Cavity ring-down spectroscopy-based baseline fluctuation suppression method and system for ring-down time

Abstract

The present application relates to the technical field of cavity ring-down spectroscopy, and specifically discloses a ring-down time baseline fluctuation suppression method and system based on cavity ring-down spectroscopy, the method comprising the following steps: in response to a target ring-down time sequence, performing signal decomposition on the target ring-down time sequence to extract baseline drift candidate components; constructing a hybrid baseline model according to the baseline drift candidate components and synchronously collected multi-dimensional feature data to predict real-time baseline drift values. Through differential calculation of multi-spectral ring-down times combined with multi-dimensional signal decomposition technology, the present application can effectively strip and suppress complex industrial interference from the underlying physical dimension; further, a hybrid baseline model fusing multi-dimensional features is constructed, and a dynamic weight adaptive filtering process based on real-time scene features is introduced, breaking the limitations of traditional static filtering in confusing true and false signals, and realizing accurate tracking and dynamic compensation of dynamic baseline drift.

Description

Technical Field

[0001] This invention relates to the field of cavity ring-down spectroscopy technology, specifically to a method and system for suppressing ring-down time baseline fluctuations based on cavity ring-down spectroscopy. Background Technology

[0002] Cavity ring-down spectroscopy, a highly sensitive absorption spectroscopy detection technique, achieves quantitative analysis of trace gases by measuring the exponential decay time of a light pulse within an optical resonant cavity. Due to its extremely high measurement accuracy and resistance to laser intensity fluctuations, it is widely used in practical applications such as continuous emission monitoring in complex industrial settings, chemical process control, and high-purity gas analysis.

[0003] However, in real and harsh industrial testing environments, cavity ring-down spectroscopy systems are inevitably subject to various strong industrial interferences, such as dust contamination, mirror erosion, mechanical vibration, and drastic fluctuations in ambient temperature and pressure. These physical interferences directly lead to an increase in the non-absorption of the optical cavity loss, manifesting as complex and nonlinear baseline drift in the ring-down time series. Traditional baseline calibration methods typically rely on periodic "zero-gas" calibration or simple static mathematical filtering, which not only disrupts the continuous measurement process in industrial settings but also makes static filtering prone to confusing the real, slowly varying gas absorption signal with low-frequency baseline drift, resulting in the incorrect filtering out of weak real gas absorption characteristics. This makes it difficult to meet the dual requirements of continuity and accuracy in industrial settings.

[0004] Even more serious is the technical problem that arises when attempting to incorporate sensor data to construct a physical compensation model for baseline drift, which involves model tearing between physical anchor failure and data-driven rigid constraints. For example, under transient extreme conditions such as localized steam surges or strong mechanical shocks, physical sensors are prone to instantaneous saturation or reading jumps, leading to severe distortion of the calculated theoretical value of the physical baseline.

[0005] If the baseline prediction model still imposes rigid static weight constraints on the distorted theoretical value, the model will be severely pulled in the opposite direction by the real data characteristics and the erroneous physical theoretical value, which will lead to the prediction algorithm failing to converge or outputting absurd baseline drift compensation amounts, ultimately causing violent fluctuations or even systemic collapse in the gas concentration inversion results. Summary of the Invention

[0006] This invention aims to at least partially solve one of the technical problems in related technologies. Therefore, the objective of this invention is to propose a method and system for suppressing the wave-down time baseline fluctuation based on cavity wave-down spectroscopy, so as to ensure the continuity and high reliability of trace gas concentration inversion.

[0007] To achieve the above objectives, a first aspect of the present invention proposes a method for suppressing the time baseline fluctuation of cavity ring-down spectroscopy, applied in a cavity ring-down spectroscopy gas detection system, comprising:

[0008] In response to the target fading time series, the target fading time series is decomposed to extract baseline drift candidate components. Specifically, this includes: performing multi-scale mode decomposition on the target fading time series to obtain multiple decomposed components; calculating the correlation between each decomposed component and synchronously acquired industrial interference feature data; and extracting the baseline drift candidate components from the multiple decomposed components based on the correlation.

[0009] Based on the baseline drift candidate components and synchronously acquired multi-dimensional feature data, a hybrid baseline model is constructed to predict the real-time baseline drift value. The hybrid baseline model includes physical baseline theoretical constraints, data-driven prediction constraints, and self-calibration benchmark constraints. Specifically, the physical baseline theoretical constraints are calculated using synchronously acquired physical parameters; the baseline drift candidate components and multi-dimensional feature data are input into the data prediction model to generate the data-driven prediction constraints; and the self-calibration benchmark constraints are obtained through a self-calibration mode. The real-time baseline drift value is predicted and output by combining the above three constraints.

[0010] A dynamic weighted adaptive filtering process based on the real-time baseline drift value is initiated to suppress baseline fluctuations in the target oscillation time series, and the filtered signal is used for gas concentration inversion.

[0011] To achieve the above objectives, a second aspect of the present invention proposes a cavity ringback-frequency oscillation suppression system based on cavity ringback-frequency spectroscopy, the system comprising:

[0012] The sequence generation module is used to acquire the first ring-down curve corresponding to the detection pulse, the second ring-down curve corresponding to the non-absorption pulse, and the ring-down curve of the reference pulse; extract the ring-down time of each ring-down curve respectively, and eliminate the non-target gas absorption interference by performing differential calculation on the ring-down time to generate the target ring-down time sequence.

[0013] The signal decomposition module is used to decompose the target fading time series in response to the target fading time series and extract baseline drift candidate components; specifically, it is used to: perform multi-scale mode decomposition on the target fading time series to obtain multiple decomposed components; calculate the correlation between each decomposed component and synchronously acquired industrial interference feature data, and extract the baseline drift candidate components based on the correlation.

[0014] The baseline prediction module is used to construct a hybrid baseline model based on the baseline drift candidate components and synchronously acquired multi-dimensional feature data to predict the real-time baseline drift value. The hybrid baseline model includes physical baseline theoretical constraints, data-driven prediction constraints, and self-calibration benchmark constraints. Specifically, the physical baseline theoretical constraints are calculated using synchronously acquired physical parameters; the baseline drift candidate components and multi-dimensional feature data are input into the data prediction model to generate the data-driven prediction constraints; and the self-calibration benchmark constraints are obtained through a self-calibration mode. The real-time baseline drift value is predicted and output by combining the above three constraints.

[0015] The filtering and output module is used to initiate a dynamic weighted adaptive filtering process based on the real-time baseline drift value to suppress baseline fluctuations in the target oscillation time series and obtain a filtered signal for gas concentration inversion.

[0016] To achieve the above objectives, a third aspect of the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory. When the computer program is executed by the processor, it implements the above-described method for suppressing the time-baseline fluctuations of the cavity ring-down spectrum.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0018] The cavity ring-down spectrum-based method and system for suppressing ring-down time baseline fluctuations in this invention can effectively isolate and suppress complex industrial interference from the underlying physical dimension by combining differential calculation of ring-down time in multiple spectral bands with multi-dimensional signal decomposition technology. Furthermore, a hybrid baseline model integrating multi-dimensional features is constructed, and a dynamic weight adaptive filtering process based on real-time scene features is introduced, which breaks the limitation of traditional static filtering that is prone to confusing true and false signals, and realizes accurate tracking and dynamic compensation of dynamic baseline drift.

[0019] In particular, to address model tearing conflicts caused by sensor distortion under extreme operating conditions, this application introduces a joint loss function and a degradation isolation mechanism into the model, which includes confidence decay and dynamic update of confidence weights. This enables the system to automatically trigger constraint degradation for fault self-healing when physical sensors fail, thereby maintaining the high robustness of baseline predictions even in extremely harsh industrial environments and thoroughly ensuring the continuity, high accuracy, and high reliability of trace gas concentration inversion. Attached Figure Description

[0020] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0021] Figure 1 This is a schematic flowchart of the method for suppressing the time baseline fluctuation of the cavity ring-down spectrum provided by the present invention.

[0022] Figure 2 This is a comparison diagram of the three-band pulse intensity ring-down curves and differential extraction sequences in the ring-down time baseline fluctuation suppression method based on cavity ring-down spectrum provided by this invention.

[0023] Figure 3 This is an empirical mode (EMD) multi-scale decomposition waveform of the target ring-down time series in the ring-down time baseline fluctuation suppression method based on cavity ring-down spectrum provided by the present invention.

[0024] Figure 4 This is a comparison of the effects of LMS adaptive filtering based on a hybrid baseline model before and after in the cavity ring-down spectrum-based ring-down time baseline fluctuation suppression method provided by this invention.

[0025] Figure 5 This is a curve of dynamic cross-correlation coefficient optimization with physical time delay compensation in the cavity ring-down spectrum-based ring-down time baseline fluctuation suppression method provided by the present invention.

[0026] Figure 6 This is a scatter plot of the proportion of changes in the dual-band projection characteristic response and the distribution of the trap interval in the cavity ring-down spectrum-based ring-down time baseline fluctuation suppression method provided by this invention.

[0027] Figure 7 This is a comparison chart of model anti-tear-break and degradation dual-path constraint baseline prediction under extreme distortion conditions in the cavity ring-down spectrum-based ring-down time baseline fluctuation suppression method provided by this invention.

[0028] Figure 8 This is a schematic diagram illustrating the implementation of the cavity ring-down spectrum-based ring-down time baseline fluctuation suppression system provided by the present invention.

[0029] Figure 9 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0030] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0031] The following describes, with reference to the accompanying drawings, a method, system, and electronic device for suppressing the time baseline fluctuations of the decay time based on cavity decay spectrum according to embodiments of the present invention.

[0032] Example 1:

[0033] This embodiment provides a method for suppressing baseline fluctuations in ring-down time based on cavity ring-down spectroscopy. This method is applied to a cavity ring-down spectroscopy gas detection system that includes a laser emitter, optical modulation components, a high-reflectivity ring-down cavity, a photodetector, and a multi-dimensional sensor array. The system's data processing unit executes the method steps disclosed in this embodiment through a hardware-software co-working architecture to solve the problems of data distortion and model tearing caused by baseline fluctuations under complex and harsh industrial conditions.

[0034] like Figure 1 As shown, the method in this embodiment specifically includes the following steps:

[0035] Step 1: Generation and preprocessing of the target oscillation time series.

[0036] First, in response to the target ring-down time series generation command issued by the data processing unit, the cavity ring-down spectral gas detection system, under the scheduling of the timing controller, triggers the underlying hardware to perform a coordinated acquisition operation of the three-band pulse sequence.

