An intelligent water quality early warning method based on multi-sensor data fusion

Abstract

The present application relates to the technical field of water quality early warning, and particularly relates to an intelligent water quality early warning method based on multi-sensor data fusion. The method comprises the following steps: obtaining and dividing multi-sensor water quality monitoring data, then performing linear coupling and residual nonlinear processing on environmental factors in the historical sequence to extract instantaneous interaction features. Through multi-scale decomposition and independent linear prediction channels, the prediction results of month, week, day and the remaining items are generated and fused to obtain the future water quality prediction sequence. Then, the prediction trajectory is formed by repeated forward reasoning, the prediction mean and confidence interval are calculated, finally the mean, trend and upper limit of the confidence interval are combined with the multi-level threshold to determine the water quality early warning information of different levels. The present application realizes the intelligent water quality early warning effect with higher precision, higher real-time performance, higher interpretability and stability through instantaneous coupling extraction of environmental factors, multi-scale trend and periodic prediction and multi-level threshold probability early warning.

Description

Technical Field

[0001] This invention relates to the field of water quality early warning technology, and in particular to an intelligent water quality early warning method based on multi-sensor data fusion. Background Technology

[0002] Water quality early warning is a key technology that uses multiple types of sensors to monitor the concentration and changing trends of pollutants in water bodies and issue warnings when indicators are abnormal. Water quality is affected by a variety of environmental factors, including water temperature, pH, dissolved oxygen, chemical oxygen demand, total phosphorus, total nitrogen, ammonia nitrogen, permanganate index, sulfides, and chlorophyll a. These factors exhibit nonlinear and dynamic coupling characteristics, and water quality also shows a trend-like and periodic change pattern over time.

[0003] However, existing water quality early warning methods suffer from the following problems: First, the instantaneous interactions and temporal evolution of environmental factors are often modeled in a mixed manner. Mainstream methods often employ complex network structures such as LSTM and Transformer, resulting in large, computationally intensive models that struggle to efficiently process two different types of dynamic information. Second, factor correlation modeling based on graph neural networks treats environmental factors as discrete nodes, neglecting the globally smooth relationships driven by continuous physical fields such as temperature and concentration fields, leading to insufficient model generalization and scalability. Third, temporal evolution models generally employ complex encoder-decoder structures, resulting in poor real-time performance and high training costs. In reality, long-term changes in water quality can usually be characterized by relatively simple trend and periodic terms; overly complex modeling can easily lead to inefficiency and overfitting. Summary of the Invention

[0004] Therefore, it is necessary to provide an intelligent water quality early warning method based on multi-sensor data fusion to solve at least one of the above-mentioned technical problems.

[0005] To achieve the above objectives, a smart water quality early warning method based on multi-sensor data fusion is proposed, the method comprising the following steps:

[0006] Step S1: Acquire multi-sensor monitoring data of the target water body; divide the multi-sensor monitoring data into sample data containing historical sequences and future target sequences;

[0007] Step S2: Perform linear coupling and residual nonlinear processing on the environmental factors of the sample data at each time step in the historical sequence, extract the instantaneous interaction features between environmental factors, and obtain the feature sequence;

[0008] Step S3: Decompose the feature sequence into monthly, weekly, daily and residual terms, and generate prediction results for the corresponding time scales through independent linear prediction channels. Then, fuse the prediction results of each scale to obtain the prediction sequence of future water quality factors.

[0009] Step S4: Perform forward inference on the sample data of the future target sequence based on the predicted sequence to obtain multiple predicted trajectories, and obtain the predicted mean and confidence interval based on the predicted trajectories.

[0010] Step S5: Based on the predicted mean, the trend of the predicted trajectory, the upper limit of the confidence interval, and the preset multi-level thresholds, a judgment is made to output the corresponding level of water quality early warning information.

[0011] The present invention has the following beneficial effects:

[0012] I. Through an innovative design that combines multi-sensor data fusion with instantaneous interaction modeling of environmental factors, this invention achieves the separate processing of "spatial correlation" and "temporal evolution" in water quality conditions. Utilizing linear coupling, residual nonlinear fusion, and frequency domain filtering mechanisms, this invention can accurately extract the instantaneous coupling features between environmental factors without relying on complex deep networks. This effectively overcomes the problems of bloated structures, training difficulties, and high computational costs caused by mixed environmental factor modeling in traditional methods. This structure not only improves the model's generalization ability and robustness but also enhances the availability of water quality early warning systems in low-power or edge computing scenarios.

[0013] Second, by employing multi-scale decomposition and independent linear prediction channels, the long-term trend and periodic components of water quality are efficiently decomposed, enabling the prediction process to fit complex time series in a lighter and more transparent manner. The multi-scale hierarchical prediction mechanism reduces the overfitting risk of traditional encoder-decoder networks and significantly improves real-time prediction performance, allowing the system to respond promptly to rapidly changing or sudden water quality anomalies. Simultaneously, this invention constructs a set of prediction trajectories through multiple forward inferences, achieving a quantitative expression of the uncertainty of future water quality conditions, making the early warning results both stable and possessing probabilistic interpretability.

[0014] Third, a multi-level threshold logic judgment mechanism is proposed, which combines key features such as predicted mean, trend, and upper limit of confidence interval for comprehensive judgment, making the early warning level more discriminative and practically operable. Furthermore, through the design of sensor deployment, physical calibration, and staggered calibration strategies, this invention can effectively improve sampling accuracy, reduce interference errors, and ensure the consistency and reliability of multi-channel acquired data. Overall, this invention achieves significant improvements in accuracy, real-time performance, interpretability, and system robustness, providing a more scientific and efficient intelligent early warning method for water environment monitoring and management. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the steps of an intelligent water quality early warning method based on multi-sensor data fusion.

[0016] Figure 2 This is a flowchart illustrating the framework of an intelligent water quality early warning method based on multi-sensor data fusion.

[0017] Figure 3 This is a structural diagram of the prediction model for an intelligent water quality early warning method based on multi-sensor data fusion proposed in this application;

[0018] Figure 4 This is a schematic diagram of the working principle of the FNO module of the intelligent water quality early warning method based on multi-sensor data fusion proposed in this application;

[0019] Figure 5 This is a schematic diagram illustrating the working principle of the multi-scale hierarchical DLinear module, which is a smart water quality early warning method based on multi-sensor data fusion according to this application.

