Water body environment pollution detection method based on marine high-precision salinity sensor

Abstract

The invention relates to the technical field of water body environment pollution detection based on a marine high-precision salinity sensor, and discloses a water body environment pollution detection method based on a marine high-precision salinity sensor. The method comprises the following steps: initializing a sensor network, adaptively calculating a neighborhood radius, constructing a fixed space neighborhood set based on an observation time sequence, extracting local statistical features (mean value, standard deviation, skewness and kurtosis) and constructing SMKI high-order moment anomaly indexes, performing baseline period analysis in combination with historical data, adaptively setting an anomaly judgment threshold value, and determining the anomaly of the sensor network. And finally, real-time anomaly detection and pollution event clustering identification of the salinity data are realized. Local neighborhood modeling improves the detection sensitivity, high-order moment indexes are introduced to enhance the accuracy of anomaly recognition, baseline modeling reduces the false report and missing report rate, and density clustering supports structured output of pollution events.

Description

Technical Field

[0001] The present invention relates to the technical field of water environment pollution detection based on a high-precision marine salinity sensor, and in particular to a water environment pollution detection method based on a high-precision marine salinity sensor. Background Art

[0002] With global climate change and overexploitation of marine resources, marine water pollution is becoming increasingly serious, posing a serious challenge to marine ecosystems and the sustainable development of human society. In recent years, marine environmental monitoring technology based on high-precision salinity sensors has garnered widespread attention, becoming an important means of detecting dissolved pollutants, assessing seawater structural stability, and assisting in early warning of marine ecological risks. However, existing pollution detection methods based on salinity sensors still have many shortcomings in practical applications and cannot meet the demand for real-time, accurate, and sensitive detection in dynamic marine environments.

[0003] Traditional water pollution detection methods typically rely on empirical or fixed thresholds for alarm generation. Typically, upper and lower alarm limits for salinity fluctuations are determined based on environmental experience or laboratory calibration. When the sensor's measured salinity value falls outside this range, an alarm is triggered. Because ocean salinity is influenced by multiple factors, such as tides, wind and waves, and temperature, salinity fluctuations exhibit complex spatiotemporal coupling. Using fixed thresholds can easily lead to high false positives and missed negatives. On the one hand, excessively low thresholds can misinterpret normal seasonal or tidal fluctuations as pollution, increasing the burden on operations and maintenance personnel. On the other hand, excessively high thresholds can make it difficult to capture early or minor pollution events, delaying source location and control. To improve the adaptability of thresholds, some studies have employed sliding window statistical methods or moving average filtering to dynamically update thresholds based on historical data. However, these methods only process first- or second-order statistics (such as mean and variance), failing to fully characterize sudden changes in distribution patterns caused by pollution events. They can easily miss pollution signals with significant non-mean deviations but sensitive tail variations. Furthermore, the selection of sliding window length and filter window still requires empirical adjustment, lacking automated, theoretically supported optimization strategies. To account for spatial correlation in multi-point sensor networks, researchers have employed spatial interpolation and kriging methods to grid-reconstruct pollution signatures, or constructed adjacency matrices based on graph theory for further graph convolution analysis. These methods can leverage spatial neighborhood information to enhance the spatial identifiability of outliers. However, given the large-scale, complex topology of the ocean, interpolation is computationally expensive and sensitive to grid resolution. While highly accurate, anomaly detection models based on graph convolution or deep learning require a large number of labeled samples and expensive computing resources, making them unsuitable for resource-constrained ocean buoys or automated observation stations. Other researchers have introduced multivariate metrics such as information entropy, fractal dimension, and multiscale analysis to identify salinity anomalies. However, most methods focus on changes in overall distribution characteristics and lack a precise characterization of local, fine-grained regional distribution patterns, making them difficult to achieve early warning of single-point or small-scale pollution events. Furthermore, these techniques often use features from multiple dimensions as input, then apply threshold judgments based on rules or model fusion. This cumbersome process and the numerous parameters make it difficult to deploy in the field in real time.

[0004] To this end, this paper aims to propose a water environment pollution detection method based on marine high-precision salinity sensors. The local distribution characteristics are captured by adaptively constructed spatial neighborhoods, and the skewness and kurtosis are constructed by third-order and fourth-order central moments to measure high-order abnormal behaviors. Then, outlier points are identified based on SMKI, and the complete evolution process of pollution events is restored through neighborhood density clustering. Summary of the Invention

[0005] The present invention provides a method for detecting water environment pollution based on a high-precision marine salinity sensor, which helps solve the problems mentioned in the above background technology.

