A multi-parameter environmental quality intelligent monitoring system and control method thereof

Abstract

The present application provides a multi-parameter environmental quality intelligent monitoring system and a control method thereof. The method includes: collecting environmental quality data, the environmental quality data including water quality parameters, air parameters and soil parameters; preprocessing the environmental quality data to obtain target environmental quality data, the preprocessing including data cleaning, standardization and time alignment; constructing a random neural network with forward regularization, and performing feature extraction and association analysis on the target environmental quality data to obtain feature data; based on the feature data, training an environmental quality assessment and early warning model to obtain an environmental quality monitoring model; deploying the environmental quality monitoring model to perform online monitoring and early warning of the environmental quality data. The present application realizes intelligent analysis and real-time early warning of environmental quality data, and improves the accuracy and efficiency of environmental monitoring.

Description

Technical Field

[0001] The present application relates to the technical field of environmental monitoring, and in particular to a multi-parameter environmental quality intelligent monitoring system and a control method thereof. Background Art

[0002] Environmental quality monitoring is an important means to ensure environmental safety and public health. At present, environmental quality monitoring mainly adopts the following technical solutions: The traditional single parameter monitoring method collects specific environmental parameters through a single sensor, but it cannot achieve multi-parameter coordinated monitoring and comprehensive analysis.

[0003] Although simple multi-parameter monitoring systems can collect multiple environmental parameters at the same time, they lack intelligent analysis capabilities, have low data utilization, and cannot fully tap the value of multi-source data.

[0004] Environmental monitoring methods based on traditional machine learning have introduced certain intelligent analysis capabilities. However, due to the relatively simple algorithm model, it is difficult to accurately identify complex environmental pollution patterns, and there is a problem of insufficient processing capabilities when faced with multi-source heterogeneous data.

[0005] The main technical problems in the existing technology are as follows: the real-time and accuracy of environmental parameter monitoring data are insufficient, making it difficult to detect and warn of environmental quality problems in a timely manner; in addition, there is a lack of deep integration and analysis capabilities for multi-source heterogeneous data, making it impossible to effectively explore the correlation between environmental parameters. Summary of the invention

[0006] In view of this, the present application provides a multi-parameter environmental quality intelligent monitoring system and a control method thereof, which solves the problems in the prior art of insufficient real-time and accuracy of environmental parameter monitoring data, and lack of deep fusion and analysis capabilities of multi-source heterogeneous data.

[0007] The present application provides a control method for a multi-parameter environmental quality intelligent monitoring system, including: Collecting environmental quality data, wherein the environmental quality data includes water quality parameters, air parameters and soil parameters; Preprocessing the environmental quality data to obtain target environmental quality data, wherein the preprocessing includes data cleaning, standardization and time alignment; Constructing a random neural network with forward regularization, and performing feature extraction and association analysis on the target environmental quality data to obtain feature data; Based on the characteristic data, an environmental quality assessment and early warning model is trained to obtain an environmental quality monitoring model; The environmental quality monitoring model is deployed to perform online monitoring and early warning of the environmental quality data.

[0008] Optionally, the preprocessing of the environmental quality data to obtain target environmental quality data includes: Identify outliers based on statistical and machine learning methods, and use preset criteria and isolation forest algorithm to perform anomaly detection; mark and process detected outliers, select deletion or correction schemes based on data characteristics, and use wavelet transform to perform noise reduction on data to obtain the target environmental quality data; or, Corresponding standardization methods are used for different types of environmental parameters to standardize the environmental quality data. The standardization methods include maximum and minimum value standardization or Z-score of the environmental quality data; time alignment of different parameter data is performed based on timestamp information, and a linear interpolation method is used to deal with the problem of inconsistent sampling frequencies. A unified data format is constructed to encapsulate the data to obtain the target environmental quality data.

[0009] Optionally, constructing a random neural network with forward regularization comprises: According to the type and number of environmental parameters, a random neural network infrastructure including an input layer, a hidden layer and an output layer is designed, wherein the input layer includes nodes corresponding to the number of environmental parameters, and the hidden layer is arranged in descending order according to a preset number; Based on the network architecture, a preset L1 regularization coefficient and an L2 regularization coefficient are introduced in the forward propagation process to control the model complexity by minimizing the regularization loss function; An incremental learning mechanism is implemented, a mini-batch method is adopted according to a preset batch size, and a preset optimizer is used to update parameters to obtain the random neural network, and the model has the characteristic of adapting to dynamic changes in environmental parameters.

[0010] Optionally, the performing feature extraction and correlation analysis on the target environmental quality data to obtain feature data includes: Based on the random neural network, the target environmental quality data is subspace-decomposed according to parameter types to obtain characteristic subspaces corresponding to different types of target environmental quality data; Based on the characteristic subspace, a plurality of groups of permutation data corresponding to each subspace are generated by a random permutation algorithm; According to the permuted data, the original environmental parameter relationship is reconstructed based on a matrix recovery algorithm of nuclear norm minimization; The reconstruction result is iteratively optimized by using an accelerated proximal gradient descent method to improve the recovery accuracy of the subspace matrix so as to determine the characteristic data.

[0011] Optionally, reconstructing the original environmental parameter relationship according to the permuted data using a matrix recovery algorithm based on nuclear norm minimization includes: Construct an optimization objective function to transform the matrix recovery problem into a combined optimization problem of nuclear norm minimization and Frobenius norm; Setting an iteration termination condition, wherein the iteration termination condition includes that the relative error is less than a preset threshold or the maximum number of iterations is reached; In each iteration, the reconstruction weights of each subspace are dynamically updated based on the adaptive weight adjustment strategy; A time series smoothing constraint is introduced to ensure that the restored data maintains time continuity, and the optimal regularization parameter is selected through cross-validation.

[0012] Optionally, the training of an environmental quality assessment and early warning model based on the characteristic data to obtain an environmental quality monitoring model includes: Based on the characteristic data, a plurality of expert sub-models are constructed, each expert sub-model being responsible for the evaluation of a corresponding environmental parameter; According to the evaluation results of the expert sub-model, a gating network is constructed, and a preset discretization selection mechanism is used to dynamically select the most suitable expert combination; Based on the selected expert combination, the number of experts activated simultaneously is limited by using a sparse technology, and the outputs of multiple experts are adaptively fused to obtain the environmental quality monitoring model.

[0013] Optionally, constructing a plurality of expert sub-models based on the feature data includes: According to the characteristic dimensions of environmental parameters, a multi-layer feedforward neural network structure is constructed for each expert model, in which the input layer dimension matches the environmental parameter dimension, and the hidden layer decreases layer by layer according to the preset ratio; Set the initial value and decay strategy of the temperature parameter to control the softness and hardness of the expert selection; A phased training strategy is adopted to train each expert model independently, train the gating network used to select experts, and perform overall end-to-end fine-tuning; The loss function is designed based on prediction error, sparsity and diversity to guide the training process of each expert model and obtain multiple expert sub-models.

[0014] Optionally, the deploying the environmental quality monitoring model to perform online monitoring and early warning on the environmental quality data includes: Compressing and quantizing the environmental quality monitoring model, and converting model parameters into fixed-point representation; Using knowledge distillation technology, knowledge of multiple expert models is transferred to a lightweight model, wherein the lightweight model is generated based on the environmental quality monitoring model; Design a model cache mechanism to store the frequently accessed parameters and intermediate results of the lightweight model in a fast cache; Based on historical data and expert experience, warning thresholds are set for different environmental parameters, and the warning thresholds are dynamically adjusted according to the model's predicted confidence in pollutants. Warning signals are issued in a timely manner by analyzing parameter change trends.

