Real-time dynamic water quality change trend analysis method based on multi-modal perception and deep learning

Abstract

The invention relates to the technical field of water quality analysis, in particular to a real-time dynamic water quality change trend analysis method based on multi-modal perception and deep learning. The method comprises the following steps: collecting a multi-modal sensing data set through a heterogeneous sensor cluster pre-deployed in a monitoring area; transmitting the multi-modal sensing data set to a preset edge computing node so as to execute dynamic Kalman filtering; identifying abnormal readings in the aligned dynamic Kalman filtering results, and performing self-calibration on the abnormal readings to obtain a sample data set; constructing a water quality change trend prediction model according to the sample data set; predicting the water quality change trend of the monitored area by using the water quality change trend prediction model; and matching and comparing the prediction result of the water quality change trend by using a preset grading early warning threshold value, and triggering an early warning mechanism in a grading manner according to the matching and comparison result. According to the invention, the scientificity and accuracy of identifying the water quality change in the complex water environment can be improved.

Description

Technical Field

[0001] This invention relates to the field of water quality analysis technology, and in particular to a method for analyzing real-time dynamic water quality change trends based on multimodal perception and deep learning. Background Technology

[0002] Traditional water quality monitoring methods mainly include chemical analysis methods such as silver nitrate titration for chloride ion determination, potassium dichromate oxidation for chemical oxygen demand (CODCr) determination, and spectrophotometry for color determination. Although these methods have high accuracy, they have significant drawbacks such as long sampling cycles, excessive human intervention, and poor real-time performance, making it difficult to meet the dynamic monitoring needs of modern water quality management.

[0003] With technological advancements, modern water quality monitoring systems can now continuously sample and analyze multiple core indicators, including pH, dissolved oxygen (DO), conductivity, turbidity, ammonia nitrogen, and chemical oxygen demand (COD), in real time. Multi-parameter water quality analyzers, by integrating electrochemical sensors and optical measurement technology, can simultaneously monitor basic indicators such as temperature, pH, turbidity, conductivity, and dissolved oxygen. Some models can also detect pollution indicators such as COD, ammonia nitrogen, total phosphorus, and total nitrogen.

[0004] Despite significant progress in modern water quality monitoring systems, several problems remain to be addressed. Firstly, data silos are prevalent, with a lack of effective fusion mechanisms for heterogeneous information from different sensors, such as spectral data, physicochemical parameters, and remote sensing images. This results in incomplete monitoring results and fails to fully leverage the complementary advantages of multimodal data. Secondly, the systems suffer from weak real-time dynamic analysis capabilities. Sensors require manual calibration and are susceptible to aging and contamination. Monitoring responses are delayed under extreme weather conditions, making timely warnings of sudden water quality anomalies difficult. Furthermore, system robustness significantly decreases when data transmission is interrupted. In addition, deep learning models are still limited to single-task processing in applications. Model training relies on high-quality labeled data and lacks generalization ability, hindering cross-regional deployment. They also lack interpretability and visualization support, which is detrimental to practical management decisions. Moreover, the evaluation system for existing water quality prediction models is incomplete, relying heavily on general error indicators and lacking specialized evaluation methods tailored to water quality characteristics. Research on the overall performance of multi-parameter equipment and high-dimensional uncertainty analysis remains insufficient. Summary of the Invention

[0005] Therefore, it is necessary for the present invention to provide a real-time dynamic water quality change trend analysis method based on multimodal perception and deep learning to solve at least one of the above-mentioned technical problems.

[0006] To achieve the above objectives, a real-time dynamic water quality change trend analysis method based on multimodal perception and deep learning includes the following steps: Step S1: Collect multimodal sensing datasets using a cluster of heterogeneous sensors pre-deployed in the monitoring area; Step S2: Transmit the multimodal sensing dataset to a preset edge computing node to perform dynamic Kalman filtering; identify abnormal readings in the aligned dynamic Kalman filtering results and perform self-calibration on the abnormal readings to obtain a sample dataset; Step S3: Construct a water quality change trend prediction model based on the sample dataset; Step S4: Use the water quality change trend prediction model to predict the water quality change trend in the monitoring area; use the preset graded early warning thresholds to match and compare the prediction results of the water quality change trend, and trigger the early warning mechanism in a graded manner according to the matching and comparison results.

[0007] This application achieves high-precision continuous monitoring of multiple key water quality indicators by fusing heterogeneous data such as spectral data, physicochemical parameters, and remote sensing imagery, significantly improving the comprehensiveness and real-time performance of monitoring. Utilizing a stacked two-layer LSTM and a multimodal weighted fusion model, it accurately extracts the temporal dependence and spatial distribution characteristics of water quality indicators, enabling high-precision prediction of future water quality trends and providing sufficient response time for sudden pollution events and algal blooms. Through sensor reliability weighting, edge computing nodes, and optimized protection design, the system's robustness is enhanced, maintaining continuous monitoring and prediction capabilities even in the event of network interruptions or single sensor failures. Furthermore, by combining visualization and physical model interfaces, a closed-loop linkage between monitoring data and the management system is achieved, automatically adjusting wastewater treatment operating parameters, improving management efficiency, and enabling scientific, accurate, and real-time dynamic analysis of water quality changes in complex aquatic environments. Attached Figure Description

[0008] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart illustrating the steps of the real-time dynamic water quality change trend analysis method based on multimodal perception and deep learning of the present invention. Figure 2 This is a schematic diagram of a heterogeneous sensor cluster scenario in an embodiment of the present invention; The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0009] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0010] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0011] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0012] To achieve the above objectives, please refer to Figures 1 to 2 This invention provides a method for real-time dynamic water quality change trend analysis based on multimodal sensing and deep learning, the method comprising the following steps: Step S1: Collect multimodal sensing datasets using a cluster of heterogeneous sensors pre-deployed in the monitoring area; In this embodiment, a heterogeneous sensor cluster is pre-deployed within the monitoring area, including an optical sensor array, an electrochemical sensor group, a machine vision module, and environmental parameter sensors. The optical sensor array integrates a multi-band spectral acquisition module covering ultraviolet (200-400nm), visible (400-700nm), and near-infrared (700-2500nm) wavelengths. It acquires the optical characteristics of water samples through dynamic scanning across the entire wavelength range, acquiring spectral images every 5 minutes at a resolution of 1024×1024 pixels to capture changes in water color and suspended solids concentration. The electrochemical sensor group consists of miniaturized pH, DO, and ORP sensors that combine ion-selective electrode methods with optical detection technology, such as ammonia nitrogen detection. The system employs Nessler's reagent spectrophotometry, combined with chemometric modeling to correct data. Key indicators such as dissolved oxygen, pH, conductivity, and ammonia nitrogen concentration are collected every 2 minutes, with accuracies of 0.01 mg / L, 0.01 mg / L, 1 μS / cm, and 0.01 mg / L, respectively. The machine vision module is equipped with a turbidity sensor featuring an 880nm high-performance LED light source. Combined with unique optical and electronic filtering technology, ambient light interference is eliminated, and water surface images are continuously acquired at 30fps for analyzing floating debris and water flow patterns. Environmental parameter sensors simultaneously collect auxiliary parameters such as water temperature, flow velocity, and rainfall, establishing a correlation model with water quality changes. Temperature, humidity, and flow velocity information are collected every 10 minutes. All collected data is timestamped within the sensors and transmitted in real-time to preset edge computing nodes via an edge communication gateway using the MQTT protocol to ensure data synchronization and form an initial multimodal sensing dataset.

