Ocean carbon sink data acquisition and real-time analysis system

Abstract

The application discloses a marine carbon sink data acquisition and real-time analysis system, obtains non-equidistant environmental parameter time series data through a sensor and forms original data, generates pretreatment time series based on curvature constraint smoothing and stability decomposition, divides time windows according to fixed steps or adaptive steps to obtain segmented sequences, generates periodic terms in a periodic detection line and obtains deperiodic residuals, recursively calculates residuals in a mutation detection line and outputs change statistics, performs closed-loop fusion on the periodic terms, the deperiodic residuals and the change statistics, dynamically adjusts smoothing parameters and drift correction windows, and generates updated time series, generates state information according to the updated time series and the change statistics and discriminates normal, suspicious or abnormal, triggers model re-calibration when discriminating as abnormal, and outputs the updated time series, the state information and the periodic terms to an edge node or a shore-based platform, so that real-time analysis on marine carbon sink data is realized.

Description

Technical Field

[0001] This invention relates to the field of marine environmental monitoring and data processing technology, and in particular to a marine carbon sink data acquisition and real-time analysis system. Background Technology

[0002] With the increasing demand for marine carbon sink monitoring, real-time sensor acquisition and edge analysis are gradually becoming important means of marine environmental data processing. Currently, monitoring platforms typically use multi-source sensors to collect parameters such as carbon dioxide concentration, pH, dissolved oxygen, and salinity, and preprocess them using conventional smoothing or filtering methods. Then, windowing segmentation and frequency analysis methods are used to identify periods and detect anomalies in the data.

[0003] Existing technologies have limitations in non-uniformly spaced sampling, noise drift separation, and multi-scale dynamic monitoring. On the one hand, traditional filtering methods lack mechanisms for dynamic adjustment of second-order curvature constraints and smoothing tension, making it difficult to simultaneously suppress drift, random walks, and white noise, resulting in insufficient stability of the preprocessed sequence. On the other hand, common window partitioning methods are mostly based on fixed step sizes, which are difficult to adapt to the non-stationary characteristics of data streams, affecting the accuracy of periodic and abrupt change detection. In addition, existing periodic and anomaly detection methods mostly operate independently, lacking a closed-loop feedback mechanism between the periodic term and the residual, making it impossible to achieve adaptive parameter updates and prone to frequency estimation bias or abrupt change identification lag. Furthermore, existing methods rely heavily on a single threshold for state discrimination, lacking a hierarchical judgment of normal, suspicious, and abnormal states, making it difficult to achieve dynamic recalibration and stable and reliable data output in complex marine environments.

[0004] Therefore, how to provide a system for the acquisition and real-time analysis of marine carbon sink data is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a marine carbon sink data acquisition and real-time analysis system. This invention utilizes sensor data acquisition, second-order curvature-constrained smoothing, Allan variance decomposition, windowed management, period detection based on weighted Lomb-Scargle and mutation detection based on Shiryaev-Roberts, as well as closed-loop fusion and state discrimination mechanisms, which have the advantages of high data processing stability, high anomaly identification accuracy and strong real-time performance.

[0006] The marine carbon sink data acquisition and real-time analysis system according to an embodiment of the present invention includes the following steps:

[0007] The data acquisition module acquires time-series data of non-uniformly spaced environmental parameters generated by sensors in the marine monitoring platform and outputs raw environmental parameter data.

[0008] The steady-state filtering model module receives raw environmental parameter data and outputs robust preprocessed time series data.

[0009] The window management module receives robust preprocessed time series data, divides it into time windows, and outputs segmented time series data.

[0010] The bistable detection model module receives segmented time series data, generates periodic terms and deperiodic residuals in the periodic detection line, and generates change statistics in the abrupt change detection line.

[0011] The closed-loop fusion module receives the periodic term, deperiodized residual, and change statistics. It inputs the deperiodized residual into the mutation detection line in the bistable detection model module, feeds back the change statistics to the stable-variable filtering model module, and outputs updated time series data.

[0012] The state determination module receives updated time series data and change statistics, and generates state information based on the input information;

[0013] The output module receives updated time series data, status information, and periodic items, and transmits the updated time series data, status information, and periodic items to edge computing nodes or shore-based analysis platforms.

[0014] Optionally, modules can be integrated using the following methods:

[0015] The data acquisition module is used to acquire time series data of non-uniformly spaced environmental parameters generated by sensors in the marine monitoring platform, and the raw environmental parameter data is output.

[0016] The module uses a steady-state filtering model to receive raw environmental parameter data, performs second-order curvature constraint smoothing based on von Draco filtering, and combines Allan variance to decompose drift components, random walk components and white noise components at a preset time scale, outputting robust preprocessed time series data.

[0017] The window management module is used to receive robust preprocessed time series data, divides it into time windows according to a preset step size, and outputs segmented time series data.

[0018] The bistable detection model module receives segmented time series data. In the periodic detection line of the bistable detection model module, the weighted Lomb-Scargle method is used to generate periodic terms and deperiodic residuals. In the mutation detection line of the bistable detection model module, the deperiodic residuals are recursively calculated based on the Shiryaev-Roberts method and the change statistics are output.

[0019] The closed-loop fusion module receives the periodic term, deperiodic residual, and change statistics. The deperiodic residual is input into the mutation detection line of the bistable detection model module, and the change statistics are fed back to the stable-variable filtering model module to adjust the smoothing parameters and drift correction window of the von Draco filter, and the updated time series data is output.

[0020] The state discrimination module receives updated time series data and change statistics, generates state information, and triggers parameter recalibration of the periodic detection line and abrupt change detection line of the bistable detection model module when the state information is an abnormal state.

[0021] The output module receives updated time series data, status information, and periodic items, and transmits these data to edge computing nodes or shore-based analysis platforms.

[0022] Optionally, generating raw environmental parameter data includes:

[0023] The data acquisition module is used to set up acquisition tasks in the marine monitoring platform and send acquisition instructions to the carbon dioxide sensor, pH sensor, dissolved oxygen sensor and salinity sensor deployed in the monitoring area. Each acquisition instruction includes a sampling timestamp, sampling parameter type and sampling accuracy parameter.

[0024] Receive response data returned by various sensors. The response data includes sensor identifier, sampling time point, parameter type and observation value, forming an observation data triplet.

[0025] All collected observation data triplets are arranged in chronological order and classified according to the sensor channels corresponding to the parameter types to generate non-equally spaced single-channel initial observation sequences.

