Intelligent water affair management system based on deep learning

Abstract

The invention discloses an intelligent water affair management system based on deep learning, and the system comprises a data collection and preprocessing module which is used for collecting the data of a water supply network and constructing a standardized water affair input tensor; the topology modeling module is used for generating a water supply pipe network diagram structure and diagram structure representation thereof; the state initialization module is used for fusing the input tensor and the graph structure to generate an initial state vector; the state evolution calculation module is used for establishing an augmented state evolution equation and carrying out adaptive integration to generate a prediction state sequence and source sink estimation; the multi-objective decision module is used for constructing a comprehensive cost function and generating a pump station and valve control instruction; the physical constraint projection module is used for executing water pressure, flow velocity and water age constraint projection and outputting an executable scheduling instruction; and the dynamic simulation feedback module is used for performing digital twin simulation and updating the model and sparse estimation weight. According to the invention, fine control and dynamic feedback optimization of the water affair system are realized, and the operation efficiency and the scheduling intelligence level are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of smart water affairs, and particularly relates to a smart water affairs management system based on deep learning. BACKGROUND

[0002] In the current smart water affairs management practice, the dispatching of the water supply system mostly relies on static rules or single target optimization strategies, and lacks the dynamic modeling capability for the complex hydrology state evolution process. The traditional dispatching method is usually based on the discrete state input of fixed time steps, ignores the coupled change law of the flow, water pressure, residual chlorine and other variables in the water supply network on the continuous time scale, and is difficult to accurately reflect the real-time evolution process of the system state under dynamic regulation. In addition, most of the existing models only stay in the basic topological relationship mapping when constructing the graph structure of the water supply network, and fail to combine the physical characteristics of the nodes and pipe sections in the graph for structure embedding expression, resulting in insufficient reasoning accuracy.

[0003] The existing neural network model mainly adopts a static feedforward structure or a sequence modeling structure when processing the state prediction problem of the water supply system, cannot model the differential change trend of the water supply state, has a lagging response when facing complex disturbances or emergencies, and is limited in stability and generalization ability. On the control instruction generation level, experience rules or optimization methods based on a single index are mostly used, it is difficult to balance the multi-objective scheduling performance while ensuring the physical constraints of system operation, and it is easy to produce problems such as overpressure, overcurrent or water quality deterioration, and lacks comprehensive consideration of the dispatching feasibility of the pump station and the valve and other execution equipment. In addition, the feedback link after dispatching execution mostly relies on experience correction, lacks digital simulation support, and cannot realize parameter dynamic calibration and model closed-loop update.

[0004] The above problems limit the stable operation and efficient dispatching of the smart water affairs system in a complex dynamic environment, and it is urgent to introduce an intelligent method with continuous modeling capability, topological structure perception capability, multi-objective regulation capability and simulation feedback capability for improvement.

[0005] Therefore, how to provide a smart water affairs management system based on deep learning is a problem that those skilled in the art need to solve. SUMMARY

[0006] An object of the present application is to provide a smart water affairs management system based on deep learning. The present application constructs a water supply network topological structure, integrates multi-source water affairs data, establishes a topological perception and physically consistent continuous time state evolution model, and has the advantages of high modeling accuracy, accurate dispatching decision and fast dynamic response.

[0007] According to the smart water affairs management system based on deep learning provided by the embodiment of the present application, the system comprises:

[0008] A data acquisition and preprocessing module is configured to acquire data of the water supply network, preprocess and construct a standardized water input tensor;

[0009] A topology modeling module is configured to construct a graph structure of the water supply network, extract node feature vectors and edge feature vectors, and generate a graph structure representation;

[0010] A state initialization module is configured to receive the standardized water input tensor and the graph structure representation, and generate a fusion state initialization vector;

[0011] A state evolution calculation module is configured to establish an augmented state evolution equation based on the fusion state initialization vector and perform adaptive integration;

[0012] A multi-objective decision-making module is configured to receive the predicted state sequence, construct a multi-objective cost function, and generate pump station power instructions and valve adjustment instructions;

[0013] A physical constraint projection module is configured to project the scheduling instructions under water pressure, flow rate, and water age constraints, and output an executable scheduling instruction set;

[0014] A dynamic simulation feedback module is configured to perform digital twin simulation and feedback to update the parameters of the neural ordinary differential equation network and the source-sink sparse estimation function.