[0015] Optionally, the adopting of knowledge distillation technology to migrate the knowledge of multiple expert models into the lightweight model includes: The original expert model is used as the teacher model to build a lightweight model with fewer parameters; Designing a distillation loss function, wherein the loss function includes a soft target loss and a hard target loss; Based on the loss function, gradually migrating the knowledge of the teacher model to the lightweight model through iterative training; The performance of the migrated lightweight model is evaluated to ensure that it is within a preset accuracy range.

[0016] The present application also provides an intelligent environmental quality monitoring device, including: A sensor array for collecting environmental quality data, wherein the environmental quality data includes water quality parameters, air parameters and soil parameters; A preprocessing module, connected to the sensor array, for preprocessing the environmental quality data to obtain target environmental quality data, wherein the preprocessing includes data cleaning, standardization and time alignment; A neural network module, connected to the preprocessing module, is used to construct a random neural network with forward regularization, and perform feature extraction and association analysis on the target environmental quality data to obtain feature data; A training module, connected to the neural network module, for training an environmental quality assessment and early warning model based on the characteristic data to obtain an environmental quality monitoring model; A deployment module is connected to the training module and is used to deploy the environmental quality monitoring model to perform online monitoring and early warning of the environmental quality data.

[0017] An embodiment of the present application also provides a computer device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above-mentioned environmental quality intelligent monitoring method.

[0018] An embodiment of the present application also provides a computer-readable storage medium, which stores computer instructions, and the computer instructions are used to enable a computer to execute the above-mentioned environmental quality intelligent monitoring method.

[0019] An embodiment of the present application also provides a computer program product, including computer instructions, which, when executed by a processor, implement the steps of the above-mentioned environmental quality intelligent monitoring method.

[0020] This application has the following technical effects: This application realizes the deep fusion analysis of multi-source heterogeneous environmental data by constructing a random neural network with forward regularization and multi-subspace matrix recovery technology, significantly improving the accuracy of environmental parameter monitoring.

[0021] This application uses a shallow expert sparse mixture model and an interpretable reinforcement learning method to achieve accurate assessment and reliable early warning of environmental quality and enhance the interpretability of the system.

[0022] This application ensures the real-time performance and warning reliability of the system and improves the efficiency of environmental monitoring through lightweight deployment and multi-level warning mechanism. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a flow chart of a control method of a multi-parameter environmental quality intelligent monitoring system provided in an embodiment of the present application;

[0024] Figure 2 is a schematic diagram of a random neural network structure provided in an embodiment of the present application; Figure 3 It is a flowchart of a multi-subspace matrix recovery method provided in an embodiment of the present application; Figure 4 It is a structural diagram of the multi-parameter environmental quality intelligent monitoring system provided in the embodiment of the present application. DETAILED DESCRIPTION

[0025] like Figure 1 As shown, the implementation method of the control method of the multi-parameter environmental quality intelligent monitoring system is as follows: S1: Collect environmental quality data, which includes water quality parameters, air parameters and soil parameters.

[0026] In a specific embodiment of the present application, the collection of environmental quality data is achieved by deploying a multi-parameter sensor array.

[0027] Specifically, the sensor array includes three subsystems: water quality sensor group, air sensor group and soil sensor group. Each subsystem is responsible for parameter collection in the corresponding field.

[0028] First, the water quality sensor group includes pH sensor, dissolved oxygen sensor, COD sensor, ammonia nitrogen sensor, total phosphorus sensor and turbidity sensor, which are directly immersed in the water body through waterproof probes for real-time monitoring. Among them, the measurement range of pH sensor is 0-14, and the accuracy reaches ±0.1; the measurement range of dissolved oxygen sensor is 0-20mg / L, and the accuracy is ±0.2mg / L; the measurement range of COD sensor is 0-2000mg / L; the measurement range of ammonia nitrogen sensor is 0-100mg / L; the measurement range of total phosphorus sensor is 0-50mg / L; the measurement range of turbidity sensor is 0-1000NTU.

[0029] Secondly, the air sensor group consists of PM2.5 sensor, PM10 sensor, SO2 sensor, NO2 sensor, CO sensor, O3 sensor and temperature and humidity sensor. These sensors are dustproof and waterproof and have automatic calibration function.

[0030] PM2.5 and PM10 sensors use the laser scattering principle, with a measurement range of 0-1000μg / m³ respectively; the gas sensor uses the electrochemical principle, with a SO2 measurement range of 0-2000ppb, a NO2 measurement range of 0-200ppb, a CO measurement range of 0-1000ppm, and an O3 measurement range of 0-1000ppb; the temperature sensor has a measurement range of -40℃ to 85℃, and the humidity sensor has a measurement range of 0-100%RH.

[0031] Again, the soil sensor group includes soil pH sensor, organic matter content sensor and heavy metal content sensor. The soil pH sensor uses the glass electrode method with a measurement range of 0-14; the organic matter content sensor is based on the principle of near-infrared spectroscopy with a measurement range of 0-10%; the heavy metal content sensor uses ion selective electrode technology and can detect the content of multiple heavy metal ions at the same time.

[0032] In addition, to ensure the reliability of data collection, all sensors are designed at industrial level, with dustproof, waterproof, anti-interference and other characteristics, and support 4G / 5G wireless communication, which can realize real-time data transmission. The sampling frequency of the sensor can be adjusted according to actual needs. Usually, the sampling interval of water quality and air parameters is 5 minutes, and the sampling interval of soil parameters is 30 minutes. At the same time, the system is also equipped with a backup power supply and data cache module to ensure the continuity of data collection in the event of network interruption or power supply abnormality.

[0033] It should be noted that the layout of the sensor array needs to take into account the geographical characteristics and environmental features of the monitoring area, and ensure the representativeness of the monitoring data through a scientific deployment plan. In actual deployment, water quality sensors are mainly installed at typical sections of key water areas; air sensors are arranged in a grid based on the division of urban functional areas and the distribution of pollution sources; soil sensors are mainly deployed in key areas such as farmland and industrial areas, forming a comprehensive environmental quality monitoring network.

[0034] S2: Preprocessing the environmental quality data to obtain target environmental quality data, wherein the preprocessing includes data cleaning, standardization and time alignment.

[0035] S2 includes: S2.1: Identify outliers based on statistical and machine learning methods, and use preset criteria and isolation forest algorithm to detect anomalies; mark and process the detected outliers, select deletion or correction schemes based on data characteristics, use wavelet transform to reduce noise on the data, and obtain the target environmental quality data.

[0036] Regarding the data cleaning process of step S2.1.

[0037] The system uses a combination of statistical and machine learning methods to detect outliers. Specifically, for each type of environmental parameter, the 3σ criterion is first used for preliminary screening, that is, data that exceeds the mean ±3 times the standard deviation is marked as a potential outlier. For example, for PM2.5 data, if the value of a certain measurement point exceeds the mean ±3 times the standard deviation of the data in the area over the past 24 hours, it will be marked as suspicious data.

[0038] The embodiment of the present application also introduces an isolation forest algorithm for deep anomaly detection. The algorithm calculates the anomaly score of each data point by constructing a decision tree.

[0039] In specific implementation, samples are randomly projected into different feature spaces. Outliers can often be isolated quickly due to the difference between their features and normal values. For example, when the dissolved oxygen data at a certain time point is detected to be seriously deviated from the historical law, the algorithm will give a higher anomaly score. The system sets the anomaly score threshold to 0.7, and data points exceeding this threshold will be marked as anomalies.