[0026] For each type of parameter observation sequence, a set of timestamps is recorded and compared with a preset reference sampling time axis to generate corresponding time interval offset information;

[0027] All classified non-equal interval single-channel initial observation sequences and their corresponding time interval offset information are encapsulated into structured data objects to form raw environmental parameter data.

[0028] Optional, stable filtering models include:

[0029] Receive raw environmental parameter data and define the input time series as X;

[0030] Based on the input time series X, a second-order curvature-constrained smoothing objective function is constructed, and the smoothing function is set as follows: The objective function is

[0031] ;

[0032] The first term is the sum of squared residuals, the second term is the curvature constraint integral, and λ is the smoothing tension parameter.

[0033] For the objective function Numerical solutions are performed to obtain a continuous smooth function. and at each timestamp t i Calculate the smoothing value and generate a smooth sequence. ;

[0034] Smooth sequence The window is divided into M windows according to a preset step size, and each window is denoted as W. k ;

[0035] For each window W k Calculate the mean to obtain the window mean sequence. And calculate the Allan variance with time scale τ based on the window mean sequence:

[0036] ;

[0037] in, This represents the difference between the means of adjacent windows;

[0038] Based on the slope characteristics of the logarithmic curves of Allan variance under different τ, the windows are divided into three categories: when the slope is -1, the window is marked as the white noise region; when the slope is +1, the window is marked as the random walk region; and when the slope is +2, the window is marked as the drift region.

[0039] For windows belonging to the drift region, a linear trend function is used for fitting, and the trend function is subtracted from the smoothed values ​​within the window to generate the drift-corrected sequence segment.

[0040] For windows that belong to the random walk region, calculate the mean shift. And subtract this drift amount from all smoothed values ​​within the window to complete the random walk compensation;

[0041] For windows belonging to the white noise region, no numerical adjustment is made; the original smoothing value is retained.

[0042] The windows, after drift correction, random walk compensation, and adjustment of the white noise region, are spliced ​​together to generate the update sequence. ;

[0043] Update sequence The output is robust preprocessed time series data.

[0044] Optionally, the window management module includes:

[0045] Receive robust preprocessed time series data ;

[0046] Set the time window step size to a constant δ t In the time interval [t1, t] n Construct continuous segmented intervals within the [framework], and define the k-th time window as [the time window]. T k =t1+(k-1)δ t And k=1,2,…,m,m=[(t n -t1) / δ t ];

[0047] W in each time window k Inside, collect all conditions T that are met. k ≤t j <T k +δ t smoothing value The collection is then stored as window data without performing default filling or interpolation.

[0048] When a certain interval [T] k ,T k +δ t If no timestamp meets the condition, the window management module directly generates an empty window. And keep the empty window occupied in the sequence;

[0049] Arrange all window data sequentially to form a window sequence, and output the window sequence as segmented time series data.

[0050] Optionally, the steps for generating the periodic term, deperiodic residuals, and change statistics include:

[0051] Receive the output segmented time series data and set the window sequence;

[0052] The window sequence is passed to the periodic detection line to generate a periodic term sequence;

[0053] Calculate the point-to-point difference between the periodic term sequence and the input sequence to generate a deperiodic residual sequence;

[0054] The mutation detection line receives the deperiodic residual sequence, performs recursive processing, and outputs a sequence of change statistics.

[0055] The periodic term sequence, deperiodic residual sequence, and change statistic sequence are used as the outputs of the bistable detection model module and transmitted to the closed-loop fusion module.

[0056] Optionally, the bistable detection model module includes:

[0057] Periodic testing line:

[0058] Receive output window sequence And read all observation points in the window in sequence;

[0059] Calculate the interval Δt between adjacent observation points k,j =t k,j -t k,j-1 Assign weight w to each point k,j =1 / Δt k,j We obtain the weighted time series;

[0060] Define the candidate frequency set Ω = ω1, ω2, ..., ω k For each frequency ω∈Ω, perform the following steps:

[0061] Traverse all observation points in the window and calculate the weighted cosine component c point by point. k,j =w k,j v k,j cos(ωt k,j ), and sum them up to get C(ω)=∑ j c k,j ;

[0062] Traverse all observation points in the window and calculate the weighted sine component s point by point. k,j =w k,j v k,j sin(ωt k,j ), and sum them up to get S(ω)=∑ j s k,j ;

[0063] Calculate the normalized denominator, including ∑ j w k,j cos 2 (ωt k,j ),∑ j w k,j sin 2 (ωt k,j );

[0064] Based on the numerator and denominator above, calculate the power spectrum value:

[0065] ;

[0066] The results are then stored in the power spectrum table.

[0067] After calculating for all candidate frequencies, the power spectrum is traversed, and the frequency ω with the highest power value is selected. * k , which is the dominant frequency of this window;

[0068] At the dominant frequency ω * k Next, a least squares equation is established, and the observed values ​​are substituted point by point to solve for the amplitude and phase parameters, generating a fitting function. ,

[0069] Each timestamp t within the window k,j Substitute the fitted function, calculate the period estimate point by point, and generate a periodic term sequence;

[0070] The residuals of the generated periodic term sequence are calculated point by point to obtain the residual sequence, and then compared with the dominant frequency ω in the frequency domain. * k Nearby components are subjected to amplitude limiting suppression, and point-by-point replacement is used to obtain a deperiodic residual sequence;

[0071] Mutation detection line:

[0072] Receive the deperiodic residual sequence output from the periodic detection line;

[0073] Initialize the Shiryaev–Roberts statistic and set an initial value R. k,0 =0;

[0074] Before the change, the model f0 is assumed to have a mean of zero and a variance of σ. 2 The model is a Gaussian distribution, and the changed model f1 is given by mean shift μ and variance σ. 2 Gaussian distribution;

[0075] Read the residual r of the periodic residual sequence point by point ' k,j Calculate the likelihood ratio: And recursively update the statistics: ;

[0076] Calculate the time interval Δt between this point and the previous sampling point. k,j And in conjunction with the global average sampling interval Δt, the correction factor ψ(Δt) is calculated. k,j )=Δt k,j / Δt;

[0077] Execution decision condition: If R k,j ·ψ(Δt k,j If ) > A, then time point t k,j Mark them as mutation points and record the statistical values;

[0078] Recursively iterate through all points within the window, saving the results point by point to form a sequence of change statistics Q. k =R k,j ;

[0079] Output: The periodic sequences of each window are concatenated in chronological order to form a global periodic sequence; the deperiodic residual sequences of each window are concatenated in chronological order to form a global deperiodic residual sequence; the change statistics sequences of each window are concatenated in chronological order to form a global change statistics sequence; the periodic sequence, deperiodic residual sequence, and change statistics sequence are combined as the output of the bistable detection model module and transmitted to the closed-loop fusion module.