[0006] The present invention provides the following technical solution: a method for detecting water environment pollution based on a high-precision marine salinity sensor, comprising:

[0007] Initialize the sensor network and adaptively calculate the spatial neighborhood radius of each sensor;

[0008] In the preset observation time series, the salinity observation value of each sensor at each time point is obtained, and the corresponding fixed spatial neighborhood set is constructed according to the spatial neighborhood radius;

[0009] Based on the spatial neighborhood set, the mean and standard deviation of salinity in the local area are calculated;

[0010] Further calculate the third-order central moment and fourth-order central moment in the local area, and derive the skewness and kurtosis values ​​respectively based on them;

[0011] The local higher-order moment index SMKI is constructed based on the skewness and kurtosis indices;

[0012] Baseline analysis is performed based on historical observation data, and an adaptive method is used to calculate the abnormality judgment threshold of each sensor;

[0013] Use the constructed SMKI indicator and threshold to perform anomaly detection on real-time observation data and identify abnormal points;

[0014] Abnormal clusters are identified through a density clustering method based on spatial neighborhood, and pollution event information is output according to the spatial-temporal correlation of abnormal clusters.

[0015] Optionally, the initializing the sensor network and adaptively calculating the spatial neighborhood radius of each sensor specifically includes:

[0016] Get the distance between sensor No. i and sensor No. p, denoted as d i,p ;

[0017] Where, i,p∈{1,2,...,N}, is the sensor number; N is the total number of sensors;

[0018] Construct an N×N dimensional real symmetric distance matrix D=[d i,p ] N×N ; Among them, d i,p =d p,i , d i,i =0;

[0019] Calculate the minimum distance between sensor No. i and all other sensors

[0020] Calculate the spatial neighborhood radius Among them, median(·) is the median function, which returns the middle value of an ordered list.

[0021] Optionally, obtaining the salinity observation value of each sensor at each time point within a preset observation time series and constructing a corresponding fixed spatial neighborhood set according to the spatial neighborhood radius specifically includes:

[0022] Set the observation time set to {t j |j=1,2,...,M}; where t j is the time point of the jth observation; M is the total number of observation times; j is the number of the observation time point;

[0023] Get the value of sensor No. i at time t j The salinity value is recorded as S i,j ;

[0024] Construct the spatial neighborhood number set of sensor i

[0025] Optionally, the calculation of the salinity mean and standard deviation in the local area based on the spatial neighborhood set specifically includes:

[0026] For any i∈{1,2,...,N}, j∈{1,2,...,M}, perform the following steps:

[0027] S101. Calculate neighborhood size Among them, |·| is the set cardinality operation;

[0028] S102, calculate the neighborhood of sensor No. i at time t j The arithmetic mean of salinity

[0029] S103, calculate the neighborhood of sensor No. i at time t j The standard deviation of salinity σ i,j :

[0030]

[0031] Optionally, further calculating the third-order central moment and the fourth-order central moment in the local area and deriving the skewness and kurtosis values ​​respectively therefrom specifically includes:

[0032] Calculate the third-order center distance

[0033] Calculating skewness

[0034] Calculate the fourth-order center distance

[0035] Calculating Kurtosis

[0036] Optionally, constructing a local high-order moment index SMKI based on the skewness and kurtosis indices specifically includes:

[0037] Constructing local high-order moment index SMKI i,j =|γ i,j |+|κ i,j |; where |·| is the absolute value.

[0038] Optionally, performing baseline analysis based on historical observation data and using an adaptive method to calculate the abnormality determination threshold of each sensor specifically includes:

[0039] Select the first B moments {t1, t2, ..., t B} as the baseline period, and 1≤B <M;

[0040] Set the abnormality judgment threshold of sensor No. i to

[0041] Optionally, the use of the constructed SMKI indicator and threshold to perform anomaly detection on the real-time observation data and identify abnormal points specifically includes:

[0042] For any j>B: If SMKI i,j >H i , then mark the sensor No. i at time t j For abnormal points.