[0037] It is important to note that, in order to effectively separate the true characteristic absorption from the broadband non-selective background loss at the physical light source level, the hardware system in this embodiment needs to continuously emit three laser pulse sequences with different wavelength characteristics. Specifically, the data processing unit acquires the first ring-down curve corresponding to the detection pulse, the second ring-down curve corresponding to the non-absorption pulse, and the ring-down curve of the reference pulse.

[0038] Here, the detection pulse is defined as a laser pulse whose emission wavelength is strictly locked to the center wavelength of the absorption peak of the target gas; the reference pulse is defined as a laser pulse whose emission wavelength is locked to the wavelength of the absorption trough of the target gas; and the non-absorption pulse is defined as a laser pulse whose emission wavelength is located at the pure baseline wavelength where neither the target gas nor the background interfering gas has any characteristic absorption.

[0039] After the photodetector receives the exponentially decaying light signals transmitted through the ring-down cavity from the three types of pulses mentioned above, the data processing unit extracts the ring-down time of each ring-down curve. The physical definition of the ring-down time for any given ring-down curve is expressed as follows:

[0040] ;

[0041] in, Indicates the time variable Real-time light intensity voltage signal collected by the photodetector. This represents the light intensity voltage signal at the instant the laser pulse is cut off, i.e., at the initial moment. This indicates the original ringback time corresponding to the curve. This represents the inherent dark current and white noise floor of the photoelectric detection system and circuit. The data processing unit extracts the first ring-down time corresponding to the detection pulse, the second ring-down time corresponding to the non-absorption pulse, and the third ring-down time corresponding to the reference pulse using a logarithmic linear fitting algorithm.

[0042] After extracting the ring-down times of the three gases, the data processing unit performs differential calculations on these ring-down times to eliminate non-target gas absorption interference and generate the target ring-down time series. This differential calculation, by evaluating the optical loss difference between the absorption wavelength and the non-absorption wavelength, directly cancels out the proportional common-mode loss caused by large dust particles adhering to the mirror or the broadband continuous absorption of water vapor at the physical level. This results in a target ring-down time series with high purity, which serves as the core input foundation for all subsequent advanced signal processing algorithms.

[0043] like Figure 2 The comparison diagram of the three-band pulse intensity ring-down curves and differential extraction sequences shown intuitively demonstrates the role of this invention in reducing common-mode interference during the physical optical signal acquisition stage.

[0044] In the upper part of the graph (Figure a), the horizontal axis represents time in microseconds (μs), and the vertical axis represents the light intensity signal in arbitrary units (au). The figure shows three exponential ring-down curves with different colors and decay rates. The red detection pulse ring-down curve, due to absorption by the target gas absorption peak, exhibits relatively rapid light intensity decay, approaching zero around the time axis value of 40. The green non-absorption pulse ring-down curve, avoiding the gas absorption band, is mainly affected by the optical cavity background loss, resulting in relatively slow decay, with its waveform extending to around the time axis value of 100. The blue reference pulse ring-down curve falls between the two. This difference in waveform characteristics demonstrates the feasibility of the system separating characteristic absorption from broadband non-selective background loss at the physical light source level.

[0045] In the lower half of the graph (b), the horizontal axis represents continuous sampling points (number of sampling points), and the vertical axis represents the extracted ring-down time in microseconds (μs). The black dashed line in the graph represents the original ring-down time series without processing, which is affected by background interference such as simulated dust pollution and temperature and pressure fluctuations, showing a waveform that fluctuates around the vertical axis value of 20. The red solid line represents the target ring-down time series extracted after three-segment difference calculation. This curve is relatively stable throughout the entire sampling period, with the value remaining around 20.

[0046] This comparison curve illustrates that by evaluating the optical loss difference between the absorption wavelength and the non-absorption wavelength, the present invention can reduce the proportional common-mode interference caused by particulate matter adhesion on the cavity mirror surface or broadband continuous absorption of water vapor at the physical source, thereby providing a highly reliable target oscillation time series for subsequent baseline model reconstruction and adaptive filtering.

[0047] Step 2: Acquisition of multi-dimensional feature data and multi-scale decomposition of signals.

[0048] While obtaining the target ring-down time series, the data processing unit must simultaneously collect multi-dimensional characteristic data of the system operating environment and its own hardware status. In this embodiment, the multi-dimensional characteristic data strictly consists of three independent datasets: physical parameters, industrial interference characteristic data, and equipment aging characteristic data. Physical parameters include at least ambient temperature, ambient pressure, the cavity mirror temperature inside the ring-down cavity, and the triaxial vibration acceleration of the equipment casing; industrial interference characteristic data includes at least dust pollution characteristic index and spatial electromagnetic interference intensity data; and equipment aging characteristic data includes at least the laser power attenuation rate of the emitted laser source.

[0049] It is important to note that, in response to the aforementioned target ring-down time series, the data processing unit performs signal decomposition on the target ring-down time series to extract baseline drift candidate components. This signal decomposition process is a multi-scale, cross-domain joint processing flow. The data processing unit first performs wavelet decomposition on the target ring-down time series to obtain low-frequency and high-frequency subsequences. Wavelet decomposition utilizes a mother wavelet function with compact support characteristics to map the one-dimensional time series to a two-dimensional phase space of time and scale, thereby stripping away the high-frequency subsequence containing high-frequency noise and retaining the low-frequency subsequence reflecting the overall slow variation trend of the baseline and the actual changes in gas concentration.

[0050] Furthermore, the data processing unit performs empirical mode decomposition (EMD) on the low-frequency subsequence, obtaining multiple EMD components. The EMD process does not require pre-setting basis functions; instead, it adaptively stabilizes the low-frequency subsequence based on its own timescale characteristics, extracting several intrinsic mode function components with different characteristic timescales—the EMD components. These components include both physical baseline drift characteristics caused by slow changes in ambient temperature and absorption characteristics resulting from actual fluctuations in the target gas concentration.

[0051] like Figure 3The empirical mode multiscale decomposition waveform of the target ring-down time series shown intuitively demonstrates the process of separating interference from the real signal in the signal separation stage of this invention. In this waveform, the four sub-graphs are named from top to bottom as Figure (a), Figure (b), Figure (d), and Figure (d), and each sub-graph has a horizontal axis sampling point marker, with the unit being (number of points), and a vertical axis amplitude or ring-down time marker, with the unit being microseconds (μs).

[0052] Figure (a) represents the input target ring-down time series, whose waveform exhibits nonlinear fluctuations around a ring-down time of 20 microseconds. These fluctuations are superimposed with physical interference from different frequency bands. To extract baseline drift features, the system performs empirical mode decomposition on the series.

[0053] Figure (b) shows the first empirical mode decomposition component, which exhibits relatively drastic waveform characteristics. This component mainly corresponds to the high-frequency interference fluctuations caused by mechanical vibration of field equipment or local dust stirring.

[0054] Figure (c) shows the second empirical mode decomposition component, whose waveform fluctuations are relatively smooth, which is consistent with the low-frequency drift phenomenon such as the slow increase of industrial ambient temperature or the gradual decrease of the reflectivity of the optical cavity mirror. In the subsequent process, by performing correlation calculation with industrial interference characteristic data, the component represented by Figure (c) can be marked as the baseline drift candidate component.

[0055] Figure (d) represents the decomposition residual component, which shows a relatively stable and slowly rising smooth curve, representing the baseline that the system retains after separating external industrial interference, reflecting the underlying physical and optical background decay time and the true concentration characteristics of the target gas.

[0056] By showcasing the decomposition of different colors and frequency band curves in different charts, this figure illustrates that the present invention, without relying on frequent calibration with external zero-point standard gas, can identify complex industrial interferences from a data dimension through empirical modal multi-scale decomposition, providing a data-driven feature source for the subsequent construction of hybrid baseline models.

[0057] To accurately eliminate true absorption components and retain interfering components, the data processing unit calculates the correlation between each empirical mode decomposition component and the industrial interference characteristic data, and marks the empirical mode decomposition components with a correlation greater than or equal to a preset correlation threshold as baseline drift candidate components. Since dust pollution or temperature creep in industrial sites are typical industrial interference characteristic data, if their fluctuation patterns show a high degree of similarity in the time domain to the fluctuation patterns of a certain empirical mode decomposition component (i.e., the correlation coefficient is greater than or equal to the preset correlation threshold), it indicates that the empirical mode decomposition component is highly likely to be a baseline fluctuation of non-true absorption induced by industrial interference. Therefore, it is extracted as a baseline drift candidate component, providing high-value data-driven features for subsequent baseline model reconstruction.

[0058] Step 3: Adaptive processing of high-frequency noise and dynamic feedback of correlation threshold.

[0059] After obtaining the high-frequency subsequence in step two, the method in this embodiment must further refine the high-frequency noise to prevent signal spikes caused by strong electromagnetic interference from compromising the subsequent model convergence. The data processing unit calculates an adaptive wavelet threshold based on the electromagnetic interference intensity data in the industrial interference characteristic data. The larger the electromagnetic interference intensity data, the denser the high-frequency spike noise doped in the signal received by the system. Therefore, the calculation of the adaptive wavelet threshold shows a positive correlation with the electromagnetic interference intensity data.

[0060] Optionally, the data processing unit performs adaptive threshold filtering on the high-frequency subsequence using the adaptive wavelet threshold. When the absolute value of some coefficients in the high-frequency subsequence is less than the adaptive wavelet threshold, they are directly set to zero or smoothed and shrunk; when the absolute value of the coefficients is greater than the adaptive wavelet threshold, they are retained. This filtering process can suppress transient electromagnetic interference pulses caused by large equipment such as motors starting and stopping on site to a great extent.

[0061] To prevent the baseline drift candidate components from being missed or misjudged due to a fixed correlation threshold, the data processing unit extracts the sharpness feature value of the first fading curve within a preset time period. The preset time period refers to the middle signal range of the first fading curve where the amplitude of the attenuated signal is between 20% and 80% of its maximum value. The sharpness feature value reflects the severity of the second derivative of the attenuation slope of the fading curve in the middle region.

[0062] The sharpness feature value is used to determine the current gas component type. If the sharpness feature value shows non-monotonic decay characteristics caused by the superposition of multiple components, it indicates that there is cross-interference of multiple absorbing gases in the current gas chamber. At this time, the data processing unit will adjust the preset correlation threshold for the next processing cycle based on the judgment result feedback. For example, the preset correlation threshold may be appropriately increased to make the signal screening conditions more stringent and prevent complex real absorbing components from being misjudged as baseline drift.

[0063] Step 4: Construction and prediction of a hybrid baseline model based on multi-source data fusion.

[0064] Since purely data-driven models are prone to getting stuck in local optima or overfitting under extreme conditions, the method in this embodiment constructs a hybrid baseline model based on the baseline drift candidate components and synchronously acquired multi-dimensional feature data to predict the real-time baseline drift value.