[0080] Optionally, the closed-loop fusion module performs the following:

[0081] Receive the output periodic term sequence, deperiodic residual sequence, and change statistic sequence, and sort them by timestamp t. j Alignment to form a synchronized data stream;

[0082] The deperiodic residual sequence R is input point by point into the mutation detection line of the bistable detection model module in chronological order, and recursive calculation is performed to generate an updated sequence of change statistics.

[0083] Update the change statistics sequence Q ‘ The input change statistic sequence Q is compared point by point at time, the differences are calculated, and an index set is built. ;

[0084] For sets Each time point t in j Perform the following steps:

[0085] In the interval [t] j -L,t j +L] extracts the smoothed value of the stable filter output. Forming local subsequences Where L is the preset time window length;

[0086] Calculate the mean square error between the local subsequence and the original observed sequence in the same interval, and then apply the mean square error and a threshold. The ratio is updated to smooth the parameter: if the error is greater than Let λ j =λ·(1+η), where η is the proportionality coefficient; otherwise, keep λ. j =λ;

[0087] The interval [t] j -L,t j +L] is set as the drift correction window, and the second-order curvature-constrained smoothing objective function is reconstructed. and use the updated λ j Perform numerical solutions on this interval;

[0088] Generate a set of corrected smooth values ​​within the drift correction window. and replace the corresponding part in the original smooth sequence;

[0089] The results corrected by all index points are concatenated with the unadjusted segments to generate the updated time series. ;

[0090] Update the time series As the output of the closed-loop fusion module.

[0091] Optionally, the execution steps of the state determination module include:

[0092] Receive the updated time series data and the corresponding change statistics sequence from the output, align them according to the timestamp order, and establish a point-by-point corresponding input set;

[0093] At each time point, the difference between the smoothed value at that time point and the previous time point is calculated to obtain the local fluctuation amount, and the local fluctuation amount and the change statistics at that time point are used together to form a discriminant vector.

[0094] For each discrimination vector, compare the local fluctuation with the first threshold, and compare the change statistic with the second threshold. When both are below their respective thresholds, it is marked as a normal state; when one of them exceeds the threshold, it is marked as a suspicious state; when both exceed the threshold, it is marked as an abnormal state.

[0095] Count consecutive suspicious and abnormal states. When the consecutive count exceeds a preset threshold, the state information for that time period is confirmed as an abnormal state.

[0096] When the status information is confirmed to be an abnormal state, a trigger signal is generated and transmitted to the bistable detection model module, which instructs the periodic detection line to recalculate the frequency fitting parameters and simultaneously instructs the mutation detection line to reinitialize the initial value of the statistics and the discrimination threshold.

[0097] After completing the above operations, output a sequence of status information arranged in chronological order. The status information is limited to three categories: normal, suspicious, and abnormal. The status information sequence is then transmitted to the output module. Beneficial effects

[0098] (1) This invention combines second-order curvature constraint smoothing with Allan variance to decompose and compensate for drift, random walk and white noise for non-equal interval sampling. It also combines fixed or adaptive step size window management to ensure consistency between local and global analysis, thereby suppressing over-smoothing and noise residue while preserving real trend and periodic information, and improving the stability and usability of preprocessing.

[0099] (2) The present invention adopts a bistable detection architecture: based on weighted Lomb-Scargle, periodic detection is performed, the tidal and day-night frequencies are extracted and the deperiodic residuals are obtained through frequency domain amplitude limiting suppression; based on Shiryaev-Roberts, the residuals are recursively statistically analyzed to realize mutation location, adapt to non-equal interval sampling conditions, and improve the accuracy and timeliness of periodic identification and anomaly detection.

[0100] (3) The present invention establishes a closed-loop fusion mechanism, feeds back the change statistics to adaptively adjust the smoothing tension coefficient and drift correction window, and performs online recalibration on the periodic detection and mutation detection parameters to reduce false alarms and false negatives caused by parameter drift and enhance the self-correction and robustness of the processing link. Attached Figure Description

[0101] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0102] Figure 1 This is a flowchart of the marine carbon sink data acquisition and real-time analysis system proposed in this invention;

[0103] Figure 2 The flowchart of the bistable detection model proposed in this invention for processing periodic detection lines and abrupt change detection lines is shown. Detailed Implementation

[0104] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0105] refer to Figure 1-2 The marine carbon sequestration data acquisition and real-time analysis system includes the following steps:

[0106] The data acquisition module acquires time-series data of non-uniformly spaced environmental parameters generated by sensors in the marine monitoring platform and outputs raw environmental parameter data.

[0107] The steady-state filtering model module receives raw environmental parameter data, performs second-order curvature constraint smoothing based on von Draco filtering, and combines Allan variance decomposition of drift components, random walk components and white noise components to output robust preprocessed time series data.

[0108] The window management module receives robust preprocessed time series data, divides it into time windows according to a fixed step size or an adaptive step size, and outputs segmented time series data.

[0109] The bistable detection model module receives segmented time series data, generates periodic terms and deperiodic residuals based on the weighted Lomb–Scargle method in the periodic detection line, and performs recursive calculations on the deperiodic residuals based on the Shiryaev–Roberts method in the mutation detection line and outputs change statistics.

[0110] The closed-loop fusion module receives the periodic term, deperiodic residual, and change statistics. It inputs the deperiodic residual into the mutation detection line in the bistable detection model module, feeds back the change statistics to the stable-variable filtering model module to adjust the smoothing parameters and drift correction window of the von Draco filter, and outputs updated time series data.

[0111] The state discrimination module receives updated time series data and change statistics, and generates state information based on the input information. When the state information is an abnormal state, it triggers the parameter recalibration of the periodic detection line and the mutation detection line in the bistable detection model module.

[0112] The output module receives updated time series data, status information, and periodic items, and transmits the updated time series data, status information, and periodic items to edge computing nodes or shore-based analysis platforms.

[0113] In this embodiment, the modules are connected through the following method:

[0114] The data acquisition module is used to acquire time series data of non-uniformly spaced environmental parameters generated by sensors in the marine monitoring platform, and the raw environmental parameter data is output.

[0115] The module uses a steady-state filtering model to receive raw environmental parameter data, performs second-order curvature constraint smoothing based on von Draco filtering, and combines Allan variance to decompose drift components, random walk components and white noise components at a preset time scale, outputting robust preprocessed time series data.

[0116] The window management module is used to receive robust preprocessed time series data, divides it into time windows according to a preset step size, and outputs segmented time series data.

