Water quality abnormal fluctuation monitoring method and system

Abstract

The invention relates to a water quality abnormal fluctuation monitoring method and system, and the method comprises the steps: data collection, data preprocessing, feature extraction, feature fusion, anomaly detection model construction, anomaly threshold setting, and anomaly detection: inputting a comprehensive feature vector into the anomaly detection model, and judging whether a monitoring section is in an abnormal fluctuation state or not. According to the invention, through cooperative work of multiple modules, comprehensive and accurate collection and analysis of water quality data are realized. Data preprocessing ensures data quality, and feature extraction and fusion provide powerful support for anomaly detection. Abnormal fluctuation of a monitored section can be accurately identified, and the water quality abnormity degree is ranked by combining the over-threshold frequency and the fluctuation amplitude. And meanwhile, the threshold is periodically updated by considering natural condition changes, so that the anomaly recognition accuracy is greatly improved. A reliable basis is provided for water quality monitoring and management, and protection and treatment of a water environment are effectively assisted.

Description

Technical Field

[0001] The present invention relates to the technical field of environmental protection information analysis and water quality monitoring, and particularly relates to a method and system for monitoring abnormal fluctuations in water quality. Background Art

[0002] With the acceleration of the global industrialization process and the advancement of the urbanization process, the problem of water body pollution has become increasingly severe. To ensure public health and ecological safety, governments and environmental protection agencies in various countries have continuously strengthened water quality monitoring efforts and constructed a wide range of online water quality monitoring networks. The online water quality monitoring system installs highly sensitive sensors at key positions in water bodies such as rivers, lakes, and reservoirs, and collects key water quality parameters such as dissolved oxygen (DO), pH value, conductivity, ammonia nitrogen (NH3-N), permanganate index (CODmn), total phosphorus (TP), and total nitrogen (TN) in real time, and transmits the data to the central database for analysis and processing. This process not only improves the efficiency of water quality management but also provides valuable data support for environmental science research.

[0003] In recent years, relevant national departments have built a large number of online water quality monitoring stations at major rivers, lakes, reservoirs, etc. across the country, and monitored the water quality of each section at regular intervals. Therefore, a huge amount of water quality monitoring data has been accumulated. These data are not only an important basis for evaluating the water environment quality but also provide rich resources for studying the laws of water quality changes and predicting future trends. However, how to effectively manage and utilize these massive data has become one of the major challenges currently faced. Traditional data analysis methods are difficult to handle such large-scale data sets, especially in capturing complex patterns and non-linear relationships, and there is an urgent need to develop more advanced data analysis technologies and tools.

[0004] Existing methods for identifying abnormal fluctuations in water quality mainly rely on simple threshold comparison, that is, comparing a single detection value with various emission standards. If the detection value is within the emission standard range, it is considered normal; if it exceeds this range, it is determined to be abnormal. Although this method is simple and easy to implement, it has many limitations in practical applications:

[0005] Fixed thresholds cannot adapt to water quality changes under different seasons and climatic conditions, and are prone to false positives or false negatives.

[0006] Lack of a dynamic adjustment mechanism, failure to fully consider the influence of historical data and environmental factors, and difficult to reflect the real situation of water quality changes.

[0007] Ignore time series characteristics, do not fully utilize the time domain and frequency domain information in time series data, and ignore the mutual correlation and dynamic change laws between water quality parameters.

[0008] Limited ability to capture complex patterns. Traditional methods are difficult to capture the complex patterns and non-linear relationships in water quality changes, especially insufficient in presenting periodic and seasonal changes over long periods and atypical fluctuations affected by special meteorological events. Summary of the Invention

[0009] (1) Technical problems to be solved

[0010] The technical problem to be solved by the present invention is to provide a method and system for monitoring abnormal fluctuations in water quality to accurately identify abnormal fluctuations in cross-section water quality in a timely manner.

[0011] (2) Technical solutions

[0012] To solve the above problems, in the first aspect, the present invention provides a method for monitoring abnormal fluctuations in water quality, including:

[0013] Data collection: Obtain the original data within the area to be measured; the original data includes meteorological data and water quality data;

[0014] Data preprocessing;

[0015] Feature extraction: Extract features from the preprocessed data to generate a time-domain change feature dataset and a frequency-domain change feature dataset;

[0016] Feature fusion: Fuse the time-domain change features and frequency-domain change features to generate a comprehensive feature vector;

[0017] Model construction: Construct an anomaly detection model and set an anomaly threshold;

[0018] Anomaly detection: Input the comprehensive feature vector into the anomaly detection model to determine whether the monitoring cross-section is in an abnormal fluctuation state.