[0015] Optionally, the modules are implemented through the following methods:

[0016] Data of the water supply network is acquired, preprocessed, and standardized to construct a water input tensor; a graph topology structure of the water supply network is constructed, node feature vectors and edge feature vectors are extracted, and a graph structure representation is generated;

[0017] The standardized water input tensor and the graph structure representation are input into a state initialization network to generate a fusion state initialization vector; the fusion state initialization vector is input as an initial state of a neural ordinary differential equation; a topology-aware drift function, a physical consistency gating function, and a source-sink sparse estimation function are constructed to establish an augmented state evolution equation; adaptive integration is performed on the continuous-time state to output a predicted state sequence and a source-sink estimation vector;

[0018] The predicted state sequence is input into a multi-target decision network, a cost function including water supply stability, energy consumption, production and sales difference and source and sink risk is constructed, and a pump station power instruction set and a valve adjustment instruction set are output; the pump station power instruction set and the valve adjustment instruction set are input into a control barrier function projector, projection calculation is performed according to the upper and lower limits of water pressure, the flow velocity interval and the water age constraint, and an executable scheduling instruction set is generated; the executable scheduling instruction set is input into a digital twin simulator for dynamic response simulation, the difference between the predicted state sequence and the simulation state sequence is calculated, and the parameters of the neural ordinary differential equation network and the gating weight of the source and sink sparse estimation function are updated according to the difference between the predicted state sequence and the simulation state sequence.

[0019] Optionally, the data of the water supply network includes flow, water pressure, residual chlorine, temperature and conductivity data; the preprocessing includes time alignment, abnormality elimination and normalization processing of the data of the water supply network.

[0020] Optionally, the step of inputting the standardized water input tensor and the graph structure representation into the state initialization network to generate a fusion state initialization vector includes:

[0021] The standardized water input tensor and the node feature vector are spliced in the node dimension to obtain a fusion node input feature set; the edge feature vector in the graph structure representation and the topological connection relationship are combined in the edge dimension to generate a fusion edge input feature set; the fusion node input feature set is input into a graph encoding network composed of a plurality of nonlinear transformation layers and residual connections to generate an initial node representation vector set; the fusion edge input feature set is input into an edge perception encoding network to perform edge feature embedding calculation to obtain an edge representation vector set; the initial node representation vector set and the edge representation vector set are jointly input into an aggregation layer constructed based on a graph attention mechanism to update the node state representation and output the fusion state initialization vector.

[0022] Optionally, the step of constructing the topological awareness drift function, the physical consistency gating function and the source and sink sparse estimation function includes:

[0023] A graph Laplacian matrix of node connection is constructed according to the graph structure representation, and an edge weight matrix is generated in combination with the pipe diameter, pipe length and roughness information contained in the edge feature vector; the fusion state initialization vector and the graph Laplacian matrix are input into a topological awareness diffusion function, weighted diffusion is performed according to the connection relationship between nodes and the edge weight, and an initial state derivative vector is output;

[0024] The initial state derivative vector is taken as a basic drift term in a neural ordinary differential equation; a flow, water pressure, and node spatial position feature in the fusion state initialization vector are used to construct a node continuity residual vector and an energy conservation residual vector; the continuity residual vector and the energy conservation residual vector are spliced into a residual feature set and input into a gating function to generate residual gating weights; the initial state derivative vector and a theoretical state derivative calculated according to physical conservation are linearly weighted and fused to generate a physically consistent state derivative vector, and the weights are determined by the residual gating weights;

[0025] The flow mutation value, water pressure fluctuation rate, and adjacent pressure difference information of each node in the fusion state initialization vector in a time window are extracted and spliced into a source-sink feature input set, and input into a sparse activation function to output a source-sink estimation vector.

[0026] The physically consistent state derivative vector and the source-sink estimation vector are taken as joint inputs of the neural ordinary differential equation, and a self-adaptive step numerical integration method is used to perform a continuous time evolution process to output a predicted state sequence and a source-sink estimation vector.

[0027] Optionally, the step of constructing the augmented state evolution equation comprises:

[0028] The state derivative vector generated according to the topology-aware drift function and the physically consistent state derivative vector generated according to the physically consistent gating function are spliced to generate an augmented state derivative vector; the augmented state derivative vector is input into the neural ordinary differential equation, combined with the source-sink estimation vector generated by the source-sink sparse estimation function to augment the state derivative; the augmented state derivative is integrated by an adaptive step numerical integrator to obtain an augmented state vector; at each time step, a gradient descent method is used to minimize the error between the augmented state derivative and the hydraulic constraints and the energy conservation constraints to ensure the physical consistency of the augmented state vector;

[0029] The augmented state vector and the predicted state sequence are input into a control decision process to generate a control instruction set. Optionally, the construction process of the multi-objective decision network comprises the following steps:

[0030] Continuous water pressure, flow rate, and water age sequences of each water supply node are extracted from the predicted state sequence to construct a state time series tensor. The state time series tensor is normalized and input into a feature cross-coding layer to extract a set of cross-coupling feature vectors. This set of cross-coupling feature vectors is then input into a four-branch objective modeling structure, which respectively constructs branches for water supply stability estimation, energy consumption estimation, production-sales difference estimation, and source-sink risk scoring. In each branch, a shared-weight feedforward subnetwork is used to calculate the single-objective cost output under the current state. All single-objective cost outputs are aggregated into a multi-objective joint cost vector at the output layer. This multi-objective joint cost vector is then input into a strategy generation unit, which, combined with the current state time series tensor, performs action decoding operations to output a set of pump station power commands and a set of valve regulation commands.