[0020] Figure 6 This is a flowchart illustrating the probabilistic prediction and hierarchical early warning process of an intelligent water quality early warning method based on multi-sensor data fusion, as described in this application.

[0021] Figure 7 This is a schematic diagram of a fully connected network for an intelligent water quality early warning method based on multi-sensor data fusion proposed in this application;

[0022] Figure 8 This is a schematic diagram of the linear layer in DLinear, a smart water quality early warning method based on multi-sensor data fusion proposed in this application.

[0023] 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

[0024] 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.

[0025] 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.

[0026] 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.

[0027] To achieve the above objectives, please refer to Figures 1 to 8 A smart water quality early warning method based on multi-sensor data fusion, the method comprising the following steps:

[0028] Step S1: Acquire multi-sensor monitoring data of the target water body; divide the multi-sensor monitoring data into sample data containing historical sequences and future target sequences;

[0029] In one embodiment, data acquisition and preprocessing are performed first. For example... Figure 2 As shown, 26,000 data points were collected hourly over three consecutive years from multiple online monitoring sensors at a reservoir. These data included 11 indicators: water temperature, pH, dissolved oxygen, total nitrogen, total phosphorus, ammonia nitrogen, light intensity, air temperature, water level, rainfall, and chlorophyll a. Linear interpolation was performed on the data, and a small number of missing values ​​were filled. The data were then divided into training, validation, and test sets in an 8:1:1 ratio. Z-score normalization was applied to all data using the mean and standard deviation of the training set. Sample construction was then performed. Figure 2 As shown, the length of the historical input sequence is set to L=168 (past 7 days), and the length of the future prediction sequence is set to H=24 (future 1 day). The data is constructed into sample pairs containing the input sequence X (dimension [168,11]) and the target sequence Y (dimension [24,1]) using the sliding window method.

[0030] Step S2: Perform linear coupling and residual nonlinear processing on the environmental factors of the sample data at each time step in the historical sequence, extract the instantaneous interaction features between environmental factors, and obtain the feature sequence;

[0031] In one embodiment, such as Figure 3 As shown, the model consists of a Fourier Neural Operator (FNO) module as a feature extractor and a multi-scale hierarchical DLinear module as a temporal predictor.

[0032] Extraction of instantaneous multiphysics interaction features of water quality environmental factors. For example... Figure 3 and Figure 4The input sequence X (dimension [L, N], i.e., [168, 11]) is fed into the FNO module for processing. The purpose of this process is to extract deep features containing the interactions between variables at each time step. Specifically, the input sequence X is processed through a fully connected network such as... Figure 7 As shown, the N-dimensional vector at each time step is upscaled to a higher-dimensional D-dimensional vector, for example, 128-dimensional vectors. This results in a vector of dimension [L, D]. The FNO module processes each of the L time steps in this vector one by one. For the D-dimensional vector z_t at the t-th time step, the following operation is performed: Figure 4 As shown: A one-dimensional Fast Fourier Transform (FFT) is performed to obtain its frequency domain representation z_hat_t. In the frequency domain, only the low-frequency k=16 modes are retained, and these modes are linearly transformed through a learnable complex weight matrix R. The shape of R is [128, 128, 16] to achieve the mixing and coupling of global information. The transformed frequency domain representation is transformed back to the D-dimensional spatial domain through an Inverse Fourier Transform (IFFT) to obtain the updated vector z_new_t. z_new_t is added to the result z_linear_t of the original z_t after a linear transformation, and then operated through the nonlinear Gaussian Error Linear Unit (GELU) activation function to obtain the D-dimensional feature vector z_new_t updated at the t-th time step. At the same time, [z_new_1, z_new_2, ..., z_new_L] obtained after processing at L time steps are combined to form the final feature sequence Z (dimension [L, D]).

[0033] Step S3: Decompose the feature sequence into monthly, weekly, daily and residual terms, and generate prediction results for the corresponding time scales through independent linear prediction channels. Then, fuse the prediction results of each scale to obtain the prediction sequence of future water quality factors.

[0034] In one embodiment, the temporal evolution of the feature sequence is predicted using a multi-scale hierarchical DLinear network. For example... Figure 3 and Figure 5As shown, the feature sequence Z (dimension [L,D]) is fed into a multi-scale hierarchical DLinear network for efficient multi-scale time series prediction. The implementation method is as follows: the input feature sequence Z is fed in parallel into multiple moving average layers with different window sizes. For example, in this embodiment, three moving average kernels are used: Moving average kernel 1 (monthly trend): window size 720 (k_month=720), used to extract the long-term trend term Z_monthly_trend; Moving average kernel 2 (weekly cycle): window size 168 (k_week=168), used to extract the weekly scale variation Z_weekly; Moving average kernel 3 (daily cycle): window size 24 (k_day=24), used to extract the daily scale variation Z_daily.

[0035] The original sequence Z is finely decomposed into four independent components through the difference operation: Z_monthly_trend=MovingAvg(Z, k_month); Z_weekly=MovingAvg(Z, k_week) -Z_monthly_trend; Z_daily=MovingAvg(Z, k_day) -MovingAvg(Z, k_week); Z_remain=Z-MovingAvg(Z, k_day).

[0036] Each decomposed scale component is equipped with a dedicated, independent linear prediction channel. Each channel contains a linear layer, such as... Figure 8 As shown, this linear layer directly maps the historical sequence (length L) at the corresponding scale to the prediction of the future (length H): long-term trend channel: pred_trend=Linear_trend(Z_monthly_trend); weekly channel: pred_weekly=Linear_weekly(Z_weekly); daily channel: pred_daily=Linear_daily(Z_daily); remaining channel: pred_remain=Linear_remain(Z_remain). The prediction results of all independent scale channels are directly added element-wise and fused into a comprehensive prediction of dimension [H,D], i.e., Prediction_combined=pred_trend+pred_weekly+pred_daily+pred_remain. For example, a fully connected layer can be used to map the fused prediction of [H,D] to the final output dimension [H,1] to obtain the prediction sequence of future chlorophyll a.

[0037] Step S4: Perform forward inference on the sample data of the future target sequence based on the predicted sequence to obtain multiple predicted trajectories, and obtain the predicted mean and confidence interval based on the predicted trajectories.

[0038] In one embodiment, the present invention employs end-to-end model training, using mean squared error (MSE) as the loss function. At each step of training, the loss between the model's final output and the true label Y is calculated. This loss is used to update the linear layer parameters of the multi-scale hierarchical DLinear network module and all learnable parameters of the FNO module (including fully connected layers, the frequency domain weight matrix R, etc.) simultaneously via backpropagation. This end-to-end training method ensures that the features extracted by the FNO module are most beneficial for subsequent linear prediction by the multi-scale hierarchical DLinear.