[0065] In this embodiment, the hybrid baseline model is not a single mathematical regression model, but a three-dimensional constraint architecture anchored by multi-source information. Its core components include: physical baseline theoretical constraints, data-driven prediction constraints, and self-calibrating benchmark constraints.

[0066] The physical baseline theoretical constraint refers to using prior knowledge reflecting the underlying optical mechanism and thermodynamic laws of the system (such as the cavity deformation equation, the mirror loss evolution law, etc.) to provide a physically reasonable numerical range for baseline prediction; the data-driven prediction constraint refers to using a data prediction model (preferably a support vector regression model in this embodiment) to perform deep feature mining on baseline drift candidate components and environmental features to capture complex nonlinear fluctuation laws; the self-calibration benchmark constraint refers to using the absolute reference point obtained by the real-time hardware calibration action at the edge to provide periodic zero-point drift correction for the model.

[0067] By mapping the above three constraints together to the joint loss function for iterative optimization, the output real-time baseline drift value can be strictly controlled by the physical mechanism and calibration anchor point on the basis of mathematical convergence, thus completely solving the technical problem of model divergence or tearing under extreme conditions.

[0068] First, based on the physical parameters, the industrial interference characteristic data, and the equipment aging characteristic data, the data processing unit calculates the theoretical value of the physical baseline drift using pre-embedded underlying physical equations. The calculation formula is shown below:

[0069] ;

[0070] in, This represents the calculated theoretical value of the physical baseline drift. This indicates the initial calibration physical cavity length of the optical cavity. This represents the dynamic cavity length deformation calculated from the triaxial vibration acceleration integral in the multi-dimensional feature data. This represents the constant speed of light in a vacuum. This indicates the initial absolute reflectivity of the endoscope during factory calibration. This represents the amount of mirror reflectance attenuation calculated from the combined results of the endoscope temperature and dust contamination characteristic index in the multi-dimensional feature data. This represents the amount of system-level response delay aging compensation introduced by the laser power attenuation rate in multi-dimensional feature data.

[0071] Subsequently, the data processing unit uses the selected baseline drift candidate components and the multi-dimensional feature data as input feature sets, and inputs them into the support vector regression model for nonlinear mapping training. The support vector regression model uses a kernel function to map low-dimensional nonlinear features to a high-dimensional space for linear regression.

[0072] It is important to note that during the training phase of the support vector regression model, the data processing unit uses a joint loss function that includes historical true baseline values, the theoretical value of the physical baseline drift, and the self-calibrated baseline value to constrain the support vector regression model, thereby outputting a highly robust real-time baseline drift value. The formula for the joint loss function is as follows:

[0073] ;

[0074] in, This represents the output penalty value of the joint loss function. This represents the real-time baseline drift value predicted by the support vector regression model during the current forward propagation. This represents the historical true baseline value stored in the system database. This represents the theoretical value of physical baseline drift derived from the physical mechanism described above. This represents the self-calibration reference value generated by the edge hardware.

[0075] also, This represents the operator for calculating the mean square error. This represents the first dynamic weight coefficient of the data-driven constraint term. This represents the second dynamic weighting coefficient of the physical theory constraint term. This represents the third dynamic weighting coefficient of the self-calibration constraint term.

[0076] Through the joint penalty constraints of these three dimensions, the model not only has the ability to mine nonlinear data, but is also stably constrained by physical formulas within a reasonable numerical range that conforms to the laws of reality, effectively preventing the model from making distorted predictions.

[0077] Step 5: Edge-end closed-loop generation mechanism for self-calibration reference value.

[0078] Specifically, the generation process of the self-calibration reference value, which plays a decisive role in the correction of deviations in the joint loss function, relies on the underlying closed-loop triggering mechanism. When the data processing unit detects that the system operating parameters have reached a preset first trigger condition, it executes the non-absorption pulse self-calibration mode. The first trigger condition typically refers to the ambient temperature gradient exceeding a set value within ten consecutive acquisition cycles. At this time, the system does not need to introduce zero-point standard gas, but directly uses pure non-absorption pulses to perform internal reference normalization to calculate the first calibration deviation, and generates the first self-calibration reference value when the first calibration deviation is less than a preset deviation threshold.

[0079] Optionally, when the system operating parameters are detected to reach a preset second trigger condition, a standard gas self-calibration mode is executed. The second trigger condition typically refers to the system's cumulative continuous operating time exceeding a preset maintenance cycle. Once triggered, the system's underlying layer automatically opens the micro-solenoid valve, releasing the built-in standard gas to replace the gas to be tested in the optical cavity, and controls the photodetector to collect the actual ring-down time. The data processing unit compares the collected actual ring-down time with the factory-preset theoretical standard absorption time to calculate the second calibration deviation, and generates a second self-calibration reference value based on the second calibration deviation.

[0080] Ultimately, the self-calibration reference value required by the aforementioned joint loss function includes both the first and second self-calibration reference values, ensuring the absolute accuracy of the baseline anchor point under long-term operation.

[0081] Step 6: Adaptive filtering and signal reconstruction based on dynamic weights.

[0082] After the hybrid baseline model outputs accurate real-time baseline drift values, the system initiates a dynamic weighted adaptive filtering process based on the real-time baseline drift values ​​to suppress non-steady-state baseline fluctuations in the target oscillation time series. The core of this process is to adaptively adjust according to the severity of the on-site working conditions.

[0083] The data processing unit extracts real-time scene features of the current detection scene, which include at least baseline fluctuation intensity and signal-to-noise ratio (SNR). Baseline fluctuation intensity is characterized by calculating the statistical variance of the target decay time series within a recent time window; the SNR is defined by the ratio of the peak value of the useful signal to the root mean square value of the background noise.

[0084] It is important to note that the data processing unit dynamically allocates the prediction weights of the hybrid baseline model and the filtering weights of the adaptive filter based on the real-time scene characteristics. If the baseline fluctuation intensity is extremely high and the signal-to-noise ratio is extremely low, indicating severe on-site interference, the system will assign a higher prediction weight to the hybrid baseline model and weaken the filtering weight that relies solely on the current data; conversely, it will increase the filtering weight.

[0085] For example, an adaptive filter is constructed based on the minimum mean square error criterion. The data processing unit adjusts the step size of the adaptive filter based on the real-time scene characteristics. The dynamic adjustment of the step size follows the principle of decreasing the step size for stability when fluctuations are severe and increasing the step size for speed when fluctuations are mild. Subsequently, the filter coefficients are updated in conjunction with the assigned filter weights, so that the convergence direction of the minimum mean square filter always approximates the true clean signal. After iterative smoothing and denoising by this adaptive filter, the system finally outputs the filtered signal that has effectively filtered out baseline drift interference.

[0086] like Figure 4 The comparison chart showing the effects of adaptive filtering before and after the implementation of the hybrid baseline model intuitively demonstrates the application effect of this invention in signal denoising and baseline fluctuation suppression. In the figure, the horizontal axis represents continuous sampling points (number of sampling points), and the vertical axis represents the extracted oscillation time value (microseconds, μs).

[0087] The black curve in the figure represents the original oscillation time series containing spikes and baseline fluctuations. Its waveform shows obvious oscillations. As mentioned in the specific implementation, the extracted value fluctuates around 25.1 microseconds at the initial sampling time, reflecting the impact of multi-source physical interference on optical signals under complex working conditions.

[0088] The blue curve in the figure represents the sequence after processing with the traditional static filtering algorithm. It can be observed that the waveform has a certain time lag in response speed and local deviation when facing sudden interference, which can easily lead to errors in the oscillation time after filtering.

[0089] In contrast, the red curve in the figure represents the target sequence processed by the hybrid baseline model that integrates physical constraints and self-calibrating weights proposed in this invention, along with the dynamic weight adaptive filtering process. This red curve smooths the fluctuations of the original signal within a shorter sampling period, and the waveform converges relatively stably to the 26.5 microsecond oscillation time baseline.

[0090] By visually comparing three different colored and shaped curves under the same coordinate system, the accompanying figure illustrates that the present invention can effectively track the real baseline drift trajectory under complex industrial interference, realize the prediction and compensation of dynamic baseline drift, and thus provide data support for the final inversion of the target gas dry concentration.

[0091] Step 7: Anomaly closed-loop classification and system disaster recovery mechanism.

[0092] After obtaining the filtered signal, to prevent system crashes caused by extreme, sudden, and destructive events, the method in this embodiment also includes a multi-dimensional anomaly protection process. The data processing unit extracts the multi-dimensional feature data of the device operation and inputs it into a single-classification model for anomaly identification. This single-classification model utilizes high-dimensional boundaries to construct a hyperspherical envelope of normal operating conditions.

[0093] If an anomaly is identified, it will be classified as an internal coupling anomaly, an industrial interference-dominated anomaly, an equipment aging-dominated anomaly, or a mixed anomaly. Internal coupling anomalies typically refer to optical path misalignment; industrial interference-dominated anomalies refer to sudden, large-scale dust blockage; and equipment aging refers to the laser's lifespan nearing its end.

[0094] The data processing unit executes anomaly handling strategies that match the classification results for different categories. For example, in the case of anomalies dominated by industrial interference, it automatically triggers a high-speed pulsed airflow to perform reverse purging cleaning of the endoscope. After the anomaly handling strategy is executed, the latest equipment status identifier is extracted and the model parameters of the hybrid baseline model are forcibly updated, thereby completing the self-healing closed loop of the entire anomaly event and ensuring the reliability of continuous measurements.

[0095] Step 8: High-precision calculation of wet gas concentration and dynamic compensation inversion of dry concentration.

[0096] After all interference suppression processes, the filtered signal is used as the final output for gas concentration inversion. First, the data processing unit calculates the wet concentration of the target gas based on the filtered signal, the absorption cross-section of the target gas, and the ambient ring-down time of the optical cavity. The physical calculation formula is as follows:

[0097] ;

[0098] in, The wet concentration volumetric number density of the target gas in its current mixed state is represented by the vacuum speed of light, denoted as . , This represents the absolute absorption cross-section parameter of the target gas at a specific characteristic wavelength. This represents the decay time value of the high-purity filtered signal output in step six. It represents the ambient decay time of the optical cavity, which is strictly calibrated under absolute vacuum or when filled with pure high-purity nitrogen.

[0099] It is important to note that industrial emissions often contain significant amounts of water vapor and corrosive acidic gases, leading to substantial errors when directly outputting the wet concentration. Therefore, the data processing unit utilizes acquired ambient temperature, ambient pressure, water vapor coupling coefficient, and corrosion characteristic parameters to perform dynamic optical path compensation on the wet concentration, thereby retrieving and outputting the dry concentration of the target gas. Dynamic optical path compensation, through the ideal gas law combined with a local water vapor partial pressure correction coefficient, eliminates the broadening effect of water vapor molecules on the light beam and the erosion and refractive index changes caused by acidic gases on the micropores of the cavity mirror reflector, ultimately deriving a standard-state dry concentration value that complies with national environmental emission standards.