[0040] In terms of outlier processing, different processing strategies are adopted according to data characteristics. For example, outliers that are obviously caused by sensor failure (such as data with a value of 0 or out of range) are directly deleted. For example, for data that is slightly deviated but has reference value, a method based on moving median is used to correct it. In addition, the system also introduces wavelet transform for data noise reduction, using db4 wavelet basis function to decompose the data into 3 layers, effectively removing high-frequency noise while retaining the effective information of the data.

[0041] or, S2.2: According to different types of environmental parameters, corresponding standardization methods are used to standardize the environmental quality data, and the standardization methods include maximum and minimum value standardization or Z-score of the environmental quality data; different parameter data are time-aligned based on timestamp information, and the problem of inconsistent sampling frequency is handled by linear interpolation method, and a unified data format is constructed to encapsulate the data to obtain the target environmental quality data; The cleaned data is normalized. For different types of environmental parameters, corresponding normalization methods are used. For parameters with clear physical meanings and ranges (such as pH value), the maximum and minimum value normalization method is used to map the data to the [0, 1] interval. The normalization formula is: x_normalized = (x - x_min) / (x_max - x_min), where x_min and x_max are the theoretical minimum and maximum values ​​of the parameter, respectively.

[0042] For parameters with different dimensions (such as PM2.5 and gas concentration), the Z-score normalization method is used to convert the data into a standard normal distribution with a mean of 0 and a standard deviation of 1. The normalization formula is: x_normalized = (x - μ) / σ, where μ is the sample mean and σ is the sample standard deviation. This method can eliminate the dimensional differences between different parameters, making the data more comparable.

[0043] In terms of time alignment, since different sensors have different sampling frequencies (such as once every 5 minutes for air parameters and once every 30 minutes for soil parameters), the system uses piecewise linear interpolation to align time series. In specific implementation, the minimum sampling interval (5 minutes) is used as the benchmark to interpolate parameters with lower sampling frequencies. The time-varying characteristics of the parameters are considered during interpolation. Local weighted average interpolation is used for parameters that change faster (such as gas concentration), and linear interpolation is used for parameters that change slower (such as soil pH).

[0044] The system encapsulates the processed data into a unified data format.

[0045] The data format includes fields such as timestamp (accurate to seconds), parameter ID, parameter value, quality mark (indicating whether the data is original value or interpolation). This unified data format facilitates subsequent feature extraction and model training. In addition, the system also saves the intermediate results of data processing and quality control logs to facilitate tracing the data processing process and quality assessment.

[0046] It should be noted that the parameters of data preprocessing (such as anomaly detection threshold, interpolation method, etc.) can be dynamically adjusted according to the actual application scenario. The system provides a parameter configuration interface, allowing professionals to optimize the processing strategy according to monitoring needs and data characteristics.

[0047] S3: Construct a random neural network with forward regularization, and perform feature extraction and association analysis on the target environmental quality data to obtain feature data.

[0048] In the specific embodiment of the present application, the construction and training of random neural networks is the core link of environmental quality monitoring. Figure 2 As shown in the figure, the network not only needs to process multi-source heterogeneous environmental data, but also has the ability of online learning and dynamic adaptation.

[0049] Among them, the specific implementation process of constructing a random neural network with forward regularization is as follows: S3.1: According to the type and number of environmental parameters, a random neural network infrastructure including an input layer, a hidden layer and an output layer is designed, wherein the input layer includes nodes corresponding to the number of environmental parameters, and the hidden layers are arranged in descending order according to a preset number.

[0050] In step S3.1, the system designs the network architecture according to the type and number of environmental parameters. The input layer is designed to have 17 nodes, corresponding to different environmental parameters, including 6 water quality parameters (pH, dissolved oxygen, COD, ammonia nitrogen, total phosphorus, turbidity), 8 air parameters (PM2.5, PM10, SO2, NO2, CO, O3, temperature, humidity) and 3 soil parameters (pH, organic matter content, heavy metal content). The hidden layer adopts a multi-layer design, specifically configured as a three-layer structure [128, 64, 32]. The number of nodes decreases layer by layer according to the preset ratio. This decreasing structure helps to gradually extract high-level features.

[0051] Among them, the ReLU activation function is used between each layer to enhance the nonlinear expression ability of the network. The design of the output layer can output environmental quality assessment results or pollutant concentration prediction values ​​according to specific task requirements.

[0052] In terms of network initialization, the weights use the improved He initialization method, that is, random sampling from the truncated normal distribution N(0, 2 / n), where n is the input dimension of the layer. This initialization method can avoid the gradient vanishing problem and accelerate network convergence. The bias term is initialized to a small positive number (usually 0.01) to break the symmetry of the network. In addition, Batch Normalization is introduced in each layer to reduce internal covariate shift and improve training stability.

[0053] S3.2: Based on the network architecture, preset L1 regularization coefficients and L2 regularization coefficients are introduced in the forward propagation process to control the model complexity by minimizing the regularization loss function.

[0054] In step S3.2, the system introduces a regularization mechanism during the forward propagation process of the network.

[0055] Specifically, two regularization strategies are implemented: L1 regularization and L2 regularization. The L1 regularization term is defined as λ1∑ wi , where λ1 is set to 0.01 to promote the sparsity of weights and achieve the effect of feature selection. The L2 regularization term is defined as λ2∑wi², where λ2 is set to 0.001 to prevent the weight value from being too large and improve the generalization ability of the model.

[0056] The loss function is designed in a combined form: L = MSE + λ1∑ wi + λ2∑wi², MSE is the mean square error loss, which is used to measure the difference between the predicted value and the true value. In actual applications, the system will assign different weight coefficients to different environmental parameters according to their importance to ensure the prediction accuracy of key parameters. For example, parameters such as PM2.5 and O3 that directly affect human health are given greater weights; auxiliary parameters such as temperature and humidity are given smaller weights.

[0057] S3.3: Implement an incremental learning mechanism, adopt a mini-batch method according to a preset batch size, and use a preset optimizer to update parameters to obtain the random neural network, and the model has the characteristic of adapting to dynamic changes in environmental parameters.

[0058] In step S3.3, the system implements an incremental learning mechanism.

[0059] First, we use mini-batch training with a batch size of 32. This method can ensure the stability of training and improve computational efficiency. The samples in each batch are selected by sliding the time window to ensure the timing and continuity of the data.

[0060] In terms of parameter updating, the system uses the Adam optimizer, and its hyperparameters are set as: β1=0.9 (exponential decay rate of first-order moment estimate), β2=0.999 (exponential decay rate of second-order moment estimate), ε=1e -8 (Numerical stability constant). The learning rate adopts a dynamic adjustment strategy, with the initial value set to 0.001, and a cosine annealing mechanism is introduced. The learning rate follows the cosine function as the training rounds change, and is updated every 1000 steps. This learning rate adjustment strategy can achieve rapid convergence in the early stages of training, and make detailed adjustments in the later stages to improve model performance.

[0061] It should be noted that in order to cope with the dynamic changes in environmental parameters, the system also implements an online update mechanism for the model. When a significant change in data distribution is detected (judged by calculating the KL divergence), the system will start the incremental learning process and use the latest data samples to fine-tune the model. To prevent catastrophic forgetting, the system will retain an experience replay buffer to store representative samples of historical data, and use both new and old data for training when the model is updated.

[0062] In addition, considering the possible computing resource limitations in practical applications, the system also implements model compression and acceleration mechanisms. Through weight pruning and quantization technology, the model parameters and computational complexity can be significantly reduced while maintaining model performance. Specifically, connections whose absolute weights are less than a preset threshold (such as 0.001) are pruned, and the remaining weights are quantized to 8-bit fixed-point representation.

