Water quality ai real-time monitoring system and method based on multi-source sensors

Abstract

The application discloses a water quality AI real-time monitoring system and method based on a multi-source sensor, relates to the technical field of environmental monitoring, and is characterized in that monitoring stations are arranged along the water flow direction in a target water area, each station is configured with a sensor group and is connected through an edge computing gateway. The gateway regularly collects sensor data, removes outliers, uniformly standardizes positive and negative indicators, and then calculates the comprehensive health score of each station. By judging whether the score significantly decreases in a continuous period and whether similar deterioration trends appear in downstream stations, the occurrence of a pollution event is determined when both conditions are met. Then, the health score spatial gradient between adjacent stations is calculated, and the interval where the pollution source is located is located according to the maximum gradient. For the identified pollution event, the decline rate and influence span are extracted to construct a comprehensive response score, and a hierarchical response mechanism is started. Meanwhile, a random forest model is trained using historical pollution event data to predict the type of pollution source in real time, providing decision support for on-site disposal.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring technology, specifically to a water quality AI real-time monitoring system and method based on multi-source sensors. Background Technology

[0002] With the rapid development of the social economy and the continuous increase in water environment pressure, real-time and accurate water quality monitoring and pollution early warning for rivers and other water bodies have become an important requirement for environmental protection and water resource management. Traditional water quality monitoring mainly relies on manual periodic sampling and site analysis, which has defects such as low monitoring frequency, limited spatiotemporal coverage, and delayed data feedback. It is difficult to capture sudden pollution events and the migration and diffusion process of pollutants in a timely manner, and cannot meet the needs of modern refined and intelligent water environment management. Furthermore, the lack of quantitative analysis of the dynamic characteristics such as the speed of pollution deterioration and the scope of impact also leads to insufficient basis for emergency response.

[0003] To address the aforementioned issues, there is an urgent need for a comprehensive water quality monitoring and analysis method that integrates high-quality data processing, reliable pollution event identification, preliminary pollution source location, quantitative emergency assessment, and intelligent identification of pollution source types. Summary of the Invention

[0004] The purpose of this invention is to provide a real-time water quality monitoring system and method based on multi-source sensors using AI, in order to solve the problems raised in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a water quality AI real-time monitoring method based on multi-source sensors, the method comprising the following steps: Step S100: Along the water flow direction in the target water area, monitoring stations are set up at fixed intervals L, with each station forming a sensor group S. k The sensors are installed at hydrologically representative depths below the water surface. Each sensor group includes at least a pH sensor, a dissolved oxygen sensor, a water temperature sensor, a conductivity sensor, and a turbidity sensor. All sensors are connected to the site's edge computing gateway via analog interfaces. The edge computing gateway integrates a wireless communication module and is responsible for data acquisition, temporary storage, and uploading to the cloud data center.

[0006] Step S200: The edge gateway synchronously reads and uploads the parameters of each sensor in each sensor group at a fixed sampling period Δt. Then, it fills in the gaps by using the sliding window statistical method, classifies them according to positive and negative indicators, and maps all sensor parameters to the [0,1] interval for standardization to eliminate differences in dimensions and numerical ranges, thus obtaining the standardized sensor parameter value z. i Ultimately, it outputs high-quality, continuous, and comparable data, providing reliable input for subsequent water quality health score calculation and pollution source tracing analysis, and building a bridge between the sensor's raw data and subsequent algorithm analysis; Step S201: Instantaneous disturbances in the water body generate anomalous data points. These anomalous data points can severely interfere with subsequent statistical analysis, so they need to be identified and removed first. A sliding window statistical method is used to judge the rationality of new data based on the statistical characteristics of recent continuous data, avoiding isolated extreme values ​​from affecting the overall analysis, while ensuring the real-time nature of data judgment. Specifically, the window length is defined as N. For each new data point included in the window, the oldest data point is discarded, always maintaining N consecutive sampled values ​​within the window. For the data sequence {X1, X2, ..., X...} collected by each sensor in the sensor group within the current window... y} Calculate the arithmetic mean μ and standard deviation σ. The arithmetic mean μ reflects the central tendency of the data within the window, and the standard deviation σ reflects the dispersion of the data within the window. When a new data point X... y Satisfying the inequality |X y When -μ∣>k×σ, it is determined to be an outlier. In the formula, the coefficient k is a preset threshold. The outlier is directly removed after being marked and does not participate in the subsequent comprehensive scoring. The data at this position is replaced with a missing identifier to ensure the integrity of the time series. Step S202: The dimensions and numerical ranges of different sensor parameters vary greatly. Directly incorporating them into the comprehensive scoring would lead to sensor parameters with larger values ​​dominating the scoring results. By uniformly mapping the values ​​of all sensor parameters to the [0,1] interval, the differences in dimensions and ranges are eliminated, making the health status of each sensor parameter comparable and providing a fair input for subsequent weighted calculations. For the positive index x... i : ; For the negative index x i : ; In the formula, x i X represents the measured value of the i-th sensor in the sensor group. max,i X represents the upper limit threshold of the i-th sensor. min,i z represents the lower threshold of the i-th sensor. i These are standardized sensor parameter values, ranging from [0,1], directly representing the health status of the sensor parameter, z. i The closer the value is to 1, the better the water quality corresponding to the sensor; the closer it is to 0, the more severe the pollution.

[0007] Step S300: For each sensor group, the standardized results of multiple independent water quality parameters are integrated into a quantitative score that can intuitively reflect the water quality health status of the location, and the comprehensive health score H of each sensor group is calculated. ; In the formula, n represents the total number of sensor parameters participating in the comprehensive scoring of the current sensor group, ε is a very small positive number to prevent calculation failure, and w i is the weight coefficient of the i-th sensor parameter, representing the importance of the sensor parameter to water quality health. The sum of the weights of all parameters is 1. H represents the comprehensive health score, H∈[0,100].