[0031] Optionally, the construction steps of the control barrier function projector include:

[0032] Initial control commands are extracted from the pump station power command set and valve regulation command set. Combined with water pressure, flow velocity, and water age information at each node in the water supply network, a set of constraints for the control commands is constructed. Based on the set upper and lower limits of water pressure, flow velocity ranges in pipe sections, and water age constraints, physical constraint boundaries for each control command are constructed, forming a constraint set containing multi-dimensional restrictions. The initial control command set and constraint set are input into a control obstacle function projector, which adjusts the initial control commands using the gradient descent method. The projected values ​​of the adjusted control commands are calculated using the control obstacle function projector, generating a set of executable scheduling commands that satisfy the physical constraints.

[0033] Optionally, the steps of the digital twin simulator performing dynamic response simulation include:

[0034] The set of executable scheduling instructions is input into the physical structure simulation diagram, the simulation start time and time step are set, and continuous time hydraulic state evolution calculation is performed based on the current state input to generate the simulation state sequence.

[0035] The nodal pressure, pipe flow velocity, and water age at each time step in the simulation state sequence are extracted to construct a set of state vectors consistent with the predicted state sequence.

[0036] Align the predicted state sequence with the simulated state sequence by time step, calculate the state difference between each corresponding time step, and generate a state difference sequence.

[0037] The network parameters of the neural ordinary differential equation in the augmented state evolution equation are optimized and updated by backpropagation of errors based on the state difference sequence.

[0038] Simultaneously, the gate weights of the source-sink sparse estimation function are adjusted and updated using difference-guided methods to generate a new set of parameter weights.

[0039] Optionally, the step of performing difference-guided adjustment and update on the gating weights of the source-sink sparse estimation function includes:

[0040] Normalize the state difference sequence to generate a set of amplitude weight coefficients, and map and match the amplitude weight corresponding to each node with the corresponding gate weight position in the source-sink sparse estimation function to construct an error localization index set.

[0041] Traverse the error location index set, extract the current weight value for the gating weight at each corresponding position, and perform a direction consistency judgment in combination with the amplitude weight coefficient. If the error direction is consistent with the gating output direction, maintain the current weight; otherwise, perform sign flipping.

[0042] The gating weight value at each position is adjusted proportionally based on the magnitude weight coefficient to obtain the updated gating weight matrix;

[0043] The updated gate weight matrix is ​​written into the corresponding position in the source-sink sparse estimation function, replacing the original weight structure, thus completing the gate weight adjustment operation of the source-sink sparse estimation function.

[0044] The beneficial effects of this invention are:

[0045] (1) Improve the accuracy of continuous time modeling: By introducing the structure of the neural ordinary differential equation and combining the topological sensing drift function and the physical consistency gating function, the precise modeling and time-series response of the dynamic state of the water supply network can be realized, overcoming the problem that traditional discrete models are difficult to accurately capture continuous state changes.

[0046] (2) Enhance the rationality of scheduling decisions: By constructing a multi-objective optimization decision network, and taking into account factors such as water supply stability, energy consumption, production and sales difference and source and sink risks, the precise control of pump stations and valves can be achieved, effectively reducing the energy consumption of pipeline operation and improving the production and sales balance efficiency.

[0047] (3) Enhance simulation closed-loop control capability: Combine the dynamic feedback output by the digital twin simulator to automatically correct the model parameters and gating weights in the state evolution equation, forming an adaptive control mechanism that integrates data-driven and physical mechanisms, significantly improving the robustness and generalization capability of the system. Attached Figure Description

[0048] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0049] Figure 1 This is a module structure diagram of a deep learning-based smart water management system proposed in this invention.

[0050] Figure 2 This is a structural diagram of a multi-objective decision network for a deep learning-based smart water management system proposed in this invention.

[0051] Figure 3 This is a schematic diagram of a digital twin simulator and feedback update process for a deep learning-based smart water management system proposed in this invention. Detailed Implementation

[0052] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0053] refer to Figures 1-3 A deep learning-based smart water management system includes:

[0054] The data acquisition and preprocessing module is used to collect data from the water supply network, preprocess it, and construct a standardized water input tensor.