[0013] Step S2: Transmit the multimodal sensing dataset to a preset edge computing node to perform dynamic Kalman filtering; identify abnormal readings in the aligned dynamic Kalman filtering results and perform self-calibration on the abnormal readings to obtain a sample dataset; In one implementation, after transmitting the multimodal sensing dataset to a preset edge computing node, a preset system sampling model is input for dynamic Kalman filtering. The system sampling model constructs state and observation equations based on historical sampling data. The state vector includes the instantaneous values ​​of each water quality indicator and its first-order rate of change. The covariance matrix is ​​initially estimated using historical data statistics. Dynamic Kalman filtering performs recursive calculations every 2-minute sampling period. First, initial state estimates are performed on the observations from each sensor. Then, the residuals between the predicted and actual observations are calculated. The residual variance is used to determine the weighting coefficients and iteratively update the system state noise covariance matrix until the residual term falls below a threshold of 0.05. Subsequently, the isolation score of each sensor is calculated through the residual sequence. Sampling points with an isolation score higher than 0.7 are identified as abnormal readings. The residual correction coefficient is calculated by comparing the covariance of sampling values ​​of the same type of sensor under the same environmental conditions. Linear regression correction is performed on the abnormal readings to generate reasonable correction values ​​and replace the original dynamic Kalman filter results. Finally, a calibrated sample dataset is formed. This dataset contains both standardized values ​​of numerical indicators (Z-score processing, mean 0, standard deviation 1) and textual environmental label embedding vectors (Word2Vec length 32) for subsequent model training.

[0014] It is worth noting that the edge computing nodes are deployed on edge servers or industrial gateways near the monitoring area, equipped with high-performance CPUs (such as 8-core 2.4GHz) and more than 4GB of memory, and NVIDIA Jetson or similar GPU acceleration units to support multimodal data processing and deep learning computation. The nodes communicate with the heterogeneous sensor cluster in real time via Ethernet or LoRa / Wi-Fi networks, using the MQTT protocol to receive data streams from optical sensors, multispectral images, machine vision frames, and environmental parameter sensors. All data undergoes timestamp verification and synchronization processing locally on the node, ensuring that the time alignment error of different modal data does not exceed ±5 seconds. During the execution of dynamic Kalman filtering, the edge nodes are responsible for: real-time reading of the observation vectors of each sensor and performing initial state estimation; calculating the residual between the predicted and actual observation values ​​and updating the system state noise covariance matrix; calculating the isolation score of the residual sequence to identify abnormal readings; and retrieving covariance features from locally stored historical data of similar sensors or from the edge database cache for linear regression correction of abnormal readings. By performing these calculations at edge nodes, the calibrated sample dataset can be generated and transmitted to the central water quality monitoring platform in real time to build a water quality change trend prediction model, while reducing data transmission latency and network load, and realizing near real-time dynamic water quality analysis.

[0015] Step S3: Construct a water quality change trend prediction model based on the sample dataset; In this embodiment, the time series of water quality parameters and the multispectral image sequences in the sample dataset are classified into different data types. Numerical water quality parameters are standardized using Z-score, and textual environmental labels are embedded into 32-dimensional vectors using Word2Vec. These vectors are then integrated with the numerical features into the embedding layer to output a unified 64-dimensional feature vector. Simultaneously, time step alignment is performed on data from different sampling frequencies. The time-dependent feature was set to 1 hour. Subsequently, a stacked two-layer LSTM framework was used for learning time-dependent features. The first LSTM layer had an input dimension of 64 and 128 hidden units to capture local short-term fluctuation features. The second LSTM layer had 64 hidden units to learn long-term dependencies. Dropout 0.3 was used between layers to prevent overfitting. The activation function was tanh. The Adam optimizer was used during training with an initial learning rate of 0.001, a batch size of 64, and a maximum number of iterations of 200. An early stopping strategy was triggered when the mean squared error of the validation set was lower than 0.05 or the change in the validation set loss was less than 0.01 for 10 consecutive iterations. The final output was a time-dependent feature matrix T×64, where T was 168 time steps, corresponding to the data collected in the most recent week, providing input for subsequent fusion with spatial distribution features.

[0016] Step S4: Use the water quality change trend prediction model to predict the water quality change trend in the monitoring area; use the preset graded early warning thresholds to match and compare the prediction results of the water quality change trend, and trigger the early warning mechanism in a graded manner according to the matching and comparison results.

[0017] In this embodiment, a pre-constructed water quality change trend prediction model is used to predict the future water quality change trend in the monitoring area. Spatial distribution features are extracted frame by frame from multispectral image sequences, including water color, suspended solids concentration, and floating debris distribution. These features are then combined with time-dependent features at the same time step to construct a joint input matrix. Dynamic weight coefficients are obtained by calculating the contribution of each feature source. The joint matrix is ​​then weighted and aggregated, and fuzzy adjustment and genetic iterative optimization are performed to generate the final prediction result. Simultaneously, the prediction result is matched and compared with preset graded warning thresholds. For example, a level 1 warning is set for dissolved oxygen below 5 mg / L, pH value deviating from 7.0 by more than ±0.5, and ammonia nitrogen exceeding 1 mg / L. If the predicted value exceeds the threshold for three consecutive time points, the corresponding level of warning is triggered. The warning information is pushed to the monitoring platform in real time. It is also possible to select to trace back the sample data that triggered the warning to the original collection sequence to verify the sensor effectiveness of the water quality indicators exceeding the limit, and to remove the sampling sequences in time abnormal sections or drift trend sections to form a closed-loop source tracing feedback, thereby realizing a complete technical closed loop for water quality monitoring and abnormal event analysis.