[0117] The bistable detection model module receives segmented time series data. In the periodic detection line of the bistable detection model module, the weighted Lomb-Scargle method is used to generate periodic terms and deperiodic residuals. In the mutation detection line of the bistable detection model module, the deperiodic residuals are recursively calculated based on the Shiryaev-Roberts method and the change statistics are output.

[0118] The closed-loop fusion module receives the periodic term, deperiodic residual, and change statistics. The deperiodic residual is input into the mutation detection line of the bistable detection model module, and the change statistics are fed back to the stable-variable filtering model module to adjust the smoothing parameters and drift correction window of the von Draco filter, and the updated time series data is output.

[0119] The state discrimination module receives updated time series data and change statistics, generates state information, and triggers parameter recalibration of the periodic detection line and abrupt change detection line of the bistable detection model module when the state information is an abnormal state.

[0120] The output module receives updated time series data, status information, and periodic items, and transmits these data to edge computing nodes or shore-based analysis platforms.

[0121] In this embodiment, generating the original environmental parameter data includes:

[0122] The data acquisition module is used to set up acquisition tasks in the marine monitoring platform and send acquisition instructions to the carbon dioxide sensor, pH sensor, dissolved oxygen sensor and salinity sensor deployed in the monitoring area. Each acquisition instruction includes a sampling timestamp, sampling parameter type and sampling accuracy parameter.

[0123] Receive response data returned by various sensors. The response data includes sensor identifier, sampling time point, parameter type, and observed value, forming an observation data triplet:

[0124] All collected observation data triplets are arranged in chronological order and classified according to the sensor channels corresponding to the parameter types to obtain non-equally spaced single-channel initial observation sequences;

[0125] For each type of parameter observation sequence, a set of timestamps is recorded and compared with a preset reference sampling time axis to generate corresponding time interval offset information;

[0126] All classified non-equal interval single-channel initial observation sequences and their corresponding time interval offset information are encapsulated into structured data objects to form raw environmental parameter data, which are then output to the steady-state filtering model module.

[0127] In this embodiment, the stable filtering model includes:

[0128] The steady-state filtering model module is used to receive the raw environmental parameter data, and the input time series is defined as X={(t1,v1),(t2,v2),…,(t…}. n ,v n )}, where t i Represents a timestamp, v i This represents the observation corresponding to the timestamp, where n represents the number of data points;

[0129] Based on the input time series X, a second-order curvature-constrained smoothing objective function is constructed, and the smoothing function is set as follows: The objective function is

[0130] ;

[0131] The first term is the sum of squared residuals, which is used to ensure that the smoothing function is close to the original observation data; the second term is the curvature constraint integral, which is used to control the second derivative of the smoothing function from producing excessive oscillations; λ is the smoothing tension parameter, which adjusts the trade-off between the two parts.

[0132] For the objective function Numerical solutions are performed to obtain a continuous smooth function. and at each timestamp t i Calculate the smoothing value and generate a smooth sequence. This serves as a smoothed output of the original time series.

[0133] Smooth sequence The system is divided into M windows according to a preset step size, and each window is denoted as M. , where m k The number of data points within the window;

[0134] For each window W k Calculate the mean to obtain the window mean sequence. And calculate the Allan variance with time scale τ based on the sequence:

[0135] ;

[0136] in, The difference between the means of adjacent windows reflects the stability fluctuation of the sequence at the time scale τ;

[0137] Based on the slope characteristics of the logarithmic curves of Allan variance under different τ, the windows are divided into three categories: when the slope is -1, the window is marked as the white noise region; when the slope is +1, the window is marked as the random walk region; and when the slope is +2, the window is marked as the drift region.

[0138] For windows belonging to the drift region, a linear trend function is used for fitting:

[0139] v trend,k (t)=a k ·t+b k

[0140] Then, the trend function is subtracted from the smoothed value within the window to obtain the drift-corrected sequence segment;

[0141] For windows that belong to the random walk region, calculate the mean shift. And subtract this drift amount from all smoothed values ​​within the window to complete the random walk compensation;

[0142] For windows belonging to the white noise region, no numerical adjustment is made; the original smoothing value is retained.

[0143] The windows, after drift correction, random walk compensation, and adjustment of the white noise region, are concatenated to generate the update sequence:

[0144] ;

[0145] Update sequence The output is robust preprocessed time series data, which is then transmitted to the window management module.

[0146] In this embodiment, the continuous smoothing function It is defined in the input timestamp interval [t1, t2]. n A real-valued function on [a_t] is used to fit and smooth the observations in the original time series; this function satisfies the following two conditions:

[0147] (1) In the closed interval [t1,t] n It has at least second-order continuous differentiability, that is... This ensures that its second derivative has mathematical meaning in the curvature constraint term and can be calculated by integration;

[0148] (2) Satisfying boundary conditions or endpoint tension penalty rules (choosing Dirichlet or Neumann boundaries depending on the specific algorithm implementation), and obtained by solving the objective functional composed of weighted least squares and curvature penalty terms, its form is as follows:

[0149] ;

[0150] In numerical implementation, the smoothing function can be constructed using natural splines, von Draco kernel interpolation, or a pseudo-inverse method based on the difference matrix, ensuring a balance between fitting residuals and global smoothness at non-equally spaced sampling points.

[0151] The von Draco filter objective function used originates from variational smoothing methods with second-order derivative constraints in mathematics, originally used to handle curve fitting problems of irregular time series. This application introduces a non-equal-interval sampling structure and dynamically adjusts the smoothing tension coefficient based on the subsequent Allan variance drift identification results, forming an improved smoothing model with a feedback mechanism. The residual term in the objective function... With curvature penalty These terms, representing the square of the observed value and the square of the observed value divided by the fourth power of time, respectively, maintain the same dimensionality as the fitted residual term after integration. λ, in a physical sense, is an adjustment coefficient for the cubic dimension of time, thus ensuring consistency in the physical dimensions of both sides of the formula. In the Allan variance part, the formula used is derived from time-frequency stability theory, describing the variance characteristics of the sampled mean sequence at a given time scale τ. Its dimension is the square of the observed value, consistent with the input data. This application uses this statistic to identify drift components, random walk components, and white noise components in the smoothing result, and then uses the trend fitting function v trend,k (t)=a k ·t+b k The drift segment is subjected to linear detrending processing. All introduced formulas are derived and applied based on existing physical and mathematical theories, and the uniformity of dimensions, logical closed loop, and engineering feasibility are ensured.