[0055] The topology modeling module is used to construct the water supply network diagram structure, extract node feature vectors and edge feature vectors, and generate a graph structure representation.

[0056] The state initialization module is used to receive standardized water input tensors and graph structure representations, and generate fused state initialization vectors.

[0057] The state evolution calculation module is used to establish augmented state evolution equations based on the fused state initialization vector and perform adaptive integration;

[0058] The multi-objective decision-making module is used to receive the predicted state sequence, construct the multi-objective cost function, and generate pump station power commands and valve adjustment commands.

[0059] The physical constraint projection module is used to project the scheduling instructions with constraints of water pressure, flow velocity, and water age, and output a set of executable scheduling instructions.

[0060] The dynamic simulation feedback module is used to perform digital twin simulations and update the parameters of the neural network of ordinary differential equations and the source-sink sparse estimation functions.

[0061] In this embodiment, the modules are interconnected using the following method:

[0062] Data from the water supply network is collected and preprocessed to construct a standardized water input tensor. A graph topology of the water supply network is then constructed, and node and edge feature vectors are extracted to generate a graph representation. In this embodiment, the graph topology construction of the water supply network is based on an abstract modeling of the connection relationships between sensor nodes and pipelines in the actual water supply area. Specifically, the following operations are included:

[0063] The locations of water pressure monitoring points, flow meters, water quality sampling points, control valves, and pumping stations with independent data acquisition capabilities within the water supply network are used as nodes in the graph structure. Each node represents a key functional unit in the network, and its unique identifier can be determined by the equipment number, spatial coordinates, or network address. The physical connections between nodes are obtained from the actual network structure diagram. Edges between nodes are constructed based on pipeline connectivity, thereby generating an initial graph structure with directional and weighted characteristics.

[0064] For the aforementioned nodes, their corresponding temporal and structural attributes are extracted to form node feature vectors. The attributes included in the node feature vectors may include, but are not limited to: the average flow rate, average water pressure, standard deviation, residual chlorine concentration, temperature, conductivity, and estimated water age within the node's historical time window, as well as environmental parameters such as topographic information, pipe burial depth, and ground temperature in the area where the node is located. The node feature vectors are used to represent the state information of the node in terms of time series and spatial attributes.

[0065] For the pipes connecting the nodes, edge feature vectors are extracted. These feature vectors reflect the pipe connection characteristics and physical properties, including pipe material, diameter, length, inner wall roughness, number of maintenance visits, presence of flow-limiting valves or check valves, and construction date. These edge features are used to guide the information propagation intensity and direction offset within the guided graph structure in subsequent modeling.

[0066] The aforementioned set of nodes and edges are organized into a unified graph structure representation. This graph structure representation employs a joint encoding method using adjacency matrices and feature matrices. The adjacency matrix represents the topological connections, while the node and edge feature matrices represent the attribute characteristics of each component in the graph structure. This graph structure representation will serve as one of the fundamental inputs in the subsequent neural network modeling process, ensuring that the network accurately represents the spatial topological characteristics and physical attributes of the water supply network. The standardized water input tensor and the graph structure representation input state are used to initialize the network, generating a fused state initialization vector. This fused state initialization vector is then used as the initial state input for the neural network's frequent differential equations. By constructing a topology-aware drift function, a physical consistency gating function, and a source-sink sparse estimation function, an augmented state evolution equation is established. Adaptive integration is performed on the continuous-time states, outputting a predicted state sequence and source-sink estimation vectors.

[0067] In this embodiment, in order to model and solve the augmented state evolution process described by the constant differential equation, an adaptive integral method is used to numerically calculate the continuous-time state.

[0068] Using the fused state initialization vector as the initial state input, an augmented state derivative expression is constructed, which includes a topology-aware drift function, a physical consistency gating function, and a source-sink sparse estimation function. This expression serves as the right-hand side of the neural ordinary differential equation and is used to describe the continuous change of the system state over time.

[0069] Based on the current state value and the augmented state derivative, an adaptive step-size integrator is used to solve the neural ordinary differential equation. During the integration process, the integrator first sets the initial step size and the error tolerance threshold, and then performs state prediction through trial steps. Finally, it compares the estimated error of this step with the set threshold to determine whether the current step size needs to be adjusted.

[0070] When the prediction error exceeds the tolerance threshold, the integrator automatically reduces the step size and repeats the calculation until the error is within an acceptable range; if the error is significantly below the threshold, the system increases the step size to improve overall computational efficiency. This adaptive adjustment process is executed independently within each time step, ensuring that the integration process has sufficient accuracy while avoiding redundant calculations.