[0063] In a specific embodiment of the present application, the training of the environmental quality assessment and early warning model is a multi-stage process that requires deep feature extraction and correlation analysis of environmental data.

[0064] Among them, feature extraction and correlation analysis are performed on the target environmental quality data to obtain feature data. The specific implementation process is as follows: S3.4: Based on the random neural network, the target environmental quality data is decomposed into subspaces according to parameter types to obtain characteristic subspaces corresponding to different types of target environmental quality data.

[0065] Based on the constructed random neural network, the system performs subspace decomposition of the target environmental quality data. Specifically, the 17-dimensional environmental parameters are decomposed into three subspaces according to the medium type: water quality subspace (6×T dimensional matrix, T is the time step), air subspace (8×T dimensional matrix) and soil subspace (3×T dimensional matrix). The rows of each subspace matrix represent different parameters, and the columns represent time series. This decomposition method not only takes into account the physical properties of the parameters, but also retains the information of the time dimension. For example, for the water quality subspace, the system will align parameters such as pH, dissolved oxygen, COD, etc. in the time dimension to form a complete monitoring sequence.

[0066] In the process of subspace division, the system also considers the correlation between parameters. By calculating the mutual information and Pearson correlation coefficient between parameters, a parameter association diagram is established. For parameters with strong correlation (correlation coefficient greater than 0.7), the system will pay special attention to their synergistic change patterns in subsequent analysis. For example, temperature and dissolved oxygen usually show a negative correlation, and this prior knowledge is used to guide the feature extraction of subspace.

[0067] S3.5: Based on the characteristic subspace, a plurality of groups of permutation data corresponding to each subspace are generated by a random permutation algorithm.

[0068] In step S3.5, the system generates multiple sets of permuted data through a random permutation algorithm to enhance the representation capability of the data.

[0069] Specifically, three replacement strategies are implemented: random sliding of time windows (±2 hours), random scaling of parameters (±10%), and Gaussian noise injection (σ=0.01). Ten sets of replacement data are generated for each subspace, thereby expanding the training sample set. This data augmentation method can improve the model's ability to adapt to spatiotemporal changes in environmental parameters.

[0070] For example, for PM2.5 data sequences, the system will generate a new sequence with a slight time offset based on the original sequence by randomly sliding the time window, which helps the model learn the time evolution law of pollutant concentration. At the same time, through random parameter scaling and noise injection, the influence of sensor measurement errors and environmental disturbances can be simulated to improve the robustness of the model.

[0071] S3.6: Based on the permuted data, the original environmental parameter relationship is reconstructed using a matrix recovery algorithm based on nuclear norm minimization.

[0072] The system uses a matrix recovery algorithm based on nuclear norm minimization to reconstruct the original environmental parameter relationship. Specifically, the optimization problem is constructed: min X_recovered - X + γ X_recovered F , in · represents the nuclear norm, · F represents the Frobenius norm, and γ is a balance parameter (typically 0.1). Minimizing the nuclear norm can effectively capture the low-rank structure of the data, while the Frobenius norm is used to control the reconstruction error.

[0073] The matrix recovery process is solved by the accelerated proximal gradient descent (APG) algorithm with a learning rate of 0.01. The algorithm alternately updates the original matrix estimate and its singular value decomposition result in each iteration until the convergence condition is reached (relative error is less than 10 -6 ) or the maximum number of iterations (1000 times). This method can effectively handle missing values ​​and noise problems in environmental data.

[0074] Specifically, they include: A1: Construct an optimization objective function and transform the matrix recovery problem into a combined optimization problem of nuclear norm minimization and Frobenius norm minimization.

[0075] In step A1, the system constructs an optimization objective function. This function consists of three parts: data fitting term, low-rank constraint term, and time series smoothing term. The data fitting term adopts the weighted bi-norm form, and the weight is dynamically adjusted according to the data reliability; the low-rank constraint is implemented by the nuclear norm to capture the main structure of the data; the time series smoothing term adopts the difference operator to ensure the time continuity of the reconstruction result.

[0076] The specific form of the optimization objective function is: F(X) = W⊙(X - D) ² F + λ X + μ X ² F , Where W is the weight matrix, D is the observed data, ⊙ represents the Hadamard product, Represents the difference operator of the time dimension, λ and μ are the weight coefficients of the low-rank constraint and smooth constraint respectively. The system determines the optimal λ and μ values ​​through cross-validation, and the typical ranges are [0.1, 1] and [0.01, 0.1] respectively.

[0077] It should be noted that the entire feature extraction and matrix recovery process is performed online. The system will continuously update the subspace decomposition and matrix reconstruction results based on the newly received monitoring data to ensure that the model can reflect changes in environmental conditions in a timely manner. At the same time, the system also maintains a feature library that stores typical feature patterns of historical data for anomaly detection and trend analysis.

[0078] A2: Setting an iteration termination condition, wherein the iteration termination condition includes that the relative error is less than a preset threshold or the maximum number of iterations is reached.

[0079] The system sets strict iteration termination conditions. Specifically, there are two aspects: one is the relative error threshold, and the other is the maximum number of iterations. The relative error is defined as the Euclidean distance between two consecutive iteration results divided by the norm of the current result, that is: ε= X k +1 - X k F / X k F , Where X k Indicates the result of the kth iteration. When ε is less than the preset threshold (usually set to 10^-5), the algorithm is considered to have reached convergence. At the same time, in order to prevent falling into a local optimum or oscillation state, the maximum number of iterations is set to 2000. In addition, the system also introduces an early stopping mechanism. When the relative error improvement of 10 consecutive iterations is less than 10 -6 When , the iteration process is terminated early. This mechanism can effectively balance the computational efficiency and optimization effect.

[0080] A3: In each iteration, the reconstruction weights of each subspace are dynamically updated based on the adaptive weight adjustment strategy.

[0081] The system implements an adjustment strategy based on adaptive weights.

[0082] Specifically, the reconstruction weight of each subspace is dynamically updated. The weight update formula is: w t+1 = w t exp(-α L t ), where w t represents the weight of the tth iteration, L t represents the reconstruction loss of the corresponding subspace, and α is the learning rate (set to 0.01).

[0083] This adaptive mechanism allows subspaces with better reconstruction results to receive higher weights, while subspaces with poorer reconstruction results have their weights reduced accordingly. For example, when a sensor for a water quality parameter temporarily fails, the system automatically reduces the weight of that parameter in the reconstruction process, thereby reducing the impact of abnormal data.

[0084] A4: Introduce time series smoothing constraints to ensure that the restored data maintains time continuity, and select the optimal regularization parameter through cross-validation.

[0085] The system introduces a time series smoothing constraint to ensure that the reconstructed data maintains time continuity. The specific implementation uses a second-order difference operator, and the constraint term is expressed as: Esmooth = D²X ² F , Where D² represents the second-order difference matrix. At the same time, the optimal regularization parameter is selected through five-fold cross validation. During the validation process, the data is divided into five parts in chronological order, four of which are used as training sets and one as validation set, and the parameter combination that minimizes the validation error is selected. The search range of the regularization parameter is set to [10^-3, 10^0], and the candidate value is generated by logarithmic uniform sampling.