[0100] For example, in one scenario, the target gas to be measured is set to a trace amount of methane. Under ideal factory conditions, the constant value of the vacuum speed of light is set to... Meters per second. The initial calibration physical cavity length of the ring-down cavity is set to 0.5 meters. The initial absolute reflectivity of the cavity mirror is relatively high during factory calibration, set to 0.99995.

[0101] After several days of continuous operation at the industrial site, the feature data collected in real time by the multi-dimensional sensor array are as follows: The dynamic cavity length deformation caused by triaxial vibration acceleration is calculated by integration as follows: The temperature of the cavity mirror increased due to the continuous baking of the high-temperature flue gas, resulting in a slight but significant decrease in reflectivity that affects measurement accuracy. The calculated decrease in mirror reflectivity was 0.00001. The system-level response delay caused by the aging of the emitted laser source after prolonged operation was calibrated to 0.05 microseconds.

[0102] Based on the derivation of the underlying physical equations in step four, the data processing unit first calculates the drift anchor point in the purely physical dimension. Substituting the specifically set numbers above into the formula for the theoretical value of the physical baseline drift:

[0103] The first step is to calculate the actual reflectivity of the endoscope after being affected by temperature and contamination:

[0104] Actual reflectivity = ;

[0105] The second step is to calculate the current actual physical length of the optical cavity after the vibration:

[0106] Actual cavity length = rice;

[0107] The third step is to calculate the optical background-limited ring-down time, which is unaffected by the light speed aging delay:

[0108] Theoretical basis decay time = Second;

[0109] Convert seconds to microseconds for easier engineering calculations. A second is approximately equal to 27.7778 microseconds.

[0110] The fourth step is to add the aging compensation amount to obtain the final theoretical value of the physical baseline drift:

[0111] ;

[0112] Meanwhile, under extremely harsh operating conditions, due to sudden dust interference, the trend prediction time derived solely from the data-driven support vector regression model (without physical and calibration constraints) is 29.5000 microseconds. Simultaneously, the most recent record of the first self-calibration baseline value generated by the system's internal calibration procedure executed using non-absorbed pulses is 28.0000 microseconds, and the historical true baseline value recorded in the historical database under stable conditions is 27.9500 microseconds.

[0113] At this point, the system enters the joint loss function iteration phase of the hybrid baseline model. Based on real-time scene characteristics, the system automatically assigns penalty weight parameters across three dimensions: Due to the intense vibration sensor readings, the system sets the second dynamic weight coefficient of the physical theory constraint term to a relatively low 0.1; because the internal self-calibration has only recently been completed and the data possesses high confidence, the system sets the third dynamic weight coefficient of the self-calibration constraint term to a relatively high 0.6; and sets the first dynamic weight coefficient of the data-driven constraint term to 0.3.

[0114] Within this constraint framework, when the support vector regression model updates its parameters through backpropagation, its predicted real-time baseline drift is guided with higher weights towards the self-calibrated baseline value of 28.0000 microseconds and the historical true baseline value of 27.9500 microseconds, rather than being biased by the absurdly deviated data-driven result of 29.5000 microseconds. After multiple rounds of loss function minimization iterations, the highly robust real-time baseline drift value output by the support vector regression model is finally 27.9850 microseconds.

[0115] The process then proceeds to an adaptive filtering procedure. It is known that at the current sampling time, the real-time extracted value of the target ring-down time series, containing numerous spikes and baseline fluctuations, is 25.1000 microseconds. Based on the accurate anchor point of 27.9850 microseconds, the filter uses a minimum mean square error iterative algorithm to perform nonlinear weighted smoothing on the current fluctuation sequence. Assuming that under the action of the converged filter coefficients, after eliminating all baseline spurious disturbances, the output filtered signal ring-down time stably converges to 26.5000 microseconds.

[0116] Finally, the gas wet concentration inversion stage is entered. It is known that the cavity's background ring-down time is constant at 28.0000 microseconds during calibration with high-purity nitrogen. The reciprocal of the product of the absolute absorption cross-section parameter of methane gas at this characteristic wavelength and the speed of light in vacuum is read as a constant factor in the instrument calibration file. (Convert units to constants appropriate for concentration dimensions). Then, use the humidity concentration calculation formula for rigorous calculation:

[0117] ;

[0118] The first step is to calculate the reciprocal difference of the ring-down times:

[0119] ;

[0120] ;

[0121] Difference = ;

[0122] The second step is to multiply by the instrument calibration constant factor to obtain the final wet concentration:

[0123] ;

[0124] The above deduction shows that without the robust correction provided by the hybrid baseline model that deeply integrates physical formulas and self-calibrating weights in this embodiment, the baseline values ​​predicted by the purely data-driven algorithm, which have large deviations, will lead to significant errors in the subsequent adaptive filtering direction. However, the solution in this embodiment, under extreme multi-source interference, relies on multi-dimensional features and a joint loss function to firmly anchor the true baseline drift trajectory, resulting in extremely high accuracy and robustness in the final retrieved target gas concentration.

[0125] Example 2:

[0126] This embodiment, as an advancement and refinement of the first embodiment above, focuses on expanding the core technical aspects of calculating the correlation between empirical mode decomposition components and industrial interference feature data and extracting baseline drift candidate components.

[0127] In real, complex industrial environments, such as blast furnace exhaust outlets or highly corrosive flues in chemical plants, the physical and optical processes of the cavity ring-down spectrometer caused by industrial interference sources inevitably involve physical phase delays and complex nonlinear coupling effects between the interference source and the actual interference characteristic data acquired by the sensor. Using only conventional static synchronization correlation calculation methods often leads to serious missed or false positives, resulting in deep confusion between the real signal and the artifact signal.

[0128] To overcome this technical obstacle, this embodiment constructs a two-dimensional verification mechanism that integrates asynchronous time-delay compensation and multi-spectral-band spatial cross-validation. The specific execution logic of this mechanism can be broken down into the following steps:

[0129] Step 1: Perform cross-correlation calculation with time delay search.

[0130] In traditional signal processing logic, directly calculating the Pearson correlation coefficient of two sets of time-series signals requires that the signals be strictly aligned on the time axis. However, in cavity ring-down spectroscopy gas detection systems, when the external environment undergoes abrupt changes, such as a gust of wind stirring up a large amount of dust particles, these dust particles travel through the sampling probe, through multiple pre-processing pipelines, and finally into the high-reflectivity ring-down cavity, adhering to the cavity mirror surface and causing a decrease in reflectivity. This series of physical and mechanical transmission processes consumes considerable time. In contrast, dust concentration sensors or vibration sensors installed externally to the equipment generate peak electrical signals immediately upon the occurrence of interference. This physical transmission time difference causes the industrial interference characteristic data in the multi-dimensional feature data to significantly lead the empirical mode decomposition components resolved from the optical signal on the time axis.

[0131] Specifically, the data processing unit performs cross-correlation calculations with time delay search on each of the empirical mode decomposition components and the industrial interference feature data. This calculation process differs from the traditional static time window alignment mode; instead, it uses a sliding time window algorithm to calculate the dynamic correlation values ​​under different delay step sizes within a preset maximum physical delay time range.

[0132] To objectively quantify this dynamic correlation, this embodiment provides the following formula for calculating cross-correlation with time-delayed search:

[0133] ;

[0134] in, This indicates that the discrete time delay step is set manually. In the case of [the specific condition], the calculated cross-correlation coefficient sequence values ​​are obtained; This indicates the total number of discrete data sampling points contained within the current sliding time analysis window; This represents the current data index sequence number, which gradually increases from one to the total number of sampling points; This represents a specific empirical mode decomposition component extracted from the target ring-down time series after performing empirical mode decomposition on the 1st... The signal amplitude at each sampling point; It represents the mathematical expectation of the arithmetic mean of all signal amplitudes of that particular empirical mode decomposition component over the entire analysis time window; This indicates a shift along the historical timeline. The numerical representation of the industrial interference feature data after each delay step at the corresponding index node, the essence of the translation operation is to compensate for the transmission delay that occurs in the physical space in the mathematical space; This represents the statistical average baseline value of the industrial interference characteristic data within the time window.

[0135] The data processing unit iterates through all reasonable delay step parameters to plot a dynamic curve of the cross-correlation coefficient as a function of the delay step. Physically, the peak of this curve precisely corresponds to the actual physical time required for external physical interference to intrude into the optical resonator. The data processing unit then extracts the correlation coefficient with the largest absolute value from all the calculation results, thus successfully obtaining the maximum aligned correlation coefficient. This process of obtaining the maximum aligned correlation coefficient fundamentally eliminates the problem of low overlap between the interference signal and feature data in the time domain due to physical phase delay, greatly restoring the true causal relationship between the interference source and the fading baseline fluctuations.

[0136] like Figure 5 The dynamic cross-correlation coefficient optimization curve with physical time delay compensation shown in the figure illustrates the dynamic time alignment process of this invention when dealing with the physical phase delay problem between industrial interference characteristic data and empirical mode decomposition components. The horizontal axis represents the discrete time delay step size of the system traversal search, in steps, and the vertical axis represents the calculated cross-correlation coefficient value, which is dimensionless.

[0137] The blue solid line in the figure represents the dynamic curve of the cross-correlation coefficient as a function of the delay step. The curve exhibits a convex peak shape, first rising and then falling. This waveform transformation reflects the physical transmission time difference experienced by interfering substances such as dust as they travel from outside the device to inside the optical resonant cavity.

[0138] As the delay step size gradually increases, the transmission lag in physical space is gradually compensated for in the mathematical calculations, and the correlation between the two sets of data on the time axis changes accordingly. The red solid dots in the figure mark the peak position of this dynamic curve. As can be seen from the corresponding values ​​on the coordinate axis, the curve reaches a peak value of approximately 0.88 when the delay step size is around 45. The red dot represents the maximum alignment correlation coefficient obtained by the system search.

[0139] This graphical, dynamic optimization process illustrates the role of this invention in reducing the probability of missed detections caused by physical phase misalignment in traditional static synchronization algorithms. By extracting the maximum alignment correlation coefficient greater than a preset correlation threshold, the system can extract the corresponding signal as a preliminary interference component with high confidence, thereby restoring the correlation between external industrial interference sources and internal fading baseline fluctuations to a certain extent, providing data support for subsequent signal stripping.

[0140] Step 2: Extract preliminary interference components based on dynamic alignment features.

