gas and enhance the reliability of the signal, overcoming the defects of some prior arts in which the recognition accuracy decreases and the robustness is poor when the environment or instrument state changes due to the dependence on a fixed spectrum library.

[0036] 3、The application effectively improves the resolution ability of different components in the mixed gas by removing the main interference spectrum in advance and using standard spectrum information to extract the target gas signal, which reduces the mutual influence between the spectral characteristics of different gases, especially for those gases with similar spectral characteristics and easy to confuse, can realize more accurate concentration inversion, solves the problem of limited distinguishing ability of the prior art when facing complex mixed gas, which easily leads to deviation of the inversion result.

[0037] 4、The application comprises a system feedback and adaptive optimization module, which monitors the stability of the final output concentration and the characteristics of the non-target residual components, and dynamically adjusts the working parameters of the front-end spectrum processing module according to these information, which enables the system to optimize itself, maintains high detection performance for a long time, and adapts to changes in environmental conditions, compared with those systems with fixed parameters and frequent manual calibration, the application improves the convenience of operation and the accuracy of continuous detection. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 The figure is the framework flow chart of the system of the application;

[0039] Figure 2 The figure is the framework flow chart of the dynamic background and interference mode online learning and reconstruction module of the application;

[0040] Figure 3 The figure is the framework flow chart of the target gas signal extraction and enhancement module of the application. DETAILED DESCRIPTION

[0041] The technical solutions in the embodiments of the application will be described clearly and completely in combination with the drawings of the application specification. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the application.

[0042] Please refer to the drawings of the application specification Figure 1 The embodiments of the application provide a high-precision device pipeline harmful gas detection system based on multi-spectrum fusion, which comprises:

[0043] A multi-spectrum data acquisition and preprocessing module is used to acquire original multi-spectrum data of the gas environment to be measured, and to preprocess the original multi-spectrum data to output preprocessed multi-spectrum data;

[0044] To achieve the high-precision detection of harmful gases in equipment pipelines based on multispectral fusion, an exemplary system can include a multispectral data acquisition and preprocessing module. The function of this module is to obtain the original multispectral data of the gas environment to be detected. The obtained original multispectral data is then preprocessed in this module. The preprocessing aims to eliminate or weaken the noise and irrelevant information in the original data, laying the foundation for subsequent accurate analysis. The module finally outputs the preprocessed multispectral data.

[0045] In some embodiments, the preprocessing operation performed by the multispectral data acquisition and preprocessing module can include, but is not limited to, at least one of the following steps. For example, dark spectrum subtraction can be performed first. This step is used to correct systematic deviations introduced by factors such as sensor dark current.

[0046] Further, the preprocessing operation can also include light source reference correction or spectral normalization. Light source reference correction uses a reference spectrum to correct light source intensity fluctuations or sensor response inconsistencies. Spectral normalization adjusts the spectral data to a uniform scale, reduces overall intensity differences caused by changes in optical path, sample concentration, etc., and highlights spectral shape features.

[0047] In addition, the preprocessing operation can also include preliminary baseline correction. This step aims to eliminate or weaken the spectral baseline changes caused by factors such as instrument drift, scattering, etc. Finally, noise filtering can also be included, such as using mean filtering, median filtering, or wavelet denoising algorithms, to smooth the spectral curve and improve the signal-to-noise ratio. These preprocessing steps work together to produce preprocessed multispectral data required for subsequent analysis.

[0048] The system can further include a dynamic background and interference mode online learning and reconstruction module. This module receives the preprocessed multispectral data output by the multispectral data acquisition and preprocessing module. Its core function is to use an online learning mechanism to identify and reconstruct the main interference spectrum caused by dynamic background and unknown interference in real time.

[0049] The dynamic background and interference mode online learning and reconstruction module, in its specific internal structure, can include an online algorithm execution unit. This unit directly receives the preprocessed multispectral data. It is responsible for executing an online learning algorithm to identify and extract the dynamic evolution law or main change component contained in the input multispectral data sequence.

[0050] Then, the module can also include a main dynamic characteristic identification and selection unit. This unit is coupled to the online algorithm execution unit. It analyzes and judges the characteristics of the dynamic evolution law or main change component output by the online algorithm execution unit, such as the energy, frequency, persistence, or background contribution of each mode. Through this process, those spectral change modes that can represent the main dynamic background and unknown interference are identified and selected.

[0051] Finally, the module can also include a dynamic background and interference spectrum reconstruction unit. This unit is coupled to the main dynamic property identification and selection unit. It utilizes the spectral variation modes determined by the main dynamic property identification and selection unit, as well as the real-time amplitudes or coefficients corresponding to these modes. By linearly combining or other reconstruction methods on these selected modes and their corresponding weights, the main interference spectrum is finally reconstructed.

[0052] In a specific embodiment, the online learning algorithm employed by the online algorithm execution unit can be the online dynamic mode decomposition (OnlineDMD) algorithm. When the OnlineDMD algorithm is employed, let the pre-processed multispectral data vector acquired at time point k be where N is the number of spectral channels.