[0039] The performance metrics of this algorithm model and other algorithm models are as follows:

[0040]

[0041] As can be seen from the table above, the FNO+ multi-scale hierarchical DLinear algorithm of this invention has the lowest MSE value for the prediction loss function in the next day, which is better than other algorithms.

[0042] Subsequently, as Figure 6 As shown, Dropout layers are inserted between the FNO module and the multi-scale hierarchical DLinear module, and between the multi-scale hierarchical DLinear module and the final fully connected layer. During training, Dropout is used normally. During inference, `model.eval()` is not called; instead, `model.train()` is maintained, keeping the Dropout layers active. After 100 forward propagations, 100 different predicted trajectory data points (pre_data) are obtained. The shape of `pre_data` is [100, 24, 1], as shown in the table below:

[0043]

[0044] Statistical calculations were performed on the data pre_data to obtain the mean and 95th percentile. The method is as follows: Calculate the mean for each time point, that is, calculate the average along the dimension of the trajectory number (axis=0). The specific calculation results are as follows:

[0045] The mean value for the first hour in the future: mean_hour_1=mean(91.5, 90.8, 92.1,…), the result is 91.8;

[0046] The mean for the next two hours: mean_hour_2=mean(92.8, 91.5, 93.0,…), the result is 92.6;

[0047] The mean value for the next 3 hours: mean_hour_3=mean(94.1, 93.5, 92.9,…), the result is 93.9.

[0048] This process is repeated for all 24-hour mean values. The result is a mean prediction sequence of length 24: [91.8, 92.6, 93.9,…].

[0049] Calculate the 95th quantile (risk ceiling) for each time point, i.e., calculate the 95th quantile along the trajectory number dimension (axis=0). For each hour, find the 95th value from the 100 predicted values, ranked from smallest to largest. For example, the 95th quantile for the first hour: Sort the 100 predicted values ​​(91.5, 90.8, 92.1,…) and find the 95th value. The result is 93.5, meaning there is a 95% probability that the concentration will not exceed 93.5. The 95th quantile for the second hour: Sort the 100 values ​​in the second column and find the 95th value. The result is 94.8. The 95th quantile for the third hour: Sort the 100 values ​​in the third column and find the 95th value. The result is 96.2.

[0050] This process is repeated for all 24 hours. The result is a 95th percentile sequence of length 24, pred_q95 (risk ceiling): [93.5, 94.8, 96.2,…].

[0051] Step S5: Based on the predicted mean, the trend of the predicted trajectory, the upper limit of the confidence interval, and the preset multi-level thresholds, a judgment is made to output the corresponding level of water quality early warning information.

[0052] In one embodiment, an early warning operation is performed based on the two sequence values ​​pred_mean and pred_q95: For example, for lakes and reservoirs, preset concentration thresholds are set as follows: T1 (safe threshold): chlorophyll a concentration less than 25 ug / L; T2 (concern threshold): chlorophyll a concentration between 65 and 100 ug / L; T3 (danger threshold): chlorophyll a concentration exceeding 160 ug / L. Then, the R1 growth rate threshold is set as follows: if the chlorophyll a concentration exceeds 5 ug / L / hour, it can be considered to be in an exponential growth period, which is a very dangerous signal.

[0053] Based on the pred_mean and pred_q95 values of these two sequences, dynamic early warning calculations are performed, calculating key information: maximum mean prediction value: max(pred_mean) = 132.5 ug / L; maximum risk upper limit: max(pred_q95) = 165.2 ug / L; trend: perform a linear regression on the last 6 hours of pred_mean […, 125.4, 131.2, 138.0], and the calculated slope value qvalue = 6.8 ug / L / hour.

[0054] Set the early warning levels: max(pred_mean) > T2 and qvalue > R1, level 3 emergency early warning, rapid outbreak of chlorophyll a concentration; max(pred_q95) > T3 or max(pred_mean) > T3, level 2 high-risk early warning, possible outbreak of chlorophyll a concentration; max(pred_mean) > T2, level 1: attention early warning, rising trend of chlorophyll a concentration; max(pred_mean) < T1, level 0: safe.

[0055] Start logical judgment, that is, judge the condition max(pred_mean) > T2 and qvalue > R1 that the above calculation results meet, and according to the logical judgment result, output and trigger a level 3 emergency early warning, rapid outbreak of chlorophyll a concentration.

[0056] Preferably, step S2 includes the following steps:

[0057] Step S21: Extract the environmental factors of the sample data at each time step in the historical sequence;

[0058] Step S22: Perform dimensionality increase processing on the environmental factors as input vectors to obtain the dimensionality-increased environmental vectors;

[0059] Step S23: Perform linear coupling and residual non-linear processing on the dimensionality-increased environmental vectors at each time step, extract the instantaneous interaction characteristics between environmental factors, and obtain updated feature vectors;

[0060] Step S24: Combine the updated feature vectors to obtain a feature sequence.

[0061] In one embodiment, after acquiring the functionalized data, a Fourier neural operator module is used to independently process the N-dimensional vector at each time step t in the input sequence X, extracting these vectors as instantaneous, nonlinear interrelationship features of the continuous field. This process first maps the original N-dimensional input vector to a higher-dimensional D-dimensional vector through a fully connected network. This not only increases the model's expressive power but also constructs a refined, continuous functional representation from discrete variables. Secondly, a Fourier transform is performed on this high-dimensional vector, converting it from the variable space or physical field space to the frequency domain, transforming it into complex vectors representing different frequencies of change. Then, a learnable complex weight matrix is ​​used to linearly transform these frequency complex vectors, obtaining new frequency domain features and achieving the mixing and coupling of global information. This captures the long-range dependencies and global coupling relationships between all vectors. Finally, an inverse Fourier transform returns the model to the spatial vector space and performs a residual connection with the original high-dimensional vector. The result of the residual connection is then processed through a nonlinear activation function to enhance the model's nonlinear modeling ability and ensure training stability.