[0018] It is worth noting that the raw water quality indicators collected by the electrochemical sensor array may be affected by factors such as temperature, pH, ion interference, and instrument drift. To improve measurement accuracy, the data for key indicators such as ammonia nitrogen and dissolved oxygen are corrected using a chemometric model. The specific steps are as follows: First, the sensor response curves at known concentrations in water samples are obtained through experimental calibration, and a multivariate correction matrix is ​​constructed, with environmental parameters such as temperature, pH, and conductivity as auxiliary variables. Then, partial least squares regression (PLS) is used to jointly model the raw sensor signals and environmental parameters, obtaining the mapping relationship between the predicted concentration and the measured signal. During real-time acquisition, the sensor output signal and corresponding environmental parameters are substituted into this model, and correction coefficients are automatically calculated to correct the raw measurements. The accurate water quality indicators corrected by the chemometric model are then output for subsequent dynamic Kalman filtering and sample dataset generation.

[0019] Of particular importance is that the heterogeneous sensor cluster includes optical sensor arrays, electrochemical sensor groups, machine vision modules, and environmental parameter sensors.

[0020] Optionally, after the graded triggering of the early warning mechanism in step S4, the following may also be included: The sample dataset corresponding to the real-time alarm triggered in the hierarchical triggering early warning mechanism is used as the dataset to be traced back, and the original data source of the dataset to be traced back is determined based on the edge computing node corresponding to the dataset to be traced back. In this embodiment, when a real-time alarm is detected in the graded triggering early warning mechanism, the sample dataset that triggered the alarm is automatically marked as the dataset to be traced back. This dataset contains the multimodal feature vectors corresponding to the time step and the water quality index values ​​after dynamic Kalman filtering correction. At the same time, the edge computing node ID and timestamp information are recorded to accurately trace the data source. The system queries the original acquisition sequence corresponding to the dataset to be traced back through the edge computing node scheduling service, including optical sensor image sequences, electrochemical sensor continuous sampling sequences, machine vision image frames and environmental parameter sequences, to ensure that the time step and sampling frequency are strictly aligned with the sample dataset, and the time synchronization error is controlled within ±5 seconds to form a closed-loop traceability basis.

[0021] Based on the initial acquisition sequences of each sensor in the original data source, the water quality indicators exceeding the limit are traced, and the traceability results are transmitted to the water quality monitoring platform.

[0022] In one embodiment, source tracing processing of water quality indicators exceeding limits is performed based on the acquisition sequence of the original data source. First, the initial acquisition sequence of each sensor within the alarm trigger period is extracted, and the corresponding sequence is located on the water quality monitoring platform based on the sensor identification information. The validity of the electrochemical sensor sequence is verified to determine whether the time interval between sampling points exceeds 1.5 times the preset period or whether the difference in short-term change rate of three consecutive sampling points exceeds a threshold of 0.05, so as to remove time abnormal segments and drift trend segments and generate the acquisition sequence to be detected. Subsequently, the real-time acquisition sequence of the target sensor within the real-time alarm trigger window (e.g., 30 minutes) is extracted, the real-time change rate after removing invalid readings is calculated, and compared with the change rate sequence of the same historical time length. If the historical change rate and real-time change rate of any water quality indicator at three or more consecutive time points exceed the preset indicator change rate threshold (e.g., ammonia nitrogen change rate 0.02 mg / L / h), then the indicator is determined to be an indicator exceeding the limit. Finally, the source tracing results, along with the corresponding sensor identifiers, timestamps, and excessive water quality index values, are packaged and transmitted in real time to the water quality monitoring platform via MQTT or REST API for abnormal event analysis, report generation, and decision support, thereby achieving closed-loop traceability and supervision of multi-source data.

[0023] Optionally, the water quality indicators exceeding the limit for source tracing include: Extract the corresponding target sensor identification information from the original data source, and locate the initial acquisition sequence in the water quality monitoring platform according to the target sensor identification information; In this embodiment, based on the sample dataset that triggers the real-time alarm, the target sensor identification information corresponding to the water quality indicators to be traced is extracted from the water quality monitoring platform database. This includes sensor ID, type, deployment location, and sampling frequency. The initial acquisition sequence is then accurately located in the database using this identification information to ensure that each sequence is completely matched with the time step in the sample dataset. The time synchronization error is controlled within ±5 seconds. At the same time, the start and end times of each sequence and sensor status information are recorded to provide a data foundation for subsequent validity verification.

[0024] The sensor validity is verified on the initial acquisition sequence, and invalid sequences are removed based on the validity verification results to obtain the acquisition sequence to be detected; In one embodiment, sensor validity verification is performed on the initial acquisition sequence. The operation includes two parts: first, detecting time anomaly segments and determining whether the time interval between adjacent sampling points exceeds 1.5 times the preset sampling period. If it does, the data in that segment is marked as the first invalid sequence; second, detecting drift trend segments and calculating the difference in short-term change rates of three consecutive sampling points by setting a sliding time window (e.g., 5 minutes). If the difference exceeds a threshold of 0.05, the data in that time segment is marked as the second invalid sequence. Subsequently, the first invalid sequence and the second invalid sequence are time-series merged to obtain the final acquisition sequence to be detected after removing invalid readings, which is used for subsequent change rate calculation.

[0025] The time window for triggering the real-time alarm is used as the observation window. The real-time acquisition sequence corresponding to the target sensor identification information within the observation window is collected, and the real-time change rate of the real-time acquisition sequence after removing invalid readings within the real-time alarm trigger window is calculated. In this embodiment, the time window for triggering the real-time alarm (e.g., 30 minutes) is used as the observation window. The real-time acquisition sequence of the corresponding target sensor within this time period is extracted from the acquisition sequence to be detected. After removing invalid readings, the real-time change rate is calculated. The calculation method is to divide the continuous time step difference by the time interval. For example, the real-time change rate of ammonia nitrogen concentration is... ,in This represents the change in ammonia nitrogen concentration. For a given time period, the real-time rate of change in dissolved oxygen is similarly calculated to ensure sensitivity to short-term abnormal changes.

[0026] Calculate the historical rate of change of each water quality indicator in the collection sequence to be detected over the same time length as the observation window. If the historical rate of change and the real-time rate of change of any water quality indicator exceed the preset indicator rate of change threshold at three or more consecutive time points, then the water quality indicator is regarded as an out-of-limit water quality indicator.

[0027] In this embodiment, the historical change rate of the collected sequence to be detected is calculated simultaneously. Historical data of the same length as the observation window (such as the sampling sequence from the previous 30 minutes to the previous week) is taken to calculate the continuous change rate matrix of each water quality indicator and compare it with the real-time change rate. If the historical change rate and real-time change rate of any water quality indicator exceed the preset indicator change rate threshold (e.g., ammonia nitrogen 0.02 mg / L / h, dissolved oxygen 0.1 mg / L / h) at three or more consecutive time points, the indicator is determined to be an out-of-limit water quality indicator. The out-of-limit result, the corresponding timestamp, sensor ID, and sequence information after invalid readings are simultaneously recorded and transmitted to the water quality monitoring platform for closed-loop alarm analysis and decision support.