[0053] The OnlineDMD algorithm aims to learn an approximate linear dynamical system operator A such that x k+1 ≈Ax k . This operator A can be efficiently updated as each new data point arrives. For example, the estimation of A or its low-rank approximation can be updated by recursive least squares or incremental singular value decomposition based methods.

[0054] By performing eigenvalue decomposition on the updated operator A, a series of DMD modes Φ = [φ1, φ2, …, φ r ] and corresponding DMD eigenvalues Λ = diag(λ1, λ2, …, λ r ) can be obtained. Here, φ i represents a coherent spectral structure, while λ i characterizes its temporal dynamic property (growth / decay rate and oscillation frequency). These DMD modes and eigenvalues constitute the “dynamic evolution law”.

[0055] In another specific embodiment, the online learning algorithm employed by the online algorithm execution unit can also be the robust principal component analysis (RobustPCA, RPCA) algorithm based on sliding window. RPCA aims to decompose the observed data matrix D into a low-rank matrix L and a sparse matrix S, i.e., D = L + S.

[0056] In this application, the observed data matrix is composed of W pre-processed multispectral data vectors x k acquired within a sliding time window W, with each column being a spectral vector. The low-rank matrix L is used to represent the slowly varying dynamic background and main interference components. The sparse matrix S is used to represent the burst signals of the target gas or part of the noise.

[0057] The RPCA problem is usually solved by solving the following convex optimization problem:

[0058] s.t.D = L + S, where, ||L||F * represents the nuclear norm (sum of singular values) of matrix L, which is used to constrain the low-rank property of L. ||S||1represents the L1-norm (sum of absolute values of all elements) of matrix S, which is used to constrain the sparsity of S. γ is a regularization parameter to balance the importance of low-rank term and sparsity term.

[0059] The sliding window based RPCA means that as new spectral data continuously enters, the data matrix DDis updated, and the RPCA decomposition is re-performed or incrementally performed. At this time, the column vectors of the low-rank matrix L obtained by decomposition or the principal components thereof constitute the "principal variation components".

[0060] Whether online DMD or sliding window based RPCA is adopted, the principal dynamic characteristic identification and selection unit receives the modes or components output by these algorithms. For example, in DMD, those DMD modes with higher energy, lower frequency or longer duration can be selected as the modes representing the background interference. In RPCA, the low-rank matrix L itself or its principal principal components can be directly regarded as the principal components of the background and interference.

[0061] Subsequently, the dynamic background and interference spectral reconstruction unit reconstructs the main interference spectrum x j (t) at the current time point using these selected spectral variation modes (such as selected DMD modes {φ j∈selected (t)} or RPCA low-rank components) and their corresponding real-time amplitude or projection coefficients. interference For example, for DMD, it can be expressed as x interference (t) = Σ j∈selected φ j a j (t), where a j (t) is the amplitude of the mode φ j at time t. For RPCA, the reconstructed interference spectrum is the corresponding column in the current window low-rank matrix L.

[0062] Through the collaborative work of the above-mentioned multi-spectral data acquisition and preprocessing module and the dynamic background and interference mode online learning and reconstruction module, the dynamically changing background and interference signals can be effectively separated from the original complex multi-spectral data, providing a more pure data basis for subsequent target gas signal extraction and concentration inversion, thereby helping to improve the precision and robustness of the detection system.

[0063] Please refer to the accompanying drawings Figure 2, a dynamic background and interference modalities online learning and reconstruction module for receiving the pre-processed multispectral data and identifying and reconstructing, in real time, a main interference spectrum caused by a dynamic background and unknown interference based on the pre-processed multispectral data using an online learning mechanism;

[0064] The system in the embodiments of the present application can further include a dynamic background and interference modalities online learning and reconstruction module. The module is coupled to the multispectral data acquisition and preprocessing module described above for receiving the pre-processed multispectral data. The core function of the module is to identify and reconstruct, in real time, a main interference spectrum caused by a dynamic background and unknown interference based on the received pre-processed multispectral data using an online learning mechanism.

[0065] In an exemplary embodiment, the dynamic background and interference modalities online learning and reconstruction module can include an online algorithm execution unit. The unit is responsible for receiving and processing the pre-processed multispectral data sequence output by the multispectral data acquisition and preprocessing module. The main task of the unit is to execute an online learning algorithm for the purpose of identifying and extracting the dynamic evolution law or main change component contained in the input multispectral data stream.

[0066] Following the unit, the module can further include a main dynamic characteristic identification and selection unit. The unit is logically connected to the output end of the online algorithm execution unit. It analyzes and judges the specific characteristics of the dynamic evolution law or main change component output by the online algorithm execution unit, such as the energy contribution of each modality, the time-frequency characteristics, the persistence, or the correlation with the background, etc. Through this process, the unit can identify and accurately select the spectral change modalities that best represent the main dynamic background and unknown interference.