[0086] S3.7: Iteratively optimize the reconstruction result using an accelerated proximal gradient descent method to improve the recovery accuracy of the subspace matrix to determine the feature data; like Figure 3 , the system uses the accelerated proximal gradient descent method to iteratively optimize the reconstruction results. In the specific implementation, an adaptive step size strategy is designed, with an initial step size of 0.1, which is dynamically adjusted according to the change of the objective function value. In order to improve the optimization efficiency, the Nesterov momentum term is introduced, and the momentum coefficient is set to 0.9. At the same time, the early stopping strategy is adopted. When the relative improvement of 5 consecutive iterations is less than 10^-4, the optimization process is terminated in advance.

[0087] S4: Based on the characteristic data, an environmental quality assessment and early warning model is trained to obtain an environmental quality monitoring model.

[0088] S4.1: Based on the characteristic data, a plurality of expert sub-models are constructed, each of which is responsible for the evaluation of corresponding environmental parameters.

[0089] The system builds multiple expert sub-models. Each expert sub-model is responsible for the evaluation of specific environmental parameters. It uses a deep neural network structure, including 4 hidden layers, with the number of nodes being [256, 128, 64, 32]. The activation function uses LeakyReLU to avoid the gradient vanishing problem.

[0090] In order to improve the generalization ability of the model, a Dropout layer is added after each hidden layer, and the dropout rate is set to 0.3. In addition, each expert model is equipped with an attention mechanism to adaptively focus on the importance of different time steps and parameters. For example, for PM2.5 prediction, the relevant expert model will focus on the change pattern of influencing factors such as wind speed and humidity.

[0091] S4.1 specifically includes: B1: Based on the characteristic dimensions of the environmental parameters, a multi-layer feedforward neural network structure is constructed for each expert model, in which the input layer dimension matches the environmental parameter dimension and the hidden layer decreases layer by layer according to the preset ratio.

[0092] In step B1, the system constructs a multi-layer feedforward neural network structure for each expert model based on the characteristic dimensions of the environmental parameters.

[0093] For the water quality expert model, the input layer dimension is 6, corresponding to 6 water quality parameters; for the air expert model, the input layer dimension is 8, corresponding to 8 air parameters; for the soil expert model, the input layer dimension is 3, corresponding to 3 soil parameters. The hidden layer adopts a layer-by-layer decreasing structure, with the number of nodes being [512, 256, 128, 64], and the number of nodes in each layer is about 1 / 2 of the previous layer. Each hidden layer is followed by a BatchNormalization layer and a Dropout layer, where the Dropout rate is set to 0.3. The activation function uses LeakyReLU with a slope parameter of 0.01 to avoid the Dead ReLU problem.

[0094] B2: Set the initial value and decay strategy of the temperature parameter to control the softness and hardness selected by the expert.

[0095] In step B2, the system sets up an adaptive temperature parameter control mechanism. The initial value of the temperature parameter τ is set to 5.0, and an exponential decay strategy is used for adjustment: τ_t = τ_0 exp(-λt), where λ is the decay rate, set to 0.001, and t is the training round.

[0096] When the temperature is high (e.g., τ > 1.0), expert selection tends to be exploratory, and the probability distribution of each expert being selected is relatively uniform; when the temperature is lowered (e.g., τ < 0.5), expert selection becomes more deterministic and tends to select the expert with the best performance. The system also sets the lower limit of temperature to 0.1 to maintain a certain degree of exploratory ability.

[0097] B3: A phased training strategy is adopted to train each expert model independently, train the gating network for selecting experts, and perform overall end-to-end fine-tuning.

[0098] In step B3, the system adopts a three-stage training strategy. In the first stage, each expert model is pre-trained independently, using the Adam optimizer, a learning rate of 0.001, a batch size of 64, and 50 training rounds. In the second stage, the gated network is trained, using the door-expert-mixture (DEM) mechanism to achieve soft selection of experts through a differentiable gating function. The gated network adopts a two-layer fully connected structure [128, 64], and the output layer uses a softmax activation function. In the third stage, end-to-end fine-tuning is performed, and the expert model and the gated network are optimized as a whole. The learning rate is reduced to 0.0001 and the training round is 20.

[0099] B4: Design loss functions based on prediction error, sparsity and diversity to guide the training process of each expert model and obtain multiple expert sub-models.

[0100] In step B4, the system designs a multi-objective loss function to guide the training of the expert model. The loss function consists of three parts: prediction error loss, sparsity loss, and diversity loss. The prediction error loss is in the form of weighted MSE: L_pred = Σ w_i(y_i - _i)², where the weights w_i are dynamically adjusted according to the importance of the parameters. The sparsity loss is achieved through L1 regularization: L_sparse = θ , in Set to 0.01. The diversity loss is implemented by maximizing the mutual information of the expert output: L_div = -Σ I(f_i,f_j), Where f_i and f_j are the outputs of different experts. The total loss function is: L_total = L_pred + L_sparse + L_div, where =1.0, =0.1, =0.05.

[0101] Through this multi-objective optimization, prediction accuracy, model simplicity and expert diversity can be guaranteed at the same time.

[0102] S4.2: According to the evaluation results of the expert sub-model, a gating network is constructed, and a preset discretization selection mechanism is used to dynamically select the most suitable expert combination.

[0103] The system constructs a dynamic selection mechanism based on a gating network.

[0104] The gating network adopts a lightweight three-layer fully connected structure. The input is the feature vector of the current environment state, and the output is the weight distribution of each expert model.

[0105] Specifically, the output of the gating network is normalized by the softmax function to obtain the selection probability of the expert model. In order to achieve discrete selection, the system uses the Gumbel-Softmax technique, with the temperature parameter τ initially set to 1.0 and gradually reduced to 0.1 according to the exponential decay law (decay rate is 0.995). This mechanism ensures that different expert combinations can be fully explored in the early stages of training, while tending to make more certain choices in the later stages of training.

[0106] The expert selection process also introduces a sparsity constraint, limiting the number of experts activated simultaneously to no more than 30% of the total number of experts. This constraint is achieved by adding an L1 regularization term to the loss function: Lsparse = λ g 1, Where g is the output vector of the gating network, and λ is the sparsity coefficient (set to 0.1). This not only improves the interpretability of the model, but also reduces computational overhead.

[0107] It should be noted that the system will regularly evaluate the performance of each expert model, including accuracy, recall, F1 score and other indicators. For expert models whose performance continues to decline, the system will trigger a retraining process and update the model parameters using the latest data. At the same time, in order to respond to sudden environmental events, the system also maintains an emergency expert model library, which contains specialized models for different types of pollution events, and can be quickly switched when an anomaly is detected.

[0108] S4.3: Based on the selected expert combination, the number of experts activated simultaneously is limited by using a sparse technology, and the outputs of multiple experts are adaptively fused to obtain the environmental quality monitoring model.

[0109] In step S4.3, the system constructs the final environmental quality monitoring model through sparse technology and adaptive fusion strategy. The specific implementation process is as follows: First, the system implements expert activation control based on the attention mechanism. The maximum number of experts activated at the same time N_max is set to 30% of the total number of experts, that is, for a total of 17 expert models, a maximum of 5 experts are activated at the same time. The activation probability is calculated as follows: p_i = softmax(a_i / τ), Where a_i is the attention score of the ith expert, and τ is the temperature parameter (set to 0.5). The system adopts the Top-K sparsification strategy, retaining only the K experts with the highest probability (K≤N_max), and the activation probability of other experts is set to 0. This mechanism can significantly reduce the computational overhead while maintaining the prediction performance.

[0110] Secondly, the system designs a dynamic weight allocation mechanism to adaptively integrate the outputs of multiple experts. For the activated experts, their weights are calculated using the following formula: w_i = α_i exp(-β e_i), Where α_i is the basic weight of the expert (output through the gating network), e_i is the expert's recent prediction error, and β is the smoothing factor (set to 0.1). The weights are normalized and used to weight the fusion expert output: y_pred = Σ(w_i y_i) / Σw_i, Where y_i is the prediction output of the ith expert. This weighting method can automatically adjust the contribution of different experts and improve the stability and accuracy of the prediction.