[0008] Step S400: Trigger a real pollution event through dual conditions, and then start the location program to analyze and lock the most likely range of the pollution source. Specifically, the system first determines the pollution event through dual conditions to avoid false triggering. The dual composite conditions include single-point mutation conditions and spatial propagation conditions. After the above dual conditions are met simultaneously, the location process is started to obtain the health score and coordinates of the stations along the water flow at the time of pollution development, calculate the deviation between the score and the background score, and eliminate the natural water quality difference. Then, the absolute value of the spatial gradient of adjacent stations is calculated to characterize the severity of water quality deterioration. Finally, the range with the largest absolute gradient value is locked, which is the most likely range of the pollution source. The result is used for subsequent analysis. The location accuracy is determined by the spacing of the sensor deployment. This result will serve as a key input for subsequent pollution dynamic feature extraction and pollution source type identification. Step S401: Identifying real pollution events and avoiding false alarms caused by sensor malfunctions, natural fluctuations in water bodies, etc., is a necessary prerequisite for starting the pollution source location process. Starting the location process requires meeting two conditions at the same time. The two conditions include the single-point mutation condition and the spatial propagation condition. The single-point mutation condition: any sensor group S k Comprehensive health score H k Within m consecutive sampling periods, its cumulative decrease value exceeds the preset threshold ΔH. th That is: H k (t)-H k (t-mΔt)≤-ΔH th In the formula, H k (t) is S k The overall health score at the current time t; The spatial propagation condition is as follows: in the current sensor group S k Downstream, at least one adjacent station has a cumulative decrease value exceeding a preset threshold ΔH over the subsequent m consecutive sampling periods. th ; Step S402: When the aforementioned dual conditions are met, the positioning procedure is initiated, specifically by: taking the comprehensive health scores of all sensor groups arranged from upstream to downstream along the water flow direction, forming a sequence H1, H2, ..., H k Simultaneously, record the spatial coordinates y1, y2, ..., y1 of each station corresponding to the direction of water flow. kThis is used to obtain spatial distribution data of water quality health scores across the entire basin at the same time, providing basic support for subsequent gradient calculations. Calculate the current overall health score and background score H for each site. bg The difference ΔH i =H i -H bg In the formula, H bg It is a comprehensive health score of the current normal state of the sensor group, for each pair of adjacent sensor groups S k and S k+1 Calculate the change in rating deviation per unit distance, i.e., the spatial gradient G. k : k=1,2,...,N-1; In the formula, the gradient G k The value is negative, and its absolute value represents the magnitude of the change from S. k To S k+1 The degree of water quality deterioration within the interval, the absolute value of the gradient |G k | Size, directly corresponding to x i To x i+1 The degree of water quality deterioration per unit distance within the interval, |G k The larger the value of |, the more drastic the water quality change within that range, and the closer it is to the pollution source. Compare the absolute values ​​of the gradients |Gk| across all N-1 intervals and find the maximum value among them: ; In the formula, k * It is the final solution that makes the absolute value of the gradient |G| the one that is found. k | The index number of the maximum interval indicates that the pollution source is located at the station S where the gradient is maximized. k * With S k * +1 Between, the corresponding spatial interval is [y k * y k * +1 The positioning accuracy of this method is determined by the sensor spacing L. The smaller the spacing, the narrower the positioning range and the higher the accuracy; the larger the spacing, the wider the positioning range and the relatively lower the accuracy.

[0009] Step S500: For pollution events, dynamic feature extraction and emergency response assessment are performed. By extracting the decline rate α and the impact span Δx, a comprehensive response score R is constructed to achieve graded emergency response and dynamic adjustment. Based on the R value, there are four response levels: Level 1 (minor pollution, recording and monitoring); Level 2 (moderate pollution, early warning and inspection); Level 3 (severe pollution, activation of the emergency plan); and Level 4 (major pollution, highest level response). The system updates the R value in real time, automatically upgrading or suggesting downgrading when conditions are met, adapting to the pollution situation. Step S501: For the identified pollution event, extract two core features: the rate of decline α and the range of influence Δx, which respectively characterize the impact intensity and spatial impact range of the pollution. The rate of decline α refers to the maximum instantaneous rate of decline of the water quality health score during the decline phase. The larger the value, the more violent the impact of the pollutants on the water body and the faster the pollution occurs. ; In the formula, H(ti) is the comprehensive health score corresponding to time ti in the i-th monitoring period, and t i+1 -t i H(t) represents the time interval between two adjacent monitoring periods, which is equal to the data acquisition period Δt. i+1 ) represents the time t of the (i+1)th monitoring cycle. i+1 The corresponding comprehensive health score; The impact span Δx refers to the spatial range of the pollution plume along the direction of water flow. The larger the value, the longer the length of the water area affected by the pollution and the wider the affected area. Δx = v × T, where v represents the average flow velocity of the target water area and T is the duration of the pollution event. The comprehensive response score R is calculated. The comprehensive response score R is a weighted linear model constructed based on two features: the rate of decline α and the impact span Δx. It is used to comprehensively quantify the severity of pollution events. The score ranges from 0 to 100, with a higher score indicating a more severe pollution event. R=W1×S α +W2×S Δx ; In the formula, W1 and W2 are preset weights, W1 + W2 = 1, S α It is the rate of descent score, S Δx It affects the span score; The S α Based on the preset descent rate threshold range [α] min ,α max Linear normalization is performed using the following formula: ; The S Δx Based on the preset influence span threshold range [Δx] min ,Δxmax Linear normalization is performed using the following formula: ; Step S502: Based on the range of the comprehensive response score R, the pollution event is divided into four response levels, each corresponding to a standardized emergency response procedure, to achieve a tiered response, specifically as follows: When R1≤R<R2, it is a Level 1 response, indicating that the current water quality is slightly polluted with a limited impact range. Monitoring is initiated and the frequency of patrols is increased. When R2≤R<R3, it is a Level 2 response, indicating that the current water quality is moderately polluted. Early warning information is sent to environmental protection duty personnel, and on-site patrols are organized. When R3≤R<R4, it is a Level 3 response, indicating that the current water quality is severely polluted. Emergency resources are mobilized to carry out pollution blocking and spread control. When R>R4, it is a Level 4 response, indicating that the current water quality is severely polluted. The emergency plan is activated, the water intake is shut down, multiple departments are coordinated to handle the situation, and comprehensive control and environmental remediation measures are taken. R1, R2, R3, and R4 are preset thresholds, where R1<R2<R3<R4.