[0028] Optionally, sensor validity verification includes: The time interval between adjacent sampling points in the initial acquisition sequence is detected. If the time interval between any two adjacent sampling points exceeds 1.5 times the preset sampling period, the segment between the two adjacent sampling points is determined to be a time abnormal segment, and the acquisition sequence in the time abnormal segment is regarded as the first invalid sequence. In one implementation, the timestamp information of each sampling point is read, and the time difference between adjacent sampling points is calculated. If any two adjacent sampling points If the sampling period exceeds 1.5 times the preset sampling period (e.g., 2 minutes) (i.e., exceeds 3 minutes), the time period is determined to be an abnormal time segment. All sampling point data in this segment are marked as the first invalid sequence, and the start and end times of the time period, sensor ID, and reason for removal are recorded in the database for subsequent tracing and statistical analysis.

[0029] The short-term rate of change of the sampled values ​​of each sampling point in the initial acquisition sequence is calculated according to the preset sliding time window and compared with the short-term rate of change of adjacent sampling points. If the difference of the short-term rate of change of three consecutive adjacent sampling points exceeds the preset fluctuation tolerance, the time interval formed by the three adjacent sampling points is determined to be the drift trend segment, and the acquisition sequence in the drift trend segment is regarded as the second invalid sequence. In this embodiment, short-term rate of change is calculated for the initial acquisition sequence. A sliding time window method (window length set to 5 minutes, step size 1 minute) can be used to calculate the short-term rate of change of adjacent points for each sampling point. ,in The difference in water quality indicators at continuous sampling points is used as the basis for identification. Then, the difference in short-term change rate of three consecutive adjacent sampling points is compared. If the difference exceeds the preset fluctuation tolerance (e.g., 0.05 mg / L or 0.05 pH units), the three sampling points are determined to constitute a drift trend segment. All sampling points in this segment are marked as the second invalid sequence, and the start and end times of the drift segment, sensor ID, and change amplitude are recorded for anomaly analysis and data removal.

[0030] Perform a time-series merging of the first invalid sequence and the second invalid sequence, and use the merged result as the invalid sequence.

[0031] In a further implementation, the first invalid sequence and the second invalid sequence are subjected to time-series merging processing. The two types of invalid sequences are sorted by timestamp, and continuous or overlapping segments are merged to form a unified invalid sequence set, used to remove unreliable data points from the original collected sequences. After merging, the invalid sequences are excluded in subsequent real-time change rate calculations, historical change rate comparisons, and the determination of water quality indicators exceeding limits, ensuring that the collected sequences to be tested contain only highly reliable valid data.

[0032] Optionally, performing dynamic Kalman filtering in step S2 includes: The sampled values ​​of each sensor in the multimodal sensing dataset are used as observation vectors and input into the preset system sampling model to perform initial state estimation and covariance setting. In this embodiment, the sampled values ​​from optical sensors, electrochemical sensors, machine vision modules, and environmental parameter sensors in the multimodal sensing dataset are combined into observation vectors. The dimension of each observation vector is the sum of the number of water quality indicators and the number of environmental parameters, such as dissolved oxygen, pH value, conductivity, ammonia nitrogen concentration, 64-dimensional spectral feature vector, turbidity index, temperature, flow velocity, etc., with a total dimension of 128. The observation vectors are input into a preset system sampling model. The system sampling model constructs state equations and observation equations based on historical sampling data. The state vector contains the current value and first-order rate of change of each water quality indicator. The covariance matrix is ​​obtained by statistically analyzing historical sampling data to obtain an initial estimate, which is used to describe the uncertainty of the system state.

[0033] Based on the initial state estimation results and the covariance setting results, the transitive Kalman recursive calculation is performed. The residual term is calculated based on the difference between the predicted value and the actual observed value in each recursive cycle of the transitive Kalman recursive calculation, and the weighting coefficient is determined based on the variance of the residual term. In this embodiment, a transitive Kalman recursive calculation is performed based on the initial state estimation results and covariance settings, with each sampling period lasting 2 minutes. The recursive calculation first predicts the current state value based on the previous state, and then calculates the residual vector between the predicted value and the actual observed value by combining the system state equation and the observation equation. Subsequently, the variance of the residual vector is calculated, and a weighting coefficient is determined based on this. This coefficient is used to balance the contributions of the predicted value and the observed value during the state update process, ensuring that the dynamic adjustment process is robust to abnormal fluctuations.

[0034] The system state noise covariance matrix in the system sampling model is iteratively corrected using weighted coefficients until the residual term is less than a preset residual threshold. In this embodiment, the system state noise covariance matrix is ​​iteratively corrected using weighted coefficients. Each time, the diagonal elements of the state noise covariance matrix are updated based on the residual variance to adjust the reliability of the prediction. When the mean square error (MSE) of the residual term is lower than the preset threshold of 0.05, the iteration is determined to be converged, thereby ensuring that the system state estimation is highly matched with the observation data and reducing outlier interference.

[0035] The system state prediction obtained from the corrected system sampling model is used as the result of dynamic Kalman filtering.

[0036] In this embodiment, after iterative convergence, the system state prediction value output by the corrected system sampling model is used as the result of dynamic Kalman filtering. This result includes the real-time prediction values ​​of each water quality index and their first-order rate of change after noise correction and anomaly correction of the input observation vector. It can be directly used as the input for subsequent anomaly reading identification, self-calibration and sample dataset generation.

[0037] Optionally, the method for obtaining the system sampling model includes: Historical sampling data from each sensor are obtained through a water quality monitoring platform to determine the dynamic relationship of water quality indicators over time in the monitoring area, and a system state equation is constructed based on this dynamic relationship. In this embodiment, historical sampling data from various heterogeneous sensors deployed within the monitoring area are acquired through a water quality monitoring platform. This includes multispectral image sequences from optical sensors, numerical sequences of dissolved oxygen, pH, conductivity, and ammonia nitrogen concentration from electrochemical sensors, and auxiliary parameters such as water temperature, flow rate, and rainfall collected by environmental parameter sensors. The sampling frequencies are 5 minutes / time for optical sensors, 2 minutes / time for electrochemical sensors, and 10 minutes / time for environmental parameter sensors. Time series analysis is performed on the historical sampling data to extract the trends, periodicity, and short-term fluctuations of each water quality indicator over time. The time dynamic relationship between water quality indicators is determined through an autoregressive moving average (ARMA) model and autocorrelation coefficient analysis. Based on this, a system state equation is constructed. The state vector includes the current value of each water quality indicator and its first and second-order rates of change, used to describe the dynamic evolution of water quality over time.