[0071] By integrating the continuous-time state variables using the aforementioned adaptive integration strategy, a complete state evolution trajectory is obtained. This trajectory includes not only estimates of the changes in key variables such as flow rate, water pressure, and water age at each node in the system over time, but also the dynamic response of the source-sink sparse estimation, which is used for scheduling command generation and digital twin simulation verification. The predicted state sequence is input into a multi-objective decision network to construct a cost function that includes water supply stability, energy consumption, production-sales difference, and source-sink risk, outputting a set of pump station power commands and a set of valve regulation commands. The set of pump station power commands and valve regulation commands is then input into a control obstacle function projector, and projection calculations are performed based on water pressure upper and lower limits, flow velocity range, and water age constraints to generate an executable scheduling command set. The executable scheduling command set is then input into a digital twin simulator for dynamic response simulation, calculating the difference between the predicted state sequence and the simulated state sequence. The parameters of the neural network and the gating weights of the source-sink sparse estimation function are updated based on the difference between the predicted and simulated state sequences.

[0072] In this embodiment, the data of the water supply network includes flow rate, water pressure, residual chlorine, temperature and conductivity data; the preprocessing includes time alignment, anomaly removal and normalization of the data of the water supply network.

[0073] In this embodiment, the step of initializing the network with the standardized water input tensor and the graph structure representation of the input state to generate the fused state initialization vector includes:

[0074] The standardized water input tensor and node feature vectors are concatenated along the node dimension to obtain a fused node input feature set. The edge feature vectors and topological connections in the graph structure representation are combined along the edge dimension to generate a fused edge input feature set. The fused node input feature set is input into a graph coding network composed of several nonlinear transformation layers and residual connections to generate an initial node representation vector set. The fused edge input feature set is input into an edge-aware coding network to perform edge feature embedding calculation to obtain an edge representation vector set. The initial node representation vector set and the edge representation vector set are jointly input into an aggregation layer constructed based on a graph attention mechanism to complete the node state representation update and output a fused state initialization vector.

[0075] In this embodiment, the steps of constructing the topology-aware drift function, the physical consistency gating function, and the source-sink sparsity estimation function include:

[0076] Based on the graph structure representation, construct the graph Laplacian matrix of node connections, and generate the edge weight matrix by combining the pipe diameter, pipe length and roughness information contained in the edge feature vector; input the fusion state initialization vector and the graph Laplacian matrix into the topology-aware diffusion function, and perform weighted diffusion according to the connection relationship between nodes and edge weights, and output the initial state derivative vector;

[0077] In this embodiment, to effectively inject water supply network structure information into the evolution process of the neural network's ordinary differential equations, a topology-aware diffusion function is constructed to generate the initial state derivative vector. The topology-aware diffusion function is executed based on the graph structure representation of the water supply network, specifically including the following steps:

[0078] Based on the aforementioned graph structure representation, the set of nodes and the set of edges are extracted, and the connection relationships of all nodes are represented as a graph Laplacian matrix. This matrix reflects the coupling relationship and connection strength of each node in the topology. The graph Laplacian matrix is ​​constructed based on the node connection state and edge weights. The edge weights are calculated jointly from the pipe diameter, pipe length, and roughness in the edge feature vector, and are used to reflect the impedance level in the hydraulic transmission path.

[0079] The initial state of each node in the fusion state initialization vector is used as input. Based on the adjacency relationships of nodes in the graph structure, the state difference between each node and its neighbors is calculated, and a weighted operation is performed using the corresponding coupling coefficients in the graph Laplacian matrix. This weighted operation is used to simulate the spatial diffusion process of hydraulic state in a water supply network, enabling the physical propagation of the state of each node within its neighborhood structure.

[0080] A node-by-node processing approach is adopted, combining the weighted diffusion results of all nodes in node order to obtain a set of diffusion vectors consistent with the initial state dimension. This set of diffusion vectors characterizes the spatial evolution trend of each node's state over time under the guidance of the graph topology, serving as the initial estimate of the state derivative in the neural ordinary differential equation system.

[0081] The output diffusion vector set of the topology-aware diffusion function is used as a component of the right-hand side of the neural network's ordinary differential equation. This is then used to fuse the physical consistency gating function with the source-sink sparse estimation function, participating in the adaptive integral solution of the continuous-time state evolution process. The initial state derivative vector is used as the basic drift term in the neural network's ordinary differential equation. Based on the flow rate, water pressure, and node spatial location characteristics in the fused state initialization vector, node continuity residual vectors and energy conservation residual vectors are constructed. The continuity residual vector and the energy conservation residual vector are concatenated to form a residual feature set, which is then input into the gating function to generate residual gating weights. The initial state derivative vector is then linearly weighted and fused with the theoretical state derivative calculated based on physical conservation, with the weights determined by the residual gating weights, to generate the physical consistency state derivative vector.