[0019] Among them, the meteorological data includes rainfall time and rainfall amount; the water quality data includes water temperature, pH value, dissolved oxygen, turbidity, conductivity, permanganate index, ammonia nitrogen, total phosphorus, and total nitrogen.

[0020] Among them, the data preprocessing includes:

[0021] Data cleaning: Delete blank values, missing values, and outliers;

[0022] Data standardization: Standardize the data to unify the dimension.

[0023] Among them, in the feature extraction:

[0024] Extract time-domain features and separately construct a time-domain change feature dataset for each water quality index;

[0025] The extracted time-domain features include the concentration mean, extreme concentration, time of occurrence of extreme concentration, concentration high and low value pulse conditions, and concentration increase and decrease change rates.

[0026] Among them, in the said feature extraction:

[0027] Obtain the time series data of water quality parameters;

[0028] Perform multi-level wavelet decomposition on the time series data of water quality parameters;

[0029] Obtain the main frequency components and energy distribution.

[0030] Among them, generate a dataset of rainfall event characteristics according to the said meteorological data. The dataset of rainfall event characteristics includes start time, end time, rainfall duration, maximum hourly rainfall, total rainfall, and daily maximum rainfall.

[0031] Among them, it also includes: counting the number of times each water quality parameter exceeds its dynamic threshold during the observation period to obtain the over-threshold frequency; recording the specific value of each over-threshold to obtain the amplitude of abnormal fluctuations;

[0032] Rank the degree of abnormal water quality fluctuations for each section according to the anomaly score.

[0033] On the other hand, the present invention provides a water quality abnormal fluctuation monitoring system, including:

[0034] A data acquisition module for obtaining the original data within the area to be measured; the original data includes meteorological data and water quality data;

[0035] A data preprocessing module for preprocessing the obtained original data;

[0036] A feature extraction module: for extracting features from the preprocessed data to generate a dataset of rainfall event characteristics, a time-domain change feature dataset, and a frequency-domain change feature dataset;

[0037] A feature fusion module: for fusing the rainfall event characteristic dataset, time-domain change features, and frequency-domain change features to generate a comprehensive feature vector;

[0038] A model construction module: for constructing an anomaly detection model and setting an anomaly threshold;

[0039] An anomaly detection module: for inputting the comprehensive feature vector into the anomaly detection model to determine whether the monitored section is in an abnormal fluctuation state.

[0040] Among them, it also includes: a user interaction and early warning management platform for displaying data and sending alarm notifications.

[0041] (III) Beneficial effects

[0042] The water quality monitoring system and method realize the comprehensive and accurate collection and analysis of water quality data through the collaborative work of multiple modules. Data preprocessing ensures data quality, and feature extraction and fusion provide strong support for anomaly detection. It can accurately identify abnormal fluctuations in the monitoring section, and combined with the anomaly recognition mechanism of the model itself and dynamic threshold adjustment, greatly improve the accuracy of anomaly recognition. It provides a reliable basis for water quality monitoring and management, and effectively contributes to the protection and treatment of the water environment. Brief Description of the Drawings

[0043] Figure 1 It is a graph showing the change of water quality data of each section over time;

[0044] Figure 2 It is a wavelet decomposition diagram of a certain section (station M);

[0045] Figure 3 It is a box plot of the change of TP in each section;

[0046] Figure 4 It is a graph showing the change of water quality of each section with daily rainfall. Detailed Implementation Manner

[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention. The formulas involved in the following process are described by taking total nitrogen as an example, and other water quality data can be equivalently processed.

[0048] The water quality monitoring system includes:

[0049] The data acquisition module: It is composed of a water quality monitoring data acquisition unit and a meteorological data acquisition unit, and is used to acquire the original data in the area to be measured. The original data includes meteorological data and water quality data. The meteorological data acquisition unit acquires meteorological data including rainfall time, rainfall amount, and rainfall intensity through integrating the data interfaces of national meteorological stations or third-party meteorological services. So as to better understand the reasons for water quality changes. The water quality monitoring data acquisition unit collects water quality data including water temperature, pH value, dissolved oxygen, turbidity, conductivity, permanganate index, ammonia nitrogen, total phosphorus, and total nitrogen by deploying a sensor network at each monitoring point. The sensor should have a self-calibration function to ensure the accuracy of the data.