[0067] Finally, the dynamic background and interference modalities online learning and reconstruction module can further include a dynamic background and interference spectrum reconstruction unit. The unit is connected to the main dynamic characteristic identification and selection unit. It uses the spectral change modalities finally determined by the main dynamic characteristic identification and selection unit, as well as the real-time amplitudes, coefficients or weights corresponding to the selected modalities at the current time. Through effective combination of the selected modalities and their corresponding weight information, such as linear superposition or other suitable reconstruction algorithms, the main interference spectrum is finally accurately reconstructed.

[0068] To be more specific, when the online algorithm execution unit uses an online dynamic modal decomposition (Online DMD) algorithm, the pre-processed multispectral data vector obtained at a discrete time point k is denoted as where N represents the number of channels or wavelength points of the spectral data. The goal of the online DMD algorithm is to learn an approximate linear dynamic system operator so that the observation sequence satisfies xk+1 ≈Ax k .

[0069] The estimate of operator A can be obtained at each new data point x. k Efficient online updates can be performed upon arrival. For example, the estimate of A itself, or its low-rank approximation, can be updated using a recursive least squares (RLS) algorithm or a method based on incremental singular value decomposition (SVD).

[0070] By performing eigenvalue decomposition on the currently updated operator A, i.e., AΦ=ΦΛ, a set of DMD modes Φ=[φ1,φ2,…,φ r ] and their corresponding DMD eigenvalues ​​Λ=diag(λ1,λ2,…,λ r Here, This represents a spectral mode with a coherent structure, while the complex eigenvalues This characterizes the time-dynamic properties of the mode, including its growth / decay rate (as defined by |λ). i |determined) and oscillation frequency (determined by arg(λ) i (This is determined by...) These DMD modes and eigenvalues ​​together constitute the "dynamic evolution law".

[0071] In another exemplary implementation, the online learning algorithm employed by the online algorithm execution unit can be a robust principal component analysis (RPCA) algorithm based on a sliding window. The basic idea of ​​RPCA is to decompose an observation data matrix D into the sum of a low-rank matrix L and a sparse matrix S, i.e., satisfying D = L + S.

[0072] In this application scenario, the observation data matrix W preprocessed multispectral data acquired within a sliding time window of fixed length W are directed to x. k The data is structured as follows: each column vector represents a spectral data point. The low-rank matrix L is used to characterize the slowly changing dynamic background and major interfering components in the data, as these components often exhibit structural similarities across multiple consecutive spectral observations. The sparse matrix S is used to characterize sudden, sparse spectral signals of the target gas caused by leaks or other reasons, or partially independent noise.

[0073] RPCA problems can usually be solved by solving the following convex optimization problems:

[0074] Where D is the observation data matrix, L is the low-rank matrix, S is the sparse matrix, and ||L|| * =∑ i σi (L) denotes the nuclear norm of the low-rank matrix L, i.e., the sum of the singular values σ i (L) of L, which is a convex approximation of the rank function, is used to enforce the low-rank structure of L; ||S||1=∑ i,j |S ij | denotes the L1 norm of the sparse matrix S, i.e., the sum of the absolute values of all elements in S, which is used to enforce the sparsity of S. γ>0 is a regularization parameter to balance the relative importance of the low-rank term and the sparsity term in the optimization objective, which is usually determined empirically or by cross-validation.

[0075] The sliding window based RPCA means that the data matrix D is updated accordingly when new spectral data enters the window while the oldest spectral data moves out of the window. Subsequently, the RPCA decomposition is re-executed for the updated D or an incremental RPCA algorithm is used for efficient update. In this process, the column vectors of the low-rank matrix L obtained from the decomposition, or the principal components obtained from L by singular value decomposition, constitute the "principal variation components".

[0076] Regardless of whether the online DMD algorithm or the sliding window based RPCA algorithm is used, the principal dynamic characteristic identification and selection unit receives the output results of these algorithms. For example, when the online DMD is used, the unit can select those DMD modes that mainly represent the background and the disturbance according to characteristics such as the energy of each DMD mode (e.g., measured by its initial amplitude or the projection energy on the data), the size of the eigenvalue (close to the unit circle indicates persistent existence), the frequency (low frequency usually corresponds to the background), and the like.

[0077] When the RPCA is used, since the low-rank matrix L is originally designed to capture the background and the main disturbance, the low-rank matrix L itself or several principal components extracted from it by further principal component analysis can be directly selected by the principal dynamic characteristic identification and selection unit as spectral variation modes representing the principal dynamic background and unknown disturbance.

[0078] Subsequently, the dynamic background and disturbance spectral reconstruction unit will use these spectral variation modes (e.g., a selected group of DMD modes {φ j} j∈selectedindices (k) obtained from the principal dynamic characteristic identification and selection unit, as well as their corresponding real-time amplitudes or projection coefficients, to reconstruct the principal disturbance spectrum at the current time, denoted as x interference (k).