[0111] Thirdly, the system introduces an expert performance evaluation and dynamic adjustment mechanism. At preset time intervals (such as 24 hours), the system evaluates the performance indicators of each expert, including prediction accuracy, recall rate, and F1 score. The performance indicators are calculated through a sliding time window (the window size is 7 days) to reflect the latest performance of the expert. For experts whose performance continues to decline, the system will trigger the following adjustment strategies: If the expert's F1 score is below a threshold (such as 0.8) for three consecutive evaluation cycles, its L2 regularization coefficient is increased to prevent overfitting.

[0112] If the expert's prediction accuracy fluctuates significantly (the standard deviation exceeds 0.1), reduce its learning rate to improve training stability.

[0113] If the expert's performance cannot be improved by the above adjustments, it will be marked as a to-be-replaced state, and a new expert model will be selected from the candidate expert library for replacement.

[0114] Finally, the system implements an adaptive tuning mechanism for model fusion. By introducing a meta-learner, the hyperparameters of the fusion strategy are dynamically adjusted. The meta-learner adopts a two-layer optimization structure, where the outer optimizer is responsible for adjusting the fusion weights and the inner optimizer is responsible for updating the expert model parameters. Specifically, the meta-learner optimizes the following objective function based on the validation set performance through the gradient descent method: L_meta = MSE(y_true,y_pred) + H(w) + R(w), Where H(w) is the entropy of the weight distribution, which is used to control the smoothness of the fusion, and R(w) is the regularization term of the weight. and They are set to 0.01 and 0.001 respectively. This adaptive tuning mechanism can automatically adjust the fusion strategy according to environmental changes and improve the environmental adaptability of the model.

[0115] Through the comprehensive application of the above mechanisms, the system can efficiently integrate the advantages of multiple expert models and build an environmental quality monitoring model with high accuracy, strong adaptability and good interpretability. At the same time, the use of sparse technology and adaptive fusion strategy ensures the computational efficiency and resource utilization of the model in actual deployment.

[0116] S5: Deploy the environmental quality monitoring model to perform online monitoring and early warning of the environmental quality data.

[0117] Specifically include: S5.1: Compress and quantize the environmental quality monitoring model, and convert the model parameters into fixed-point representation.

[0118] In step S5.1, the system uses a multi-level compression and quantification strategy to optimize the environmental quality monitoring model.

[0119] First, the model is compressed using structured pruning technology, which includes the following steps: Calculate the weight importance score: The importance of each weight parameter is evaluated through the Fisher information matrix. The score calculation formula is: s_i = w_i L / w_i , where w_i is the weight parameter and L is the loss function.

[0120] Set pruning threshold: Determine the weight retention threshold based on the preset compression rate (such as 70%), and set the weight parameters with importance scores lower than the threshold to zero.

[0121] Implement channel pruning: For convolutional layers and fully connected layers, remove the corresponding input and output channels and reconstruct the network structure.

[0122] Secondly, the model parameters are quantized and the floating-point numbers are converted into 8-bit fixed-point representation.

[0123] Determine the quantization range: Determine the dynamic range [min, max] of weights and activation values ​​through statistical analysis; Calculate the quantization factor: scale = (max - min) / (2^8 - 1); Perform quantization mapping: q = round((x - min) / scale), where x is the original value and q is the quantized integer value; Implement retraining: Use quantization-aware training methods to simulate quantization effects during training and optimize model parameters.

[0124] S5.2: Using knowledge distillation technology, the knowledge of multiple expert models is transferred to a lightweight model, where the lightweight model is generated based on the environmental quality monitoring model.

[0125] S5.2 may also include: C1: Use the original expert model as the teacher model to build a lightweight model with fewer parameters.

[0126] In step C1, the system uses the original expert model as a teacher model and builds a lightweight student model. The student model adopts the following design principles: Network depth optimization: The 8-layer structure of the original model is simplified to 4 layers, with the layer configuration of [256, 128, 64, 32]; Module reconstruction: Use depth-wise separable convolution instead of standard convolution to reduce the number of parameters and computational complexity; Bottleneck layer design: A 1×1 convolutional layer is introduced in each module for channel compression with a compression ratio of 4:1; Residual connection: retain key cross-layer connections and maintain feature transfer efficiency.

[0127] C2: Design a distillation loss function, which includes soft target loss and hard target loss; In step C2, the system designs a composite distillation loss function.

[0128] L_distill = α L_soft + (1-α) L_hard + β L_struct, in: L_soft is the soft target loss, and the KL divergence is used to measure the difference in the output distribution of the teacher model and the student model: L_soft = KL(σ(z_t / T, σ(z_s / T)) Where z_t and z_s are the logits outputs of the teacher and student models respectively, and T is the temperature parameter (set to 4); L_hard is the hard target loss, and the mean square error is used to measure the gap with the true label: L_hard = MSE(y_true,y_pred); L_struct is the structural loss, which is used to maintain the similarity of intermediate features: L_struct = F_t W - F_s ², Among them, F_t and F_s are the intermediate features of the teacher and student models, and W is the adaptation matrix; α and β are balance coefficients, which are set to 0.5 and 0.1 respectively.

[0129] C3: Based on the loss function, the knowledge of the teacher model is gradually transferred to the lightweight model through iterative training.

[0130] In step C3, the system adopts a progressive knowledge transfer strategy: Initialization phase: Use the pre-trained weights of the teacher model to initialize the corresponding layer of the student model; Feature alignment: First, only optimize the matching of the middle layer features and freeze the parameters of other layers; Output distillation: unfreeze all layer parameters and optimize both soft and hard target losses; Joint fine-tuning: Add a structural loss term to perform end-to-end overall optimization.

[0131] The optimization process uses the cosine annealing learning rate strategy: lr = lr_base (1 + cos(π t / T)) / 2, Where lr_base is the base learning rate (set to 0.001), and T is the total number of training rounds.

[0132] C4: Evaluate the performance of the migrated lightweight model to ensure that it is within the preset accuracy range.

[0133] In step C4, the system established a complete performance evaluation system: Accuracy Assessment: Calculate the average relative error (ARE): ARE < 5% is required; Evaluate R² score: R² > 0.95 is required; Verify the coverage of the prediction confidence interval: the coverage of the 90% confidence interval is required to be > 85%.

[0134] Efficiency evaluation: Model size: The compressed size must be less than 20% of the original model; Inference latency: single prediction time is required to be less than 10ms; Memory usage: runtime memory required to be < 100MB.

[0135] Robustness evaluation: Noise stability: After adding ±5% random noise to the input, the performance does not drop by more than 10%; Long-term stability: After 7 days of continuous operation, the performance fluctuation range does not exceed 3%.

[0136] If the lightweight model fails to meet any of the above evaluation indicators, the system will automatically trigger the model optimization process, including adjusting the network structure, updating the distillation strategy, or retraining until the preset performance requirements are met.

[0137] S5.3: Design a model cache mechanism to store the frequently accessed parameters and intermediate results of the lightweight model in a fast cache.