[0043] Optionally, identifying abnormal clusters by a density clustering method based on spatial neighborhood and outputting pollution event information according to the spatial-temporal correlation of the abnormal clusters specifically includes:

[0044] Extraction time t j The abnormal sensors are numbered and form a set A j ={i|SMKI i,j >H i};

[0045] Perform the following steps to perform point-by-point iterative neighborhood clustering:

[0046] S2100, i∈A j All are marked as ungrouped;

[0047] S2200, take an ungrouped point i in turn, create a new contaminated cluster C = {i}, and mark i as grouped;

[0048] S2300. For each cluster member p∈C, perform the following steps:

[0049] S2310, traverse the remaining ungrouped points q∈A j :

[0050] S2311, if dp,q ≤r, then add q to C and mark it as grouped;

[0051] S2400, repeat step S2300 until there are no new members in cluster C; record C as a final cluster;

[0052] S2500, repeat steps S2100 to S2400 until all points are grouped;

[0053] If time t j-1 Cluster C j-1 With t j Cluster C j There is a number intersection, that is That is, it is regarded as a continuation of an event and its time intervals are merged;

[0054] For each final event cluster C, output:

[0055] Sensor list {i|i∈C}; time label interval between first appearance and last appearance [t min ,t max ]; Peak index

[0056] The present invention has the following beneficial effects:

[0057] 1. By constructing a real symmetric distance matrix, the spatial relationship between sensors is quantified, and a strategy using the "minimum distance median" as the adaptive determination of the neighborhood radius is proposed. First, this method does not rely on prior knowledge or manually set neighborhood radius, which solves the problem of detection performance degradation caused by unreasonable neighborhood settings in traditional methods and enhances the system's adaptability to sensor networks with different distribution densities. Second, by using the minimum distance median as the global scale estimation method, it effectively reduces the interference of abnormal nodes or dense clusters on the radius setting, and has good robustness. Compared with existing methods based on fixed radius or empirical values, this method automatically adjusts the neighborhood size, so that it can still maintain high area coverage and detection accuracy in the case of uneven spatial distribution or node damage. In addition, this neighborhood partitioning mechanism lays a solid foundation for subsequent local statistical analysis and spatial clustering.

[0058] 2. By uniformly constructing the set of observation moments and the set of spatial neighborhoods, a stable data input structure is formed, providing clear boundaries for subsequent algorithm design. The time series information and the spatial neighborhood are constructed synchronously, so that the entire data structure has "time-space synchronization". This mechanism is different from the existing methods that only focus on time series or spatial discrete points, and provides higher resolution for processing time-space coupled data. In addition, the construction of the spatial neighborhood is based on the adaptive radius of step 1, so that the neighborhood of each sensor can truly reflect its local water body characteristics, thereby avoiding local distortion caused by "too wide" or "too narrow" neighborhoods. Compared with global processing strategies such as traditional sliding windows, this method adopts a local observation mechanism, which is more suitable for handling local abnormal pollution events, especially in situations with dispersed pollution sources in large water areas.

[0059] 3. Based on the data within a fixed spatial neighborhood, the mean and standard deviation of the local salinity are calculated to form basic statistical features. The statistical construction of "spatial localization" enables each node to have personalized background modeling capabilities at the moment of observation, avoiding the generalization error caused by the use of global thresholds. Traditional detection methods based on the global mean or sliding window standard deviation often ignore local characteristics and easily misidentify natural fluctuations as anomalies or miss weak anomalies. This method reflects the local fluctuation range through the neighborhood standard deviation, provides a standardized benchmark for high-order anomaly indicators (skewness, kurtosis), and ensures that the indicator calculation has a consistent scale. Improve the sensitive detection capability of small-scale pollution mutations and adapt to the diverse changes in the complex ocean dynamic environment.

[0060] 4. Introducing skewness and kurtosis to analyze the local data distribution pattern is one of the technical cores of the entire method. Compared with the traditional strategy of relying solely on mean / variance to identify anomalies, skewness characterizes the asymmetry of the distribution, and kurtosis reflects extreme values ​​or spike anomalies, which can capture weak, non-mean mutations. Introducing higher-order moments in statistics into sensor network data analysis improves the ability to identify "distribution variation" type anomalies. At the same time, skewness and kurtosis have dimensionless characteristics, support cross-time and cross-node comparative analysis, and improve the versatility of the model. Compared with existing methods, the use of this higher-order moment not only improves detection accuracy, but also greatly expands the types of detectable anomalies (such as mild anomalies, tail pollution, etc.), which is of great value in early warning of pollution.

[0061] 5. SMKI is a comprehensive anomaly indicator proposed in this solution, combining the absolute values ​​of skewness and kurtosis to form a single anomaly metric. Simplifying the two high-order features into a unified standard simplifies subsequent judgment logic and improves processing efficiency; at the same time, it avoids complex multi-dimensional judgments in high-dimensional feature space, improving real-time performance. Theoretically, SMKI combines asymmetry and peak characteristics to comprehensively characterize local "distribution extremes," taking into account both amplitude and morphology. Compared to traditional methods such as the z-score, which rely solely on first- or second-order statistics, SMKI has stronger expressive power. In actual application scenarios, it can accurately identify various types of water anomalies, such as "pollution spikes," "diffusion offsets," and "local high-frequency fluctuations," providing possibilities for the classification and identification of pollution events.