[0050] Among them, the meteorological data can be expressed as: the time series data of rainfall amount P=(p1, p2, p i ,..., p n ), where p iDenote the rainfall at the $i$-th time point, including timestamp information. For example, $P = [{'time': '2024-01-01 08:00:00', 'rainfall_amount': 0.5 \}, {'time': '2024-01-01 09:00:00', 'rainfall_amount': 0.2 \},...]

[0051] Water quality data: The time series data of total phosphorus (TP) is $TN=(TP_1, TP_2, TP i, ..., TP m ), where $TP i Denote the total phosphorus concentration at the $i$-th time point (unit: mg / L), including timestamp information. $TP = [{'time': '2024-01-01 08:00:00', 'total_phosphorus': 3.3 \}, {'time': '2024-01-01 12:00:00', 'total_phosphorus': 3.5 \},...]

[0052] Data preprocessing module: It includes a data cleaning unit and a data standardization unit, which are used to preprocess the acquired raw data. The data cleaning unit deletes blank values, missing values, and outliers from the acquired raw data; for outliers, statistical methods such as Z-score, box plot, etc. are used for identification and it is decided whether to retain them according to the actual situation. Blank values refer to that a certain field in the data record has no information filled. Missing values refer to that some data that should exist in the data record fail to be collected or recorded, usually represented by a specific symbol NA. Outliers refer to data points that deviate significantly from the normal range compared with other observed values, which may be caused by measurement errors, data entry errors, or special events. The data standardization unit standardizes the data to unify the dimension. It enables data from different sources to be comparable with each other and reduces the impact of outliers on model training.

[0053] Standardize the total phosphorus data to unify the dimension. The standardization formula is:

[0054]

[0055] where $\mu$ is the mean of the total phosphorus data and $\sigma$ is the standard deviation of the total phosphorus data.

[0056] Feature extraction module: Extract features from the preprocessed data to generate a rainfall event feature dataset, a time-domain change feature dataset, and a frequency-domain change feature dataset;

[0057] Specifically, it includes:

[0058] Rainfall event feature dataset: $Q=(q_1, q_2, q_3,..., qn ), where n can be set to 6, q1 = start time, q2 = end time, q3 = rainfall duration, q4 = maximum hourly rainfall, q5 = total rainfall, q6 = maximum daily rainfall. Construct the rainfall feature dataset R during the evaluation period based on the rainfall feature dataset of each rainfall event: R = (r1, r2,..., r n ), where n can be set to 3, r1 = number of rainfall events, r2 = total rainfall, r3 = maximum hourly rainfall.

[0059] Extract time-domain features and construct a time-domain change feature dataset for each water quality indicator separately;

[0060] The extracted time-domain features include five categories of evaluation indicators: mean concentration, extreme concentration, time of occurrence of extreme concentration, high and low value pulse conditions of concentration, and concentration increase and decrease rate. Each category of indicators can include several sub-categories of indicators. For example, construct an evaluation system for the water quality indicator total phosphorus (TP). This system will include multiple indicators to describe and quantify the change of total phosphorus over time.

[0061] Time-domain change feature set S:

[0062] Mean concentration: Calculate the average concentration at different time scales, such as 1-day average concentration, 3-day average concentration, 7-day average concentration, 30-day average concentration, etc.

[0063]

[0064] Extreme concentration: Calculate the minimum and maximum concentrations.

[0065] Minimum concentration = min(x'1, x'2,…, x' m )

[0066] Maximum concentration = max(x'1, x'2,…, x' m )

[0067] Time of occurrence of extreme concentration: Record the time points at which the minimum and maximum concentrations occur.