[0079] Specifically, if the DMD method is used, the principal disturbance spectrum at time k can be reconstructed by a linear combination of the selected background modes:

[0080] xinterference (k) =∑ j∈selected indices φ j b j (k), where b j (k) is the modal φ j Amplitude at k time; x interference (k) is the reconstructed main interference spectrum vector at k time, with dimension N x 1; φ j is the jth selected DMD (Dynamic Modal Decomposition) modal vector, representing a coherent spectral structure, with dimension N x 1; b j (k) is the DMD modal φ j Amplitude or coefficient at k time; ∑ j∈selected indices represents the summation of all selected DMD modals representing background and interference. These amplitudes can be estimated online by projecting the original data onto the DMD modals or other ways.

[0081] If the RPCA method is used, the main interference spectrum at k time can be directly taken from the column vector in the low-rank matrix L k decomposed from the data matrix D k at k time, i.e.:

[0082] x interference (k) = (L k ) :,c , where (L k ) :,c represents the cth column of the low-rank matrix L k obtained at k time (corresponding to the current sliding window); x interference (k) is the reconstructed main interference spectrum vector at k time.

[0083] Through the accurate operation of the above online learning and reconstruction module of dynamic background and interference modals, the system can effectively separate the main spectral components of the dynamically changing background and unknown interference from the continuously collected multi-spectral data in real time. This provides a more pure data basis for the subsequent target gas signal extraction and concentration inversion steps, thereby helping to significantly improve the detection accuracy, sensitivity and robustness of the entire harmful gas detection system in complex environments.

[0084] Please refer to the attached Figure 3 , the target gas signal extraction and enhancement module is used to receive the pre-processed multi-spectral data and the main interference spectrum, to obtain the residual spectrum by subtracting the main interference spectrum from the pre-processed multi-spectral data, and to perform orthogonal projection processing on the residual spectrum using the preset target gas standard spectrum information, to extract and enhance the target gas characteristic spectrum component;

[0085] The system in the embodiments of the present application can further include a target gas signal extraction and enhancement module. The upstream input ends of the module are coupled to the multispectral data acquisition and preprocessing module and the dynamic background and interference modal online learning and reconstruction module respectively. It is configured to receive the preprocessed multispectral data output by the former and the main interference spectrum output by the latter. Its core function is to extract and enhance the characteristic spectral component of the target gas through specific processing.

[0086] Specifically, the module first obtains a residual spectrum by subtracting the main interference spectrum from the preprocessed multispectral data. Subsequently, it performs orthogonal projection processing on the residual spectrum using preset target gas standard spectral information. In this way, the module can effectively extract the characteristic spectral signal of the target gas and enhance the intensity of the signal relative to residual noise and interference to some extent, and finally output the target gas characteristic spectral component.

[0087] In an exemplary embodiment, the target gas signal extraction and enhancement module can specifically include a residual spectrum calculation unit. The input ends of the unit are connected to the output ends of the multispectral data acquisition and preprocessing module and the output ends of the dynamic background and interference modal online learning and reconstruction module respectively. It is used to receive the preprocessed multispectral data, denoted as (wherein k is a time index, and N is the number of spectral channels), and the reconstructed main interference spectrum, denoted as

[0088] The operation performed by the residual spectrum calculation unit is to subtract the reconstructed main interference spectrum from the preprocessed multispectral data. The difference obtained in this way is the residual spectrum, denoted as x residual (k). Its calculation process can be represented as:

[0089] x residual (k) = x k - x interference (k), wherein x residual (k) is the residual spectrum vector calculated at k time, with a dimension of N x 1; x k is the preprocessed multispectral data vector at k time, with a dimension of N x 1; x interference (k) is the main interference spectrum vector reconstructed at k time.

[0090] The residual spectrum x residual (k) theoretically mainly contains the spectral signal of the target gas and part of the residual interference and noise that has not been completely removed. The residual spectrum calculated by the unit is output to the subsequent processing unit.

[0091] Further, the target gas signal extraction and enhancement module can further comprise a target gas standard spectrum subspace construction unit. The function of this unit is to construct a mathematical subspace, i.e. a target gas feature subspace, which can effectively represent the spectral characteristics of one or more target gases, according to a pre-set target gas standard spectrum database. The database stores the standard reference spectra of various target gases that are expected to be detected.

[0092] Suppose the target gas standard spectrum database contains the standard spectra of M target gases, which can be arranged into a matrix where each column is the standard spectrum vector of the jth target gas. The target gas standard spectrum subspace construction unit then uses this matrix S to construct the target gas feature subspace.

[0093] A commonly used method is to perform singular value decomposition (SVD) on the standard spectrum matrix S: S = U∑V T where and are orthogonal matrices, is a diagonal matrix whose diagonal elements are singular values. Select the P left singular vectors in U corresponding to the larger singular values (e.g. greater than a certain threshold or the top P according to energy proportion), to form an orthogonal basis matrix The column vectors of this matrix U S span the target gas feature subspace. Alternatively, the Gram-Schmidt orthogonalization method can also be used to extract the orthogonal basis from S. The unit outputs the constructed subspace basis U S .