[0062] 6. By selecting historical observation period data (baseline period) to set the abnormal threshold, it is completely based on historical normal state data and does not introduce any prior model parameters. It is a typical "data-driven adaptive learning" strategy. The length of the baseline period can be flexibly selected to support periodic modeling of different ocean dynamic environments (such as diurnal tidal cycles); at the same time, the maximum value strategy is used to construct the threshold, which naturally has high safety redundancy. Compared with the common fixed threshold and empirical coefficient setting methods, the adaptive mechanism can better cope with the systematic changes brought about by dynamic backgrounds (such as temperature differences, wind speeds, and water flow changes) while ensuring the control of false alarm rates. In addition, the threshold is completely driven by data, is transparent and explainable, and is suitable for deployment in unmanned ocean intelligent monitoring platforms.

[0063] 7. Combining SMKI with adaptive thresholds, this step enables real-time outlier identification, a key decision-making step in algorithm execution. Using a "point-by-point judgment + local threshold" approach, it overcomes the real-time and accuracy bottlenecks of traditional "overall window statistics" methods. Anomaly detection eliminates the need to wait for a window to complete, enabling immediate responses to sudden events and significantly reducing pollution response latency. This approach is particularly suitable for scenarios such as sudden leak monitoring and early identification of red tides.

[0064] 8. Combining abnormal points into "pollution clusters" through density clustering methods based on spatial neighborhoods, and then constructing complete pollution events through time series merging, is a key transition link from "point anomalies" to "structural events." Unsupervised clustering based on neighborhood propagation is adopted, which does not rely on external parameters such as the number of clusters or shape assumptions, and completely self-organizes to identify pollution forms. In real-world scenarios, pollution sources often manifest as multiple points changing simultaneously and spreading over time. Traditional methods that rely solely on threshold alarms are unable to track the entire incident. This method achieves dynamic restoration of the pollution "start-diffusion-end" process through synchronous tracking in the spatiotemporal domain, providing comprehensive data support for subsequent decisions such as source tracing, alarm response, and closed containment. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0066] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0067] Example, see Figure 1 , a method for detecting water environment pollution based on a high-precision marine salinity sensor, comprising:

[0068] Initialize the sensor network and adaptively calculate the spatial neighborhood radius of each sensor;

[0069] In the preset observation time series, the salinity observation value of each sensor at each time point is obtained, and the corresponding fixed spatial neighborhood set is constructed according to the spatial neighborhood radius;

[0070] Based on the spatial neighborhood set, the mean and standard deviation of salinity in the local area are calculated;

[0071] Further calculate the third-order central moment and fourth-order central moment in the local area, and derive the skewness and kurtosis values ​​respectively based on them;

[0072] The local higher-order moment index SMKI is constructed based on the skewness and kurtosis indices;

[0073] Baseline analysis is performed based on historical observation data, and an adaptive method is used to calculate the abnormality judgment threshold of each sensor;

[0074] Use the constructed SMKI index and threshold to perform anomaly detection on real-time observation data and identify abnormal points;

[0075] Abnormal clusters are identified through a density clustering method based on spatial neighborhood, and pollution event information is output according to the spatial-temporal correlation of abnormal clusters.

[0076] From sensor initialization and adaptive neighborhood construction to local statistical feature extraction, anomaly indicator calculation, threshold adaptive setting, real-time anomaly identification, and finally cluster output of pollution events, a closed-loop detection system is formed. By comprehensively integrating information in both spatial and temporal dimensions, the problems of strong data dynamics, hidden local pollution events, and inapplicability of global models in marine environments are solved. Local neighborhood calculations are used to improve sensitivity, high-order moment indicators are introduced to enhance anomaly identification accuracy, baseline modeling is used to avoid false alarms and missed alarms, and finally, quantitative output of pollution events is achieved through cluster identification. Compared with existing static threshold alarms or sliding window methods, this solution takes into account detection accuracy, timely response, and structural interpretability, and is particularly suitable for water quality anomaly monitoring tasks in such a changeable and dynamic scenario as the marine environment.