[0068] Minimum concentration time = argmin(x'1, x'2,…, x' m )

[0069] Maximum concentration time = arg max(x'1, x'2,…, x' m )

[0070] High and low concentration pulse conditions: Use the sliding window method to extract local extrema point by point. Set the window size to 5 (i.e., the interval is 5 data points), find the local maximum and local minimum within each window, and calculate the number of high-concentration and low-concentration pulses. For each sub-interval, sequentially determine whether each data point meets the conditions for the local maximum or minimum. Use the argrelextrema function in the scipy package in Python. Adjust the window size as needed and set the window parameter to 5. Record the local extremum points and their values within each sub-interval

[0071] Rate of change of concentration increase and decrease: Calculate the rate of change of concentration increase and decrease

[0072]

[0073] In addition, obtain the time series data of water quality parameters, perform multi-level wavelet decomposition on it, preferably use the Daubechies wavelet function, extract the approximation coefficients and detail coefficients, and then analyze the main frequency components and energy distributions at different scales. Perform multi-level wavelet decomposition on the time series data of water quality parameters; it can preferably be the Daubechies wavelet function, extract the approximation coefficients and detail coefficients, and then analyze the main frequency components and energy distributions at different scales

[0074] Frequency domain change feature set P: Perform multi-level wavelet decomposition on the total nitrogen time series data using the Daubechies wavelet function. Extract the approximation coefficients and detail coefficients, and analyze the main frequency components and energy distributions at different scales

[0075] DWT(T) = {C(A), C(D)}

[0076] Among them, C(A) is the approximation coefficient (Approximation Coefficients), representing the characteristics of the signal in the low-frequency part. C(D) is the detail coefficient (Detail Coefficients), representing the characteristics of the signal in the high-frequency part

[0077] Feature fusion module: Used to fuse rainfall features, time domain change features, and frequency domain change features to generate a comprehensive feature vector; fuse the rainfall feature R, time domain change feature set S, and frequency domain change feature set P to generate a comprehensive feature vector X. X = R ∪ S ∪ P

[0078] Model construction module: Used to construct an anomaly detection model and set an anomaly threshold; Machine learning or statistical methods such as clustering, support vector machines, neural networks, etc. can be selected to construct the anomaly detection model. In this case, the IsolationForest method is used for construction. Initialize an IsolationForest model. For a data point x, the path length h(x) in the isolation tree is defined as the number of edges passed from the root node to the leaf node containing the data point. For a dataset containing n samples, its average path length c(n) can be approximately calculated by the harmonic number. The definition of the harmonic number H(i) is:

[0079]

[0080] where γ≈0.5772 is the Euler-Mascheroni constant.

[0081] The calculation formula for the average path length c(n) is:

[0082]

[0083] The anomaly score calculated by the decision_function method of IsolationForest in scikit-learn is based on the average path length of the data point in multiple isolation trees. Let the path lengths of the data point x in T isolation trees be h1(x), h2(x), …… h T (x), then the average path length is:

[0084]

[0085] The calculation formula for the anomaly score S(x) is:

[0086]

[0087] That is, the lower the score, the more likely the data point is an anomaly point. In scikit-learn, the score calculated by decision_function will be offset, and the offset is represented by the offset_ attribute. The actual calculation formula for the anomaly score is:

[0088] score(x) = s(x) - offset

[0089] The comprehensive feature vector R of the input is trained using model.fit() in Python to construct a boundary for distinguishing normal data from abnormal data. After training, scores = model.decision_function() can be used to calculate the scores. The 5th percentile of the scores is used as the threshold threshold = np.percentile(scores, 5). Through the above steps, the optimal value of the anomaly threshold can be determined. This threshold enables the anomaly detection model to have the highest accuracy on historical data and can adapt to changes in natural conditions. The dynamic threshold is adjusted according to a sliding window: the data in the most recent period (such as the most recent 30 days) is used to calculate the new threshold.

[0090] Anomaly detection module: used to input the comprehensive feature vector R into the anomaly detection model to determine whether the monitoring section is in an abnormal fluctuation state. Longitudinal comparison and transverse comparison are performed on each section separately. Longitudinal comparison is to compare and analyze the anomaly scores of each section at different times with the dynamic threshold determined in step three; transverse comparison is to count the number of times each water quality parameter exceeds its dynamic threshold during the observation period; record the specific values of each time the threshold is exceeded to obtain the amplitude of the abnormal fluctuation; determine the key abnormal fluctuation sections in the region through section ranking.

[0091] Threshold comparison calculation:

[0092] Calculate the anomaly score for each site.