[0094] Finally, the target gas signal extraction and enhancement module can further comprise an orthogonal projection and signal enhancement unit. The input terminals of this unit are connected to the output terminal of the residual spectrum calculation unit and the output terminal of the target gas standard spectrum subspace construction unit, respectively. It receives the residual spectrum x residual (k) and the basis matrix U S of the target gas feature subspace.

[0095] The core operation of the orthogonal projection and signal enhancement unit is to orthogonally project the received residual spectrum x residual (k) onto the target gas feature subspace. Through this projection operation, the components of the residual spectrum that are most relevant to the spectral characteristics of the target gas can be extracted, and those noise and interference components that are orthogonal to the target gas feature subspace can be suppressed, thereby achieving the extraction and enhancement of the target gas feature spectral components.

[0096] The calculation formula of the target gas feature spectral components x target (k) is:

[0097] wherein, is the orthogonal projection operator of the projection to the target gas feature subspace; x target (k) is the target gas feature spectrum component vector extracted and enhanced at k moment; U S is the orthogonal basis matrix of the target gas feature subspace constructed by the target gas standard spectrum library; x residual (k) is the residual spectrum vector at k moment; x target (k) is one of the main results output by this unit, which will be passed to the subsequent multi-gas concentration inversion module.

[0098] In addition, in some embodiments, the orthogonal projection and signal enhancement unit is also configured to calculate a non-target residual component orthogonal to the target gas feature subspace. The non-target residual component x non-target (k) represents the part of the residual spectrum that cannot be explained by the target gas feature subspace, mainly including residual noise and other unknown interference.

[0099] The calculation of the non-target residual component x non-target (k) can be obtained by subtracting its projection on the target gas feature subspace from the residual spectrum: x non-target (k) = x residual (k) - x target (k) or, it can also be obtained by projecting to the complementary space orthogonal to the target gas feature subspace:

[0100] wherein, I is an N x N identity matrix, x non-target (k) is the non-target residual component vector calculated at k moment; x residual (k) is the residual spectrum vector at k moment; is the orthogonal projection operator of the projection to the complementary space orthogonal to the target gas feature subspace. This non-target residual component x non-target (k) can be output to the system feedback and adaptive optimization module for evaluating the system performance and adjusting the parameters.

[0101] Through the cooperative work of the above-mentioned units, the target gas signal extraction and enhancement module can effectively separate and enhance the feature signal of the target gas from the residual spectrum after preliminary interference suppression, while also providing information about the non-target component, providing key support for the high precision and adaptive optimization of the entire detection system.

[0102] The multi-gas concentration inversion module is configured to receive the target gas characteristic spectral component and preset target gas standard spectral information, and to calculate the concentration of one or more target harmful gases in the gas environment to be measured based on the target gas characteristic spectral component and the preset target gas standard spectral information.

[0103] The system in the embodiment of the present application can further include a multi-gas concentration inversion module. The input end of the module is coupled to the output end of the target gas signal extraction and enhancement module. The module is configured to receive the target gas characteristic spectral component output by the module, and to receive or pre-store preset target gas standard spectral information. Based on the two types of information, the core task of the module is to calculate the concentration of one or more target harmful gases in the gas environment to be measured by inversion.

[0104] The target gas characteristic spectral component is denoted as , where N is the number of spectral channels. The preset target gas standard spectral information is usually represented as a standard spectral library matrix , where M is the number of target harmful gases. j Each column s j represents the standard spectral response of the jth target harmful gas at a unit concentration (or other known concentration).

[0105] Based on the basic principles of spectroscopy, such as the extended form of the Beer-Lambert law, which indicates that the absorbance (or spectral characteristic quantity related thereto) of a mixture is a linear superposition of the absorbance of each component, the target gas characteristic spectral component x target (k) can be approximately represented as a linear combination of the standard spectral library matrix S and the gas concentration vector , where c j (k) is the concentration of the jth target harmful gas at time k.

[0106] In an exemplary embodiment, the multi-gas concentration inversion module can use the classical least squares (CLS) method for concentration inversion. The CLS method aims to find a concentration vector c(k) that minimizes the sum of squared residuals between the reconstructed spectrum Sc(k) and the observed target gas characteristic spectral component x target (k). That is, to solve the following optimization problem:

[0107] , where c(k) is the target harmful gas concentration vector to be solved at time k; x target (k) is the target gas characteristic spectral component vector extracted and enhanced at time k; S is the target gas standard spectral library matrix; is the square of the L2 norm of the vector (i.e., the sum of the squares of the vector elements).

[0108] If matrix S T If S is invertible, then the analytical solution to the above least squares problem is:

[0109] c(k)=(S T S) -1 S T x target (k), where c(k) is the target harmful gas concentration vector calculated at time k; S is the target gas standard spectral library matrix; S T This is the transpose of matrix S; (S T S) -1 Let S be a matrix T The inverse matrix of x; target (k) is the vector of characteristic spectral components of the target gas extracted and enhanced at time k;

[0110] The concentrations of various target harmful gases can be calculated using the above formula.