[0010] Step S600: Construct a multi-dimensional feature vector F and correlate it with the verified real pollution source type label L. true Together, they are stored as labeled samples in the cloud database. Then, the random forest algorithm is used, with the feature vector F as input and the category label L as output, to divide the dataset for training and validation of the model. When a new pollution event is confirmed, the system automatically extracts the feature vector and inputs it into the deployed model. Finally, the category with the highest probability is used as the prediction result L. When a pollution event occurs, the system provides the prediction result of the pollution source type simultaneously to assist on-site personnel. Step S601: For pollution events determined in real time in history, construct a multi-dimensional feature vector F as the model input. The vector F integrates information from multiple sources, F=[S α S Δx k * , |G k * ∣], in the formula, ∣G k * | is the gradient G calculated in step S402. k The gradient value of the interval with the largest absolute value directly represents the severity of the sudden change in water quality near the pollution source; a labeled sample (F, L) is generated for pollution events determined in real time in history. true All labeled samples are stored in a cloud-based sample database. Among the labeled samples, L... true These are labels indicating the actual types of pollution sources, used to identify the types of pollution. Step S602: Using the Random Forest algorithm, with the feature vector F as input and the pollution source type label L as output, the model is randomly divided into training, validation, and test sets according to a preset ratio for training. When a new pollution event is confirmed, F is input into the deployed Random Forest algorithm. Each decision tree within the model makes an independent prediction, and finally, the probability distribution P of all pollution source type labels is output through a voting mechanism. j =[P1,P2,...,P m ], where P j This represents the probability that the event belongs to the j-th type of pollution source. The type with the highest probability value is taken as the pollution source type L. When a pollution event occurs, the system not only issues a level 4 response, but also provides the pollution source type prediction result L to assist pollution disposal personnel.

[0011] A water quality AI real-time monitoring system based on multi-source sensors includes a monitoring network deployment module, a data acquisition and preprocessing module, a health score fusion module, a pollution event determination and source tracing module, an event assessment and graded response module, and a pollution source type prediction module. The monitoring network deployment module is the physical and network foundation of the system. It is responsible for deploying multiple monitoring stations along the water flow direction. Each station integrates multiple water quality sensors to form a sensor group. Data is aggregated and preliminarily processed through an edge computing gateway to build a distributed real-time sensing network covering the target water area. The data acquisition and preprocessing module is responsible for real-time data acquisition and cleaning. The edge gateway reads and uploads data from each sensor group at fixed intervals. Outliers are removed by sliding window and statistical methods. Then, all sensor parameters are divided into positive and negative indicators. The corresponding normalization functions are applied to the positive and negative indicators respectively, and all data are uniformly mapped to the [0,1] interval to provide standardized and high-quality data input for subsequent fusion calculations. The data acquisition and preprocessing module includes an abnormal data detection and cleaning unit and a multi-indicator classification and standardization unit. The abnormal data detection and cleaning unit is based on a sliding window. It calculates the mean and standard deviation of the data sequence within the window, determines whether a new data point is an outlier by using a preset threshold, marks and removes the identified outliers, and replaces them with empty identifiers to ensure the accuracy and robustness of the data stream. The multi-index classification and standardization unit classifies all sensor parameters into positive and negative indices based on the positive and negative impacts of the indices on water quality health. It then applies corresponding normalization functions to the positive and negative indices respectively, mapping all data uniformly to the [0,1] interval to eliminate dimensional differences.

[0012] The health score fusion module calculates a comprehensive health score for each sensor group using preprocessed and standardized multi-source sensor parameters and a preset fusion algorithm (such as weighted average). This score reflects the overall water quality health status of each local monitoring point in real time and intuitively. The pollution event determination and source tracing module confirms the occurrence of pollution events by setting single-point mutation conditions and spatial propagation conditions to avoid false alarms. After confirming a pollution event, it calculates the spatial gradient based on the changes in the comprehensive health scores of each station along the water flow direction. By locating the interval with the largest absolute value of the gradient, it locks the range of pollution sources, realizing the upgrade from "alarm" to "source tracing". The pollution event determination and tracing module includes a dual-condition determination unit and a spatial gradient positioning unit; The dual-condition determination unit sets two conditions to confirm a real contamination event. These two conditions include a single-point mutation condition and a spatial propagation condition. The single-point mutation condition is: any sensor group S... k The comprehensive health score, within a preset continuous m sampling period, shows a cumulative decrease exceeding a preset threshold; the spatial propagation condition is: in the current sensor group S... k Downstream, one or more adjacent monitoring stations have a cumulative decrease value exceeding a preset threshold in the subsequent m consecutive sampling periods; The spatial gradient localization unit is activated after a pollution event is confirmed, calculating the change in health score per unit distance between each pair of adjacent stations along the water flow direction, and the spatial gradient G. k By comparing the absolute values ​​of the gradients across all intervals, the interval corresponding to the maximum absolute value of the gradient is identified as the pollution source interval.

[0013] The event assessment and graded response module is responsible for assessing the severity of confirmed pollution events and triggering corresponding emergency procedures. By quantifying two key characteristics of the sensor group's decline rate and the scope of impact in a pollution event, a comprehensive response score is constructed. Based on the comprehensive response score, the event is divided into four response levels, thereby initiating differentiated handling strategies ranging from enhanced inspections to multi-departmental collaboration. The event assessment and graded response module includes a response feature extraction and scoring unit and a multi-level response strategy execution unit; The response feature extraction and scoring unit extracts the decline rate and impact span from the identified pollution event data. The decline rate refers to the maximum instantaneous decline rate of the water quality health score during the decline phase, and the impact span refers to the spatial impact range of the pollution plume along the water flow direction. A comprehensive response score is calculated by weighting and used as a basis for classification. The multi-level response strategy execution unit classifies pollution events into response levels from Level 1 to Level 4 based on the comprehensive response score and four preset threshold ranges. Each response level corresponds to a set of standardized emergency operation procedures, and the system automatically triggers corresponding level warning information, resource allocation, and control measures instructions.