[0141] After accurately obtaining the maximum alignment correlation coefficient corresponding to each empirical mode decomposition component, the system enters the preliminary qualitative screening stage. The data processing unit extracts the empirical mode decomposition components whose maximum alignment correlation coefficient is greater than or equal to the preset correlation threshold as preliminary interference components.

[0142] Specifically, the preset correlation threshold is a statistical decision boundary set by the system based on historical big data experience to distinguish between random noise and strong coupling interference. If the maximum alignment correlation coefficient of a certain empirical mode decomposition component fails to reach the decision boundary, the system considers that the component is likely caused by inherent thermal noise, shot noise, or weak gas concentration fluctuations completely independent of the currently monitored industrial interference source, and therefore will not be included in the candidate list for interference rejection.

[0143] It is important to note that by using the maximum alignment correlation coefficient as a threshold, the system effectively reduces the risk of missing relevant signals. In previous techniques, severe oscillations in the optical cavity baseline, originally induced by strong dust or electromagnetic interference, were often mistakenly identified as genuine, intense absorption by an unknown gas due to slight time axis misalignment, resulting in alarms indicating severe exceedances in concentration inversion results. This embodiment consolidates and reliably extracts empirical mode decomposition components that meet the maximum alignment correlation coefficient threshold as preliminary interference components, providing precise targets for subsequent, deeper spatial multi-spectral feature identification.

[0144] Step 3: Precise extraction of projection feature values ​​of the dual-band optical cavity response.

[0145] Specifically, while time-domain delay search and correlation comparison may resolve the risk of missed detections due to phase misalignment, they can easily induce another, more insidious, misjudgment crisis. When the target gas experiences extremely slow, low-frequency leakage or gradual concentration changes within the monitoring pipeline, this low-frequency, gentle fluctuation trend caused by real gas absorption is highly likely to exhibit a high degree of mathematical similarity to, within a specific monitoring period, the slowly rising boiler ambient temperature or the gradually accumulating dust index on the detector surface. This phenomenon is known as accidental low-frequency common potential characteristic in signal processing. If the system operates solely based on the preliminary conclusions of step two, it will inevitably erase this real, slowly changing gas absorption signal as baseline drift.

[0146] To eliminate the misjudgment caused by this accidental low-frequency common potential, this embodiment introduces a multi-spectral spatial cross-validation mechanism based on physical spectral characteristics. The data processing unit extracts the projection feature values ​​of the preliminary interference component onto the first and second ring-down curves, respectively. As defined in Embodiment 1, the first ring-down curve is excited by a detection pulse whose wavelength is tightly locked to the characteristic absorption peak of the gas to be measured, while the second ring-down curve is excited by a non-absorption pulse whose wavelength deliberately avoids all known gas absorption bands.

[0147] Optionally, the calculation of the projection eigenvalue aims to quantify the physical loss weight of this initial interference component at different laser wavelengths. The data processing unit performs this projection analysis using an inner product operation model, and the mathematical formula for calculating the projection eigenvalue is defined as follows:

[0148] ;

[0149] in, The scalar form of the projected eigenvalue represents the final calculated output of the system. When the calculation is performed on the first oscillation curve, it is specifically instantiated as the first projected eigenvalue, and when the calculation is performed on the second oscillation curve, it is specifically instantiated as the second projected eigenvalue. This represents the total length of the decay points effectively acquired by the system's analog-to-digital converter within a single pulse decay cycle. Integer index control variables for iterating through each point; This represents the discrete amplitude sequence of the initial interference components in the time domain, which has already been processed by the time alignment algorithm. This represents the discrete amplitude sequence of the time-domain decay envelope of a specific fading curve itself. During separate calculations, it is alternately replaced by data from either the first or second fading curve. By summing the absolute values ​​of the products of the corresponding points of the two curves, the system can accurately determine the proportion of energy influence that the initial interference component occupies at the absorption and non-absorption wavelengths.

[0150] Step 4: Calculation of response change ratio and rigorous verification of broadband background fluctuation range.

[0151] After acquiring the projected characteristic values ​​for two different wavelength pulses, the system enters the final logical decision-making stage to determine whether the component should be retained or discarded. The data processing unit calculates the response change ratio of the projected characteristic value between the first and second ring-down curves.

[0152] For example, the calculation model for the response change ratio is expressed as the quotient of the division of the aforementioned two independent projected eigenvalues, as shown in the following formula:

[0153] ;

[0154] in, This represents the numerical value of the response change that reveals the essence of physical absorption; The first projected characteristic value is obtained by integrating the initial interference component at the first ring-down curve, i.e., the absorption wavelength. The second projected characteristic value represents the result of integrating the initial interference component at the second ring-down curve, i.e., at the non-absorption wavelength.

[0155] It is important to note that, according to the fundamental principles of molecular spectroscopy, if the current baseline fluctuations are entirely caused by Mie scattering from dust particles, broad-spectrum indiscriminate absorption by water vapor, or physical-mechanical deformation of the optical cavity mirror reflector, the optical characteristics exhibited by these industrial interferences are typical of broad-spectrum non-selective loss. This means that regardless of whether the detection pulse is at an absorption wavelength or a non-absorption wavelength, the additional photon attenuation loss caused by these interferences is almost perfectly proportional.

[0156] Under this physical mechanism, no matter how large the absolute amplitude of the initial disturbance component is, the calculated first projected characteristic value and the second projected characteristic value should be extremely close, so that the value of the response change ratio should infinitely approach the value of one.

[0157] Conversely, if this initial interference component unfortunately encounters a random low-frequency common potential characteristic, its true identity is actually the slow increase in the concentration of the target gas. Based on the physical fact that the target gas only strongly absorbs the absorption wavelengths and is transparent to non-absorption wavelengths, the first projected eigenvalue will exhibit a huge value, while the second projected eigenvalue will only contain weak background noise. In this case, the calculated response change ratio will show a huge imbalance, deviating drastically from the first value.

[0158] Specifically, after obtaining the response change ratio, the data processing unit marks the corresponding preliminary interference component as the baseline drift candidate component when the response change ratio is within a preset background fluctuation range. The preset background fluctuation range is a closed numerical range centered at the value 1, extending upwards and downwards with a small physical tolerance.

[0159] For example, in most general-purpose optical path designs, this preset background fluctuation range is set between 0.85 and 1.15. Only when the response change ratio falls strictly within this numerical trap range, which represents broadband non-selective loss, can the system determine with a high degree of confidence that the component is actually a baseline drift caused by environmental interference.

[0160] like Figure 6The scatter plot of the response change ratio of the dual-band projection feature and the distribution of the trap intervals shown intuitively demonstrate the spatial cross-validation mechanism of this invention in identifying accidental low-frequency common potential features. In the figure, the horizontal axis represents the sample number of different initial interference components, in the unit of (serial number), and the vertical axis represents the calculated response change ratio value, which is in (dimensionless) unit.

[0161] The rectangular shaded area in the figure, defined by two black horizontal dashed lines and filled with light gray, represents the system's preset broadband background fluctuation trap range, with a value range between 0.85 and 1.15.

[0162] The blue circular data points scattered within the light gray trap area in the figure represent that their projected characteristic values ​​at the absorption wavelength and the non-absorption wavelength are relatively close, which conforms to the proportional attenuation physical mechanism of broadband non-selective loss such as dust particle scattering or mirror physical deformation. Therefore, the system can determine the preliminary interference components corresponding to these blue circular data points as candidate components of baseline drift caused by environmental interference with a high degree of confidence.

[0163] Conversely, the red triangular data points scattered outside the light gray trap area in the figure show a response change ratio that deviates from the value of one. This reflects the physical characteristic that the target gas has strong absorption of specific absorption wavelengths but weak absorption of non-absorption wavelengths. Based on this, the system identifies it as a slow increase in the concentration of the real gas.

[0164] The accompanying figure illustrates the distribution pattern of different colors and shapes of scattered points within a specific range, demonstrating that the present invention can effectively distinguish and retain low-frequency real gas signals from complex industrial backgrounds by leveraging dual-band physical spectral characteristics, thus providing a data foundation for subsequent concentration inversion.

[0165] Through this series of extremely rigorous and interconnected time delay synchronization compensation and dual-band ratio verification logic, the data processing unit, using physical laws as a benchmark, accurately and flawlessly eliminated the true absorption components that had accidental low-frequency common potential characteristics with the industrial interference feature data. This not only greatly expanded the system's ability to sensitively detect low-frequency, slowly varying trace gas leaks, but also ensured that the data set ultimately marked as baseline drift candidate components possessed absolute purity and reliability, providing an extremely solid underlying foundation for constructing a hybrid baseline model in subsequent embodiments.

[0166] Example 3:

[0167] Based on the aforementioned embodiments, the method in this embodiment focuses on system disaster recovery and model self-healing mechanisms when sensors in industrial fields experience hardware failures or severe data distortions.

[0168] In extremely harsh industrial monitoring environments, such as exhaust pipelines of high-temperature and high-pressure reactors or flues of large coal washing plants with strong vibrations, the hardware sensors responsible for collecting the physical parameters of the system will inevitably encounter localized transient extreme environmental damage.

[0169] For example, an unexpected, minute leak of high-temperature, high-pressure steam could cause a temperature sensor attached to the outer shell of a cavity to output an abnormal temperature jump signal of several thousand degrees Celsius within milliseconds. Alternatively, at the moment of startup of large, heavy-duty equipment, a massive mechanical shock wave could cause a piezoelectric ceramic vibration sensor to instantly enter a dead zone state with saturation. If these physical sensors output erroneous extreme values ​​that severely violate thermodynamic laws or material mechanical limits, the theoretical values ​​of physical baseline drift calculated from these parameters will exhibit absurdly precipitous jumps. If the support vector regression model still rigidly uses static, fixed loss function weights, these significantly biased theoretical predictions will generate large backward penalty gradients, causing the model's optimization direction in the feature space to deviate severely, ultimately leading to model divergence or outputting severely distorted baseline compensation values.

[0170] To address this model divergence problem at its root, this embodiment introduces a model constraint degradation and fault physical isolation mechanism based on dynamic data confidence assessment. This security isolation mechanism specifically includes the following steps:

[0171] Step 1: Extraction of real-time gradients of physical parameters and dynamic determination of sensor distortion status.

[0172] To prevent erroneous physical data from destroying the hybrid baseline model, the system's data processing unit must be capable of rapidly identifying anomalies and blocking the propagation of erroneous data within a very short time. The data processing unit extracts the real-time gradients of each physical parameter. Physical parameters mainly refer to continuous physical quantities reflecting the external environment or hardware state of the cavity ring-down spectrometer, such as cavity mirror temperature and vibration acceleration. The extraction of real-time gradients aims to quantify the drastic changes of these physical parameters within extremely short discrete sampling time intervals, representing the instantaneous evolution rate of the physical quantity in the time dimension.