[0038] An observation equation is constructed based on the correspondence between historical sampling data and multimodal sensing datasets; In this embodiment, an observation equation is constructed based on the correspondence between historical sampling data and the multimodal sensing dataset. Specifically, the observed value at each time step in the historical sampling data is paired with the spectral features, machine vision image indicators, and environmental parameter feature vectors at the corresponding time step in the multimodal sensing dataset to form a linear mapping relationship between the observation vector and the state vector. The least squares method is used to fit and obtain the observation matrix H, which describes how the system state vector is mapped to the actual observable data.

[0039] Based on historical sampling data, the system state noise covariance matrix and observation noise covariance matrix are statistically analyzed, and joint modeling calculations are performed by combining the system state equation and observation equation to obtain the system sampling model.

[0040] In this embodiment, historical sampling data is preprocessed by converting multispectral image sequences from optical sensor arrays into spectral feature vectors (e.g., 64-dimensional principal component features), numerical sequences of dissolved oxygen, pH, conductivity, and ammonia nitrogen from electrochemical sensors, and parameters such as water temperature, flow rate, and rainfall collected by environmental parameter sensors, and then performing unified time step alignment processing. =2 minutes), ensuring that all data sources are comparable at the same time point. Then, the short-term fluctuation variance of each water quality indicator and environmental parameter is calculated. Specifically, the root mean square of the indicator value changes is calculated within a sliding time window (window length set to 5 time steps), and this is used as the diagonal element of the observation noise covariance matrix R to reflect the uncertainty and noise level of each observation. If the indicator exhibits significant periodic fluctuations, a weighted average of the variances can be applied to reduce the periodic impact. Next, covariance statistics are performed on each water quality indicator and its first and second-order rates of change in the system state vector. Specifically, the first-order rate of change of each indicator is first calculated based on historical data. and second-order rate of change ( (These are the changes in indicators). Then, the variance of these rate-of-change sequences and the covariance between different indicators are calculated to form the diagonal and off-diagonal elements of the state noise covariance matrix Q. The diagonal elements reflect the state uncertainty of a single indicator, while the off-diagonal elements reflect the dynamic correlation between indicators. For example, changes in dissolved oxygen and pH may be positively correlated. This statistical result can accurately describe the dynamic fluctuation characteristics of the system state.

[0041] In a further embodiment, the system state equation, observation equation, and statistically obtained covariance matrices Q and R are combined for joint modeling calculation. Joint modeling employs an iterative optimization method, such as the Expectation-Maximization (EM) algorithm: In the E-step, Kalman filtering prediction is performed on historical observation data based on the current Q and R, calculating the state prediction error and residuals; in the M-step, the state noise covariance matrix Q and observation noise covariance matrix R are updated based on the residuals to minimize the prediction error. Iteration continues until Q and R converge or the residual mean square error is below a preset threshold (e.g., 0.05), yielding the final system sampling model. This model includes not only optimized state and observation equation matrices but also calibrated state noise covariance Q and observation noise covariance R, which can be used for subsequent dynamic Kalman filtering, providing initial state prediction, covariance setting, and observation mapping for each sampling period.

[0042] Optionally, the abnormal readings identified in step S2 in the aligned dynamic Kalman filter result include: Based on the aligned dynamic Kalman filter results, the difference between the predicted value and the actual observed value of each sensor within the preset observation time window is extracted, and the residual sequence is calculated; In this embodiment, firstly, based on the aligned dynamic Kalman filter results, each sensor is placed within a preset observation time window (e.g., The predicted system state values ​​(based on 30 sampling points over 1 hour) are compared point-by-point with the actual observed values ​​to calculate the residual sequence. The residual sequence represents the deviation between the predicted and actual observed values, specifically calculated as: Residual = Actual Observed Value - Kalman Predicted Value. Each residual sequence forms a vector along the time dimension.

[0043] The variance of the residual series is calculated, and the interval of three standard deviations of the variance of the residual series is taken as the normal fluctuation range. If the absolute value of the residual sequence of any sensor in two consecutive sampling periods exceeds the normal fluctuation range, then the sensor is determined to be a sensor to be verified. In a further embodiment, statistical analysis is performed on the residual sequence to calculate its variance. and in intervals of three times the standard deviation of the variance (±3) This range serves as the normal fluctuation range, used to distinguish between reasonable and abnormal fluctuations. If the absolute value of the residual of any sensor in the residual sequence exceeds this normal fluctuation range in two consecutive sampling periods (e.g., 4 consecutive minutes, since the sampling period is 2 minutes / time), then the sensor is determined to be a sensor to be calibrated. This criterion can effectively eliminate occasional noise points and ensure the accuracy of the selection of sensors to be calibrated.

[0044] The isolation score of the residual sequence of the sensor to be calibrated is calculated using edge computing nodes, and the sampling points of the sensor to be calibrated with isolation scores higher than the preset isolation score threshold are identified as abnormal readings.

[0045] In this embodiment, the edge computing node further calculates the isolation score of the residual sequence of the sensor to be calibrated. The isolation score is evaluated based on the relative density of the residual distribution and the difference between neighboring residuals. For example, an isolation score is calculated using a local outlier factor method. For each sampling point, the ratio of its local residual density to the residual density of its neighboring points is calculated and normalized to the range of 0 to 1. If the isolation score is higher than a preset threshold (e.g., 0.7), the sampling point is determined to be an abnormal reading. This scoring mechanism can effectively identify local anomalies or drift trends of a single sensor and eliminate false anomalies caused by system prediction errors.

[0046] Optionally, performing self-calibration in step S2 includes: Extract sample values ​​from sensors of the same type as the sensor to be calibrated using a water quality monitoring platform; In this embodiment, the database is queried through the water quality monitoring platform to extract historical sampling values ​​of sensors of the same type as the sensor to be calibrated. For example, if the sensor to be calibrated is a dissolved oxygen sensor, the sampling sequences of all dissolved oxygen sensors at the same site or in the vicinity are extracted. The sampling frequency is uniformly set to 2 minutes / time, and the sequences are time-aligned to ensure that each sequence can be compared at the same time point.

[0047] Based on the monitoring condition characteristics of the sensor to be calibrated, sensors with similar monitoring condition characteristics are selected from the same type of sensors, and the sampled value of the sensor is used as the sampling reference value. In a further embodiment, based on the monitoring condition characteristics of the sensor to be calibrated (including sensor model, deployment depth, water body location, and environmental parameters such as water temperature and flow velocity range), sensors with similar monitoring condition characteristics are selected from the same type of sensors, and their sampling sequences are used as sampling reference values. Specific selection criteria can include a water temperature difference of less than 2°C, a flow velocity difference of less than 0.1 m / s, and identical sensor models, ensuring that the reference sequence is as close as possible to the sensor to be calibrated in terms of environmental and equipment conditions, thereby guaranteeing the accuracy of calibration.