[0093] Anomaly score i = model output(R i )

[0094] Longitudinal comparison to determine whether it exceeds the set anomaly threshold.

[0095]

[0096] Transverse comparison to determine whether it is an abnormal fluctuation section within the region.

[0097] Assume that there are n observation time points during the observation period, the water quality parameter is P, and its value at the i-th observation time point is V i , the dynamic threshold is T P , and the section name is S. Let C be the number of times the water quality parameter P is lower than the dynamic threshold T P .

[0098]

[0099] For each case of exceeding the threshold, the abnormal fluctuation amplitude A j (where j represents the j-th case of exceeding the threshold):

[0100]

[0101] wherein is the corresponding value at the j-th time exceeding the threshold.

[0102] The user interaction and early warning management platform is used to display data and send alarm notifications. It provides an intuitive user interface to show real-time water quality status, historical trends, abnormal events, and early warning details.

[0103] In view of the seasonal and interannual changes in natural conditions, the present invention can be optionally updated regularly with the analysis model and percentile threshold to ensure that the threshold is adjusted according to the actual water quality situation, improving the accuracy of anomaly recognition.

[0104] Embodiment 1: The TP abnormal fluctuation monitoring method provided by the embodiment of the present invention includes:

[0105] Data collection: Obtain the original data of a certain area in 2024; the original data includes 53,731 hourly meteorological data of 6 meteorological stations throughout the year in 2024 and 34,006 hourly water quality data of 17 national control sections; the meteorological data includes monitoring time and rainfall; the water quality data includes station name, monitoring time, and total phosphorus (once every 4 hours).

[0106] Data preprocessing includes: data cleaning, deleting blank values, filling missing values, and deleting outliers using the Z-score method; data standardization, performing standardization processing on the data to unify the dimension. The preprocessed TP water quality data is 33,957 pieces. The variation of water quality data of each section over time is as Figure 1 and it is difficult to directly judge which section has a more significant fluctuation degree.

[0107] Feature extraction: Set a sliding window (window length is 1 month, step size is 1 day), and each time the extraction range is the data of the previous month. Starting from February 1, 2024, perform feature extraction on the preprocessed data to generate a time-domain change feature dataset and a frequency-domain change feature dataset;

[0108] Specifically include: In the said feature extraction:

[0109] (1) Extract time-frequency domain features. Taking TP as an example, construct a time-frequency domain change feature dataset for the TP index of each section in each time window separately, and perform multi-level wavelet decomposition on the time series data of water quality parameters; obtain 22 feature parameters including concentration mean, extreme concentration, time of extreme concentration occurrence, concentration high and low value pulse situation, concentration increase and decrease rate, obtaining main frequency components, and energy distribution, etc. Among them, wavelet decomposition uses the wavedec method in the pywt library to perform multi-level wavelet decomposition on the input signal and automatically confirm the decomposition level. This data is basically decomposed into 4 - 5 levels, and the decomposition result of a single station is asFigure 2 As shown, the wavelet decomposition diagram of site M.

[0110] (2) Generate a dataset of rainfall event characteristics based on the meteorological data, and extract the number of rainfall events, total rainfall, and maximum hourly rainfall within each time window.

[0111] (3) Feature fusion: Fuse the rainfall characteristics, time-domain variation characteristics, and frequency-domain variation characteristics according to the start and end times of the window to generate a comprehensive feature vector, forming a dataset with 25×4856 rows in total.

[0112] Perform Z-score standardization on the feature fusion data for each section. The specific calculation formula is: where x is the original data, μ is the mean of the dataset, and σ is the standard deviation of the dataset. Use the standardized data to train the Isolation Forest anomaly detection model, and set the model parameters as contamination = 0.1, random_state = 42, and output the anomaly score. The lower the score, the more abnormal it is; since the shortest time of the sliding window is 1 month, the anomaly score has data starting from February 1st. Calculate the dynamic threshold using the 5% quantile of the anomaly scores in the previous 1 month, threshold = window_data['anomaly_score'].quantile(0.05), then the threshold comparison can start from March 1st.

[0113] The results of the number of times each section exceeds the dynamic threshold per month are shown in the following table:

[0114]

[0115] As Figures 3 - 4 , it can be seen that June and July are the months with the most times that TP exceeds the dynamic threshold, indicating that TP fluctuates greatly between May and July. There are also fluctuations in some sections in September and November, which is basically consistent with the Figure 1 rule of the line chart before.