[0014] The pollution source type prediction module uses machine learning technology to provide auxiliary decision-making for on-site handling of pollution incidents. By constructing a historical sample library containing the spatiotemporal characteristics of events and training a random forest classification model, when a new event occurs, the model can predict the type of pollution source based on the event characteristics, providing key clues for rapid and accurate handling. The pollution source type prediction module includes a historical feature sample database construction unit and a random forest online prediction unit; The historical feature sample library construction unit extracts multi-dimensional feature vectors F for confirmed pollution events in history. The multi-dimensional feature vectors F include the rate of decline, the span of influence, the location of the pollution source, and the maximum spatial gradient. These are paired with real pollution source type labels to form labeled samples, which are stored in a cloud database for model training and iterative optimization. The online random forest prediction unit deploys a trained random forest classification model. When a new pollution event is confirmed, it immediately extracts the feature vector F of the pollution event and inputs it into the model. The model outputs the probability distribution of each type of pollution source and outputs the type with the highest probability as the prediction result, providing on-site personnel with auxiliary judgment of the pollution source type and helping to formulate targeted disposal plans.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. Timely and accurate pollution incident identification: By setting dual judgment conditions of "single-point mutation" and "spatial propagation", false alarms caused by single-point data anomalies or random fluctuations are avoided, improving the accuracy of pollution incident identification and enabling timely activation of response mechanisms in the early stages of pollution spread.

[0016] 2. High accuracy in locating pollution sources: By using spatial gradient analysis, the gradient change of health scores between adjacent monitoring stations is calculated, and the interval with the largest absolute gradient value is identified as the possible range of pollution sources. This enables rapid and accurate spatial location of pollution sources, which helps on-site personnel to quickly carry out source tracing and disposal.

[0017] 3. Scientific Response Decision-Making: A comprehensive response score is constructed based on the rate of decline and the scope of impact. Based on the score, pollution events are classified into response levels, which realizes the quantitative assessment and graded response of the severity of pollution, making the allocation of emergency resources more reasonable and the response measures more targeted. Attached Figure Description

[0018] Figure 1This is a schematic diagram illustrating the steps of applying the present invention to a water quality AI real-time monitoring system based on multi-source sensors; Figure 2 This is a schematic diagram of the structure of the water quality AI real-time monitoring method based on multi-source sensors according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example: Figures 1-2 As shown, this invention provides a technical solution, which sets up 5 monitoring stations (numbered S1 to S5) along the water flow direction in the downstream section of River A, with each station deploying a sensor group, and adopts a water quality AI real-time monitoring method based on multi-source sensors. The method includes the following steps: Step S100: Along the water flow direction in the target water area, monitoring stations are set up at fixed intervals L, with each station forming a sensor group S. k The sensors are installed at hydrologically representative depths below the water surface. Each sensor group includes at least a pH sensor, a dissolved oxygen sensor, a water temperature sensor, a conductivity sensor, and a turbidity sensor. All sensors are connected to the site's edge computing gateway via analog interfaces. The edge computing gateway integrates a wireless communication module and is responsible for data acquisition, temporary storage, and uploading to the cloud data center.

[0021] Step S200: The edge gateway synchronously reads and uploads the parameters of each sensor in each sensor group at a fixed sampling period Δt. Then, it fills in the gaps by using the sliding window statistical method, classifies them according to positive and negative indicators, and maps all sensor parameters to the [0,1] interval for standardization to eliminate differences in dimensions and numerical ranges, thus obtaining the standardized sensor parameter value z. i Ultimately, it outputs high-quality, continuous, and comparable data, providing reliable input for subsequent water quality health score calculation and pollution source tracing analysis, and building a bridge between the sensor's raw data and subsequent algorithm analysis; Step S201: Instantaneous disturbances in the water body generate anomalous data points. These anomalous data points can severely interfere with subsequent statistical analysis, so they need to be identified and removed first. A sliding window statistical method is used to judge the rationality of new data based on the statistical characteristics of recent continuous data, avoiding isolated extreme values ​​from affecting the overall analysis, while ensuring the real-time nature of data judgment. Specifically, the window length is defined as N. For each new data point included in the window, the oldest data point is discarded, always maintaining N consecutive sampled values ​​within the window. For the data sequence {X1, X2, ..., X...} collected by each sensor in the sensor group within the current window... y} Calculate the arithmetic mean μ and standard deviation σ. The arithmetic mean μ reflects the central tendency of the data within the window, and the standard deviation σ reflects the dispersion of the data within the window. When a new data point X... y Satisfying the inequality |X y When -μ∣>k×σ, it is determined to be an outlier. In the formula, the coefficient k is a preset threshold. The outlier is directly removed after being marked and does not participate in the subsequent comprehensive scoring. The data at this position is replaced with a missing identifier to ensure the integrity of the time series. Step S202: The dimensions and numerical ranges of different sensor parameters vary greatly. Directly incorporating them into the comprehensive scoring would lead to sensor parameters with larger values ​​dominating the scoring results. By uniformly mapping the values ​​of all sensor parameters to the [0,1] interval, the differences in dimensions and ranges are eliminated, making the health status of each sensor parameter comparable and providing a fair input for subsequent weighted calculations. For the positive index x... i : ; For the negative index x i : ; In the formula, x i X represents the measured value of the i-th sensor in the sensor group. max,i X represents the upper limit threshold of the i-th sensor. min,i z represents the lower threshold of the i-th sensor. i These are standardized sensor parameter values, ranging from [0,1], directly representing the health status of the sensor parameter, z. i The closer the value is to 1, the better the water quality corresponding to the sensor; the closer it is to 0, the more severe the pollution.

[0022] Step S300: For each sensor group, the standardized results of multiple independent water quality parameters are integrated into a quantitative score that can intuitively reflect the water quality health status of the location, and the comprehensive health score H of each sensor group is calculated. ; In the formula, n represents the total number of sensor parameters participating in the comprehensive scoring of the current sensor group, ε is a very small positive number to prevent calculation failure, and w i is the weight coefficient of the i-th sensor parameter, representing the importance of the sensor parameter to water quality health. The sum of the weights of all parameters is 1. H represents the comprehensive health score, H∈[0,100].