[0138] In the cache mechanism implemented by the system, a three-level cache structure is used to optimize computing efficiency and resource utilization. The L1 cache capacity is 2MB, which is used to store the most commonly used model parameters; the L2 cache capacity is 8MB, which stores the less commonly used parameters; the L3 cache capacity is 32MB, which stores other parameters. The cache update adopts the LRU (Least Recently Used) elimination algorithm, where the L1 cache update cycle is 100ms, the L2 cache update cycle is 500ms, and the L3 cache update cycle is 2000ms. At the same time, the system implements a pre-fetch mechanism based on parameter access mode, which can load the parameters that may be needed into higher-level caches in advance.

[0139] For caching of intermediate results, the system first analyzes the computational graph to identify computationally intensive nodes. Then, a hash table is created for the intermediate results of repeated computations, with the fingerprint of the input feature as the key value and the corresponding computational result as the value. At the same time, the cache validity period is set. For time-sensitive intermediate results, a sliding time window is used for updating, and the window size is dynamically adjusted according to the parameter change rate.

[0140] In terms of cache management, the system implements a cache preheating mechanism to preload historically frequently accessed parameters at startup. By establishing a parameter access frequency statistical matrix, the system can dynamically adjust cache space allocation. In addition, a cache synchronization strategy is designed to ensure data consistency in a distributed deployment environment.

[0141] For the environmental early warning mechanism, the system first sets the benchmark threshold based on the statistical characteristics of historical data. In terms of water quality parameters, the normal range of pH value is 6.0-9.0, dissolved oxygen is not less than 5mg / L, COD does not exceed 40mg / L, and ammonia nitrogen does not exceed 2.0mg / L. In terms of air parameters, PM2.5 does not exceed 75μg / m³, PM10 does not exceed 150μg / m³, SO2 does not exceed 150μg / m³, and NO2 does not exceed 80μg / m³. In terms of soil parameters, the normal range of pH value is 5.5-8.5, and the heavy metal content refers to the national standard.

[0142] S5.4: Set warning thresholds for different environmental parameters based on historical data and expert experience, dynamically adjust warning thresholds based on the model's confidence in pollutant predictions, and issue warning signals in a timely manner by analyzing parameter change trends.

[0143] The system uses a dynamic threshold adjustment mechanism to adjust the warning threshold according to the prediction confidence. The adjustment formula is: T_adjusted = T_base (1 + k (1 - confidence)), where T_base is the base threshold, confidence is the prediction confidence, and k is the adjustment coefficient. The default value is 0.2.

[0144] Taking seasonal factors into consideration, the formula is: T_final = T_adjusted S_factor is used for compensation, where S_factor is a seasonal adjustment factor obtained based on historical data statistics. In addition, the system also sets differentiated threshold systems for different functional areas.

[0145] In terms of trend analysis, the system has built a multi-scale time series analysis model. Exponential smoothing is used for short-term trends (within 1 hour), ARIMA model is used for medium-term trends (within 24 hours), and wavelet decomposition is used for long-term trends (within 7 days). The warning level is divided into four levels: Level 1 warning (red) indicates that the parameter exceeds the adjustment threshold and the upward trend is significant; Level 2 warning (orange) indicates that the parameter is close to the threshold and there is an upward trend; Level 3 warning (yellow) indicates that the parameter fluctuates abnormally but does not exceed the threshold; Level 4 warning (blue) indicates that the parameter fluctuates abnormally.

[0146] The system has designed a composite warning rule to generate warning signals. When the parameter value exceeds the final threshold and the trend slope is greater than 0.1, the first-level warning is triggered; when the parameter value exceeds 90% of the final threshold and the trend slope is greater than 0.05, the second-level warning is triggered; when the parameter value exceeds 80% of the final threshold and the variance exceeds the threshold, the third-level warning is triggered. The warning information includes the warning time and location, the exceeded parameters and their values, the warning level and duration, possible pollution source analysis, and recommended measures.

[0147] To ensure the effectiveness of early warning, the system has established a complete evaluation index system, including early warning accuracy (number of correct warnings / total number of warnings), missed reporting rate (number of missed events / total number of events that should be warned), and early warning lead time (the difference between the warning time and the time when the event occurred). The system regularly evaluates the early warning effect every month, dynamically adjusts the early warning parameters based on the evaluation results, and establishes an early warning case library for model training and experience accumulation to ensure the continuous optimization and improvement of the early warning mechanism.

[0148] The present application also provides a multi-parameter environmental quality intelligent monitoring system. Figure 4 ,include: A sensor array for collecting environmental quality data, wherein the environmental quality data includes water quality parameters, air parameters and soil parameters; A preprocessing module, connected to the sensor array, for preprocessing the environmental quality data to obtain target environmental quality data, wherein the preprocessing includes data cleaning, standardization and time alignment; A neural network module, connected to the preprocessing module, is used to construct a random neural network with forward regularization, and perform feature extraction and association analysis on the target environmental quality data to obtain feature data; A training module, connected to the neural network module, for training an environmental quality assessment and early warning model based on the characteristic data to obtain an environmental quality monitoring model; A deployment module is connected to the training module and is used to deploy the environmental quality monitoring model to perform online monitoring and early warning of the environmental quality data.

[0149] The present invention also provides a computer system suitable for implementing a multi-parameter environmental quality intelligent monitoring method, wherein the computer system includes a central processing unit (CPU), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) or a program loaded from a storage part to a random access memory (RAM), such as executing the method described in the above embodiment. Various programs and data required for system operation are also stored in the RAM. The CPU, ROM and RAM are connected to each other through a bus. The input / output (I / O) interface is also connected to the bus.

[0150] The following components are connected to the I / O interface: an input part including a keyboard, a mouse, etc.; an output part including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker; a storage part including a hard disk, etc.; and a communication part including a network interface card such as a LAN (Local Area Network) card, a modem, etc. The communication part performs communication processing via a network such as the Internet. The drive 310 is also connected to the I / O interface as needed. Removable media, such as magnetic disks, optical disks, magneto-optical disks, semiconductor memories, etc., are installed on the drive as needed so that the computer program read therefrom can be installed into the storage part as needed.

[0151] In particular, according to an embodiment of the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present application includes a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a computer program for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication part, and / or installed from a removable medium. When the computer program is executed by a central processing unit (CPU), various functions defined in the system of the present application are executed.

[0152] It should be noted that the computer-readable medium shown in the embodiment of the present application may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (Erasable Programmable Read Only Memory, ERPOM), a flash memory, an optical fiber, a portable compact disk read-only memory (Compact Disc Read-Only Memory, CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0153] In the present application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, apparatus, or device. In the present application, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries a computer-readable computer program. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium, which may send, propagate, or transmit a program for use by or in combination with an instruction execution system, apparatus, or device. The computer program contained on the computer-readable medium may be transmitted using any suitable medium, including but not limited to: wireless, wired, etc., or any suitable combination of the above.

[0154] The flowchart and block diagram in the accompanying drawings illustrate the possible architecture, functions and operations of the system, method and computer program product according to various embodiments of the present application. Wherein, each box in the flowchart or block diagram can represent a module, a program segment, or a part of the code, and the above-mentioned module, program segment, or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flowchart, and the combination of boxes in the block diagram or flowchart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0155] The units involved in the embodiments of the present invention may be implemented by software or hardware, and the units described may also be arranged in a processor. The names of these units do not, in some cases, limit the units themselves.

[0156] According to one aspect of the present application, a computer program product or a computer program is provided, the computer program product or the computer program comprising computer instructions, the computer instructions being stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the methods provided in the above-mentioned various optional implementations.

[0157] As another aspect, the present application also provides a computer-readable medium, which may be included in the electronic device described in the above embodiment; or may exist independently without being assembled into the electronic device. The above computer-readable medium carries one or more programs, and when the above one or more programs are executed by an electronic device, the electronic device implements the method described in the above embodiment.