[0082] In this embodiment, to achieve consistency constraints between data-driven prediction results and hydraulic laws, a physical consistency gating function is constructed to dynamically adjust the fusion ratio between topology-sensing diffusion results and theoretical state changes. The operation process of the physical consistency gating function includes the following steps:

[0083] From the predicted state sequence corresponding to the current integration time of the constant differential equation, the inflow, outflow and current water pressure values ​​of each pipeline node are extracted, and a node-level mass conservation constraint expression is constructed based on the connection relationship of each node in the graph structure. Furthermore, based on the node water pressure and the pipe segment connection direction, combined with the pipe length and roughness parameters in the edge features, the local energy conservation deviation is calculated. The continuity residual vector and the energy residual vector are formed respectively as the basis for measuring the degree of deviation of the physical constraint.

[0084] The continuous residual vector and the energy residual vector are concatenated to construct a set of physical residual features, which serves as the input for generating gating weights. This input reflects the degree to which the current state deviates from the theoretical model in a physical sense and is used to dynamically control the weight allocation between data-driven terms and physical terms.

[0085] The aforementioned set of physical residual features is input into a residual gating function, which consists of two linear transformation layers and a set of normalization functions. The function outputs a gating weight vector corresponding to the state derivative of each node. The gating weight vector ranges from 0 to 1, representing the weight of the physical term in relation to the state derivative at the current time step.

[0086] Using the gated weight vector as a fusion coefficient, the state derivative vector output by the topology-aware diffusion function is linearly weighted and fused with the theoretical state derivative vector calculated based on the continuity and energy conservation equations to generate a fused physically consistent state derivative vector. This vector is then used as one of the right-hand side terms of the neural network's ordinary differential equation to participate in the continuous-time integral evolution process. The flow mutation value, water pressure fluctuation rate, and neighborhood pressure difference information of each node within the time window are extracted from the fused state initialization vector and concatenated into a source-sink feature input set. This set is then input into a sparse activation function and outputs a source-sink estimation vector.

[0087] In this embodiment, to achieve dynamic estimation of potential abnormal source and sink behaviors in the water supply network, a sparse activation function is constructed for the source and sink sparse estimation function. This function is used to compress, filter, and extract the saliency of abnormal state responses of nodes. The construction and use of this sparse activation function includes the following steps:

[0088] The flow rate change rate, water pressure fluctuation amplitude, and water pressure difference between adjacent nodes are extracted from the initialization vector of the fusion state. These three types of indicators are then concatenated according to the node arrangement order to form a source-sink anomaly feature vector set. This set serves as the input to the source-sink estimation function to determine whether there is leakage, abnormal water injection, or other abnormal water use behavior in the local area.

[0089] The set of source-sink anomaly feature vectors is input into a sparse activation function, which consists of two parts: a linear mapping layer, which transforms the input vector into a response intensity vector with the same number of nodes; and a sparse mapping layer, which performs amplitude compression and nonlinear gating on the response intensity. The sparse mapping layer adopts a SoftShrink function structure with an adjustable temperature parameter, and performs the following processing for each input response value: when the response value is less than a given sparse threshold, it is suppressed to zero; when the response value is greater than the threshold, it is scaled and retained proportionally, thereby achieving the filtering of weak unstructured perturbations and the enhancement of significant source-sink signals.

[0090] The processed response vector is output as the source-sink estimation vector, where each dimension corresponds to the source-sink intensity estimate of a node in the pipeline network. Nodes with a source-sink estimation vector value of zero are considered to have no source-sink abnormal behavior, while nodes with a value greater than zero are marked as potential leak points or illegal water intake points. This output vector, together with the state derivative, serves as the state augmentation input to the neural ordinary differential equation.

[0091] In the adaptive step-size integration process of the ordinary differential equation, the time derivatives of the state evolution vector and the source-sink estimation vector are solved jointly to achieve continuous modeling and dynamic updating of source-sink behavior. The temperature parameter in the sparse activation function is dynamically adjusted during training to control the sparsity and adapt to different pipeline structures and data fluctuation characteristics.

[0092] The physical consistency state derivative vector and the source-sink estimation vector are used as joint inputs to the neural ordinary differential equation. An adaptive step-size numerical integration method is adopted to perform a continuous-time evolution process, and the predicted state sequence and the source-sink estimation vector are output.

[0093] In this embodiment, the steps for constructing the augmented state evolution equation include:

[0094] An augmented state derivative vector is generated by concatenating the state derivative vector generated by the topology-aware drift function and the physically consistent state derivative vector generated by the physical consistency gating function. This augmented state derivative vector is then input into the neural ordinary differential equation and, combined with the source-sink estimation vector generated by the source-sink sparse estimation function, augments the state derivative. The augmented state derivative is then integrated using an adaptive step-size numerical integrator to obtain the augmented state vector. At each time step, the gradient descent method is used to minimize the error between the augmented state derivative and the hydraulic and energy conservation constraints, ensuring the physical consistency of the augmented state vector.