[0023] Step S400: Trigger a real pollution event through dual conditions, and then start the location program to analyze and lock the most likely range of the pollution source. Specifically, the system first determines the pollution event through dual conditions to avoid false triggering. The dual composite conditions include single-point mutation conditions and spatial propagation conditions. After the above dual conditions are met simultaneously, the location process is started to obtain the health score and coordinates of the stations along the water flow at the time of pollution development, calculate the deviation between the score and the background score, and eliminate the natural water quality difference. Then, the absolute value of the spatial gradient of adjacent stations is calculated to characterize the severity of water quality deterioration. Finally, the range with the largest absolute gradient value is locked, which is the most likely range of the pollution source. The result is used for subsequent analysis. The location accuracy is determined by the spacing of the sensor deployment. This result will serve as a key input for subsequent pollution dynamic feature extraction and pollution source type identification. Step S401: Identifying real pollution events and avoiding false alarms caused by sensor malfunctions, natural fluctuations in water bodies, etc., is a necessary prerequisite for starting the pollution source location process. Starting the location process requires meeting two conditions at the same time. The two conditions include the single-point mutation condition and the spatial propagation condition. The single-point mutation condition: any sensor group S k Comprehensive health score H k Within m consecutive sampling periods, its cumulative decrease value exceeds the preset threshold ΔH. th That is: H k (t)-H k (t-mΔt)≤-ΔH th In the formula, H k (t) is S k The overall health score at the current time t; The spatial propagation condition is as follows: in the current sensor group S k Downstream, at least one adjacent station has a cumulative decrease value exceeding a preset threshold ΔH over the subsequent m consecutive sampling periods. th ; Step S402: When the aforementioned dual conditions are met, the positioning procedure is initiated, specifically by: taking the comprehensive health scores of all sensor groups arranged from upstream to downstream along the water flow direction, forming a sequence H1, H2, ..., H k Simultaneously, record the spatial coordinates y1, y2, ..., y1 of each station corresponding to the direction of water flow. kThis is used to obtain spatial distribution data of water quality health scores across the entire basin at the same time, providing basic support for subsequent gradient calculations. Calculate the current overall health score and background score H for each site. bg The difference ΔH i =H i -H bg In the formula, H bg It is a comprehensive health score of the current normal state of the sensor group, for each pair of adjacent sensor groups S k and S k+1 Calculate the change in rating deviation per unit distance, i.e., the spatial gradient G. k : k=1,2,...,N-1; In the formula, the gradient G k The value is negative, and its absolute value represents the magnitude of the change from S. k To S k+1 The degree of water quality deterioration within the interval, the absolute value of the gradient |G k | Size, directly corresponding to x i To x i+1 The degree of water quality deterioration per unit distance within the interval, |G k The larger the value of |, the more drastic the water quality change within that range, and the closer it is to the pollution source. Compare the absolute values ​​of the gradients |Gk| across all N-1 intervals and find the maximum value among them: ; In the formula, k * It is the final solution that makes the absolute value of the gradient |G| the one that is found. k | The index number of the maximum interval indicates that the pollution source is located at the station S where the gradient is maximized. k * With S k * +1 Between, the corresponding spatial interval is [y k * y k * +1 The positioning accuracy of this method is determined by the sensor spacing L. The smaller the spacing, the narrower the positioning range and the higher the accuracy; the larger the spacing, the wider the positioning range and the relatively lower the accuracy.

[0024] Step S500: For pollution events, dynamic feature extraction and emergency response assessment are performed. By extracting the decline rate α and the impact span Δx, a comprehensive response score R is constructed to achieve graded emergency response and dynamic adjustment. Based on the R value, there are four response levels: Level 1 (minor pollution, recording and monitoring); Level 2 (moderate pollution, early warning and inspection); Level 3 (severe pollution, activation of the emergency plan); and Level 4 (major pollution, highest level response). The system updates the R value in real time, automatically upgrading or suggesting downgrading when conditions are met, adapting to the pollution situation. Step S501: For the identified pollution event, extract two core features: the rate of decline α and the range of influence Δx, which respectively characterize the impact intensity and spatial impact range of the pollution. The rate of decline α refers to the maximum instantaneous rate of decline of the water quality health score during the decline phase. The larger the value, the more violent the impact of the pollutants on the water body and the faster the pollution occurs. ; In the formula, H(ti) is the comprehensive health score corresponding to time ti in the i-th monitoring period, and t i+1 -t i H(t) represents the time interval between two adjacent monitoring periods, which is equal to the data acquisition period Δt. i+1 ) represents the time t of the (i+1)th monitoring cycle. i+1 The corresponding comprehensive health score; The impact span Δx refers to the spatial range of the pollution plume along the direction of water flow. The larger the value, the longer the length of the water area affected by the pollution and the wider the affected area. Δx = v × T, where v represents the average flow velocity of the target water area and T is the duration of the pollution event. The comprehensive response score R is calculated. The comprehensive response score R is a weighted linear model constructed based on two features: the rate of decline α and the impact span Δx. It is used to comprehensively quantify the severity of pollution events. The score ranges from 0 to 100, with a higher score indicating a more severe pollution event. R=W1×S α +W2×S Δx ; In the formula, W1 and W2 are preset weights, W1 + W2 = 1, S α It is the rate of descent score, S Δx It affects the span score; The S α Based on the preset descent rate threshold range [α] min ,α max Linear normalization is performed using the following formula: ; The S Δx Based on the preset influence span threshold range [Δx] min ,Δxmax Linear normalization is performed using the following formula: ; Step S502: Based on the range of the comprehensive response score R, the pollution event is divided into four response levels, each corresponding to a standardized emergency response procedure, to achieve a tiered response, specifically as follows: When R1≤R<R2, it is a Level 1 response, indicating that the current water quality is slightly polluted with a limited impact range. Monitoring is initiated and the frequency of patrols is increased. When R2≤R<R3, it is a Level 2 response, indicating that the current water quality is moderately polluted. Early warning information is sent to environmental protection duty personnel, and on-site patrols are organized. When R3≤R<R4, it is a Level 3 response, indicating that the current water quality is severely polluted. Emergency resources are mobilized to carry out pollution blocking and spread control. When R>R4, it is a Level 4 response, indicating that the current water quality is severely polluted. The emergency plan is activated, the water intake is shut down, multiple departments are coordinated to handle the situation, and comprehensive control and environmental remediation measures are taken. R1, R2, R3, and R4 are preset thresholds, where R1<R2<R3<R4.