[0048] Calculate the covariance characteristics of the sampled reference values, and determine the residual correction coefficients based on the covariance characteristics and the residual distribution of the observation sequence corresponding to the sensor to be calibrated. In this embodiment, covariance analysis is performed on the sampled reference values ​​to calculate the covariance matrix at each time step, which describes the variation characteristics of the reference sequence and the correlation between indicators. The covariance characteristics of the reference sequence are compared with the residual distribution of the sensor to be calibrated. Based on the degree to which the residuals deviate from the normal fluctuations of the reference sequence, a residual correction coefficient is generated. This coefficient can be adjusted between 0.7 and 1.0. A coefficient close to 1 indicates almost complete correction for anomalous readings with small deviations, while a coefficient close to 0.7 indicates correction for anomalous readings with large deviations. This coefficient is used to adjust the amplitude of anomalous readings of the sensor to be calibrated to make it consistent with the reference sequence under similar environmental conditions.

[0049] Perform linear regression correction on outlier readings based on residual correction coefficients to generate reasonable correction values; In this embodiment, linear regression correction is performed on the abnormal readings based on the residual correction coefficient. Specifically, a linear mapping function is used to scale the residuals to a reasonable range, generating a reasonable correction value. The linear regression employs the least squares fitting method, and the fitting formula is as follows: ;in The corrected sensor sample value. These are the original observations. The residual value represents the difference between the dynamic Kalman filter prediction and the actual observed value. This is the residual correction factor (usually taken as 0.7~1.0 to balance the correction range and fidelity), ensuring that the corrected value conforms to the physically possible range.

[0050] It is worth noting that the observed values ​​are the actual sampled values ​​of each sensor in the multimodal sensing dataset.

[0051] Replace the appropriate correction values ​​with the corresponding values ​​in the aligned dynamic Kalman filter results to obtain the sample dataset.

[0052] In a further embodiment, the generated reasonable correction value is replaced with the corresponding position in the aligned dynamic Kalman filter result to complete the abnormal reading calibration and form the final sample dataset. Optionally, the water quality change trend prediction model constructed in step S3 includes: The sample dataset was divided into data types to obtain multispectral image sequences and water quality parameter time series; In this embodiment, when performing data type classification on the sample dataset, the multimodal sensing data is first divided into multispectral image sequences and water quality parameter time series based on the acquisition format and field characteristics of the sensor data. The multispectral image sequences are derived from the optical sensor array and machine vision module, with an image resolution of 1024×1024 pixels, and one frame is acquired every 5 minutes to capture information on water color, suspended solids concentration, and floating debris distribution. The water quality parameter time series are derived from the electrochemical sensor group and environmental parameter sensors, including key indicators such as dissolved oxygen, pH value, conductivity, ammonia nitrogen concentration, temperature, and flow rate, and are sampled every 2 to 10 minutes, and have undergone standardization and embedding processing.

[0053] Determining the spatial distribution variation characteristics of water bodies based on multispectral image sequences; In this embodiment, a 3D convolutional neural network (3D-CNN) is used to process each frame of the image to extract spatial features such as the color distribution of the water surface, the concentration of suspended matter, and the location of floating objects. Specifically, a 1024×1024 pixel image can be input into a pre-trained ResNet-50 network, and the 512-dimensional feature vector of the intermediate layer is used as the spatial representation of each frame of the image; subsequently, time steps are performed on consecutive image frames. Feature alignment and moving average processing are used to generate a continuous spatial distribution variation matrix of water bodies, which is used to characterize the spatial dynamic evolution of water bodies in the short term.

[0054] A pre-defined stacked two-layer LSTM framework is used to process the time series of water quality parameters in order to extract the time-dependent features of water quality indicators. In a further embodiment, when processing water quality parameter time series using a pre-defined stacked two-layer LSTM framework, the water quality parameter time series is input into the two-layer LSTM model. The first LSTM layer has an input dimension of 64, corresponding to the embedded feature dimension, and 128 hidden units, used to capture local short-term fluctuation features. The second LSTM layer receives the output of the first layer, has 64 hidden units, and is used to learn long-term global temporal dependencies. A random deactivation strategy with a dropout rate of 0.3 is used between the two layers, and the activation function is tanh. The model training batch size is set to 64, the initial learning rate is 0.001, and adaptive adjustments are performed based on changes in the validation set loss. The maximum number of training epochs is 200, and the model is considered converged when the validation set mean squared error (MSE) is below 0.05. Through this two-layer LSTM, the dynamic evolution features of each water quality indicator in the time dimension can be extracted, generating a time dependency feature matrix of dimension T×64, where T is the number of time steps (e.g., 168 hours corresponds to one week of data). Multimodal fusion of spatial distribution variation characteristics and time-dependent characteristics of water bodies is performed to construct a water quality change trend prediction model.

[0055] In this embodiment, the spatial feature matrix and the temporal dependent feature matrix are first registered frame by frame, so that the spatial and temporal features corresponding to the same time node are combined into a joint input matrix. Then, dynamic weight coefficients are calculated based on the contribution of each feature source in the sample dataset, and the joint input matrix is ​​weighted and aggregated to obtain a weighted fused feature matrix. Finally, fuzzy adjustment and genetic iterative optimization are performed on the weighted fused matrix to couple the spatial features of water bodies with the time series features in the short and long term trends, constructing the final water quality change trend prediction model, which can output the predicted change trend values ​​of each water quality indicator in the next 1 to 24 hours.

[0056] It is worth noting that the example operation parameters can be 512 dimensions for each frame of image extraction using convolutional features, and a moving average window of 3 frames; LSTM time steps =1h, batch size 64, dropout rate 0.3, maximum number of training rounds 200; multimodal fusion dynamic weight adjustment iterations are 50 times, and the fuzzy adjustment coefficient is set to 0.1.

[0057] Of particular importance, the time series of water quality parameters treated include: Perform data preprocessing on the time series of water quality parameters, and then divide the preprocessing results into test sets; In this embodiment, the data type of the water quality parameter time series is identified. Based on the original record format uploaded by the sensors, the data type of each parameter field is determined. Continuous variables such as dissolved oxygen and pH value are identified as numerical data, while semantic tags such as equipment operating conditions and weather descriptions are identified as text data. For numerical data, Z-score normalization is used to ensure that the mean and standard deviation of each index value are 0, thus eliminating scale bias between different physical quantity units. For text data, word segmentation and vectorization are performed based on a preset Word2Vec embedding model, mapping each semantic tag to a word vector representation of length 32. Subsequently, the numerical processing results and text embedding results are integrated to form classification data, which is then converted into a unified feature dimension through an embedding layer. The embedding layer adopts a fully connected structure with an output dimension of 64, and a tanh activation function is added at the output to enhance the nonlinear feature expression capability. After the embedding layer output, feature concatenation and time step alignment operations are performed to align features of different sampling frequencies with time steps. Synchronization takes 1 hour. Finally, the training and test sets are divided in an 8:2 ratio, with the data from the most recent 7 days allocated to the test set for subsequent model validation and generalization evaluation.