[0062] By performing the aforementioned Fourier neural operator processing flow at each time step of the input sequence X, the original input sequence X (dimension L x N) is transformed into a new, more feature-rich sequence Z (dimension L x D). Compared to the original input sequence X, the vector at each time point in sequence Z is no longer an isolated, unprocessed sensor reading. It is a high-dimensional state representation after deep nonlinear transformation. Through a multi-physics coupling process, it captures and fuses the instantaneous and complex interdependencies between all N original variables. For example, the vector Z_t at time t contains not only information about the water temperature, pH, dissolved oxygen, total phosphorus, and total nitrogen at that time, but also the patterns of how these variables interact and influence water quality as continuous fields such as temperature and concentration fields. The feature sequence Z extracted by the Fourier neural operator module provides a more information-dense and clearer input for the subsequent multi-scale hierarchical DLinear network module, thereby making subsequent predictions more efficient.

[0063] Preferably, step S23 includes:

[0064] Perform a one-dimensional fast Fourier transform on the environment vector after dimensionality increase at each time step to obtain the frequency domain of the environment vector;

[0065] The low-frequency environment vector frequency domain is filtered based on a preset frequency domain threshold, and the corresponding environment vector is linearly transformed according to the low-frequency environment vector frequency domain to obtain the linearly transformed environment vector frequency domain.

[0066] The frequency domain of the linearly transformed environment vector is transformed into the spatial domain through the inverse Fourier transform to obtain the updated environment vector.

[0067] Based on the nonlinear fusion of the updated environment vector and the environmental vector residual after dimensionality increase, an updated feature vector is obtained to represent the environmental coupling characteristics at the t-th time step.

[0068] In one embodiment, for the upgraded environment vector at each time step in the historical sequence, a one-dimensional Fast Fourier Transform (FFT) is first performed on it in the first sub-step of step S23 to convert the time-domain environment vector signal into a frequency-domain representation, thereby obtaining the spectral characteristics of the environment vector at that time step. This frequency-domain representation contains the amplitude and phase information of environmental factors at different frequencies, which can be used to analyze the periodic changes and rapid fluctuations of environmental factors.

[0069] Subsequently, in the second sub-step of step S23, the low-frequency environmental vector frequency domain is filtered according to a preset frequency domain threshold. The low-frequency component mainly reflects the long-term trend and slow change characteristics of environmental factors. The environmental vector corresponding to the low-frequency frequency domain data is linearly transformed. By adjusting the amplitude or weighting combination, the linearly transformed environmental vector frequency domain is obtained, which enhances the contribution of the low-frequency signal to the environmental coupling characteristics and suppresses high-frequency noise interference.

[0070] In the third sub-step of step S23, the frequency domain of the linearly transformed environment vector is converted back to the spatial domain using an inverse Fourier transform (IFFT) to obtain the updated environment vector. This vector, while retaining the original temporal information of the environmental factors, incorporates adjustments for low-frequency trends from the frequency domain analysis, thus more accurately reflecting the overall dynamic changes of the environmental factors.

[0071] Finally, in the fourth sub-step of step S23, the updated environment vector and the upgraded original environment vector are subjected to residual nonlinear fusion processing. Residual nonlinear fusion includes nonlinear mapping or weighted correction of the difference between the updated vector and the original vector to capture the nonlinear dependencies between environmental factors. Through this processing, an updated feature vector representing the environmental coupling characteristics at time step t is obtained, which can fully reflect the instantaneous interactions between environmental factors and provide high-quality input for subsequent feature sequence generation and prediction models.

[0072] Preferably, step S3 includes the following steps:

[0073] Step S31: Decompose the feature sequence according to the monthly scale, weekly scale, daily scale and the remaining terms to obtain the corresponding decomposed components;

[0074] Step S32: Based on the corresponding decomposed components, linear prediction channels are allocated through a multi-scale hierarchical DLinear network to obtain independent linear prediction channels;

[0075] Step S33: Perform hierarchical prediction on the corresponding decomposition components based on the independent linear prediction channels to obtain hierarchical prediction results, and add the hierarchical prediction results element by element to obtain the prediction sequence of future water quality factors.

[0076] In one embodiment, after extracting the feature sequence Z containing complex coupling relationships between variables using a Fourier neural operator module, a multi-scale hierarchical DLinear network module is used for time series prediction. Water quality time series evolution is composed of superimposed patterns at different time scales, such as daily cycles, weekly cycles, and longer-term seasonal trends. Therefore, to accurately capture these hierarchical patterns, the input feature sequence Z (dimension L x D) undergoes multi-scale time series decomposition. Using multiple moving average kernels with different window sizes (24 hours, 168 hours, and 720 hours), the feature sequence Z is decomposed into multiple independent parts: a residual term Z_remain representing high-frequency noise, a component Z_daily representing daily fluctuations, a component Z_weekly representing weekly changes, and a long-term trend term Z_monthly_trend representing monthly and longer-term scales.

[0077] After completing the temporal decomposition, the multi-scale hierarchical DLinear network module assigns an independent linear prediction channel to each component at a different scale, achieving hierarchical prediction through a divide-and-conquer strategy. One Linear_daily layer is dedicated to learning and predicting the future evolution of the Z_daily component, one Linear_weekly layer focuses on the future evolution of the Z_weekly component, one Linear_monthly_trend layer focuses on the future evolution of the Z_monthly_trend component, and one Linear_remain layer focuses on the future evolution of the Z_remain component. Finally, the prediction results of the daily, weekly, long-term trend, and remainder terms from the independent scale channels are directly added element-wise and fused into a final comprehensive prediction for the next H time steps. This hierarchical structure of multi-scale decomposition-parallel linear prediction-fusion not only enhances the interpretability of the prediction results but also accurately captures the core evolutionary patterns of the system state at different time scales in an efficient and robust manner.

[0078] Preferably, step S4 includes the following steps:

[0079] Step S41: Repeat the forward inference on the sample data of the future target sequence based on the predicted sequence to obtain the future predicted sequence;

[0080] Step S42: Summarize the future prediction sequences obtained from multiple forward inferences in chronological order to form a set of prediction trajectories with dimension M×H×1, where M is the number of inferences and H is the time step length of the future prediction.

[0081] Step S43: Perform time-by-time statistics on the predicted trajectory set along the forward inference dimension to obtain the predicted mean, variance, and required confidence quantile for each future time step, where the quantiles include, but are not limited to, the 5th quantile and the 95th quantile.

[0082] Step S44: Based on the mean, variance and their confidence quantiles, generate confidence intervals to characterize the uncertainty of future evolution.