[0025] Step S600: Construct a multi-dimensional feature vector F and correlate it with the verified real pollution source type label L. true Together, they are stored as labeled samples in the cloud database. Then, the random forest algorithm is used, with the feature vector F as input and the category label L as output, to divide the dataset for training and validation of the model. When a new pollution event is confirmed, the system automatically extracts the feature vector and inputs it into the deployed model. Finally, the category with the highest probability is used as the prediction result L. When a pollution event occurs, the system provides the prediction result of the pollution source type simultaneously to assist on-site personnel. Step S601: For pollution events determined in real time in history, construct a multi-dimensional feature vector F as the model input. The vector F integrates information from multiple sources, F=[S α S Δx k * , |G k * ∣], in the formula, ∣G k * | is the gradient G calculated in step S402. k The gradient value of the interval with the largest absolute value directly represents the severity of the sudden change in water quality near the pollution source; a labeled sample (F, L) is generated for pollution events determined in real time in history. true All labeled samples are stored in a cloud-based sample database. Among the labeled samples, L... true These are labels indicating the actual types of pollution sources, used to identify the types of pollution. Step S602: Using the Random Forest algorithm, with the feature vector F as input and the pollution source type label L as output, the model is randomly divided into training, validation, and test sets according to a preset ratio for training. When a new pollution event is confirmed, F is input into the deployed Random Forest algorithm. Each decision tree within the model makes an independent prediction, and finally, the probability distribution P of all pollution source type labels is output through a voting mechanism. j =[P1,P2,...,P m ], where P j This represents the probability that the event belongs to the j-th type of pollution source. The type with the highest probability value is taken as the pollution source type L. When a pollution event occurs, the system not only issues a level 4 response, but also provides the pollution source type prediction result L to assist pollution disposal personnel.

[0026] A water quality AI real-time monitoring system based on multi-source sensors includes a monitoring network deployment module, a data acquisition and preprocessing module, a health score fusion module, a pollution event determination and source tracing module, an event assessment and graded response module, and a pollution source type prediction module. The monitoring network deployment module is the physical and network foundation of the system. It is responsible for deploying multiple monitoring stations along the water flow direction. Each station integrates multiple water quality sensors to form a sensor group. Data is aggregated and preliminarily processed through an edge computing gateway to build a distributed real-time sensing network covering the target water area. The data acquisition and preprocessing module is responsible for real-time data acquisition and cleaning. The edge gateway reads and uploads data from each sensor group at fixed intervals. Outliers are removed by sliding window and statistical methods. Then, all sensor parameters are divided into positive and negative indicators. The corresponding normalization functions are applied to the positive and negative indicators respectively, and all data are uniformly mapped to the [0,1] interval to provide standardized and high-quality data input for subsequent fusion calculations. The data acquisition and preprocessing module includes an abnormal data detection and cleaning unit and a multi-indicator classification and standardization unit. The abnormal data detection and cleaning unit is based on a sliding window. It calculates the mean and standard deviation of the data sequence within the window, determines whether a new data point is an outlier by using a preset threshold, marks and removes the identified outliers, and replaces them with empty identifiers to ensure the accuracy and robustness of the data stream. The multi-index classification and standardization unit classifies all sensor parameters into positive and negative indices based on the positive and negative impacts of the indices on water quality health. It then applies corresponding normalization functions to the positive and negative indices respectively, mapping all data uniformly to the [0,1] interval to eliminate dimensional differences.

[0027] The health score fusion module calculates a comprehensive health score for each sensor group using preprocessed and standardized multi-source sensor parameters and a preset fusion algorithm (such as weighted average). This score reflects the overall water quality health status of each local monitoring point in real time and intuitively. The pollution event determination and source tracing module confirms the occurrence of pollution events by setting single-point mutation conditions and spatial propagation conditions to avoid false alarms. After confirming a pollution event, it calculates the spatial gradient based on the changes in the comprehensive health scores of each station along the water flow direction. By locating the interval with the largest absolute value of the gradient, it locks the range of pollution sources, realizing the upgrade from "alarm" to "source tracing". The pollution event determination and tracing module includes a dual-condition determination unit and a spatial gradient positioning unit; The dual-condition determination unit sets two conditions to confirm a real contamination event. These two conditions include a single-point mutation condition and a spatial propagation condition. The single-point mutation condition is: any sensor group S... k The comprehensive health score, within a preset continuous m sampling period, shows a cumulative decrease exceeding a preset threshold; the spatial propagation condition is: in the current sensor group S... k Downstream, one or more adjacent monitoring stations have a cumulative decrease value exceeding a preset threshold in the subsequent m consecutive sampling periods; The spatial gradient localization unit is activated after a pollution event is confirmed, calculating the change in health score per unit distance between each pair of adjacent stations along the water flow direction, and the spatial gradient G. k By comparing the absolute values ​​of the gradients across all intervals, the interval corresponding to the maximum absolute value of the gradient is identified as the pollution source interval.

[0028] The event assessment and graded response module is responsible for assessing the severity of confirmed pollution events and triggering corresponding emergency procedures. By quantifying two key characteristics of the sensor group's decline rate and the scope of impact in a pollution event, a comprehensive response score is constructed. Based on the comprehensive response score, the event is divided into four response levels, thereby initiating differentiated handling strategies ranging from enhanced inspections to multi-departmental collaboration. The event assessment and graded response module includes a response feature extraction and scoring unit and a multi-level response strategy execution unit; The response feature extraction and scoring unit extracts the decline rate and impact span from the identified pollution event data. The decline rate refers to the maximum instantaneous decline rate of the water quality health score during the decline phase, and the impact span refers to the spatial impact range of the pollution plume along the water flow direction. A comprehensive response score is calculated by weighting and used as a basis for classification. The multi-level response strategy execution unit classifies pollution events into response levels from Level 1 to Level 4 based on the comprehensive response score and four preset threshold ranges. Each response level corresponds to a set of standardized emergency operation procedures, and the system automatically triggers corresponding level warning information, resource allocation, and control measures instructions.

[0029] The pollution source type prediction module uses machine learning technology to provide auxiliary decision-making for on-site handling of pollution incidents. By constructing a historical sample library containing the spatiotemporal characteristics of events and training a random forest classification model, when a new event occurs, the model can predict the type of pollution source based on the event characteristics, providing key clues for rapid and accurate handling. The pollution source type prediction module includes a historical feature sample database construction unit and a random forest online prediction unit; The historical feature sample library construction unit extracts multi-dimensional feature vectors F for confirmed pollution events in history. The multi-dimensional feature vectors F include the rate of decline, the span of influence, the location of the pollution source, and the maximum spatial gradient. These are paired with real pollution source type labels to form labeled samples, which are stored in a cloud database for model training and iterative optimization. The online random forest prediction unit deploys a trained random forest classification model. When a new pollution event is confirmed, it immediately extracts the feature vector F of the pollution event and inputs it into the model. The model outputs the probability distribution of each type of pollution source and outputs the type with the highest probability as the prediction result, providing on-site personnel with auxiliary judgment of the pollution source type and helping to formulate targeted disposal plans.