[0077] Initializing the sensor network and adaptively calculating the spatial neighborhood radius of each sensor specifically includes:

[0078] Get the distance between sensor No. i and sensor No. p, denoted as d i,p ;

[0079] Where, i,p∈{1,2,...,N}, is the sensor number; N is the total number of sensors;

[0080] Construct an N×N dimensional real symmetric distance matrix D=[d i,p ] N×N ; Among them, d i,p =d p,i , d i,i =0;

[0081] Providing the pairwise distance relationship between all sensors, laying the foundation for subsequent neighborhood division and spatial statistics;

[0082] Calculate the minimum distance between sensor No. i and all other sensors Evaluate the minimum spacing between each sensor to characterize the typical spatial distribution between measurement points;

[0083] Calculate the spatial neighborhood radius Among them, median(·) is the median function, which returns the middle value of the ordered list. The neighborhood radius is automatically generated according to the median of the minimum distance of all sensors, avoiding manual setting and improving the robustness of the algorithm.

[0084] By calculating the minimum distance between sensors and using the median of the minimum distances of all sensors to set the spatial neighborhood radius, the neighborhood parameters can be adaptively set. In this way, the problems of inadaptability and local distortion caused by relying on empirical parameters or global fixed values ​​in traditional neighborhood division methods are solved. The advantage of this method lies in its "adaptability" and "robustness": using the median in the distance matrix as a representative scale, it can not only automatically reflect the spatial density changes of sensors in actual deployment, but also cope with the situation of partial node failures or irregular distribution, ensuring that the neighborhood construction is real and stable. This design effectively improves the representativeness and accuracy of subsequent statistical indicators, lays a stable spatial structure foundation for the entire pollution detection system, and is a key prerequisite for realizing localized detection and distributed sensitive detection capabilities.

[0085] The method of obtaining the salinity observation value of each sensor at each time point within a preset observation time series and constructing a corresponding fixed spatial neighborhood set according to the spatial neighborhood radius specifically includes:

[0086] Set the observation time set to {t j |j=1,2,...,M}; where t j is the time point of the jth observation; M is the total number of observation times; j is the number of the observation time point;

[0087] Get the value of sensor No. i at time t j The salinity value is recorded as S i,j ;

[0088] Clarify the time and measurement input format required by the algorithm to provide data sources for subsequent statistical calculations;

[0089] Construct the spatial neighborhood number set of sensor i Determine the local neighborhood range of each sensor and provide an object set for subsequent local statistics.

[0090] The organization of observation time is clarified, and based on the previously calculated spatial neighborhood radius, a fixed spatial neighborhood set is constructed for each sensor. This structure organizes data into a combination of "space-time points + neighborhood structure," providing an efficient data access framework for subsequent calculations. By constructing a fixed neighborhood set, the problem of ineffective statistical analysis due to the spatial discreteness and non-uniform dimensions of the original observation data is resolved. The benefit of this operation is that every moment and every sensor has contextual neighborhood support, thus enabling true local statistical analysis. This approach provides the necessary data support for higher-order statistics such as local mean, standard deviation, skewness, and kurtosis, ensuring the feasibility of the entire high-order moment anomaly analysis method. At the same time, the introduction of a fixed neighborhood also greatly improves data processing efficiency, making the algorithm suitable for deployment in resource-limited edge computing nodes or embedded devices.

[0091] The calculation of the salinity mean and standard deviation in the local area based on the spatial neighborhood set specifically includes:

[0092] For any i∈{1,2,...,N}, j∈{1,2,...,M}, perform the following steps:

[0093] S101. Calculate neighborhood size Among them, |·| is the set cardinality operation; it calculates the number of sensors in the neighborhood and provides a normalization factor for subsequent average and variance calculations;

[0094] S102, calculate the neighborhood of sensor No. i at time t j The arithmetic mean of salinity Get the neighborhood of sensor No. i at time t j The average level of salinity, describing the local background concentration;

[0095] S103, calculate the neighborhood of sensor No. i at time t j The standard deviation of salinity σ i,j :

[0096] Quantify the fluctuation of neighborhood salinity and provide a standardized benchmark for skewness and kurtosis calculations.

[0097] By calculating the mean and standard deviation of the salinity observations in the neighborhood of each sensor, a basic statistical framework for evaluating the local environmental status was constructed. The "mean" is used to characterize the local background level, and the "standard deviation" is used to characterize the degree of local fluctuation, providing a standardized quantitative basis for subsequent high-order analysis and anomaly detection. It effectively solves the problem that sensor readings in the marine environment fluctuate violently due to factors such as local fluid dynamics and temperature gradients, resulting in the failure of traditional global models. Compared with the global mean or moving average method, the local mean / standard deviation system has stronger sensitivity, especially in capturing regional pollutant leaks, local red tide anomalies and other phenomena. In addition, the introduction of the standard deviation allows the subsequent skewness and kurtosis to be normalized to a dimensionless form, improving the comparability of indicators, and is a basic link in achieving high-order anomaly detection.