[0111] In another exemplary implementation, when there is a strong linear correlation between standard spectra (i.e., multicollinearity), leading to S T When S approaches singularity or ill-conditioned conditions, the CLS method may become unstable. In such cases, the Ridge Regression algorithm can be used for multiple gas concentration inversion modules. Ridge Regression improves the stability of the solution by adding a regularization term (the squared L2 norm of the concentration) to the least-squares objective function. Its optimization objective is:

[0112] Where c(k) is the target harmful gas concentration vector to be solved at time k; x target (k) is the vector of characteristic spectral components of the target gas extracted and enhanced at time k; S is the matrix of the standard spectral library of the target gas. Let α be the square of the L2 norm of the concentration vector c(k), where α > 0 is the ridge parameter, or regularization parameter, used to control the strength of the regularization term. The solution to the ridge regression is:

[0113] c(k)=(S T S+αI) -1 S T x target c(k), where I is an M×M identity matrix; c(k) is the target harmful gas concentration vector calculated at time k; S is the target gas standard spectral library matrix; S T Let S be the transpose of matrix S; α is the ridge parameter or regularization parameter, a positive scalar; (S T S+αI) -1 For matrix (S) T The inverse matrix of S+αI); xtarget (k) is the extracted and enhanced target gas feature spectrum component vector at time k. The selection of the ridge parameter a is crucial for the performance of the model, which can be determined by cross-validation or other methods.

[0114] Further, in more complex scenarios, such as when the spectral data is noisy or there are many collinear variables, the multi-gas concentration inversion module can also use the Partial Least Squares Regression (PLS) algorithm. PLS is a multivariate statistical analysis method that establishes a regression model by extracting latent structures (i.e., latent variables or factors) from the independent variables (standard spectrum library S or features derived therefrom) and dependent variables (concentrations c(k)).

[0115] The PLS algorithm first extracts a number of pairs of latent variables (score vectors) from the standard spectrum library S (or X train ) and the corresponding known concentrations Y train . Let the latent variable scores of S be T = [t1, …, t A ] (A is the number of latent variables), and the latent variable scores of the concentrations be U = [u1, …, u A ]. PLS establishes a regression relationship in these latent variable spaces, for example, U = TB inner + F, where B inner is the internal regression coefficient matrix and F is the residual.

[0116] For a new target gas feature spectrum component x target (k), first project it into the latent variable space of S to obtain its score vector t new (k). Then use the established internal regression model to predict the concentration score u new (k), and finally convert it to the actual concentration value c(k) through the load in the concentration space. The entire process can be summarized as a general regression coefficient matrix B PLS , such that: c(k) = B PLS x target (k), where B PLS is learned through the PLS algorithm in the calibration phase. PLS algorithm can effectively handle high-dimensional, collinear data and has good prediction performance.

[0117] The multi-gas concentration inversion module finally outputs the calculated concentration vector c(k) of one or more target harmful gases in the gas environment to be measured. These concentration information is the final output result of the gas detection system, which can be used for subsequent leakage alarm, environmental assessment or process control. By using the above at least one inversion algorithm, the module can realize accurate conversion from enhanced spectral features to gas concentrations, and is a key component of the entire high-precision detection system.

[0118] Also included are:

[0119] a system feedback and adaptive optimization module for generating a feedback signal according to the statistical properties of the non-target residual component or the stability and reasonability of the concentration of one or more target harmful gases in the gas environment under test, and dynamically adjusting the mode selection strategy of the dynamic background and interference modal online learning and reconstruction module or the calibration parameters of the target gas standard spectral library in the target gas signal extraction and enhancement module or the parameters of the concentration inversion model in the multi-gas concentration inversion module based on the feedback signal.

[0120] The system in the embodiments of the present application can further include a system feedback and adaptive optimization module. The function of this module is to generate a feedback signal according to specific performance indicators during the operation of the system, and dynamically adjust the parameters or strategies of other key modules in the system based on this feedback signal, in order to continuously optimize the overall detection performance of the system and its adaptability to environmental changes.

[0121] In an exemplary embodiment, the system feedback and adaptive optimization module can include a performance evaluation and feedback signal generation unit. The input end of this unit is configured to receive performance indicator information from other modules. Specifically, it can receive the non-target residual component x non-target (k) output by the target gas signal extraction and enhancement module. At the same time, it can also receive the concentration c(k) of one or more target harmful gases in the gas environment under test calculated by the multi-gas concentration inversion module.

[0122] The performance evaluation and feedback signal generation unit analyzes the received information. For example, it can calculate the statistical properties of the non-target residual component x non-target (k), such as its energy (e.g. ), variance, or whether there are certain expected or unexpected features in its spectral structure. The presence of higher energy or specific structure can indicate that the background subtraction is not complete or there are unmodeled interferences.

[0123] At the same time, the unit can also evaluate the stability and reasonability of the concentration c(k). For example, monitor whether the concentration value is within the physically allowed range, whether its rate of change over time is too fast or too slow, or whether there are persistent, unreasonable deviations or oscillations. By comprehensively analyzing these performance indicators, the unit generates a feedback signal. The feedback signal can be one or more scalar values, status codes, or structured data containing specific adjustment instructions.