[0083] In one embodiment, an end-to-end training paradigm is employed based on the predicted sequence, combining the preceding Fourier neural operator module with the subsequent multi-scale hierarchical DLinear network module into a seamless, unified deep learning network. During model training, the root mean square error (RMSE) is used as the loss function to measure the difference between the model's final predicted output and the true target value. Once this final error is calculated, the backpropagation algorithm is initiated, propagating the error signal from the model's output layer, layer by layer, until it reaches the input layer at the very front. This process ensures that gradients can flow unimpeded throughout the network, while simultaneously updating the weights of the linear layers in the multi-scale hierarchical DLinear module and all learnable parameters in the Fourier neural operator module, including the learnable complex weight matrix in the frequency domain.

[0084] This end-to-end training mechanism is highly efficient. Direct supervision from the final target forces the Fourier neural operator modules to learn not generic, aimless features, but highly task-oriented, deeply coupled representations most beneficial for subsequent predictions. Therefore, end-to-end training ensures that the various modules in the model can co-evolve and cooperate closely, ultimately outputting a predicted sequence Y_pred (with dimensions H x 1).

[0085] Simultaneously, the Monte Carlo Dropout method is employed to probabilistically predict the future evolution of the target. Dropout layers are inserted at the outputs of the Fourier neural operator module and the multi-scale hierarchical DLinear module. During model training, these Dropout layers randomly disable a portion of neurons according to their preset probabilities to prevent overfitting. During the model inference and prediction phase, instead of switching the model to evaluation mode (eval()) to disable Dropout as in conventional inference prediction, the model remains in training mode (train()), ensuring that the Dropout layers remain active during prediction. In this state, multiple (e.g., 100) independent forward propagations are performed on the same set of input data. Since the combination of neurons randomly disabled in each forward propagation is different, the model produces multiple slightly different but equally reasonable future predictions. This set of multiple results forms a probability distribution of future evolution, from which the prediction mean at each future time point and the quantiles at the confidence level (e.g., 5% and 95% quantiles) can be calculated, thus constructing the confidence interval for the prediction. This provides crucial information for subsequent intelligent early warning.

[0086] Preferably, step S5 includes the following steps:

[0087] Step S51: Extract key dimension information from the probabilistic prediction results, including the predicted mean sequence, the trend of the mean sequence, and the upper limit of the confidence interval;

[0088] Step S52: Logically compare the predicted mean sequence, trend, and upper limit of the confidence interval with the preset multi-level concentration threshold and growth rate threshold to obtain a logical judgment result, wherein the logical comparison includes:

[0089] An emergency warning is issued when the predicted mean is ≥2.0 mg / L and the mean sequence change trend is ≥0.15 mg / L·h.

[0090] When the predicted mean is <2.0 mg / L but the upper limit of the confidence interval is ≥2.0 mg / L, it is judged as a high-risk warning;

[0091] When the predicted mean is between 1.0 and 1.5 mg / L and the trend is <0.10 mg / L·h, or the upper limit of the confidence interval is between 1.5 and 2.0 mg / L, it is judged as a warning of concern.

[0092] When the predicted mean is <1.0 mg / L, the trend is <0.05 mg / (L·h), and the upper limit of the confidence interval is <1.5 mg / L, it is considered a safe state;

[0093] Step S53: Generate water quality early warning information of the corresponding level based on the judgment result.

[0094] In one embodiment, after obtaining a probabilistic prediction of the future evolution of a target using the Monte Carlo Dropout method, including the predicted mean and confidence interval, a dynamic hierarchical early warning module transforms this numerical information into intuitive and actionable decision support for managers. This module provides intelligent early warning based on three dimensions: value, trend, and probability. It abandons the rigid model of traditional single-threshold alarms and can more finely distinguish different risk warning situations. The module first extracts information from the probabilistic prediction results in three key dimensions: "value," i.e., the predicted mean sequence, representing the most likely future concentration level; "trend," i.e., the changing trend obtained by calculating the rate of change of the mean sequence; and "probability," i.e., the risk situation requiring vigilance represented by the upper limit of the confidence interval (e.g., the 95th percentile). Secondly, the information in these three dimensions is logically compared with preset multi-level concentration thresholds and growth rate thresholds. For example, the triggering condition for an "emergency" warning may not simply be the predicted mean exceeding a certain threshold, but rather the result of the predicted mean exceeding a certain threshold and an abnormally rapid growth trend. A "high-risk" warning may be triggered when the predicted mean has not reached the danger line, but the risk ceiling has been exceeded.

[0095] Specifically, the predicted mean sequence, trend, and upper limit of the confidence interval are logically compared with preset multi-level concentration thresholds and growth rate thresholds to determine the water quality risk level. The specific logical comparison process includes: when the predicted mean is greater than or equal to 2.0 mg / L and the trend of the mean sequence is greater than or equal to 0.15 mg / L·h, it is determined as an emergency warning; when the predicted mean is less than 2.0 mg / L but the upper limit of the confidence interval is greater than or equal to 2.0 mg / L, it is determined as a high-risk warning; when the predicted mean is between 1.0 and 1.5 mg / L and the trend is less than 0.10 mg / L·h, or the upper limit of the confidence interval is between 1.5 and 2.0 mg / L, it is determined as a warning of concern; when the predicted mean is less than 1.0 mg / L and the trend is less than 0.05 mg / (L·h) and the upper limit of the confidence interval is less than 1.5 mg / L, it is determined as a safe state. Through the above logical determination, future water quality risks can be classified and identified, ensuring timely response to potential water quality anomalies.

[0096] Preferably, acquiring multi-sensor monitoring data of the target water body also includes:

[0097] Multiple sets of sensor data acquisition nodes are deployed in the target water body. Each set of nodes includes a dissolved oxygen electrode, a pH electrode, a thermometer, a turbidity meter, and a conductivity meter. The nodes are fixed in the water body by buoys, conduits, or underwater supports.

[0098] Multi-channel physical sampling is performed using a built-in sampling pump or water flow introduction device in the sensor. By adjusting the length and flow rate of the sampling pipeline, vertical and horizontal multi-channel measurement data can be obtained.

[0099] The measurement time interval of each sensor is adjusted based on the vertical and horizontal multi-channel measurement data, and the sensors that have been deployed are physically calibrated on site to collect multi-sensor monitoring data.