[0098] The further calculation of the third-order central moment and the fourth-order central moment in the local area and the skewness and kurtosis values ​​​​are obtained accordingly, specifically including:

[0099] Calculate the third-order center distance Measure the skewness of neighborhood salinity distribution; extract the skewness information of local salinity distribution and preliminarily capture the asymmetry of the distribution;

[0100] Calculating skewness The third-order central moment is standardized into dimensionless skewness to facilitate comparison between different regions or time periods;

[0101] Calculate the fourth-order center distance Measure the kurtosis or tail thickness of the neighborhood salinity distribution; extract the kurtosis information of the local salinity distribution and measure the tail thickness or peak degree;

[0102] Calculating Kurtosis After subtracting 3, it is aligned with the normal distribution; the fourth-order central moment is converted into excess kurtosis and aligned with the kurtosis level of the normal distribution to highlight the abnormal tail characteristics.

[0103] It proposes in detail how to calculate the third-order and fourth-order central moments based on neighborhood data, and derive skewness and kurtosis based on this, as core indicators for measuring the distribution morphology of local data. Through this step, the system can identify the asymmetry (skewness) and abnormal spikes (kurtosis) of the distribution, and realize the identification of weak anomalies or extreme pollution behaviors that are difficult to capture with traditional mean / variance. It solves the problem of "Gaussian assumption failure" caused by traditional anomaly detection algorithms based only on first-order or second-order statistics, and significantly improves the sensitivity of detection to asymmetric or fat-tailed pollution events. For example, in the scenario of marine micro-leakage or sudden diffusion, the pollution signal may not significantly increase the average value, but its distribution will show an obvious skew or concentration trend. At this time, skewness / kurtosis is the key to identification. By combining these two high-order statistical indicators, not only the detection dimension is enhanced, but also a solid foundation is laid for subsequent indicator fusion and cluster analysis.

[0104] Construct the local high - order moment index SMKI according to the skewness and kurtosis indexes, which specifically includes:

[0105] Construct the local high - order moment index SMKI i,j = |γ i,j | + |κ i,j |; where, |·| is the absolute value;

[0106] Combine the absolute values of skewness and kurtosis to form a single metric, which directly reflects the degree of abnormality of the neighborhood high - order distribution.

[0107] Propose a construction method for the SMKI index. By adding the absolute values of skewness and kurtosis, a unified high - order anomaly detection index is formed. This design overcomes the problem of limited coverage of a single index by integrating the information of two distribution characteristics, realizes a wider anomaly detection ability, improves the real - time performance and computational efficiency of the detection algorithm. Compared with multi - dimensional vector discrimination, this single index can be used for quick threshold judgment, which is suitable for embedded environments and online real - time detection scenarios; improve the accuracy of anomaly recognition. Skewness is sensitive to distribution skewness, and kurtosis is sensitive to anomaly concentration. Through the unified index, the combined anomalies of the two can be captured.

[0108] Perform baseline period analysis according to historical observation data, and use an adaptive method to calculate the anomaly determination threshold for each sensor, which specifically includes:

[0109] Select the first B moments {t1, t2,..., t B} as the baseline period, and 1 ≤ B < M; define a pollution - free reference period for generating the adaptive threshold; if the system or environment has an obvious intra - day cycle (such as tides, diurnal temperature differences, etc.), the baseline period should cover at least one complete cycle. For example, if the salinity sensor samples once an hour and the diurnal cycle is 24 hours, then B ≥ 24 can be taken. The larger B is, the more historical maximum values are included, the higher the threshold, and the lower the sensitivity of the system to minor anomalies; conversely, the smaller B is, the lower the threshold, the easier it is to trigger an alarm, but the more likely false alarms are to occur.

[0110] Set the anomaly determination threshold for the i - th sensor as Directly take the maximum value of the index within the baseline period as the threshold without any external parameters, ensuring that the threshold is completely determined by historical normal data.

[0111] By introducing a baseline period mechanism, historical data is used to construct a dynamic, parameter-free anomaly detection threshold, thus avoiding the subjectivity and low adaptability problems caused by manual settings. The maximum SMKI value observed during the baseline period is directly used for the current threshold setting to ensure the accuracy and stability of detection. It solves the problems of false alarms and missed alarms caused by dynamic changes in the environment (such as tides, temperature, sunshine, etc.) on the fixed threshold detection system, and significantly improves the adaptability and practical application range of the detection algorithm. The system has the ability to "learn" and can automatically adjust the sensitivity according to the historical status of each sensor to implement a "locally adapted" monitoring strategy. It is a key component for achieving highly reliable online monitoring and alarms.