[0124] Then, the system feedback and adaptive optimization module can further include a parameter adjustment and model updating unit. This unit is coupled to the performance evaluation and feedback signal generation unit for receiving the feedback signal generated by it. Based on this feedback signal, the unit is responsible for dynamically adjusting and optimizing the working parameters or internal models of one or more other modules in the system.

[0125] Specifically, the parameter adjustment and model updating unit can dynamically adjust the modal selection strategy of the dynamic background and interference modal online learning and reconstruction module. For example, if the feedback signal indicates that the non-target residual component x non-target (k) contains significant dynamic components that should have been removed by the background reconstruction module, the threshold for selecting dynamic modes can be adjusted. For example, when using online DMD, the energy or frequency threshold for filtering background modes can be adjusted; when using sliding window-based RPCA, the target rank of the low-rank matrix L or the regularization parameter γ can be adjusted to capture more background changes.

[0126] Further, the parameter adjustment and model updating unit can also dynamically adjust the calibration parameters of the target gas signal extraction and enhancement module in the target gas standard spectrum library. For example, if the feedback signal indicates that even after background subtraction, the non-target residual component x non-target (k) still contains components similar in shape but slightly offset or deformed in shape to some target gas standard spectrum s j , or some gas concentration c j (k) shows systematic deviations in the long run, which may indicate that the corresponding spectrum s j in the original standard spectrum library S needs to be calibrated.

[0127] This calibration can be a small wavelength shift, peak shape adjustment or intensity scaling of the standard spectrum s j . For example, a calibration model Δs j can be established to correct the standard spectrum: s j,calibrated = f cal (s j , θ j ), where f cal is a calibration function and θ j is a set of calibration parameters for gas j. The feedback signal can drive the adjustment of θ j . The updated s j,calibrated will replace the original s j for subsequent target gas feature subspace construction and concentration inversion.

[0128] In addition, the parameter adjustment and model updating unit can also dynamically adjust the parameters of the concentration inversion model in the various gas concentration inversion modules. For example, if the feedback signal indicates that the calculated concentration c(k) exhibits excessive noise or instability, the value of the regularization parameter a can be appropriately increased to enhance the stability of the model when using ridge regression. Conversely, if the concentration response is too smooth, resulting in insufficient tracking of real changes, a can be reduced.

[0129] If partial least squares regression (PLS) is used, the feedback signal can be used to adjust the number of latent variables A used by the model. For example, if the model exhibits overfitting (sensitivity to noise in the concentration), the number of latent variables can be reduced; if underfitting (insufficient response to real concentration changes), the number of latent variables can be increased. In some cases, if the system performance continues to be poor, the feedback signal can even trigger a retraining of the entire PLS model (i.e., the regression coefficient matrix B PLS ) using recent accumulated, validated data.

[0130] Through this closed-loop feedback and adaptive optimization mechanism, the system can continuously adjust and improve its internal models and parameters according to actual operating effects and environmental changes. This helps to maintain and improve the accuracy, robustness, and adaptability to complex dynamic environments of gas detection, thereby achieving more reliable and efficient harmful gas monitoring.

[0131] The multi-spectrum fusion-based high-precision detection method for harmful gases in equipment pipelines described below can be mutually referenced with the multi-spectrum fusion-based high-precision detection system for harmful gases in equipment pipelines described above.

[0132] The multi-spectrum fusion-based high-precision detection method for harmful gases in equipment pipelines includes the following steps:

[0133] Obtain the original multi-spectrum data of the gas environment to be measured;

[0134] Preprocess the original multi-spectrum data to obtain preprocessed multi-spectrum data;

[0135] Based on the preprocessed multi-spectrum data, use an online learning mechanism to identify and reconstruct the main interference spectrum caused by dynamic background and unknown interference in real time;

[0136] Subtract the reconstructed main interference spectrum from the preprocessed multi-spectrum data to obtain a residual spectrum;

[0137] Further, using the pre-set target gas standard spectrum information, the residual spectrum is subjected to orthogonal projection processing to extract and enhance the target gas feature spectrum component;

[0138] Finally, based on the enhanced target gas characteristic spectral component and the target gas standard spectral information, the concentration of one or more target harmful gases in the to-be-detected gas environment is obtained through inversion calculation.

[0139] The one or more target harmful gases include at least one of a hydrocarbon gas, a sulfur compound gas, a nitrogen oxide gas, carbon monoxide, chlorine, and ammonia.

[0140] The method of the embodiment can be used to execute the system embodiment described above, and has similar principles and technical effects, which will not be described here again.