[0030] Example: Set the number of continuous sampling periods m=3h, and the cumulative decrease threshold ΔH th =0.3, the health score of the S3 site decreased by 0.35 cumulatively within 3 consecutive hours, and the downstream S4 and S5 also showed a cumulative decrease of >0.3 in the following 3 hours, which met the dual conditions and triggered the pollution location process; Set the rating sequence for each station as H=[0.85,0.82,0.50,0.45,0.40], and the location coordinates as y=[0,1.2,2.5,3.8,5.0] (unit: km). Calculate the gradient G between adjacent stations. k The maximum absolute value of the gradient was found to occur between S2 and S3, indicating that the pollution source is located in the interval y∈[1.2km,2.5km]. Set the descent rate threshold range [a] min ,a max = [0.05, 0.30], affecting the span threshold range [Δx] min ,Δx max= [0.5, 3.0], response grading thresholds R1=0.3, R2=0.5, R3=0.7, R4=0.9, calculated and measured descent rate a=0.25, influence span S Δx =2.0km, after normalization and weighted calculation, the comprehensive response score R=0.65. Since R3≤R<R4, a level 3 response is triggered: mobilize emergency resources and carry out pollution blocking and diffusion control. For this pollution incident, a feature vector F was extracted and set to F=[0.25,2.0,2,0.28]. F was input into a trained random forest model, and the output probability distribution was set as follows: industrial wastewater: 0.6, agricultural non-point source pollution: 0.3, domestic sewage: 0.1. The final predicted pollution source was industrial wastewater, which helped on-site personnel to take targeted measures.

[0031] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A water quality AI real-time monitoring method based on multi-source sensors, characterized in that: The method includes the following steps: Step S100: Set up monitoring stations along the water flow direction in the target water area, with each monitoring station forming a sensor group S. k All sensors in the sensor group are connected to the edge computing gateway of the monitoring site; Step S200: Read and upload sensor parameters from the sensor group at a fixed sampling period through the edge gateway, classify them according to positive and negative indicators, and uniformly map them to the [0,1] interval for standardization processing; Step S300: For each sensor group, calculate a comprehensive health score based on the standardized sensor parameters; Step S400: Determine the pollution event based on dual conditions. If both conditions are met simultaneously, start the location process, calculate the absolute value of the spatial gradient between adjacent stations, and lock the range with the largest absolute gradient value as the range where the pollution source is located. Step S500: For pollution events, extract the decline rate and impact span, and construct a comprehensive response score; Step S600: Using the random forest algorithm, with feature vector F as input and category label L as output, the algorithm provides prediction results of pollution source types when a pollution event occurs.

2. The water quality AI real-time monitoring method based on multi-source sensors according to claim 1, characterized in that: Step S400 includes the following steps: Step S401: The dual conditions include two conditions: single-point mutation condition and spatial propagation condition; The single-point mutation condition: any sensor group S k The comprehensive health score, within a preset continuous sampling period of m cycles, shows a cumulative decrease exceeding a preset threshold ΔH. th ; The spatial propagation condition is as follows: in the current sensor group S k Downstream, if one or more adjacent monitoring stations have a cumulative decrease exceeding a preset threshold ΔH over the subsequent m consecutive sampling periods. th ; Step S402: When the aforementioned dual conditions are met, the positioning procedure is initiated, specifically by: taking the comprehensive health scores of all sensor groups arranged from upstream to downstream along the water flow direction, forming a sequence H1, H2, ..., H k The H k Corresponding to the sensor group S k The comprehensive health score records the spatial coordinates y1, y2, ..., y of each station corresponding to the direction of water flow. k The y k Corresponding to the sensor group S k Spatial location coordinates; Calculate the difference between the current comprehensive health score and the background score for each station, where the background score is the current sensor group S. k The comprehensive health score under normal conditions is calculated for each pair of adjacent sensor groups S. k and S k+1 Calculate its spatial gradient G k The spatial gradient G k For the adjacent sensor group S k and S k+1 Change in scoring deviation per unit distance; Compare the absolute values ​​of the gradients |G across all N-1 intervals. k | Find the maximum value among them, and you will find that the pollution source is located at the site S that maximizes the gradient. k * With S k * +1 Between, the corresponding spatial interval is [y k * y k * +1 ].

3. The water quality AI real-time monitoring method based on multi-source sensors according to claim 1, characterized in that: Step S500 includes the following steps: Step S501: For the identified pollution events, extract the decrease rate α and the impact span S. Δx The rate of decline α refers to the maximum instantaneous rate of decline of the water quality health score during the decline phase, based on a preset rate of decline threshold range [α]. min ,α max Linear normalization is performed, and the influence span S Δx This refers to the spatial extent of the pollution plume's influence along the water flow direction, based on a preset influence span threshold range [Δx]. min ,Δx max Perform linear normalization; based on the descent rate α and the influence span S Δx The constructed weighted linear model is used to calculate the overall response score R; Step S502: Based on the range of values ​​for the comprehensive response score R, the pollution event is classified into four response levels, specifically as follows: When R1≤R<R2, it is a Level 1 response, indicating that the current water quality is slightly polluted with a limited impact range. Monitoring is initiated and the frequency of patrols is increased. When R2≤R<R3, it is a Level 2 response, indicating that the current water quality is moderately polluted. Early warning information is sent to environmental protection duty personnel, and on-site patrols are organized. When R3≤R<R4, it is a Level 3 response, indicating that the current water quality is severely polluted. Emergency resources are mobilized to carry out pollution blocking and spread control. When R>R4, it is a Level 4 response, indicating that the current water quality is severely polluted. The emergency plan is activated, the water intake is shut down, multiple departments are coordinated to handle the situation, and comprehensive control and environmental remediation measures are taken. R1, R2, R3, and R4 are preset thresholds, where R1<R2<R3<R4.