[0116] For the amplitude of the anomaly score of each section exceeding the dynamic threshold, calculate the difference (anomaly score - dynamic threshold), and rank the results. The smaller the difference, the lower the ranking, indicating the greater the abnormal fluctuation amplitude. The results are shown in the following table:

[0117]

[0118]

[0119] Comparing with the box plots of the TP changes of each section in June and July, site H is indeed the section with smaller fluctuations. The means and ranges of variation of sections such as site M, site E, and site A are relatively large, and the ranking results are basically accurate.

[0120] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that not every embodiment only contains an independent technical solution. In the case where there is no conflict between the solutions, the various technical features mentioned in each embodiment can be combined in any way to form other embodiments that can be understood by those skilled in the art.

[0121] In addition, without departing from the scope of the present invention, modifying the technical solutions recorded in the foregoing embodiments, or equivalently replacing some of the technical features, does not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for monitoring abnormal fluctuations in water quality, characterized in that Including: Data collection: Obtain the original data within the area to be measured; the original data includes meteorological data and water quality data; Data preprocessing; Feature extraction: Extract features from the preprocessed data to generate a time-domain change feature dataset and a frequency-domain change feature dataset; Feature fusion: Fuse the time-domain change features and frequency-domain change features to generate a comprehensive feature vector; Model construction: Construct an anomaly detection model; Anomaly detection: Input the comprehensive feature vector into the anomaly detection model to determine whether the monitoring section is in an abnormal fluctuation state.

2. The water quality abnormal fluctuation monitoring method according to claim 1, wherein The meteorological data includes rainfall time and rainfall amount; the water quality data includes water temperature, pH value, dissolved oxygen, turbidity, conductivity, permanganate index, ammonia nitrogen, total phosphorus, and total nitrogen.

3. The water quality abnormal fluctuation monitoring method according to claim 2, wherein, The data preprocessing includes: Data cleaning: Delete blank values, missing values, and outliers; Data standardization: Standardize the data to unify the dimension.

4. The water quality abnormal fluctuation monitoring method according to claim 3, characterized in that In the feature extraction: Extract time-domain features and separately construct a time-domain change feature dataset for each water quality index; The extracted time-domain features include concentration mean, extreme concentration, time of extreme concentration occurrence, concentration high and low value pulse conditions, and concentration increase and decrease change rates.

5. The water quality anomaly fluctuation monitoring method according to claim 4, characterized in that In the feature extraction: Obtain the time series data of water quality parameters; Perform multi-level wavelet decomposition on the time series data of water quality parameters; Obtain the main frequency components and energy distribution.

6. The water quality abnormal fluctuation monitoring method according to claim 5, characterized in that, Generate a rainfall event feature dataset according to the meteorological data, and the rainfall event feature dataset includes start time, end time, rainfall duration, maximum hourly rainfall amount, total rainfall amount, and daily maximum rainfall amount.

7. The water quality anomaly fluctuation monitoring method according to claim 6, characterized in that, Also included: Count the number of times each water quality parameter exceeds its dynamic threshold during the observation period to obtain the over-threshold frequency; Record the specific values of each over-threshold to obtain the amplitude of abnormal fluctuations; Rank the degree of water quality fluctuation anomalies for each section according to the abnormal fluctuation scores.

8. Water quality abnormal fluctuation monitoring system, characterized in that, Including: A data collection module for obtaining the original data within the area to be measured; the original data includes meteorological data and water quality data; A data preprocessing module for preprocessing the obtained original data; A feature extraction module: for extracting features from the preprocessed data to generate a rainfall event feature dataset, a time-domain change feature dataset, and a frequency-domain change feature dataset; A feature fusion module: for fusing the rainfall event feature dataset, time-domain change features, and frequency-domain change features to generate a comprehensive feature vector; A model construction module: for constructing an anomaly detection model and setting an anomaly threshold; An anomaly detection module: for inputting the comprehensive feature vector into the anomaly detection model to determine whether the monitoring section is in an abnormal fluctuation state.

9. The water quality anomaly fluctuation monitoring system according to claim 8, characterized in that, Also included: A user interaction and early warning management platform for displaying data and sending alarm notifications.

Images