[0095] An adaptive step-size numerical integrator is used to solve the augmented state evolution equation in the neural ordinary differential equation. The integrator automatically adjusts the calculation step size according to the rate of change of the current state to improve computational efficiency while ensuring accuracy.

[0096] At each time step, the integrator determines the step size based on the current state derivative and error estimate. If the error at the current step size exceeds a set threshold, the integrator reduces the step size to increase computational accuracy; if the error is small, the integrator increases the step size to improve computational speed. In this way, the integrator can flexibly adapt to changes in the system state, especially under rapid or abrupt changes.

[0097] The integrator uses local error estimation and global error control to ensure that the error remains within an acceptable range throughout the integration process. Whenever the calculation error exceeds the predetermined range, the integrator adjusts the step size based on the error estimate and recalculates the state until the error is reduced to a suitable level.

[0098] An adaptive step-size numerical integrator works in conjunction with a neural network of ordinary differential equations (NCEs). By dynamically adjusting the step size at each time step, it ensures the stability and accuracy of model predictions, especially under long-step predictions and complex system state changes, effectively reducing computational resource waste. The augmented state vector and the predicted state sequence are input into the control decision process to generate a set of control commands. During the numerical solution process, error control and time step adjustment are combined to optimize the solution accuracy of the augmented state evolution equations, ensuring the stability and accuracy of long-term predictions.

[0099] In this embodiment, the construction process of the multi-objective decision network includes the following steps:

[0100] Continuous water pressure, flow rate, and water age sequences of each water supply node are extracted from the predicted state sequence to construct a state time series tensor. The state time series tensor is normalized and input into a feature cross-coding layer to extract a set of cross-coupling feature vectors. This set of cross-coupling feature vectors is then input into a four-branch objective modeling structure, which respectively constructs branches for water supply stability estimation, energy consumption estimation, production-sales difference estimation, and source-sink risk scoring. In each branch, a shared-weight feedforward subnetwork is used to calculate the single-objective cost output under the current state. All single-objective cost outputs are aggregated into a multi-objective joint cost vector at the output layer. This multi-objective joint cost vector is then input into a strategy generation unit, which, combined with the current state time series tensor, performs action decoding operations to output a set of pump station power commands and a set of valve regulation commands.

[0101] Among them, the action decoding operation refers to concatenating the multi-objective joint cost vector and the predicted state time series tensor into a joint input, generating continuous control variables for pump station power and valve regulation through a two-branch feedforward neural network, performing range pruning and discretization mapping on the above continuous control variables, and outputting a set of pump station power commands containing pump station start-stop status, operating frequency, and power setpoint, and a set of valve regulation commands containing valve opening value and regulation rate.

[0102] In this embodiment, the construction steps of the control barrier function projector include:

[0103] Initial control commands are extracted from the pump station power command set and valve regulation command set. Combined with water pressure, flow velocity, and water age information at each node in the water supply network, a set of constraints for the control commands is constructed. Based on the set upper and lower limits of water pressure, flow velocity ranges in pipe sections, and water age constraints, physical constraint boundaries for each control command are constructed, forming a constraint set containing multi-dimensional restrictions. The initial control command set and constraint set are input into a control obstacle function projector. The control obstacle function projector adjusts the initial control commands using the gradient descent method to ensure that the control commands are feasible within the constraint range. The control obstacle function projector calculates the adjusted projection values ​​of the control commands, generating a set of executable scheduling commands that satisfy the physical constraints. The set of executable scheduling commands that satisfy the physical constraints refers to the set of commands that can be actually executed in the water supply network after adjustment by the control obstacle function projector. This set ensures that each command is executed under physical constraints, avoiding system instability or equipment damage caused by unreasonable control commands. Specifically, the scheduling command set includes:

[0104] Pump station power command set: This set contains the power setting values ​​of each pump station under physical constraints; specifically, the power command of the pump station is limited to the rated power range of the pump station, which is determined by the maximum and minimum power of the pump station; when the water supply system requires more flow, the pump station power command will increase, and vice versa, to ensure that the pump station always operates within a safe and efficient power range.

[0105] Valve regulation command set: This set contains regulation commands for each valve, indicating the opening value of each valve. Each valve's regulation command is constrained by its design specifications and actual operational limitations (such as maximum and minimum opening degrees). By adjusting the valve opening, the flow rate and water pressure in the pipeline network are ensured to operate stably within safe ranges, while avoiding excessive valve closure or opening, which could cause pipeline damage or pressure instability.