[0152] In this embodiment, the window management module includes:

[0153] The window management module is used to receive robust preprocessed time series data, and the series is set as follows: ,in, For timestamp t i The corrected smoothing value is given below, where n is the number of data points;

[0154] Set the time window step size to a constant δ t In the time interval [t1, t] n Construct continuous segmented intervals within the [framework], and define the k-th time window as [the time window]. T k =t1+(k-1)δ t And k=1,2,…,m,m=[(t n -t1) / δ t ];

[0155] W in each time window k Inside, collect all conditions T that are met. k ≤t j <T k +δ t smoothing value The collection is then stored as window data without performing default filling or interpolation.

[0156] When a certain interval [T] k ,T k +δ t If no timestamp meets the conditions, the window management module will directly generate an empty window. And keep the empty window occupied in the sequence;

[0157] Arrange all window data sequentially to form a window sequence {W1, W2, ..., W...} m};

[0158] The window sequence is then output as segmented time series data to the bistable detection model module.

[0159] In this embodiment, the steps for generating periodic terms, deperiodic residuals, and change statistics include:

[0160] The bistable detection model module is used to receive the output segmented time series data and set the window sequence.

[0161] The window sequence is passed to the periodic detection line of the bistable detection model module, and the periodic detection line generates a periodic term sequence.

[0162] Calculate the point-to-point difference between the periodic term sequence and the input sequence to generate a deperiodized residual sequence, and use this residual sequence as the input to the mutation detection line;

[0163] The mutation detection line of the bistable detection model module receives the residual sequence, performs recursive processing on it, and outputs a sequence of change statistics.

[0164] The periodic term sequence, deperiodic residual sequence, and change statistic sequence are used as the outputs of the bistable detection model module and transmitted to the closed-loop fusion module.

[0165] In this embodiment, the bistable detection model module includes:

[0166] Periodic testing line:

[0167] Receive output window sequence And read all observation points in the window in sequence;

[0168] Calculate the interval Δt between adjacent observation points k,j =t k,j -t k,j-1 Assign weight w to each point k,j =1 / Δt k,j We obtain the weighted time series;

[0169] Define the candidate frequency set Ω = ω1, ω2, ..., ω k For each frequency ω∈Ω, perform the following steps:

[0170] Traverse all observation points in the window and calculate the weighted cosine component c point by point. k,j =w k,j v k,j cos(ωt k,j ), and sum them up to get C(ω)=∑ j c k,j ;

[0171] Traverse all observation points in the window and calculate the weighted sine component s point by point. k,j =w k,j v k,j sin(ωt k,j ), and sum them up to get S(ω)=∑ j s k,j ;

[0172] Calculate the normalized denominator, including ∑ j w k,j cos 2 (ωt k,j ),∑ j w k,j sin 2 (ωt k,j );

[0173] Based on the numerator and denominator above, calculate the power spectrum value:

[0174] ;

[0175] The results are then stored in the power spectrum table.

[0176] After calculating for all candidate frequencies, the power spectrum is traversed, and the frequency ω with the highest power value is selected. * k , which is the dominant frequency of this window;

[0177] At the dominant frequency ω * k Next, a least squares equation is established, and the observed values ​​are substituted point by point to solve for the amplitude and phase parameters, generating a fitting function. ,

[0178] Each timestamp t within the window k,j Substitute the fitted function, calculate the period estimate point by point, and generate a periodic term sequence;

[0179] The residuals of the generated periodic term sequence are calculated point by point to obtain the residual sequence, and then compared with the dominant frequency ω in the frequency domain. * k Nearby components are subjected to amplitude limiting suppression, and point-by-point replacement is used to obtain a deperiodic residual sequence;

[0180] Mutation detection line:

[0181] Receive the deperiodic residual sequence output from the periodic detection line;

[0182] Initialize the Shiryaev–Roberts statistic and set an initial value R. k,0 =0;

[0183] Before the change, the model f0 is assumed to have a mean of zero and a variance of σ.2 The model is a Gaussian distribution, and the changed model f1 is given by mean shift μ and variance σ. 2 Gaussian distribution;

[0184] Read the residual r of the periodic residual sequence point by point ' k,j Calculate the likelihood ratio: And recursively update the statistics: ;

[0185] Calculate the time interval Δt between this point and the previous sampling point. k,j And in conjunction with the global average sampling interval Δt, the correction factor ψ(Δt) is calculated. k,j )=Δt k,j / Δt;

[0186] Execution decision condition: If R k,j ·ψ(Δt k,j If ) > A, then time point t k,j Mark them as mutation points and record the statistical values;

[0187] Recursively iterate through all points within the window, saving the results point by point to form a sequence of change statistics Q. k =R k,j ;

[0188] Output: The periodic sequences of each window are concatenated in chronological order to form a global periodic sequence; the deperiodic residual sequences of each window are concatenated in chronological order to form a global deperiodic residual sequence; the change statistics sequences of each window are concatenated in chronological order to form a global change statistics sequence; the periodic sequence, deperiodic residual sequence, and change statistics sequence are combined as the output of the bistable detection model module and transmitted to the closed-loop fusion module.

[0189] In this embodiment, the closed-loop fusion module performs the following:

[0190] It receives the periodic term sequence, the deperiodic residual sequence, and the change statistics sequence, and aligns them by timestamp to form a synchronous data stream;

[0191] The deperiodic residual sequence is input point by point into the mutation detection line of the bistable detection model module in chronological order, and recursive calculation is performed to generate an updated sequence of change statistics.

[0192] Update the change statistics sequence Q ‘ The input change statistic sequence Q is compared point by point at time, the differences are calculated, and an index set is built. ;

[0193] For sets Each time point t in j Perform the following steps:

[0194] In the interval [t] j -L,t j +L] extracts the smoothed value of the stable filter output. Forming local subsequences Where L is the preset time window length;

[0195] Calculate the mean square error between the local subsequence and the original observed sequence in the same interval, and then apply the mean square error and a threshold. The ratio is updated to smooth the parameter: if the error is greater than Let λ j =λ·(1+η), where η is the proportionality coefficient; otherwise, keep λ. j =λ;

[0196] The interval [t] j -L,t j +L] is set as the drift correction window, and the second-order curvature-constrained smoothing objective function is reconstructed. and use the updated λ j Perform numerical solutions on this interval;

[0197] Generate a set of corrected smooth values ​​within the drift correction window. and replace the corresponding part in the original smooth sequence;

[0198] The results corrected by all index points are concatenated with the unadjusted segments to generate the updated time series. ;

[0199] Update the time series As the output of the closed-loop fusion module, it is transmitted to the state discrimination module.