[0173] To accurately perform this extraction, the data processing unit executes gradient calculation using the backward finite difference algorithm. The specific calculation formula is defined as follows:

[0174] ;

[0175] in, It represents a sequence of real-time changes in the absolute value of gradients calculated for a specific physical parameter; This represents the discrete sampled value of the physical parameter read by the system's high-precision analog-to-digital converter at the latest sampling clock cycle. It represents the discrete sampled value of the same physical parameter that the system latched in the previous historical sampling clock clock clock immediately preceding the current sampling clock clock on the time axis; It represents the absolutely constant time interval step between two adjacent sampling clock cycles set by the system.

[0176] By calculating the absolute value of the difference between the current value and the value at the previous moment and dividing it by the time step, the system can accurately and sensitively capture the instantaneous rate of change of any physical parameter.

[0177] After calculating the gradient sequence of all physical parameters, the system's data processing unit determines that the corresponding physical sensor is in a data distortion state when the real-time gradient exceeds a preset threshold. The preset threshold is the highest legal rate of change limit burned into non-volatile memory before the system leaves the factory, based on strict physical laws and the physical response limits of the sensor's semiconductor components. Taking a temperature sensor as an example, due to the physical limitations of the specific heat capacity and thermal conductivity of the metal probe's packaging material, the actual temperature rise rate it senses can never exceed the theoretical thermodynamic limit of several hundred degrees Celsius per second. Once the calculated real-time gradient ignores this objective physical limit and significantly exceeds it, the system has sufficient logical support to immediately determine that the physical sensor generating the data is experiencing hardware-level data distortion, such as electromagnetic interference breakdown or a short circuit.

[0178] Specifically, after establishing the sensor's data distortion state, the data processing unit immediately triggers a penalty generation mechanism and generates a corresponding confidence decay factor. The confidence decay factor is the output value of a nonlinear scalar mapping function whose value range is strictly limited to zero and one; its core purpose is to quantify the residual reliability of the current sensor data. For physical sensors determined to be in a distorted state, the data processing unit calculates the confidence decay factor based on the severity of exceeding a preset threshold. The specific mathematical model for calculation is defined as follows:

[0179] ;

[0180] in, The value of the confidence attenuation factor corresponding to a specific distorted physical sensor, representing the calculated output; a constant. This represents the system's preset penalty coefficient for steep attenuation. The larger the value, the lower the system's tolerance for out-of-limit data, and the more severe the attenuation process. This represents the real-time gradient value exceeding the limit calculated above; This represents the preset limit threshold constant that is strictly defined for this physical parameter.

[0181] When the real-time gradient is exactly equal to the limit threshold, the confidence decay factor remains at the highest full score value of 1. However, once the real-time gradient exceeds the limit threshold by an explosive rate, due to the strong suppression characteristics of the exponential decay function, the confidence decay factor will decrease nonlinearly and rapidly, converging towards the zero limit at an extremely fast speed. This prepares the data for effective isolation of the sensor data from the model constraints and subsequent numerical quantization.

[0182] Step 2: Self-calibration timeliness assessment and continuous evolution generation of confidence weights.

[0183] It is important to note that in the joint loss function system, in addition to anchoring based on physical formulas, the self-calibration benchmark value generated periodically by the built-in micro standard gas capsule or non-absorption pulses is also a core pillar supporting the support vector regression model to avoid long-period baseline drift.

[0184] However, the validity of self-calibration reference values ​​is not static; they inherently possess a physical decay characteristic, accumulating uncertainty over time. In harsh industrial measurement processes, where automatic calibrations may be spaced for hours or even tens of hours, the continuous deposition of microscopic aerosol particles within the optical cavity and the threshold current creep caused by the prolonged operation of high-energy laser diodes can lead to the once absolutely accurate self-calibration reference values ​​gradually losing their representative guidance for current real-world conditions. Therefore, a dynamic trust assessment mechanism with diminishing time-dependent validity must be introduced for self-calibration reference values.

[0185] Specifically, the data processing unit obtains the update timestamp of the self-calibration reference value. The update timestamp refers to the absolute physical point in time when the system's underlying clock chip successfully executed and completed the automatic calibration closed-loop process most recently, and securely wrote the latest calibration result into the system's valid storage register. This timestamp has a globally unique system clock tick attribute and is the absolute time scale starting point for measuring the freshness of the self-calibration data.

[0186] For example, the data processing unit queries the system's high-precision real-time clock and calculates the time difference between the current moment and the updated timestamp. The derivation formula for the time difference is expressed as follows:

[0187] ;

[0188] in, This represents the current calculated absolute time difference variable value; This represents the global absolute physical time value when the data processing unit is currently executing the feature calculation task; This represents the update timestamp value of the self-calibration reference value extracted from the system register as described above. By linearly subtracting the two, the system obtains the cumulative duration of the system's operation in the dead zone since the last reference refresh.

[0189] Optionally, the data processing unit generates a confidence weight that decreases over time based on the time difference. The confidence weight is also a dimensionless penalty coefficient, designed to dynamically adjust the model's dependence on historical calibration data during backpropagation. To align with the physical and statistical laws governing the accumulation of interference in complex industrial environments, the confidence weight is generated based on a smooth and gradual nonlinear decay model. The specific formula for calculating the confidence weight is defined as follows:

[0190] ;

[0191] in, This represents the dynamic value of the confidence weight for the self-calibration benchmark value ultimately generated by the system; a constant. This represents the initial highest-weighted reference constant assigned by the system to the self-calibration reference value at the instant calibration is completed, i.e., when the time difference is zero; constant. The constant represents the time decay hysteresis coefficient, which is manually adjusted based on the severity of the on-site environment or derived from the system's long-term big data self-learning. For working conditions with extremely severe dust pollution, this constant will be adjusted to be relatively large in order to accelerate the decay of old calibration data. This represents the numerical value of the absolute time difference variable calculated above.

[0192] Through this nonlinear rational fractional decay model, the newly generated self-calibration baseline value has the highest basic weight constraint. As time goes by and no new calibration action refreshes the timestamp, its weight influence in the joint loss function of the model will show a gradual decrease.

[0193] Step 3: Dynamic interweaving and reconstruction of model constraint weights in the joint loss function.

[0194] After successfully quantifying the confidence decay factor for sudden physical distortions and the confidence weight for slow time-dependent failures, the system enters the deep-water zone of reconstructing the underlying optimization environment of the hybrid baseline model. The data processing unit uses the confidence decay factor and the confidence weights to dynamically update the weights of the physical constraint terms corresponding to the theoretical values ​​of the physical baseline drift and the calibration constraint terms corresponding to the self-calibration benchmark values ​​in the joint loss function.

[0195] Specifically, in the traditional support vector regression training framework, the weight ratios between different loss terms are often fixed and rigid hyperparameters set at the factory. This static architecture is simply unable to withstand the sudden storms of industrial environments. However, in the architecture of this embodiment, updating the weights becomes a high-frequency, real-time computational task. The data processing unit dynamically refreshes these core weight parameters through a forward multiplication network, with the mathematical logic defined as follows:

[0196] ;

[0197] ;

[0198] In the two sets of dynamic weight refresh formulas mentioned above, This represents the updated weight of the physical constraint term, which is generated after the data health assessment and is specifically used for the physical theory deviation penalty term in the direct multiplication joint loss function. Represents the updated calibration constraint weights, specifically used for the self-calibration deviation penalty term in the direct multiplication joint loss function, generated after time decay assessment; constant. Represents the static fundamental equilibrium coefficients originally set by the system to account for physical constraints; a constant. This represents the static basic balance coefficient originally set by the system for self-calibration constraints; This represents the trust decay factor output in step one; This represents the confidence weight output in step two.

[0199] Through the aforementioned multiplicative mapping reconstruction, when the physical sensor data is intact and the self-calibration data is sufficiently fresh, both attenuation factors remain high. The hybrid baseline model receives comprehensive and multi-dimensional abundant data source constraints, exhibiting excellent prediction accuracy. However, if an extremely severe destructive event occurs, causing serious distortion of one of the physical sensor data, the system can instantly and rapidly weaken the weight of the physical constraint term in the joint loss function to near zero through a non-linearly decreasing confidence attenuation factor. This dynamic weight update closed-loop mechanism effectively isolates the absurd and distorted physical theoretical values ​​from severely interfering with the hyperplane optimization direction of the support vector regression model from the underlying dimension of the mathematical optimization space.

[0200] Step 4: Seamless and smooth switching between the ultimate circuit breaking of the underlying logic and the dual-path constraint mode.

[0201] While theoretically, dynamically decaying weights can significantly mitigate the negative impact of erroneous physical data, in the extremely rigorous mathematical optimization process, even the smallest residual higher-order error gradient penalty can potentially cause subtle distortions in the model's feature space over thousands of iterations. To improve the reliability of industrial applications, the system is equipped with anomaly isolation and mode degradation mechanisms.

[0202] Specifically, the data processing unit monitors the dynamically updated weights of the physical constraint terms in real time. When the weights of the physical constraint terms fall below a preset failure threshold, the prediction mode of the support vector regression model is switched to a dual-constraint mode dominated by the historical true baseline value and the self-calibration benchmark value. The preset failure threshold is an extremely small safety tolerance limit parameter that approaches zero, symbolizing the system's negativity towards the physical sensors.

[0203] It's also important to note that when the weights of the physical constraints breach this failure threshold, the data processing unit interrupts its program execution and switches the model architecture. At this point, the entire data link containing the erroneous physical parameters, along with the corresponding physical baseline drift calculation module, is completely isolated at the hardware or software level. The system forcibly truncates the massive multidimensional joint loss function within the support vector regression model, constructing a clean and isolated array for calculating the degraded loss function. The mathematical model of the loss function under the dual-path constraint mode is extremely simplified and redefined as follows:

[0204] ;

[0205] in, This represents the total penalty loss value of the downgraded model specifically generated under the dual-path constraint mode; The conservative weight constants represent historical trends that are deliberately amplified by the system to maintain the model's enormous inertia and stability. The baseline predictor values ​​that the model attempts to provide at the current iteration tick; Represents an unquestionable historical true baseline value constant stored in a deep radiation resistance database; The updated calibration constraint weights, which, although decaying over time, remain the only objective and real-world anchor point of the system; This represents the last remaining high-quality self-calibration reference constant.