[0058] The test set partitioning results are input into a stacked two-layer LSTM framework to construct a two-layer LSTM model; In this embodiment, after data partitioning, preprocessed training set samples are input into a stacked two-layer LSTM framework to establish a time-dependent prediction model. This framework comprises two layers of sequentially stacked Long Short-Term Memory (LSTM) units. The first LSTM layer has an input dimension of 64, corresponding to the feature dimension of the aforementioned embedding layer output, and 128 hidden units to capture local short-term fluctuation features. The second LSTM layer receives the output of the first layer, with 64 hidden units, and is used to learn long-term global temporal dependencies. A random deactivation strategy with a dropout rate of 0.3 is used between the two LSTM layers to prevent overfitting, and the tanh function is used for both layers. A fully connected layer is connected to the output, mapping the LSTM output to the water quality indicator prediction space. The number of output nodes corresponds to the number of water quality indicators (e.g., 6 dimensions corresponding to six monitoring parameters).

[0059] Perform forward and backward propagation on the two-layer LSTM model, iteratively optimize and construct the two-layer LSTM model until the validation set loss is lower than the preset loss threshold. In this embodiment, a forward and backward propagation training process is performed on the two-layer LSTM. During the forward propagation phase, the input sequence at each time step is progressively fed into the network, and the hidden state and output prediction value are calculated sequentially. During the backward propagation phase, the mean squared error (MSE) between the predicted output and the actual observed sequence is used as the loss function, and the Adam optimizer is used for gradient backward updates. The initial learning rate is set to 0.001 and is adaptively adjusted based on the rate of change of the validation set loss. If the decrease in validation set loss is less than 0.01 over 10 consecutive iterations, an early stopping strategy is automatically triggered to avoid overfitting. The model training batch size is set to 64, and the maximum number of iterations is set to 200. When the validation set loss falls below a preset threshold of 0.05, the model is considered converged, the iteration is terminated, and the parameter weights are fixed. The model weights are initialized using a Glorot uniform distribution, and gradient clipping (threshold 5) is performed after each weight update during optimization to prevent gradient explosion.

[0060] The time-dependent characteristics of water quality indicators were determined using an optimized two-layer LSTM model.

[0061] In this embodiment, an optimized and convergent stacked two-layer LSTM model is used to perform inference operations on the test set samples. This model receives the input sequence at each time step and outputs the corresponding hidden layer state vector. By expanding the hidden state matrix of the two LSTM layers, a time-dependent feature matrix of dimension T×64 is obtained, where T is the number of time steps (e.g., 168 hours corresponds to one week of data). Each column of the matrix represents the dynamic evolution characteristics of a specific water quality indicator over time, and each row corresponds to the multidimensional feature response at the monitoring time. After mapping through a fully connected layer, the generated output vector is the fusion of the predicted results and time-dependent features for each water quality indicator, used to characterize the temporal coupling relationship of water quality indicators under short-term and long-term trends. This output serves as a key input in the multimodal fusion stage, providing a high-dimensional time-dependent foundation for the subsequent joint prediction of spatial and temporal features.

[0062] Optionally, performing multimodal fusion includes: The spatial distribution variation characteristics and temporal dependence characteristics of water bodies are registered frame by frame, and the features corresponding to the same time node are constructed into a joint input matrix. In this embodiment, when performing frame-by-frame registration of the spatial distribution variation characteristics and temporal dependence characteristics of water bodies, the spatial feature matrix extracted from the multispectral image sequence and the temporal dependence feature matrix output by the dual-layer LSTM are first aligned by timestamp. The spatial feature vector corresponding to each time node t has a dimension of 512, and the temporal dependence feature vector has a dimension of 64. A joint input vector with a total dimension of 576 is generated through a concatenation operation. This operation is performed on all time nodes of the entire observation period T (e.g., 168 hours) to obtain the joint input matrix with a size of T×576, which is used for subsequent multimodal fusion processing.

[0063] Calculate dynamic weight coefficients based on the contribution of each feature source in the sample dataset; In another embodiment, the weighted variance analysis method is used to statistically analyze the variance contribution of spatial features and time-dependent features in the prediction of target water quality indicators at each time point. The contribution is standardized to obtain weight coefficients. The initial value of the weight of spatial features is set to 0.6, and the initial value of the weight of time-dependent features is set to 0.4. In each training iteration, the weight coefficients are dynamically adjusted according to the changes in prediction error, with an adjustment step size of 0.05, to ensure that the two types of features contribute according to their actual importance proportions during the fusion process.

[0064] A weighted aggregation is performed on the joint input matrix based on the dynamic weight coefficients, and fuzzy adjustment and genetic iterative optimization are performed based on the weighted aggregation results to construct a water quality change trend prediction model.

[0065] In a further embodiment, each joint input vector is linearly weighted and summed according to its weight coefficients to generate a weighted fusion vector. Subsequently, the full-cycle weighted fusion vector matrix is ​​input into the fuzzy adjustment and genetic iterative optimization framework. Fuzzy adjustment performs fuzzy smoothing on the prediction results by defining a membership function for changes in water quality indicators, with the fuzziness coefficient set to 0.1 to suppress the interference of short-term noise fluctuations on trend judgment. In genetic iterative optimization, individuals are feature combinations of the weighted fusion vector, the population size is 50 per generation, the number of iterations is 50, the crossover probability is 0.8, and the mutation probability is 0.1. The optimal fusion feature combination is selected using a fitness function (based on the prediction error MSE).

[0066] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, 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 the equivalents of the application are intended to be included within the invention.