[0200] In this embodiment, the execution steps of the state determination module include:

[0201] Receive updated time series data and corresponding change statistics sequences, align them according to timestamp order, and establish a point-by-point corresponding input set;

[0202] At each time point, the difference between the smoothed value at that time point and the previous time point is calculated to obtain the local fluctuation amount, and the local fluctuation amount and the change statistics at that time point are used together to form a discriminant vector.

[0203] For each discrimination vector, compare the local fluctuation with the first threshold, and compare the change statistic with the second threshold. When both are below their respective thresholds, it is marked as a normal state; when one of them exceeds the threshold, it is marked as a suspicious state; when both exceed the threshold, it is marked as an abnormal state.

[0204] Count consecutive suspicious and abnormal states. When the consecutive count exceeds a preset threshold, the state information for that time period is confirmed as an abnormal state.

[0205] When the status information is confirmed to be an abnormal state, a trigger signal is generated and transmitted to the bistable detection model module. This signal is used to recalculate the frequency fitting parameters of the command cycle detection line and simultaneously reinitialize the initial value of the statistics and the discrimination threshold of the command mutation detection line.

[0206] After completing the above operations, output a sequence of status information arranged in chronological order. The status information is limited to three categories: normal, suspicious, and abnormal. This sequence is then transmitted to the output module.

[0207] Example:

[0208] To verify the engineering applicability and stability of the marine carbon sink data acquisition and real-time analysis system proposed in this invention in complex marine environments, a nearshore sea area was selected as a pilot area, and dynamic monitoring of blue carbon capacity was carried out based on a marine ecological observation station. This area is a typical tidal dynamic-dominated shallow sea ecosystem, and marine environmental parameters (such as dissolved oxygen, carbon dioxide, salinity, and pH) exhibit diurnal periodic fluctuations and high-frequency abrupt changes, making it a key area for blue carbon research for a long time.

[0209] In traditional monitoring schemes, the data collected by the observation platform has significant non-uniform intervals and multi-source heterogeneity. For example, due to power consumption limitations, the CO2 sensor can only collect a maximum of 80 records per day, while salinity and pH data are subject to large fluctuations due to hydrodynamic interference, and are prone to drift and jumps. This results in poor performance of processing methods based on average filtering and smoothing curves, often causing periodic identification failures and abnormal response delays.

[0210] To address the aforementioned issues, the observation station adopted the complete system proposed in this invention, comprising a data acquisition module, a stable filtering model module, a window management module, a bistable detection model module, a closed-loop fusion module, a state discrimination module, and an output module. All modules are embedded in the buoy platform using an edge computing architecture, and the results are synchronized in real time to a shore-based server for long-term archiving and display.

[0211] In the specific implementation process, the data acquisition module is first configured with a non-uniformly spaced sampling structure. CO2, pH, DO (dissolved oxygen), and salinity sensors collect environmental data according to the set tasks. Each record includes information such as timestamp, sampling parameters, accuracy configuration, and channel number, forming raw triplet observation data. After format standardization and time alignment, this data is used to construct the initial non-uniformly spaced time series.

[0212] After entering the data preprocessing stage, the steady-state filtering model module of this invention is invoked, and a von Draco filter is used to complete the second-order curvature constraint smoothing calculation under non-uniform sampling points. The smoothing tension parameter is initialized to λ0=0.4, and the continuous time series function is obtained by jointly solving the fitting residual term and curvature penalty term of the objective function. Based on this, according to the Allan variance theory model, window partitioning and variance calculation are performed on the smoothed sequence to extract the slope of change at each time scale, thereby automatically identifying the drift segment, white noise segment, and random walk segment, and applying three operations in the corresponding segment: linear detrending, mean shift adjustment, or original value preservation, finally outputting robust smoothed time series data.

[0213] The data is processed by the window management module, which slices it into windows at a fixed 5-minute interval. Each window segment retains all non-interpolated data for subsequent periodic and mutation joint analysis. During processing, when there is a large amount of missing data in some windows, the module retains empty window labels to prevent interpolation from introducing false signals.

[0214] Periodicity identification and abrupt change detection are performed in parallel by the bistable detection model module. The periodicity detection line uses a weighted Lomb-Scargle spectral analysis method to extract the dominant frequency from the non-equally spaced time series within each window, extracting tidal and diurnal components. The frequency set range is set to 0.05~0.5Hz, and the power spectrum peaks are used to fit the optimal sine and cosine periodic terms, thereby generating a periodic term sequence. The difference between the periodic terms and the actual observations forms the deperiodized residual data, which is then fed into the abrupt change detection line. This line internally calls the Shiryaev-Roberts statistical recursive structure, setting the initial model to N(0,σ0). 2 The mutated model is N(μ1,σ1). 2 The system updates statistics via online likelihood ratio and outputs abrupt change signals based on a dynamic threshold. In actual settings, the drift identification window length is T=15 minutes, and the mutation identification delay does not exceed 3 sampling points.

[0215] The closed-loop fusion module, as one of the key technologies of this invention, receives information on periodic terms, deperiodic residuals, and mutation statistics, and aligns them according to timestamps. Upon identifying a mutation point, the module automatically enters a parameter recalibration mechanism, calculating a new value for the smoothing tension λ by combining the global mean square error and the current local error. Simultaneously, it reconstructs the objective function and performs local second-order smoothing within the identification window. This corrected result replaces the original result after splicing, forming the final updated time series.

[0216] To enhance the system's autonomous operation capability, this invention also introduces a state discrimination module. This module calculates the local change amplitude Δx and the mutation statistics S at each time point, compares them with static thresholds θ1 and θ2 respectively, and combines them into a state vector. In implementation, the state is divided into "normal," "suspicious," and "abnormal." When the state continuously exceeds the set threshold, the system triggers the periodic detection line to refit the dominant frequency, and the mutation detection line to reinitialize the statistical recursive conditions, thus forming a complete adaptive feedback mechanism.

[0217] The entire data analysis process runs on an edge computing device based on the ARM Cortex-A72 architecture, with a data throughput rate of up to 3MB / min and CPU utilization not exceeding 60%, ensuring stable online operation 24 hours a day. The final results are packaged and transmitted to the shore base station server through the output module, generating monitoring reports, graphical trends, and anomaly alarm records.

[0218] During a three-month continuous deployment period (May 15, 2024 to August 15, 2024), the system of this invention and traditional methods were compared in experiments at 10 buoy monitoring points, collecting a total of approximately 860,000 data points. The core performance comparisons are as follows:

[0219] Table 1 Comparison of the effectiveness of the system of the present invention and traditional methods in nearshore carbon sink monitoring tasks.