[0206] In this extremely streamlined dual-path constraint mode, the support vector regression model completely cuts off any gradient information interaction with the distorted physical world outside. Instead, the model adopts a self-sustaining operation concept similar to trajectory extrapolation or inertial navigation in the maritime field. It relies solely on the low-frequency historical inertial data accumulated over a long period of time by the system and the valuable absolute position anchor points left over from the last calibration, combined with the still healthy empirical mode decomposition component feature set, to maintain a stable optimization direction in the multidimensional feature space and calculate a safe and conservative baseline drift trajectory.

[0207] This dual-path constraint mode switching essentially isolates the distorted physical baseline drift theoretical value from any potential contamination or interference to the real-time baseline drift value output. By sacrificing some of the sensitivity of the high-frequency response in exchange for the system's robustness under large-scale sensor failures, the system can continuously provide the backend with stable, reliable, conservative but not absurd gas concentration inversion data until the engineering maintenance personnel arrive on site to troubleshoot the sensor hardware-level fault and manually reset the system.

[0208] like Figure 7 The comparison chart showing the model's anti-tear fuse failure and degraded dual-path constrained baseline prediction under extreme distortion conditions intuitively demonstrates the anomaly isolation and mode switching capabilities of this invention when facing physical sensor hardware-level failures. In the chart, the horizontal axis represents the continuous operating time of the system in cycles, and the vertical axis represents the real-time baseline prediction value output by the hybrid baseline model in microseconds (μs).

[0209] The black dashed line in the figure represents the comprehensive safety baseline formed by combining the historical true baseline value stored in the system database with the recent self-calibration baseline value, and its value remains around 27.9 microseconds.

[0210] Before the runtime reaches 100, the system is in normal monitoring mode, and the predicted trajectories are all based on this benchmark. When the runtime reaches 100, the system simulates the extreme failure condition of abnormal jump signals in the output of physical sensors.

[0211] The blue dotted line in the figure represents the predicted trajectory of the support vector regression model using traditional static weight constraints. It can be observed that due to the deviation caused by the abnormal physical sensor, a large inverse penalty gradient was caused. The blue curve deviated significantly after the failure occurred, and the predicted value rose to more than 35 microseconds in a short period of time, showing the phenomenon that the failure of the physical anchor point caused the model to deviate from the optimization direction.

[0212] In contrast, the solid red line in the figure represents the predicted trajectory after the introduction of the confidence decay and degradation mechanism in this invention. When a sensor malfunctions, the system detects a real-time gradient exceeding the limit, triggering a decrease in the weight of the physical constraint term to reduce the impact of abnormal physical data, and switches the model to a dual-constraint mode dominated by historical true baseline values ​​and self-calibration benchmark values. Therefore, the solid red line does not deviate significantly from the blue curve, but continues to maintain a relatively stable range around 27.9 microseconds due to historical inertia and the constraints of the calibration anchor point.

[0213] By comparing the blue deviation curve and the red stable curve in the same coordinate system, the attached figure illustrates that the isolation mechanism of the present invention can reduce the interference of abnormal physical parameters on the model, and still help to ensure the stable output of baseline prediction under extreme failure conditions such as large-scale sensor failure, thereby improving the reliability of gas concentration inversion.

[0214] Example 4:

[0215] like Figure 8 As shown, this embodiment provides a ring-down time baseline fluctuation suppression system based on cavity ring-down spectroscopy. This embodiment serves as a physical device and system-level architecture illustration of the methods disclosed in Embodiments 1 to 3 above, focusing on demonstrating how the aforementioned mathematical deduction and anomaly prevention mechanisms can be practically implemented through deep collaboration of specific industrial-grade hardware devices, electronic components, and underlying embedded software logic.

[0216] It is important to note that, as pointed out in the background section, traditional cavity ring-down spectroscopy detection equipment is highly susceptible to dust contamination, strong mechanical vibration, corrosive gas corrosion, and sudden electromagnetic interference when facing harsh industrial environments such as blast furnace exhaust gas and chemical flue gas, leading to severe baseline drift. Furthermore, traditional static filtering and frequent zero-gas calibration schemes not only interrupt continuous measurements but also face the technical bottleneck of model tearing between physical anchor point failure under extreme conditions and data-driven rigid constraints.

[0217] The system disclosed in this embodiment is designed to completely solve the aforementioned conflict-related technical problems. By combining multi-dimensional hardware perception with a heterogeneous edge computing platform, it achieves the beneficial effects of dynamic baseline accurate tracking, fault self-healing isolation, and high-precision continuous gas concentration inversion even in extremely harsh environments.

[0218] Specifically, the cavity ring-down spectrum-based time-baseline fluctuation suppression system disclosed in this embodiment is deeply applied in the aforementioned cavity ring-down spectrum gas detection system. Its hardware base mainly includes a broadband tunable semiconductor laser, an acousto-optic modulator, a high-reflectivity optical resonant cavity, a double-shielded avalanche photodiode (APD) detector, an industrial-grade multidimensional sensor array, and an edge heterogeneous computing platform (including a field-programmable gate array FPGA and a high-performance ARM processor) as the core.

[0219] From the perspective of system functional logic architecture, the system is characterized by including: a sequence generation module, a signal decomposition module, a baseline prediction module, and a filtering and output module.

[0220] For example, the sequence generation module is used to acquire the first ring-down curve corresponding to the detection pulse, the second ring-down curve corresponding to the non-absorption pulse, and the ring-down curve of the reference pulse; extract the ring-down time of each ring-down curve respectively, and eliminate the non-target gas absorption interference by performing differential calculation on the ring-down time to generate the target ring-down time sequence.

[0221] In practical industrial applications, the operation of this sequence generation module relies not only on underlying software algorithms but also on sophisticated opto-mechanical-electronic hardware coordination. A high-precision timing controller within the edge computing platform sends nanosecond-level trigger commands, driving a tunable semiconductor laser to switch rapidly between three different center wavelengths and controlling an acousto-optic modulator to instantly cut off the beam. The transmitted beam passing through the optical resonant cavity is captured by a double-shielded avalanche photodiode detector, converted into a weak electrical signal, and then digitized into discrete decay curves by an ultra-low noise transimpedance amplifier and a high-speed analog-to-digital converter (ADC).

[0222] Leveraging its powerful parallel processing capabilities, the FPGA chip performs fast logarithmic linear fitting on a massive number of sampling points to extract the decay time of each pulse. By directly performing differential calculations on the decay times of the absorption wavelength and the non-absorption wavelength at the FPGA hardware logic level, the system eliminates common-mode background interference caused by particulate matter adhering to the cavity mirror surface or broadband continuous absorption of water vapor at the physical source before the optical signal enters the advanced algorithm layer. This provides subsequent modules with a high-purity and high-reliability target decay time series.

[0223] For example, the signal decomposition module is used to decompose the target decay time series in response to the target decay time series and extract baseline drift candidate components.

[0224] In practical applications, because signal decomposition involves matrix operations with high time complexity, this module is primarily deployed in the DSP (Digital Signal Processing) coprocessor unit of a high-performance ARM processor. Upon receiving the target waning time series, the coprocessor unit executes wavelet transform instructions and the Empirical Mode Decomposition (EMD) algorithm at high speed.

[0225] It is also important to note that, in order to address the technical problem of confusion between true and false signals caused by physical phase delay and accidental low-frequency common potential, this signal decomposition module not only performs a single mathematical decomposition, but also retrieves industrial interference characteristic data sent by microelectromechanical systems (MEMS) sensors or external dust concentration meters on the bus. The signal decomposition module opens a sliding time window in memory and performs hardware-level cross-correlation calculations with time delay search to compensate for the physical time consumed by dust entering the optical cavity circuit.

[0226] Subsequently, the module uses the projection feature values ​​of the dual-band optical cavity response for spatial cross-validation. Only when the response change ratio calculated from the projection feature values ​​extracted at different wavelengths falls into the hardware preset range representing broadband non-selective loss, will the coprocessing unit confirm that the component is not a slow leak of real gas and securely solidify and extract it as a baseline drift candidate component, which will then be sent to the next level network.

[0227] For example, the baseline prediction module is used to construct a hybrid baseline model based on the baseline drift candidate components and synchronously acquired multi-dimensional feature data to predict the real-time baseline drift value.

[0228] At the equipment level, multi-dimensional feature data is collected in real time and synchronously by an industrial-grade sensor array widely distributed inside and outside the testing chassis: a PT100 high-precision platinum resistance thermometer is placed close to the outer wall of the cavity to collect the temperature of the cavity mirror, a high-frequency vibration sensor monitors the mechanical impact of the mounting base, and an optical power meter monitors the output power attenuation of the laser in real time. The signals collected by these hardware components are aggregated to the baseline prediction module via an industrial fieldbus.

[0229] Furthermore, the module's hardware deeply integrates an edge-end dual-mode self-calibration solenoid valve assembly. Inside the chassis, there is a miniature standard gas capsule with a volume of less than one milliliter and a corrosion-resistant flow path switching valve controlled by a high-pressure pulse. When the system runs for an extended period and triggers the self-calibration condition, the baseline prediction module opens the solenoid valve through the control I / O port, automatically completing the replacement of the optical cavity gas and the recalibration of the reference value, generating an extremely valuable self-calibration reference value.

[0230] The baseline prediction module allocates an independent monitoring thread within the processor. When an industrial sensor (such as a temperature probe) short-circuits due to a high-temperature steam leak, causing the real-time gradient uploaded via the analog-to-digital conversion interface to momentarily exceed the physical limit threshold, the baseline prediction module immediately determines that the sensor is in a data distortion state and instantly generates a confidence decay factor close to zero. Under this hardware interruption mechanism, the baseline prediction module forcibly disconnects the physical computation link associated with the distorted sensor in the Support Vector Regression (SVR) model, seamlessly switching the model to a dual-constraint degradation mode supported by a historical database and a micro gas cylinder self-calibration anchor point. This hardware-software combined isolation mechanism essentially prevents catastrophic disruption of the core baseline prediction algorithm by external physical hardware damage, outputting stable and reliable real-time baseline drift values.

[0231] For example, the filtering and output module is used to initiate a dynamic weighted adaptive filtering process based on the real-time baseline drift value to suppress baseline fluctuations in the target oscillation time series and obtain a filtered signal for gas concentration inversion.

[0232] In terms of hardware implementation, the filtering and output module relies on the main control processor's computing core to perform adaptive filtering based on the least mean square (LMS) criterion. Its step size and weights are dynamically adjusted in real time according to the on-site signal-to-noise ratio. After obtaining an extremely smooth filtered signal that has eliminated all drift interference, the module executes the concentration calculation equation and, in conjunction with the readings from the system's built-in pressure and temperature sensors, performs dynamic optical path and temperature-pressure compensation calculations, ultimately retrieving the target gas dry concentration that conforms to standard conditions.