[0100] In one embodiment, multiple sets of sensor data acquisition nodes are first deployed in the target water body. Each set of nodes includes a dissolved oxygen electrode, a pH electrode, a thermometer, a turbidimeter, and a conductivity meter. Each set of nodes is fixed in the water body by a buoy, conduit, or underwater support to ensure that the sensors maintain a stable position during data acquisition and cover the key measurement areas of the water body. By rationally arranging the nodes, data from the surface, middle, and bottom layers of the water body can be acquired, thereby enabling monitoring of different water layers.

[0101] Subsequently, multi-channel physical sampling is performed using the built-in sampling pumps or water flow introduction devices of each sensor. By adjusting the sampling pipeline length and water flow velocity, water samples can be collected at different depths and horizontal positions in the target water body to obtain multi-channel measurement data in both vertical and horizontal directions.

[0102] After acquiring vertical and horizontal multi-channel measurement data, the measurement time intervals of each sensor are adjusted based on the measurement characteristics of different strata to ensure that key water quality parameters can be collected at a higher frequency during periods of high variability. Simultaneously, on-site physical calibration is performed on the deployed sensors, including standard solution calibration, zero-point calibration, and range calibration, to ensure the accuracy and consistency of the output data from each sensor. Once calibration is complete, continuous and reliable multi-sensor monitoring data can be obtained for subsequent water quality analysis, stratified assessment, and intelligent early warning.

[0103] Preferably, on-site physical calibration of the deployed sensors includes:

[0104] Each sensor was calibrated using standard solutions. By sequentially introducing standard solutions of different concentrations into the sensor sampling ports, the corresponding output values ​​were obtained, thereby establishing the sensor response curve.

[0105] During the calibration process, the sampling flow rate and circulation direction of the sensors are physically adjusted, and adjacent sensors are calibrated in an alternating sequence.

[0106] The sensor sampling port was rinsed on-site, and the sensor output results were obtained after rinsing.

[0107] Multi-node cross-validation is performed on the sensor output results to obtain multi-sensor monitoring data.

[0108] In one embodiment, each sensor is first calibrated using standard solutions. Specifically, this involves sequentially introducing standard solutions of different concentrations into the sampling port of each sensor and recording the corresponding output values ​​to establish a response curve for each sensor. This step ensures the measurement accuracy of the sensors under different concentration conditions and provides a basic reference for subsequent data correction.

[0109] During calibration, to prevent cross-interference between adjacent sensors, the sampling flow rate and circulation direction of the sensors are physically adjusted, and each sensor is calibrated in an alternating sequence. This maintains sampling stability while avoiding signal interference between adjacent sensors, improving the reliability and repeatability of calibration data.

[0110] After calibration, the sensor sampling port was rinsed on-site to remove residual liquid or contaminants. During rinsing, the water flow rate and circulation method were controlled to ensure the sampling port was thoroughly cleaned. Subsequently, the sensor output results after rinsing were acquired to verify the stability and accuracy of the sensor under actual water conditions.

[0111] Finally, multi-node cross-validation was performed on the output results of each sensor. Measurement data from different sensors and different layers were compared and verified, outliers were removed, and highly reliable multi-sensor monitoring data was generated. Through the above steps, the on-site physical calibration of the deployed sensors was completed, ensuring that their output data accurately reflects the true water quality of the target water body.

[0112] Preferably, calibrating adjacent sensors by physically adjusting the sampling flow rate and circulation direction of the sensors and using an alternating sequence includes:

[0113] Obtain the sensor's water inlet structure and natural flow path;

[0114] The sampling flow rate of the sensor is physically regulated based on the water inlet structure of the sensor, and the circulation direction of the sensor is adjusted through the natural flow path of the sensor to obtain the first adjustment data of the sensor.

[0115] Based on the first adjustment data of the sensors, the startup sequence of adjacent sensors is staggered to obtain the second adjustment data of the sensors;

[0116] The sensor and adjacent sensors are calibrated based on the first and second adjustment data of the sensor.

[0117] In one embodiment, after the sensor deployment is completed, the water inlet structure information and its natural flow path of each sensor are first obtained to determine the flow direction and distribution characteristics of water inside the sensor.

[0118] Based on the acquired inlet water structure data, operators or the automatic control system physically adjust the sampling flow rate of the sensors. This can be done, for example, by adjusting the sampling pump rate or valve opening to control the water flow entering the sensors and keep the sampling flow rate within a predetermined range. Simultaneously, the internal circulation direction of the sensors is adjusted according to the natural flow path, ensuring that the water flow fully covers the sensor detection unit along the designed path, thereby guaranteeing sampling uniformity. After completing these adjustments, initial adjustment data for each sensor is generated for subsequent calibration reference.

[0119] Subsequently, based on the initial sensor adjustment data, the startup sequence of adjacent sensors is staggered. For example, odd-numbered sensors are started first, followed by even-numbered sensors, to avoid water flow interference or signal crosstalk between adjacent sensors. This staggered startup strategy generates second adjustment data, providing a stable measurement environment for calibration.

[0120] Finally, combining the first and second adjustment data, calibration was performed on each sensor and its adjacent sensors. During calibration, standard solutions of different concentrations were sequentially introduced into the sensor sampling ports, and the output signals were recorded. By cross-comparing the response curves of each sensor, accurate calibration of multiple sensors was completed, thereby ensuring that the collected water quality data has high accuracy and repeatability.

[0121] Preferably, staggering the startup sequence of adjacent sensors based on the first adjustment data of the sensors includes:

[0122] Based on the first adjustment data of the sensor, the startup sequence of adjacent sensors is staggered so that adjacent sensors do not enter the sampling state at the same time point, and the first and second startup sensors are confirmed.

[0123] The water flow trace perception and discrimination data of the second start-up sensor are confirmed based on the first start-up sensor.

[0124] By analyzing the response time of the water flow trace perception and discrimination data, the difference in response sensitivity between the first and second start-up sensors under real water flow environment was determined.

[0125] Secondary sensor adjustment data is generated by using differences in response sensitivity to characterize the activation characteristics of adjacent sensors.

[0126] In one embodiment, after physically adjusting the sampling flow rate and circulation direction of a single sensor and generating first adjustment data, the startup sequence of adjacent sensors is staggered using this first adjustment data. Specifically, according to the design scheme, adjacent sensors do not enter the sampling state simultaneously at the same time point, thereby reducing water flow interference and measurement errors, and clearly distinguishing between the first startup sensor and the second startup sensor.

[0127] After confirming the first activation sensor, the second activation sensor is used to sense and identify the water flow traces generated by the first activation sensor, and the corresponding response data is recorded. This water flow trace sensing and identification data is used to analyze the mutual influence between adjacent sensors in a real water flow environment.