[0220] As can be seen from Table 1, the system of this invention achieves a significant and quantifiable performance leap in key aspects: First, relying on the joint preprocessing of second-order curvature constraint smoothing and Allan variance, the drift residual mean square error is reduced from 0.094 ppm. 2 Reduced to 0.020 ppm 2 First, the reduction in carbon flux was approximately 78.7%, directly improving the modelability of the original sequence. Second, the periodic identification based on weighted Lomb–Scargle improved the accuracy of the main frequency fitting from 85.3% to 97.2%, a relative improvement of 13.9%, effectively reducing the interference of tidal and diurnal terms on trend interpretation. Third, the mutation detection using Shiryaev–Roberts recursive statistics improved the identification accuracy from 69.4% to 94.1%, a relative improvement of 35.5%, significantly reducing anomaly misses and false alarms. Under the closed-loop fusion drive, the parameter readjustment delay was shortened from 51 minutes to 12 minutes, a reduction of approximately 76.5%, shifting anomaly handling from "post-event correction" to "online adaptation". Finally, the end-to-end data availability increased from 82.2% to 96.8%, a relative improvement of 17.8%, meaning that under the same deployment, it can provide longer-term, more continuous, and more reliable effective sample support for carbon flux estimation and early warning models, thus forming synergistic gains in stability, accuracy, and responsiveness.

[0221] In summary, this invention, through a modular, end-to-end data processing system, achieves stable modeling, periodic extraction, and anomaly detection of non-equidistant time series while ensuring the rationality of the physical mechanism. It breaks through the processing bottleneck of existing systems under high-frequency disturbances, drift superposition, and periodic backgrounds, and provides an engineering-feasible and high-performance solution for blue carbon monitoring and marine ecological early warning.

[0222] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A marine carbon sequestration data acquisition and real-time analysis system, characterized in that, include: The data acquisition module acquires time-series data of non-uniformly spaced environmental parameters generated by sensors in the marine monitoring platform and outputs raw environmental parameter data. The steady-state filtering model module receives raw environmental parameter data and outputs robust preprocessed time series data. The window management module receives robust preprocessed time series data, divides it into time windows, and outputs segmented time series data. The bistable detection model module receives segmented time series data, generates periodic terms and deperiodic residuals in the periodic detection line, and generates change statistics in the abrupt change detection line. The closed-loop fusion module receives the periodic term, deperiodized residual, and change statistics. It inputs the deperiodized residual into the mutation detection line in the bistable detection model module, feeds back the change statistics to the stable-variable filtering model module, and outputs updated time series data. The state determination module receives updated time series data and change statistics, and generates state information based on the input information; The output module receives updated time series data, status information, and periodic items, and transmits the updated time series data, status information, and periodic items to edge computing nodes or shore-based analysis platforms. The data acquisition module is used to acquire time series data of non-uniformly spaced environmental parameters generated by sensors in the marine monitoring platform, and the raw environmental parameter data is output. The module uses a steady-state filtering model to receive raw environmental parameter data, performs second-order curvature constraint smoothing based on von Draco filtering, and combines Allan variance to decompose drift components, random walk components and white noise components at a preset time scale, outputting robust preprocessed time series data. The window management module is used to receive robust preprocessed time series data, divides it into time windows according to a preset step size, and outputs segmented time series data. The bistable detection model module receives segmented time series data. In the periodic detection line of the bistable detection model module, the weighted Lomb-Scargle method is used to generate periodic terms and deperiodic residuals. In the mutation detection line of the bistable detection model module, the deperiodic residuals are recursively calculated based on the Shiryaev-Roberts method and the change statistics are output. The closed-loop fusion module receives the periodic term, deperiodic residual, and change statistics. The deperiodic residual is input into the mutation detection line of the bistable detection model module, and the change statistics are fed back to the stable-variable filtering model module to adjust the smoothing parameters and drift correction window of the von Draco filter, and the updated time series data is output. The state discrimination module receives updated time series data and change statistics, generates state information, and triggers parameter recalibration of the periodic detection line and abrupt change detection line of the bistable detection model module when the state information is an abnormal state. The output module receives updated time series data, status information, and periodic items, and transmits these data to edge computing nodes or shore-based analysis platforms.

2. The marine carbon sink data acquisition and real-time analysis system according to claim 1, characterized in that, The generated raw environmental parameter data includes: The data acquisition module is used to set up acquisition tasks in the marine monitoring platform and send acquisition instructions to the carbon dioxide sensor, pH sensor, dissolved oxygen sensor and salinity sensor deployed in the monitoring area. Each acquisition instruction includes a sampling timestamp, sampling parameter type and sampling accuracy parameter. Receive response data returned by various sensors. The response data includes sensor identifier, sampling time point, parameter type and observation value, forming an observation data triplet. All collected observation data triplets are arranged in chronological order and classified according to the sensor channels corresponding to the parameter types to generate non-equally spaced single-channel initial observation sequences. For each type of parameter observation sequence, a set of timestamps is recorded and compared with a preset reference sampling time axis to generate corresponding time interval offset information; All classified non-equal interval single-channel initial observation sequences and their corresponding time interval offset information are encapsulated into structured data objects to form raw environmental parameter data.

3. The marine carbon sink data acquisition and real-time analysis system according to claim 1, characterized in that, Steady-state filtering models include: Receive raw environmental parameter data, and define the input time series as follows: ; Input time series Based on this, a second-order curvature-constrained smoothing objective function is constructed, and the smoothing function is set as follows: The objective function is: ； The first term is the sum of squared residuals from the fit, and the second term is the curvature constraint integral. To smooth out tension parameters; For the objective function Numerical solutions are performed to obtain a continuous smooth function. and at each timestamp Calculate the smoothing value and generate a smooth sequence. ; Smooth sequence Divided according to the preset step size There are 1 window, each window is denoted as _ . ; For each window Calculate the mean to obtain the window mean sequence. And the time scale is calculated based on the window mean sequence. Allan variance: ； in, This represents the difference between the means of adjacent windows; According to Allan variance in different Based on the slope characteristics of the logarithmic curve, the window is divided into three categories: when the slope is... When the slope is [value missing], the window is marked as a white noise region; when the slope is [value missing], the window is When the slope is 0, the window is marked as a random walk region; when the slope is 0, the window is marked as a random walk region. At that time, the window is marked as a drift area; For windows belonging to the drift region, a linear trend function is used for fitting, and the trend function is subtracted from the smoothed values ​​within the window to generate the drift-corrected sequence segment. For windows that belong to the random walk region, calculate the mean shift. And the mean shift is subtracted from all smoothed values ​​within the window to complete the random walk compensation; For windows belonging to the white noise region, no numerical adjustment is made; the original smoothing value is retained. The windows, after drift correction, random walk compensation, and adjustment of the white noise region, are spliced ​​together to generate the update sequence. ; Update sequence The output is robust preprocessed time series data.