[0233] Finally, to meet the access requirements of modern industrial IoT and distributed control systems (DCS / PLC), the hardware terminals of the filtering and output modules are equipped with abundant industrial communication interfaces with electromagnetic isolation protection. The high-precision gas concentration data, system health status (including whether degradation constraints have been triggered, sensor trust scores, etc.), and calibration logs will be transmitted in real-time, securely, and uninterrupted to the central control room screen of chemical plants or steel mills via an isolated 4-20mA analog output channel, an RS485 digital communication bus (supporting Modbus RTU protocol), and a LoRa wireless RF module.

[0234] In summary, the system provided in this embodiment four, through opto-mechatronics co-design and heterogeneous edge computing architecture, can not only distinguish between true and false signals in harsh dust and vibration environments, but also achieve fault self-healing under extreme failure conditions of sensor hardware damage, and has considerable industrial promotion value and feasibility.

[0235] Example 5:

[0236] Corresponding to the above embodiments, the present invention also proposes an electronic device.

[0237] like Figure 9 The diagram shows a structural schematic of an electronic device according to the present invention. The electronic device 100 includes a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, for example, via a bus 102. Optionally, the electronic device 100 may further include a transceiver 104. It should be noted that in practical applications, the transceiver 104 is not limited to one unit, and the structure of this electronic device 100 does not constitute a limitation on the embodiments of the present invention.

[0238] Processor 101 may be a CPU, a general-purpose processor, a DSP, an ASIC, an FPGA, or other programmable logic device, transistor logic device, hardware component, or any combination thereof. It may implement or execute the various exemplary logic blocks, modules, and circuits described in connection with this disclosure. Processor 101 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0239] Bus 102 may include a pathway for transmitting information between the aforementioned components. Bus 102 may be a PCI bus or an EISA bus, etc. Bus 102 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 9 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0240] The memory 103 is used to store a computer program corresponding to the cavity ring-down spectrum-based ring-down time baseline fluctuation suppression method of the above embodiments of the present invention. The computer program is controlled and executed by the processor 101. The processor 101 is used to execute the computer program stored in the memory 103 to implement the content shown in the foregoing method embodiments.

[0241] Among them, electronic devices 100 include, but are not limited to: mobile terminals such as laptops and PADs (tablet computers) and fixed terminals such as desktop computers. Figure 9 The electronic device 100 shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention.

[0242] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for suppressing the time baseline fluctuation of cavity ring-down spectroscopy, applied in a cavity ring-down spectroscopy gas detection system, characterized in that, include: In response to the target fading time series, the target fading time series is decomposed to extract baseline drift candidate components. Specifically, this includes: performing multi-scale mode decomposition on the target fading time series to obtain multiple decomposed components; calculating the correlation between each decomposed component and synchronously acquired industrial interference feature data; and extracting the baseline drift candidate components from the multiple decomposed components based on the correlation. Based on the baseline drift candidate components and synchronously acquired multi-dimensional feature data, a hybrid baseline model is constructed to predict the real-time baseline drift value. The hybrid baseline model includes physical baseline theoretical constraints, data-driven prediction constraints, and self-calibration benchmark constraints. Specifically, the physical baseline theoretical constraints are calculated using synchronously acquired physical parameters; the baseline drift candidate components and multi-dimensional feature data are input into the data prediction model to generate the data-driven prediction constraints; and the self-calibration benchmark constraints are obtained through a self-calibration mode. The real-time baseline drift value is predicted and output by combining the above three constraints. A dynamic weighted adaptive filtering process based on the real-time baseline drift value is initiated to suppress baseline fluctuations in the target oscillation time series, and the filtered signal is used for gas concentration inversion, specifically including: Extract real-time scene features of the current detection scene, wherein the real-time scene features include at least baseline fluctuation intensity and signal-to-noise ratio; Based on the real-time scene characteristics, the prediction weights of the hybrid baseline model and the filtering weights of the adaptive filter are dynamically allocated; The step size of the adaptive filter is adjusted based on the real-time scene characteristics, and the filter coefficients are updated in combination with the allocated filter weights to output the filtered signal.

2. The method according to claim 1, characterized in that, The generation process of the target decay time series includes: Acquire the first ring-down curve corresponding to the detection pulse, the second ring-down curve corresponding to the non-absorption pulse, and the ring-down curve of the reference pulse. The decay time of each decay curve is extracted, and the interference of non-target gas absorption is eliminated by differential calculation of the decay time to generate the target decay time series. The step of performing signal decomposition on the target ring-down time series to extract baseline drift candidate components includes: Perform wavelet decomposition on the target ring-down time series to obtain low-frequency subsequences and high-frequency subsequences; Empirical mode decomposition (EMD) is performed on the low-frequency subsequence to obtain multiple EMD components; Calculate the correlation between each of the empirical mode decomposition components and the industrial interference feature data, and mark the empirical mode decomposition components with a correlation greater than or equal to a preset correlation threshold as the baseline drift candidate components.

3. The method according to claim 2, characterized in that, The multi-dimensional feature data includes physical parameters, industrial interference feature data, and equipment aging feature data; The construction of the hybrid baseline model to predict real-time baseline drift values ​​includes: The baseline drift candidate components and the multi-dimensional feature data are used as input feature sets and input into the support vector regression model for training. The support vector regression model is constrained by a joint loss function that includes the historical true baseline value, the theoretical value of the physical baseline drift, and the self-calibration benchmark value, and the real-time baseline drift value is output.

4. The method according to claim 3, characterized in that, After obtaining the high-frequency subsequence, the process further includes: Calculate the adaptive wavelet threshold based on the electromagnetic interference intensity data in the industrial interference characteristic data; The high-frequency subsequence is filtered using the adaptive wavelet threshold. The sharpness feature value of the first decay curve during a preset time period is extracted, and the current gas component type is determined by combining the sharpness feature value. Based on the determination result, the preset correlation threshold for the next processing cycle is adjusted.

5. The method according to claim 3, characterized in that, The step of calculating the correlation between each of the empirical mode decomposition components and the industrial interference feature data, and marking the empirical mode decomposition components with a correlation greater than or equal to a preset correlation threshold as the baseline drift candidate components, includes: Perform cross-correlation calculations with time-delay search on each of the empirical mode decomposition components and the industrial interference feature data to obtain the maximum alignment correlation coefficient; The empirical mode decomposition components whose maximum alignment correlation coefficient is greater than or equal to the preset correlation threshold are extracted as preliminary interference components. The projection feature values ​​of the initial interference component on the first fading curve and the second fading curve are extracted respectively; Calculate the response change ratio of the projected feature value between the first fading curve and the second fading curve, and when the response change ratio is within a preset background fluctuation range, mark the corresponding preliminary interference component as the baseline drift candidate component.

6. The method according to claim 3, characterized in that, The process of generating the self-calibration reference value includes: When the system operating parameters are detected to reach the preset first trigger condition, the non-absorbed pulse self-calibration mode is executed, the first calibration deviation is calculated, and when the first calibration deviation is less than the preset deviation threshold, the first self-calibration reference value is generated. When the system operating parameters are detected to reach the preset second trigger condition, the standard gas self-calibration mode is executed. The actual decay time is collected by releasing the built-in standard gas, the second calibration deviation is calculated, and a second self-calibration reference value is generated based on the second calibration deviation. The self-calibration reference value includes the first self-calibration reference value and the second self-calibration reference value.

7. The method according to claim 3, characterized in that, The method of constraining the support vector regression model using a joint loss function that includes historical true baseline values, the theoretical value of physical baseline drift, and the self-calibrated benchmark value, and outputting the real-time baseline drift value, includes: Extract the real-time change gradient of each physical parameter, and when the real-time change gradient is greater than a preset limit threshold, determine that the corresponding physical sensor is in a data distortion state and generate a corresponding confidence decay factor. Obtain the update timestamp of the self-calibration benchmark value, calculate the time difference between the current time and the update timestamp, and generate a confidence weight that decreases over time based on the time difference; Using the confidence decay factor and the confidence weight, the weights of the physical constraint terms corresponding to the theoretical value of the physical baseline drift and the weights of the calibration constraint terms corresponding to the self-calibration benchmark value in the joint loss function are dynamically updated. When the weight of the physical constraint term is lower than the preset failure threshold, the prediction mode of the support vector regression model is switched to a dual-path constraint mode dominated by the historical true baseline value and the self-calibration benchmark value.

8. The method according to claim 1, characterized in that, The obtained filtered signal is used for gas concentration inversion, including: The wet concentration of the target gas is calculated based on the filtered signal, the absorption cross section of the target gas, and the ambient ring-down time of the optical cavity. Using the acquired ambient temperature, ambient pressure, water vapor coupling coefficient, and corrosion erosion characteristic parameters, dynamic optical path compensation is performed on the wet concentration to invert and output the dry concentration of the target gas.

9. A system for suppressing ringback time baseline fluctuations based on cavity ringback spectrum, characterized in that, include: The sequence generation module is used to acquire the first ring-down curve corresponding to the detection pulse, the second ring-down curve corresponding to the non-absorption pulse, and the ring-down curve of the reference pulse. The decay time of each decay curve is extracted, and the interference of non-target gas absorption is eliminated by differential calculation of the decay time to generate the target decay time series. The signal decomposition module is used to decompose the target decay time series in response to the target decay time series and extract baseline drift candidate components; specifically, it is used to perform multi-scale mode decomposition on the target decay time series to obtain multiple decomposed components. The correlation between each of the decomposed components and the synchronously acquired industrial interference feature data is calculated, and the baseline drift candidate components are extracted based on the correlation. The baseline prediction module is used to construct a hybrid baseline model based on the baseline drift candidate components and synchronously acquired multi-dimensional feature data to predict the real-time baseline drift value. The hybrid baseline model includes physical baseline theoretical constraints, data-driven prediction constraints, and self-calibration benchmark constraints. Specifically, the physical baseline theoretical constraints are calculated using synchronously acquired physical parameters; the baseline drift candidate components and multi-dimensional feature data are input into the data prediction model to generate the data-driven prediction constraints; and the self-calibration benchmark constraints are obtained through a self-calibration mode. The real-time baseline drift value is predicted and output by combining the above three constraints. The filtering and output module is used to initiate a dynamic weighted adaptive filtering process based on the real-time baseline drift value to suppress baseline fluctuations in the target oscillation time series, and obtain a filtered signal for gas concentration inversion, specifically for: Extract real-time scene features of the current detection scene, including at least baseline fluctuation intensity and signal-to-noise ratio; dynamically allocate prediction weights of the hybrid baseline model and filtering weights of the adaptive filter based on the real-time scene features; adjust the step size of the adaptive filter based on the real-time scene features, and update the filtering coefficients in combination with the allocated filtering weights to output the filtered signal.

Images