[0128] Subsequently, by analyzing the response time of the water flow trace sensing and discrimination data, the difference in response sensitivity between the first and second activation sensors under actual water flow conditions was evaluated. This analysis quantifies the response delay and interference of adjacent sensors, providing data support for a staggered activation strategy.

[0129] Finally, based on the differences in response sensitivity, second adjustment data is generated to characterize the startup characteristics of adjacent sensors. This data is used to guide the startup sequence of sensors during actual calibration operations, ensuring that adjacent sensors can obtain a stable sampling environment during staggered startup, thereby guaranteeing the measurement accuracy and reliability of multiple sensors.

[0130] Of particular importance is that the water flow trace sensing and discrimination data of the second start-up sensor, confirmed by the first start-up sensor, includes:

[0131] The initial flow direction of the water in the target area is determined by the activation time of the first activation sensor and the water flow trigger state.

[0132] The sampling response time of the second start-up sensor is compared with the initial flow direction to determine whether the second start-up sensor is in the path of the water flow.

[0133] Based on the advance or lag of the second starting sensor's response in the water flow path, the strength of its perception of water flow traces is confirmed.

[0134] The second activation sensor generates water flow trace perception and discrimination data based on the perception strength characteristics.

[0135] In one embodiment, during the staggered activation of adjacent sensors, when the first sensor enters the sampling state, its sampling behavior generates a measurable water flow triggering effect on the surrounding water body. By recording the activation time of the first sensor and the water flow triggering state formed in its internal flow channel after activation, the initial flow direction of the water in the target water body area can be determined. This initial flow direction serves as the basis for determining whether the water flow will affect the second sensor.

[0136] Subsequently, based on the initial flow direction, the sampling response time of the second activation sensor in the corresponding direction is compared to determine whether the second activation sensor is located on the water flow path of the first activation sensor. When the sampling response time of the second activation sensor is consistent with the water flow direction and shows a significant response change, it can be determined that it is on the water flow path; otherwise, it is determined that the second activation sensor is less affected by the water flow.

[0137] Furthermore, based on whether the second initiation sensor is located along the water flow path and its response advance or lag within that path, the strength of the second initiation sensor's perception of water flow traces can be determined. For example, if the second initiation sensor shows a significant advance response before the water flow arrives, its perception capability is considered strong; while if its response is significantly delayed, its perception capability is considered weak.

[0138] Finally, based on the determined sensing strength characteristics, the water flow trace sensing and discrimination data of the second start-up sensor is constructed and used in the subsequent sensor start-up optimization and calibration process to ensure that the multi-sensor network maintains high-precision collaborative measurement performance in the on-site flowing water environment.

[0139] Of particular importance is that the water flow trace perception and discrimination data generated based on the perception strength characteristics of the second activation sensor includes:

[0140] By utilizing the water flow contact continuity reflected in the sensing strength characteristics, the stability of the response of the second activation sensor during the water flow process can be distinguished.

[0141] Based on the difference between the response stability and the response amplitude of the first start-up sensor, the relationship between the flow depth of the second start-up sensor in the water flow path is determined.

[0142] The sensing levels of the second activation sensor are sorted according to the depth of the current to obtain the corresponding water flow trace recognition level;

[0143] The final judgment state of the second start sensor regarding the existence of water flow is determined by the water flow trace recognition level, and water flow trace perception and judgment data of the second start sensor is generated.

[0144] In one embodiment, after identifying the sensing strength characteristics of the second activation sensor, the stability of the second activation sensor's response during the water flow process is first distinguished using the water flow contact continuity reflected in these characteristics. Specifically, when the water flow continuously acts on the sensor, its output signal usually exhibits a stable response with good time continuity; if the continuity is weak, its output signal often shows intermittent or jittery changes. Thus, the sensor's response stability can be classified as stable, transient, or weakly sensing.

[0145] Subsequently, based on the stability of the response of the second activation sensor and in combination with the difference in response amplitude of the first activation sensor, the depth of the second activation sensor in the water flow path is determined. For example, when the response amplitudes of the second activation sensor and the first activation sensor are close and change synchronously, it can be considered that it is in the main path of the water flow; if the response amplitude is significantly lower than that of the first activation sensor, it can be considered that it is only in the edge region of the water flow; if the response is significantly weak or the change is not obvious, it is determined that it is only exposed to a weak water flow.

[0146] After determining the depth of the current flow, the sensing levels of the second activation sensor are ranked according to these different depths. The ranking method includes classifying the sensors into different levels of water flow trace recognition, such as "strong sensing," "medium sensing," and "weak sensing." This level not only reflects its sensitivity to the presence of water flow but also represents its reference value for judging water flow trajectories in actual aquatic environments.

[0147] Finally, the generated water flow trace recognition level is used to further determine the final judgment state of the second activation sensor regarding the existence of water flow. When the recognition level is high, it is determined that it can truly reflect the passage of water flow; if the level is low, it is determined that it is located in a non-primary water flow path or an area with weak water flow influence. Based on the above judgment state, water flow trace perception and discrimination data of the second activation sensor is generated.

[0148] 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.