[0067] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A real-time dynamic water quality change trend analysis method based on multi-modal perception and deep learning, characterized in that, The method comprises the following steps: Step S1: collecting a multi-modal perception data set through a heterogeneous sensor cluster pre-deployed in a monitoring area; Step S2: transmitting the multi-modal perception data set to a preset edge computing node to perform dynamic Kalman filtering; identifying abnormal readings in the aligned dynamic Kalman filtering result, and performing self-calibration on the abnormal readings to obtain a sample data set; Step S3: constructing a water quality change trend prediction model according to the sample data set; Step S4: predicting the water quality change trend of the monitoring area by using the water quality change trend prediction model; and matching and comparing the prediction result of the water quality change trend by using a preset hierarchical warning threshold, and triggering a warning mechanism according to the matching and comparison result. 2.The real-time dynamic water quality change trend analysis method based on multi-modal perception and deep learning according to claim 1, wherein, After the hierarchical warning mechanism is triggered in step S4, the method further comprises: taking the sample data set corresponding to the triggered real-time alarm in the hierarchical warning mechanism as a to-be-traced data set, and determining the original data source of the to-be-traced data set according to the edge computing node corresponding to the to-be-traced data set; tracing the out-of-limit water quality index according to the initial collection sequence of each sensor in the original data source, and transmitting the tracing result to a water quality monitoring platform. 3.The real-time dynamic water quality change trend analysis method based on multi-modal perception and deep learning according to claim 2, characterized in that, The tracing of the out-of-limit water quality index comprises: extracting the corresponding target sensor identification information from the original data source, and positioning the initial collection sequence in the water quality monitoring platform according to the target sensor identification information; performing sensor effectiveness verification on the initial collection sequence, and removing invalid sequences according to the effectiveness verification result to obtain a to-be-detected collection sequence; taking a time window of the real-time alarm as an observation window, collecting real-time collection sequences corresponding to the target sensor identification information in the observation window, and calculating the real-time change rate of the real-time collection sequences after removing invalid readings in the real-time alarm triggering window; calculating the historical change rate of each water quality index in the to-be-detected collection sequence in the same time length as the observation window, and if the historical change rate of any water quality index at three or more consecutive time points exceeds a preset index change rate threshold, the water quality index is taken as an out-of-limit water quality index.

4. The real-time dynamic water quality change trend analysis method based on multi-modal perception and deep learning according to claim 3, characterized in that, The sensor effectiveness verification comprises: detecting the time interval between adjacent sampling points in the initial collection sequence, if the time interval between any two adjacent sampling points exceeds 1.5 times of the preset sampling period, the section between the two adjacent sampling points is determined as a time abnormal section, and the collection sequence in the time abnormal section is taken as a first invalid sequence; calculating the short-term change rate of the sampling value of each sampling point in the initial collection sequence according to a preset sliding time window, and comparing it with the short-term change rate of adjacent sampling points, if the difference between the short-term change rates of three consecutive adjacent sampling points exceeds a preset fluctuation tolerance, the time section composed of the three adjacent sampling points is determined as a drift trend section, and the collection sequence in the drift trend section is taken as a second invalid sequence; performing time sequence merging on the first invalid sequence and the second invalid sequence, and taking the time sequence merging result as the invalid sequence. 5.The real-time dynamic water quality change trend analysis method based on multi-modal perception and deep learning according to claim 1, wherein, The dynamic Kalman filtering in step S2 comprises: inputting the sampling values of each sensor in the multi-modal perception data set as an observation vector into a preset system sampling model to perform initial state estimation and covariance setting; According to the initial state estimation result and the covariance setting result, a transfer Kalman recursion calculation is performed, a residual term is calculated according to a difference between a predicted value and an actual observation value in each recursion period in the transfer Kalman recursion calculation, and a weighting coefficient is determined according to a variance of the residual term; The system state noise covariance matrix in the system sampling model is iteratively corrected by using the weighting coefficient until the residual term is less than a preset residual threshold value; A system state predicted value obtained by the corrected system sampling model is taken as a dynamic Kalman filtering result. 6.The real-time dynamic water quality change trend analysis method based on multi-modal perception and deep learning according to claim 5, wherein, The method for obtaining the system sampling model comprises: acquiring historical sampling data of each sensor through a water quality monitoring platform, determining a time variation dynamic relationship of a water quality index in a monitoring area, and constructing a system state equation based on the time variation dynamic relationship; constructing an observation equation according to a corresponding relationship between the historical sampling data and a multi-modal perception data set; statistically obtaining a system state noise covariance matrix and an observation noise covariance matrix based on the historical sampling data, and performing joint modeling calculation in combination with the system state equation and the observation equation to obtain the system sampling model. 7.The real-time dynamic water quality change trend analysis method based on multi-modal perception and deep learning according to claim 1, wherein, The step S2 of identifying the abnormal readings in the aligned dynamic Kalman filtering result comprises: extracting a difference between a predicted value and an actual observation value of each sensor within a preset observation time window according to the aligned dynamic Kalman filtering result, and calculating a residual sequence; statistically obtaining a variance of the residual sequence, and taking a three times standard deviation interval of the variance of the residual sequence as a normal fluctuation interval; if the absolute value of the residual sequence of any sensor in two continuous sampling periods exceeds the normal fluctuation interval, the sensor is determined as a sensor to be checked; calculating an isolation score of the residual sequence of the sensor to be checked by using the edge computing node, and determining a sampling point of the sensor to be checked whose isolation score is higher than a preset isolation score threshold value as an abnormal reading. 8.The real-time dynamic water quality change trend analysis method based on multi-modal perception and deep learning according to claim 1, wherein, The step S2 of performing self-calibration comprises: extracting a sampling value of a sensor of the same type as the sensor to be checked through the water quality monitoring platform; screening a sensor similar in monitoring condition characteristics from the sensors of the same type according to the monitoring condition characteristics of the sensor to be checked, and taking the sampling value of the sensor as a sampling reference value; calculating a covariance characteristic of the sampling reference value, and determining a residual correction coefficient according to the covariance characteristic and a residual distribution of a corresponding observation sequence of the sensor to be checked; performing linear regression correction on the abnormal reading according to the residual correction coefficient to generate a reasonable correction value; replacing the reasonable correction value to a corresponding value in the aligned dynamic Kalman filtering result to obtain a sample data set. 9.The real-time dynamic water quality change trend analysis method based on multi-modal perception and deep learning according to claim 1, wherein, The step S3 of constructing the water quality change trend prediction model comprises: performing data type division on the sample data set to obtain a multi-spectral image sequence and a water quality parameter time sequence; determining a water body spatial distribution change characteristic based on the multi-spectral image sequence; processing the water quality parameter time sequence by using a preset stacked double-layer LSTM framework to extract a time dependence characteristic of a water quality index; performing multi-modal fusion on the water body spatial distribution change characteristic and the time dependence characteristic to construct the water quality change trend prediction model.

10. The real-time dynamic water quality change trend analysis method based on multi-modal perception and deep learning according to claim 9, characterized in that, The multi-modal fusion comprises: The water body spatial distribution change characteristics and time dependent characteristics are frame by frame registered, and the characteristics corresponding to the same time node are constructed as a joint input matrix; Dynamic weight coefficients are calculated according to the contribution degrees of the feature sources in the sample data set; A water quality change trend prediction model is constructed by performing weighted aggregation on the joint input matrix according to the dynamic weight coefficients, and performing fuzzy adjustment and genetic iteration optimization according to the weighted aggregation result.

Images