4. The marine carbon sink data acquisition and real-time analysis system according to claim 1, characterized in that, The window management module includes: Receive robust preprocessed time series data ; Set the time window step to a constant. In the timeline interval Construct continuous segmented intervals within the interval, and define the first segment. The time window is ,in ,and , ; In each time window Inside, collect all those that meet the conditions. smoothing value The collected smooth values ​​that meet the conditions are stored as window data without performing default filling or interpolation. When a certain interval If no timestamp meets the condition, the window management module directly generates an empty window. And keep the empty window occupied in the sequence; Arrange all window data sequentially to form a window sequence, and output the window sequence as segmented time series data.

5. The marine carbon sink data acquisition and real-time analysis system according to claim 1, characterized in that, The steps for generating periodic terms, deperiodic residuals, and change statistics include: Receive the output segmented time series data and set the window sequence; The window sequence is passed to the periodic detection line to generate a periodic term sequence; Calculate the point-to-point difference between the periodic term sequence and the input sequence to generate a deperiodic residual sequence; The mutation detection line receives the deperiodic residual sequence, performs recursive processing, and outputs a sequence of change statistics. The periodic term sequence, deperiodic residual sequence, and change statistic sequence are used as the outputs of the bistable detection model module and transmitted to the closed-loop fusion module.

6. The marine carbon sink data acquisition and real-time analysis system according to claim 1, characterized in that, The bistable detection model module includes: Periodic testing line: Receive output window sequence And read all observation points in the window in sequence; Calculate the interval between adjacent observation points Assign weights to each point We obtain the weighted time series; Set candidate frequency set For each frequency Perform the following steps: Traverse all observation points in the window and calculate the weighted cosine components point by point. And accumulate to get ; Traverse all observation points in the window and calculate the weighted sine component point by point. And accumulate to get ; Calculate the normalized denominator, including ; Calculate the power spectrum value: ； The results are then stored in the power spectrum table. After calculating for all candidate frequencies, the power spectrum is traversed, and the frequency with the highest power value is selected. , as the dominant frequency of the window; In dominant frequency Next, a least squares equation is established, and the observed values ​​are substituted point by point to solve for the amplitude and phase parameters, generating a fitting function. , Each timestamp in the window Substitute the fitted function, calculate the period estimate point by point, and generate a periodic term sequence; The residuals of the generated periodic term sequence are calculated point by point to obtain the residual sequence, and then compared with the dominant frequency in the frequency domain. Nearby components are subjected to amplitude limiting suppression, and point-by-point replacement is used to obtain a deperiodic residual sequence; Mutation detection line: Receive the deperiodic residual sequence output from the periodic detection line; Initialize the Shiryaev–Roberts statistic and set initial values. ; Setting the model before the change With a mean of zero and a variance of Gaussian distribution, setting the model after the change The mean shift is variance is Gaussian distribution; Read the residuals of the deperiodic residual sequence point by point Calculate the likelihood ratio: And recursively update the statistics: ; Calculate the time interval between the point and the previous sampling point. And combined with the global average sampling interval Calculate the correction factor ; Execution judgment condition: If Then the time point Mark them as mutation points and record the statistical values; By iteratively applying the results to all points within the window and saving the results point by point, a series of change statistics is formed. ; Output: The periodic sequences of each window are concatenated in chronological order to form a global periodic sequence; the deperiodic residual sequences of each window are concatenated in chronological order to form a global deperiodic residual sequence; the change statistics sequences of each window are concatenated in chronological order to form a global change statistics sequence; the periodic sequence, deperiodic residual sequence, and change statistics sequence are combined as the output of the bistable detection model module and transmitted to the closed-loop fusion module.

7. The marine carbon sink data acquisition and real-time analysis system according to claim 1, characterized in that, The closed-loop fusion module execution includes: Receive the output periodic term sequence, deperiodic residual sequence, and change statistics sequence, and sort them by timestamp. Alignment to form a synchronized data stream; Deperiodic residual sequence The mutation detection line of the bistable detection model module is input point by point in chronological order, and recursive calculation is performed to generate an updated sequence of change statistics. Update the change statistics series The sequence of changes in the input statistics Compare each item one by one according to time points, calculate the difference, and build an index set. ; For sets Each point in time Perform the following steps: In the interval Smoothed value of stable filter output is extracted internally. Forming local subsequences ,in The preset time window length is used; the mean square error between the local subsequence and the original observation sequence in the same interval is calculated, and the mean square error is calculated based on a threshold. The ratio is updated to smooth the parameter: if the error is greater than Then let ,in It is a proportionality coefficient; otherwise, it remains unchanged. ; the interval Set as the drift correction window and reconstruct the second-order curvature-constrained smoothing objective function. and use the updated Perform numerical solutions over the interval; generate a set of corrected smooth values ​​within the drift correction window. and replace the corresponding part in the original smooth sequence; The results corrected by all index points are concatenated with the unadjusted segments to generate the updated time series. ; Update the time series As the output of the closed-loop fusion module.

8. The marine carbon sink data acquisition and real-time analysis system according to claim 1, characterized in that, The execution steps of the state determination module include: Receive the updated time series data and the corresponding change statistics sequence from the output, align them according to the timestamp order, and establish a point-by-point corresponding input set; At each time point, the difference between the smoothed value at the time point and the previous time point is calculated to obtain the local fluctuation amount, and the local fluctuation amount and the change statistics at the time point are used together to form a discriminant vector. For each discrimination vector, compare the local fluctuation with the first threshold, and compare the change statistic with the second threshold. When both are below their respective thresholds, it is marked as a normal state; when one of them exceeds the threshold, it is marked as a suspicious state; when both exceed the threshold, it is marked as an abnormal state. Count consecutive suspicious and abnormal states. When the consecutive count exceeds a preset threshold, the overall state information of the time period corresponding to the consecutive suspicious or abnormal states is confirmed as an abnormal state. When the status information is confirmed to be an abnormal state, a trigger signal is generated and transmitted to the bistable detection model module, which instructs the periodic detection line to recalculate the frequency fitting parameters and simultaneously instructs the mutation detection line to reinitialize the initial value of the statistics and the discrimination threshold. Output a sequence of status information arranged in chronological order. The status information is limited to three categories: normal, suspicious, and abnormal. The status information sequence is then transmitted to the output module.

Images