[0112] The method of using the constructed SMKI indicator and threshold to detect anomalies in real-time observation data and identify abnormal points specifically includes:

[0113] For any j>B: If SMKI i,j >H i , then mark the sensor No. i at time t j It is an abnormal point; it is judged online whether the current distribution pattern exceeds the historical normal range to realize real-time pollution alarm.

[0114] A specific discrimination logic for identifying abnormal points based on the relationship between SMKI and threshold values ​​in a real-time monitoring framework is proposed. At each moment and each node, it is possible to independently determine whether an abnormality exists, without waiting for window data or post-processing judgment, which greatly improves the detection response speed. Through this real-time mechanism, typical problems such as "delayed response" and "discontinuous alarm" in traditional systems are solved. Especially in the early stages of pollution spread, it can identify abnormalities in the first place, trigger early warning and response mechanisms, and prevent pollution from expanding. At the same time, this method realizes the integration of complex indicators and simple decision-making, providing feasibility for deployment on resource-constrained platforms such as offshore buoys and autonomous surface vehicles, and helping to promote the practical implementation of intelligent water environment monitoring technology.

[0115] The method of identifying abnormal clusters by using a density clustering method based on spatial neighborhood and outputting pollution event information based on the spatial-temporal correlation of abnormal clusters specifically includes:

[0116] Extraction time t j The abnormal sensors are numbered and form a set A j ={i|SMKI i,j >H i Summarize all abnormal sensor numbers at the current moment to prepare for cluster identification;

[0117] Perform the following steps to perform point-by-point iterative neighborhood clustering:

[0118] S2100, i∈A j All are marked as ungrouped;

[0119] S2200, take an ungrouped point i in turn, create a new contaminated cluster C = {i}, and mark i as grouped;

[0120] S2300. For each cluster member p∈C, perform the following steps:

[0121] S2310, traverse the remaining ungrouped points q∈A j :

[0122] S2311, if d p,q ≤r, then add q to C and mark it as grouped;

[0123] S2400, repeat step S2300 until there are no new members in cluster C; record C as a final cluster;

[0124] S2500, repeat steps S2100 to S2400 until all points are grouped;

[0125] The outliers are divided into pollution clusters according to the spatial neighborhood relationship, and the spatial range of each cluster is clarified;

[0126] If time t j-1 Cluster C j-1 With t j Cluster C j There is a number intersection, that is That is, it is regarded as a continuation of an event and its time intervals are merged; the evolution of the same pollution cluster is tracked across time to achieve a closed loop of event sequence;

[0127] For each final event cluster C, output:

[0128] Sensor list {i|i∈C}; time label interval between first appearance and last appearance [t min ,t max ]; Peak index Generate complete pollution incident reports to provide accurate information for decision-making and disposal.

[0129] Through the density clustering algorithm, abnormal points in the spatial neighborhood are organized into pollution clusters, and the evolution process of pollution events is identified in combination with temporal relationships. The use of an unsupervised iterative algorithm based on neighborhood propagation effectively avoids the reliance of traditional clustering on the number of clusters or structural assumptions. It solves the core problem of scattered distribution of abnormal points and the difficulty in determining whether they constitute pollution events, and significantly improves the ability of pollution identification from "point identification" to "event-level understanding". It can output specific pollution sensors, duration, and intensity indicators, provide regulatory agencies with actionable reports and decision-making basis, and provide structured basic data for pollution tracing and dynamic response. It is a bridge link for the solution to achieve the connection from technical system to business system.

[0130] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0131] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A method for detecting water environment pollution based on a high-precision marine salinity sensor, characterized in that: include: Initialize the sensor network and adaptively calculate the spatial neighborhood radius of each sensor; In the preset observation time series, the salinity observation value of each sensor at each time point is obtained, and the corresponding fixed spatial neighborhood set is constructed according to the spatial neighborhood radius; Based on the spatial neighborhood set, the mean and standard deviation of salinity in the local area are calculated; Further calculate the third-order central moment and fourth-order central moment in the local area, and derive the skewness and kurtosis values ​​respectively based on them; The local higher-order moment index SMKI is constructed based on the skewness and kurtosis indices; Baseline analysis is performed based on historical observation data, and an adaptive method is used to calculate the abnormality judgment threshold of each sensor; Use the constructed SMKI index and threshold to perform anomaly detection on real-time observation data and identify abnormal points; Abnormal clusters are identified through a density clustering method based on spatial neighborhood, and pollution event information is output according to the spatial-temporal correlation of abnormal clusters.