[0141] Although the embodiments of the present application have been shown and described, it is to be understood that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the present application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A high-precision detection system for hazardous gases in equipment pipelines based on multispectral fusion, characterized in that, include: The multispectral data acquisition and preprocessing module is used to acquire the raw multispectral data of the gas environment to be tested, and to preprocess the raw multispectral data to output preprocessed multispectral data. The dynamic background and interference mode online learning and reconstruction module is used to receive the preprocessed multispectral data and, based on the preprocessed multispectral data, use an online learning mechanism to identify and reconstruct the main interference spectra caused by dynamic background and unknown interference in real time. The target gas signal extraction and enhancement module is used to receive the preprocessed multispectral data and the main interference spectrum, obtain the residual spectrum by subtracting the main interference spectrum from the preprocessed multispectral data, and perform orthogonal projection processing on the residual spectrum using preset target gas standard spectral information to extract and enhance the characteristic spectral components of the target gas. A multi-gas concentration inversion module is used to receive the characteristic spectral components of the target gas and the preset standard spectral information of the target gas, and to calculate the concentration of one or more target harmful gases in the gas environment to be tested based on the characteristic spectral components of the target gas and the preset standard spectral information of the target gas.

2. The high-precision detection system for hazardous gases in equipment pipelines based on multispectral fusion according to claim 1, characterized in that, The dynamic background and interference mode online learning and reconstruction module specifically includes: An online algorithm execution unit is used to execute an online learning algorithm on the preprocessed multispectral data to identify and extract the dynamic evolution patterns or main changing components in the data; The main dynamic characteristic identification and selection unit is used to identify and select the spectral change modes that represent the main dynamic background and unknown interference based on the characteristics of the dynamic evolution law or main change components output by the online algorithm execution unit. The dynamic background and interference spectrum reconstruction unit is used to reconstruct the main interference spectrum by utilizing the spectral change mode and its corresponding real-time amplitude determined by the main dynamic characteristic identification and selection unit.

3. The high-precision detection system for hazardous gases in equipment pipelines based on multispectral fusion according to claim 2, characterized in that, The online learning algorithm used by the online algorithm execution unit is at least one of the following: online dynamic mode decomposition algorithm or robust principal component analysis algorithm based on sliding window.

4. The high-precision detection system for hazardous gases in equipment pipelines based on multispectral fusion according to claim 1, characterized in that, The target gas signal extraction and enhancement module specifically includes: The residual spectral calculation unit is used to subtract the reconstructed main interference spectrum from the preprocessed multispectral data to obtain the residual spectrum; The target gas standard spectral subspace construction unit is used to construct the target gas feature subspace based on a preset target gas standard spectral database. The orthogonal projection and signal enhancement unit is used to orthogonally project the residual spectrum onto the target gas feature subspace to extract and enhance the target gas feature spectral components.

5. The high-precision detection system for hazardous gases in equipment pipelines based on multispectral fusion according to claim 4, characterized in that, The orthogonal projection and signal enhancement unit is also configured to calculate non-target residual components that are orthogonal to the target gas feature subspace.

6. The high-precision detection system for hazardous gases in equipment pipelines based on multispectral fusion according to claim 1, characterized in that, The multi-gas concentration inversion module employs at least one of the classical least squares method, ridge regression, or partial least squares regression algorithm to invert the concentration of the characteristic spectral components of the target gas.

7. The high-precision detection system for hazardous gases in equipment pipelines based on multispectral fusion according to claim 1, characterized in that, The preprocessing operations in the multispectral data acquisition and preprocessing module include at least one of dark spectrum subtraction, light source reference correction or spectral normalization, preliminary baseline correction, and noise filtering.

8. The high-precision detection system for hazardous gases in equipment pipelines based on multispectral fusion according to claim 5, characterized in that, Also includes: The system feedback and adaptive optimization module is used to generate a feedback signal based on the statistical characteristics of the non-target residual components or the stability and rationality of the concentration of one or more target harmful gases in the gas environment to be measured, and dynamically adjust the mode selection strategy of the dynamic background and interference mode online learning and reconstruction module, the calibration parameters of the target gas standard spectral library in the target gas signal extraction and enhancement module, or the parameters of the concentration inversion model in the multi-gas concentration inversion module based on the feedback signal.

9. A high-precision detection method for harmful gases in equipment pipelines based on multispectral fusion, characterized in that, The high-precision detection system for hazardous gases in equipment pipelines based on multispectral fusion as described in any one of claims 1-8 comprises the following steps: Acquire raw multispectral data of the gas environment to be tested; The original multispectral data is preprocessed to obtain preprocessed multispectral data; Based on the preprocessed multispectral data, an online learning mechanism is used to identify and reconstruct the main interference spectra caused by dynamic background and unknown interference in real time. The reconstructed main interfering spectra are subtracted from the preprocessed multispectral data to obtain the residual spectra; Furthermore, the residual spectrum is orthogonally projected using the preset target gas standard spectral information to extract and enhance the characteristic spectral components of the target gas. Finally, based on the enhanced target gas characteristic spectral components and the target gas standard spectral information, the concentration of one or more target harmful gases in the gas environment to be tested is calculated by inversion.

10. The high-precision detection method for harmful gases in equipment pipelines based on multispectral fusion according to claim 9, characterized in that, The one or more target harmful gases include at least one of hydrocarbon gases, sulfur-containing compound gases, nitrogen oxide gases, carbon monoxide, chlorine, and ammonia.

Images