[0149] 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 smart water quality early warning method based on multi-sensor data fusion, characterized in that, Includes the following steps: Step S1: Acquire multi-sensor monitoring data of the target water body; divide the multi-sensor monitoring data into sample data containing historical sequences and future target sequences; Step S2: Perform linear coupling and residual nonlinear processing on the environmental factors of the sample data at each time step in the historical sequence to extract the instantaneous interaction features between environmental factors and obtain the feature sequence; wherein, step S2 includes the following steps: Step S21: Extract environmental factors from the sample data at each time step in the historical sequence; Step S22: Use the environmental factors as input vectors to perform dimensionality upscaling to obtain the dimensionality-upgraded environmental vector; Step S23: Perform linear coupling and residual nonlinear processing on the upgraded environment vectors in each time step, extract the instantaneous interaction features between environmental factors, and obtain the updated feature vector; Step S24: Combine the updated feature vectors to obtain the feature sequence; wherein, step S23 includes: Perform a one-dimensional fast Fourier transform on the environment vector after dimensionality increase at each time step to obtain the frequency domain of the environment vector; The low-frequency environment vector frequency domain is filtered based on a preset frequency domain threshold, and the corresponding environment vector is linearly transformed according to the low-frequency environment vector frequency domain to obtain the linearly transformed environment vector frequency domain. The frequency domain of the linearly transformed environment vector is transformed into the spatial domain through the inverse Fourier transform to obtain the updated environment vector. Based on the nonlinear fusion of the updated environment vector and the environmental vector residual after dimensionality increase, an updated feature vector is obtained to represent the environmental coupling characteristics at the t-th time step. Step S3: Decompose the feature sequence into monthly, weekly, daily, and residual terms, and generate prediction results for the corresponding time scales through independent linear prediction channels. Then, fuse the prediction results at each scale to obtain the prediction sequence of future water quality factors. Step S3 includes the following steps: Step S31: Decompose the feature sequence according to the monthly scale, weekly scale, daily scale and the remaining terms to obtain the corresponding decomposed components; Step S32: Based on the corresponding decomposed components, linear prediction channels are allocated through a multi-scale hierarchical DLinear network to obtain independent linear prediction channels; Step S33: Perform hierarchical prediction on the corresponding decomposition components based on the independent linear prediction channel to obtain the hierarchical prediction results, and add the hierarchical prediction results element by element to obtain the prediction sequence of future water quality factors; Step S4: Perform forward inference on the sample data of the future target sequence based on the predicted sequence to obtain multiple predicted trajectories, and obtain the predicted mean and confidence interval based on the predicted trajectories. Step S5: Based on the predicted mean, the trend of the predicted trajectory, the upper limit of the confidence interval, and the preset multi-level thresholds, a judgment is made to output the corresponding level of water quality early warning information.

2. The intelligent water quality early warning method based on multi-sensor data fusion according to claim 1, characterized in that, Step S4 includes the following steps: Step S41: Repeat the forward inference on the sample data of the future target sequence based on the predicted sequence to obtain the future predicted sequence; Step S42: Summarize the future prediction sequences obtained from multiple forward inferences in chronological order to form a set of prediction trajectories with dimension M×H×1, where M is the number of inferences and H is the time step length of the future prediction. Step S43: Perform time-by-time statistics on the predicted trajectory set along the forward inference dimension to obtain the predicted mean, variance, and required confidence quantile for each future time step, where the quantiles include, but are not limited to, the 5th quantile and the 95th quantile. Step S44: Based on the mean, variance and their confidence quantiles, generate confidence intervals to characterize the uncertainty of future evolution.

3. The intelligent water quality early warning method based on multi-sensor data fusion according to claim 1, characterized in that, Step S5 includes the following steps: Step S51: Extract key dimension information from the probabilistic prediction results, including the predicted mean sequence, the trend of the mean sequence, and the upper limit of the confidence interval; Step S52: Logically compare the predicted mean sequence, trend, and upper limit of the confidence interval with the preset multi-level concentration threshold and growth rate threshold to obtain a logical judgment result, wherein the logical comparison includes: An emergency warning is issued when the predicted mean is ≥2.0 mg / L and the mean sequence change trend is ≥0.15 mg / L·h. When the predicted mean is <2.0 mg / L but the upper limit of the confidence interval is ≥2.0 mg / L, it is judged as a high-risk warning; When the predicted mean is between 1.0 and 1.5 mg / L and the trend is <0.10 mg / L·h, or the upper limit of the confidence interval is between 1.5 and 2.0 mg / L, it is judged as a warning of concern. When the predicted mean is <1.0 mg / L, the trend is <0.05 mg / L·h, and the upper limit of the confidence interval is <1.5 mg / L, it is considered a safe state; Step S53: Generate water quality early warning information of the corresponding level based on the judgment result.

4. The intelligent water quality early warning method based on multi-sensor data fusion according to claim 1, characterized in that, Acquiring multi-sensor monitoring data of the target water body also includes: Multiple sets of sensor data acquisition nodes are deployed in the target water body. Each set of nodes includes a dissolved oxygen electrode, a pH electrode, a thermometer, a turbidity meter, and a conductivity meter. The nodes are fixed in the water body by buoys, conduits, or underwater supports. Multi-channel physical sampling is performed using a built-in sampling pump or water flow introduction device in the sensor. By adjusting the length and flow rate of the sampling pipeline, vertical and horizontal multi-channel measurement data can be obtained. The measurement time interval of each sensor is adjusted based on the vertical and horizontal multi-channel measurement data, and the sensors that have been deployed are physically calibrated on site to collect multi-sensor monitoring data.

5. The intelligent water quality early warning method based on multi-sensor data fusion according to claim 4, characterized in that, On-site physical calibration of the deployed sensors includes: Each sensor was calibrated using standard solutions. By sequentially introducing standard solutions of different concentrations into the sensor sampling ports, the corresponding output values ​​were obtained, thereby establishing the sensor response curve. During the calibration process, the sampling flow rate and circulation direction of the sensors are physically adjusted, and adjacent sensors are calibrated in an alternating sequence. The sensor sampling port was rinsed on-site, and the sensor output results were obtained after rinsing. Multi-node cross-validation is performed on the sensor output results to obtain multi-sensor monitoring data.

6. The intelligent water quality early warning method based on multi-sensor data fusion according to claim 5, characterized in that, Calibration of adjacent sensors by physically adjusting the sampling flow rate and circulation direction and using an alternating sequence includes: Obtain the sensor's water inlet structure and natural flow path; The sampling flow rate of the sensor is physically regulated based on the water inlet structure of the sensor, and the circulation direction of the sensor is adjusted through the natural flow path of the sensor to obtain the first adjustment data of the sensor. Based on the first adjustment data of the sensors, the start-up order of adjacent sensors is staggered to obtain the second adjustment data of the sensors; The sensor and adjacent sensors are calibrated based on the first and second adjustment data of the sensor.

7. The intelligent water quality early warning method based on multi-sensor data fusion according to claim 6, characterized in that, The staggered startup sequence of adjacent sensors based on the initial adjustment data from the sensors includes: Based on the first adjustment data of the sensor, the startup sequence of adjacent sensors is staggered so that adjacent sensors do not enter the sampling state at the same time point, and the first and second startup sensors are confirmed. The water flow trace perception and discrimination data of the second start-up sensor are confirmed based on the first start-up sensor. By analyzing the response time of the water flow trace perception and discrimination data, the difference in response sensitivity between the first and second start-up sensors under real water flow environment was determined. Secondary sensor adjustment data, which characterizes the activation characteristics of adjacent sensors, is generated by the difference in response sensitivity.

Images