2. The method for detecting water environment pollution based on a high-precision marine salinity sensor according to claim 1, characterized in that: Initializing the sensor network and adaptively calculating the spatial neighborhood radius of each sensor specifically includes: Get the distance between sensor No. i and sensor No. p, denoted as d i,p ; Where, i,p∈{1,2,...,N}, is the sensor number; N is the total number of sensors; Construct an N×N dimensional real symmetric distance matrix D=[d i,p ] N×N ; Among them, d i,p =d p,i , d i,i =0; Calculate the minimum distance between sensor No. i and all other sensors Calculate the spatial neighborhood radius Among them, median(·) is the median function, which returns the middle value of an ordered list.

3. The method for detecting water environment pollution based on a high-precision marine salinity sensor according to claim 2, characterized in that: The method of obtaining the salinity observation value of each sensor at each time point within a preset observation time series and constructing a corresponding fixed spatial neighborhood set according to the spatial neighborhood radius specifically includes: Set the observation time set to {t j |j=1,2,...,M}; where t j is the time point of the jth observation; M is the total number of observation times; j is the number of the observation time point; Get the value of sensor No. i at time t j The salinity value is recorded as S i,j ; Construct the spatial neighborhood number set of sensor i 4. The method for detecting water environment pollution based on a high-precision marine salinity sensor according to claim 3, characterized in that: The calculation of the salinity mean and standard deviation in the local area based on the spatial neighborhood set specifically includes: For any i∈{1,2,...,N}, j∈{1,2,...,M}, perform the following steps: S101. Calculate neighborhood size Among them, |·| is the set cardinality operation; S102, calculate the neighborhood of sensor No. i at time t j The arithmetic mean of salinity S103, calculate the neighborhood of sensor No. i at time t j The standard deviation of salinity σ i,j :

5. The method for detecting water environment pollution based on a high-precision marine salinity sensor according to claim 4, characterized in that: The further calculation of the third-order central moment and the fourth-order central moment in the local area and the skewness and kurtosis values ​​​​are obtained accordingly, specifically including: Calculate the third-order center distance Calculating skewness Calculate the fourth-order center distance Calculating Kurtosis 6. The method for detecting water environment pollution based on a high-precision marine salinity sensor according to claim 5, characterized in that: The construction of the local high-order moment index SMKI based on the skewness and kurtosis indicators specifically includes: Constructing local high-order moment index SMKI i,j =|γ i,j |+|κ i,j |; where |·| is the absolute value.

7. The method for detecting water environment pollution based on a high-precision marine salinity sensor according to claim 6, characterized in that: The baseline analysis is performed based on historical observation data, and an adaptive method is used to calculate the abnormality determination threshold of each sensor, specifically including: Select the first B moments {t1, t2, ..., t B } as the baseline period, and 1≤B <M; Set the abnormality judgment threshold of sensor No. i to 8. The method for detecting water environment pollution based on a high-precision marine salinity sensor according to claim 7, characterized in that: The method of using the constructed SMKI indicator and threshold to detect anomalies in real-time observation data and identify abnormal points specifically includes: For any j>B: If SMKI i,j >H i , then mark the sensor No. i at time t j For abnormal points.

9. The method for detecting water environment pollution based on a high-precision marine salinity sensor according to claim 8, characterized in that: The method of identifying abnormal clusters by using a density clustering method based on spatial neighborhood and outputting pollution event information based on the spatial-temporal correlation of abnormal clusters specifically includes: Extraction time t j The abnormal sensors are numbered and form a set A j ={i|SMKI i,j >H i }; Perform the following steps to perform point-by-point iterative neighborhood clustering: S2100, i∈A j All are marked as ungrouped; S2200, take an ungrouped point i in turn, create a new contaminated cluster C = {i}, and mark i as grouped; S2300. For each cluster member p∈C, perform the following steps: S2310, traverse the remaining ungrouped points q∈A j : S2311, if d p,q ≤r, then add q to C and mark it as grouped; S2400, repeat step S2300 until there are no new members in cluster C; record C as a final cluster; S2500, repeat steps S2100 to S2400 until all points are grouped; If time t j-1 Cluster C j-1 With t j Cluster C j There is a number intersection, that is That is, it is regarded as a continuation of an event and its time intervals are merged; For each final event cluster C, output: Sensor list {i|i∈C}; time label interval between first appearance and last appearance [t min ,t max ]; Peak index

Images