4. The water quality AI real-time monitoring method based on multi-source sensors according to claim 1, characterized in that: Step S600 includes the following steps: Step S601: For pollution events identified in history, construct a multi-dimensional feature vector F as the model input, wherein the vector F = [S α S Δx k * , |G k * |], the |G k * | is the gradient G calculated in step S402. k The gradient values ​​of the interval with the largest absolute values ​​are used to generate a labeled sample (F, L) for pollution events determined in real time in history. true All labeled samples are stored in a cloud-based sample database, and L in the labeled samples true These are labels indicating the actual types of pollution sources, used to identify the types of pollution. Step S602: Using the Random Forest algorithm, with the feature vector F as input and the pollution source type label L as output, when a new pollution event is confirmed, F is input into the deployed Random Forest algorithm, and the probability distribution P of all pollution source type labels is output. j =[P1,P2,...,P m ], where P j This represents the probability that the event belongs to the j-th type of pollution source. The type with the highest probability value is taken as the pollution source type L. When a pollution event occurs, the pollution source type prediction result L is provided to assist pollution disposal personnel.

5. The water quality AI real-time monitoring method based on multi-source sensors according to claim 1, characterized in that: Step S200 includes the following steps: Step S201: Read and upload sensor parameters from the sensor group at a fixed sampling period through the edge gateway. Define a window length of N. For each new data point included in the window, discard the oldest data point, always maintaining N consecutive sampled values ​​within the window. For the data sequence {X1, X2, ..., X...} collected by each sensor in the sensor group within the current window... y } Calculate the arithmetic mean μ and standard deviation σ when a new data point X y Satisfying the inequality |X y When -μ∣>k×σ, it is determined to be an outlier. In the formula, the coefficient k is a preset threshold. The outlier is directly removed after being marked, and the data at that position is replaced by a vacancy identifier. Step S202: Divide all sensor parameters into positive and negative indices, apply the corresponding normalization function to the positive and negative indices respectively, and map all sensor parameters uniformly to the [0,1] interval.

6. A water quality AI real-time monitoring system based on multi-source sensors, characterized in that: The system includes a monitoring network deployment module, a data acquisition and preprocessing module, a health score fusion module, a pollution event determination and source tracing module, an event assessment and graded response module, and a pollution source type prediction module. The monitoring network deployment module is responsible for setting up multiple monitoring stations along the waterway in the direction of water flow. Each station constitutes a sensor group, and data is aggregated through an edge computing gateway. The data acquisition and preprocessing module removes outliers using a sliding window and statistical methods. Then, it divides all sensor parameters into positive and negative indicators, applies corresponding normalization functions to the positive and negative indicators respectively, and maps all data to the [0,1] interval. The health score fusion module calculates a comprehensive health score for each sensor group using preprocessed and standardized multi-source sensor parameters. The pollution event determination and tracing module confirms the occurrence of a pollution event by setting single-point mutation conditions and spatial propagation conditions. After confirming the pollution event, it calculates the spatial gradient based on the changes in the comprehensive health scores of each station along the water flow direction, and locks the pollution source range by locating the interval with the largest absolute value of the gradient. The event assessment and graded response module constructs a comprehensive response score by quantifying the rate of decrease and the scope of impact of the sensor group in a pollution event, and classifies the event into four response levels based on the comprehensive response score. The pollution source type prediction module constructs a historical sample library containing the spatiotemporal characteristics of events and trains a random forest classification model. When a new event occurs, the model can predict the type of pollution source based on the event characteristics.

7. The water quality AI real-time monitoring system based on multi-source sensors according to claim 6, characterized in that: The data acquisition and preprocessing module includes an abnormal data detection and cleaning unit and a multi-indicator classification and standardization unit. The abnormal data detection and cleaning unit is based on a sliding window. It calculates the mean and standard deviation of the data sequence within the window, determines whether a new data point is an outlier by using a preset threshold, marks and removes the identified outliers, and replaces them with a vacancy identifier. The multi-index classification and standardization unit classifies all sensor parameters into positive and negative indices based on the positive and negative impacts of the indices on water quality health. It then applies corresponding normalization functions to the positive and negative indices respectively, mapping all data uniformly to the [0,1] interval to eliminate dimensional differences.

8. The water quality AI real-time monitoring system based on multi-source sensors according to claim 6, characterized in that: The pollution event determination and tracing module includes a dual-condition determination unit and a spatial gradient positioning unit; The dual-condition determination unit sets two conditions to confirm a real contamination event. These two conditions include a single-point mutation condition and a spatial propagation condition. The single-point mutation condition is: any sensor group S... k The cumulative decrease in the comprehensive health score exceeds a preset threshold within a preset m consecutive sampling periods; The spatial propagation condition is as follows: in the current sensor group S k Downstream, one or more adjacent monitoring stations have a cumulative decrease value exceeding a preset threshold in the subsequent m consecutive sampling periods; The spatial gradient localization unit is activated after a pollution event is confirmed, calculating the change in health score per unit distance between each pair of adjacent stations along the water flow direction, and the spatial gradient G. k By comparing the absolute values ​​of the gradients across all intervals, the interval corresponding to the maximum absolute value of the gradient is identified as the pollution source interval.

9. The water quality AI real-time monitoring system based on multi-source sensors according to claim 6, characterized in that: The event assessment and graded response module includes a response feature extraction and scoring unit and a multi-level response strategy execution unit; The response feature extraction and scoring unit extracts the decline rate and impact span from the identified pollution event data. The decline rate refers to the maximum instantaneous decline rate of the water quality health score during the decline phase, and the impact span refers to the spatial impact range of the pollution plume along the water flow direction. The comprehensive response score is calculated by weighting. The multi-level response strategy execution unit classifies pollution events into response levels from Level 1 to Level 4 based on the comprehensive response score and by comparing it with four preset threshold ranges. Each response level corresponds to a set of standardized emergency operation procedures.

10. The water quality AI real-time monitoring system based on multi-source sensors according to claim 6, characterized in that: The pollution source type prediction module includes a historical feature sample database construction unit and a random forest online prediction unit; The historical feature sample library construction unit extracts multi-dimensional feature vectors F for confirmed pollution events in history. The multi-dimensional feature vectors F include the rate of decline, the span of influence, the location of the pollution source, and the maximum spatial gradient. These feature vectors are paired with real pollution source type labels to form labeled samples, which are then stored in a cloud database. The online random forest prediction unit deploys a trained random forest classification model. When a new pollution event is confirmed, it immediately extracts the feature vector F of the pollution event and inputs it into the model. The type with the highest probability is output as the prediction result, providing on-site personnel with auxiliary judgment on the type of pollution source.

Images