[0158] It should be noted that, although several modules or units of the equipment for action execution are mentioned in the above detailed description, this division is not mandatory. In fact, according to the embodiments of the present application, the features and functions of two or more modules or units described above can be embodied in one module or unit. On the contrary, the features and functions of one module or unit described above can be further divided into being embodied by multiple modules or units.

[0159] Through the description of the above implementation methods, it is easy for those skilled in the art to understand that the example implementation methods described here can be implemented by software, or by combining software with necessary hardware. Therefore, the technical solution according to the implementation method of the present application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, including several instructions to enable a computing device (which can be a personal computer, a server, a touch terminal, or a network device, etc.) to execute the method according to the implementation method of the present application.

[0160] Although the present specification is disclosed as above, the present invention is not limited thereto. Any person skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, so the protection scope of the present invention shall be subject to the scope defined by the claims.

Claims

1. A control method for a multi-parameter environmental quality intelligent monitoring system, characterized in that: include: Collecting environmental quality data, wherein the environmental quality data includes water quality parameters, air parameters and soil parameters; Preprocessing the environmental quality data to obtain target environmental quality data, wherein the preprocessing includes data cleaning, standardization and time alignment; Constructing a random neural network with forward regularization, and performing feature extraction and association analysis on the target environmental quality data to obtain feature data; Based on the characteristic data, an environmental quality assessment and early warning model is trained to obtain an environmental quality monitoring model; The environmental quality monitoring model is deployed to perform online monitoring and early warning of the environmental quality data.

2. The control method according to claim 1, characterized in that: The preprocessing of the environmental quality data to obtain target environmental quality data includes: Identify outliers based on statistical and machine learning methods, and use preset criteria and isolation forest algorithm to perform anomaly detection; mark and process detected outliers, select deletion or correction schemes based on data characteristics, and use wavelet transform to perform noise reduction on data to obtain the target environmental quality data; or, Corresponding standardization methods are used for different types of environmental parameters to standardize the environmental quality data. The standardization methods include maximum and minimum value standardization or Z-score of the environmental quality data; time alignment of different parameter data is performed based on timestamp information, and a linear interpolation method is used to deal with the problem of inconsistent sampling frequencies. A unified data format is constructed to encapsulate the data to obtain the target environmental quality data.

3. The control method according to claim 1, characterized in that: The construction of a random neural network with forward regularization includes: According to the type and number of environmental parameters, a random neural network infrastructure including an input layer, a hidden layer and an output layer is designed, wherein the input layer includes nodes corresponding to the number of environmental parameters, and the hidden layer is arranged in descending order according to a preset number; Based on the network architecture, a preset L1 regularization coefficient and an L2 regularization coefficient are introduced in the forward propagation process to control the model complexity by minimizing the regularization loss function; An incremental learning mechanism is implemented, a mini-batch method is adopted according to a preset batch size, and a preset optimizer is used to update parameters to obtain the random neural network, and the model has the characteristic of adapting to dynamic changes in environmental parameters.

4. The control method according to claim 1, characterized in that: The feature extraction and correlation analysis of the target environmental quality data to obtain feature data includes: Based on the random neural network, the target environmental quality data is subspace-decomposed according to parameter types to obtain characteristic subspaces corresponding to different types of target environmental quality data; Based on the characteristic subspace, a plurality of groups of permutation data corresponding to each subspace are generated by a random permutation algorithm; Reconstructing the original environmental parameter relationship based on the permuted data and a matrix recovery algorithm based on nuclear norm minimization; The reconstruction result is iteratively optimized by using an accelerated proximal gradient descent method to improve the recovery accuracy of the subspace matrix so as to determine the characteristic data.

5. The control method according to claim 4, characterized in that: The method of reconstructing the original environmental parameter relationship based on the permuted data and a matrix recovery algorithm based on nuclear norm minimization includes: Construct an optimization objective function to transform the matrix recovery problem into a combined optimization problem of nuclear norm minimization and Frobenius norm; Setting an iteration termination condition, wherein the iteration termination condition includes that the relative error is less than a preset threshold or the maximum number of iterations is reached; In each iteration, the reconstruction weights of each subspace are dynamically updated based on the adaptive weight adjustment strategy; A time series smoothing constraint is introduced to ensure that the restored data maintains time continuity, and the optimal regularization parameter is selected through cross-validation.

6. The control method according to claim 1, characterized in that: The environmental quality assessment and early warning model is trained based on the characteristic data to obtain an environmental quality monitoring model, including: Based on the characteristic data, a plurality of expert sub-models are constructed, each expert sub-model being responsible for the evaluation of a corresponding environmental parameter; According to the evaluation results of the expert sub-model, a gating network is constructed, and a preset discretization selection mechanism is used to dynamically select the most suitable expert combination; Based on the selected expert combination, the number of experts activated simultaneously is limited by using a sparse technology, and the outputs of multiple experts are adaptively fused to obtain the environmental quality monitoring model.

7. The control method according to claim 6, characterized in that: The step of constructing a plurality of expert sub-models based on the feature data includes: According to the characteristic dimensions of environmental parameters, a multi-layer feedforward neural network structure is constructed for each expert model, in which the input layer dimension matches the environmental parameter dimension, and the hidden layer decreases layer by layer according to the preset ratio; Set the initial value and decay strategy of the temperature parameter to control the softness and hardness of the expert selection; A phased training strategy is adopted to train each expert model independently, train the gating network used to select experts, and perform overall end-to-end fine-tuning; The loss function is designed based on prediction error, sparsity and diversity to guide the training process of each expert model and obtain multiple expert sub-models.

8. The control method according to claim 1, characterized in that: The deploying the environmental quality monitoring model to perform online monitoring and early warning of the environmental quality data includes: Compressing and quantizing the environmental quality monitoring model, and converting model parameters into fixed-point representation; Using knowledge distillation technology, knowledge of multiple expert models is transferred to a lightweight model, wherein the lightweight model is generated based on the environmental quality monitoring model; Design a model cache mechanism to store the frequently accessed parameters and intermediate results of the lightweight model in a fast cache; Based on historical data and expert experience, warning thresholds are set for different environmental parameters, and the warning thresholds are dynamically adjusted according to the model's predicted confidence in pollutants. Warning signals are issued in a timely manner by analyzing parameter change trends.

9. The control method according to claim 8, characterized in that: The knowledge distillation technology is used to migrate the knowledge of multiple expert models to the lightweight model, including: The original expert model is used as the teacher model to build a lightweight model with fewer parameters; Designing a distillation loss function, wherein the loss function includes a soft target loss and a hard target loss; Based on the loss function, gradually transferring the knowledge of the teacher model to the lightweight model through iterative training; The performance of the migrated lightweight model is evaluated to ensure that it is within a preset accuracy range.

10. A multi-parameter environmental quality intelligent monitoring system, characterized in that: include: A sensor array for collecting environmental quality data, wherein the environmental quality data includes water quality parameters, air parameters and soil parameters; A preprocessing module, connected to the sensor array, for preprocessing the environmental quality data to obtain target environmental quality data, wherein the preprocessing includes data cleaning, standardization and time alignment; A neural network module, connected to the preprocessing module, is used to construct a random neural network with forward regularization, and perform feature extraction and association analysis on the target environmental quality data to obtain feature data; A training module, connected to the neural network module, for training an environmental quality assessment and early warning model based on the characteristic data to obtain an environmental quality monitoring model; A deployment module is connected to the training module and is used to deploy the environmental quality monitoring model to perform online monitoring and early warning of the environmental quality data.

Images