[0106] Executability check: Each instruction undergoes an executability check after generation to ensure that it meets system requirements without violating any physical constraints. For pump station power instructions, the system checks whether they exceed the pump station's power limit. For valve adjustment instructions, the system checks whether they exceed the valve's adjustment range. If an instruction exceeds the constraint range, the system will automatically adjust the instruction to restore it to the allowable range.

[0107] Command optimization and adjustment: After the projector calculates and meets the physical constraints, the set of scheduling commands will be further optimized and adjusted to ensure that it can respond efficiently to the actual system needs during the scheduling process. For example, when the water supply pressure is insufficient, the power command of the pump station may be increased, or the valve opening command may be appropriately increased to quickly make up for the insufficient flow in the system.

[0108] Through the above process, the final set of executable scheduling instructions ensures the stability and safety of the system under various operating conditions, while meeting the actual needs of the water supply network, including requirements for flow, water pressure, and energy consumption. These instructions can be directly sent to the control system to drive pump stations and valves to operate according to the optimal strategy, thereby optimizing the overall operating efficiency of the system.

[0109] In this embodiment, the steps of the digital twin simulator performing dynamic response simulation include:

[0110] Collect static structural information and operational boundary parameters of the water supply network, and construct a physical structure simulation diagram that includes node locations, hydraulic connection relationships, pipe diameters and lengths, and control boundaries of pump stations and valves;

[0111] The set of executable scheduling instructions is input into the physical structure simulation diagram, the simulation start time and time step are set, and continuous time hydraulic state evolution calculation is performed based on the current state input to generate the simulation state sequence.

[0112] The nodal pressure, pipe flow velocity, and water age at each time step in the simulation state sequence are extracted to construct a set of state vectors consistent with the predicted state sequence.

[0113] Align the predicted state sequence with the simulated state sequence by time step, calculate the state difference between each corresponding time step, and generate a state difference sequence.

[0114] The network parameters of the neural ordinary differential equation in the augmented state evolution equation are optimized and updated by backpropagation of errors based on the state difference sequence.

[0115] Simultaneously, the gate weights of the source-sink sparse estimation function are adjusted and updated using difference-guided methods to generate a new set of parameter weights.

[0116] In this embodiment, the step of performing difference-guided adjustment and update of the gating weights of the source-sink sparse estimation function includes:

[0117] Normalize the state difference sequence to generate a set of amplitude weight coefficients, and map and match the amplitude weight corresponding to each node with the corresponding gate weight position in the source-sink sparse estimation function to construct an error localization index set.

[0118] Traverse the error location index set, extract the current weight value for the gating weight at each corresponding position, and perform a direction consistency judgment in combination with the amplitude weight coefficient. If the error direction is consistent with the gating output direction, maintain the current weight; otherwise, perform sign flipping.

[0119] The gating weight value at each position is adjusted proportionally based on the magnitude weight coefficient to obtain the updated gating weight matrix;

[0120] The updated gate weight matrix is ​​written into the corresponding position in the source-sink sparse estimation function, replacing the original weight structure, thus completing the gate weight adjustment operation of the source-sink sparse estimation function.

[0121] Example 1:

[0122] To verify the feasibility of this invention in practice, it was applied to a smart water management scenario within a large water supply area. This area comprises multiple pumping stations, storage tanks, main pipelines, and end-user nodes, exhibiting a complex hydraulic structure and drastically changing user demands. Traditional scheduling methods struggle to balance supply and demand in real time, resulting in high energy consumption, large pressure fluctuations, and pipeline water age exceeding limits. To address these issues, the proposed smart water management system, driven by neural network constant differential equations, was deployed on an edge server within the area, enabling real-time modeling and optimization of multi-source water data.

[0123] In this scenario, the system first deploys various types of sensors to collect real-time data on flow rate, water pressure, temperature, residual chlorine, and conductivity within the pipe network. Edge computing devices then perform preprocessing operations on the data, including time alignment, outlier removal, and normalization. The standardized data is then used to abstract the spatial structure through a graph topology modeling module, constructing a water supply graph structure containing 512 nodes and 1864 edges. Node feature vectors include dimensions such as flow rate, water pressure, and node type, while edge feature vectors include parameters such as pipe diameter, length, and material.

[0124] A standardized water input tensor and graph structure representation input state initialization module generates a fused state initialization vector. This vector is then input to an augmented neural network frequent differential equation module, which constructs a topology-aware drift function, a physical consistency gating function, and a source-sink sparse estimation function. An adaptive step-size integration operation is performed in the continuous time domain to predict the evolution of flow rate, water pressure, and water age over the next two hours. The predicted state sequence is processed in a multi-objective decision-making module, which constructs four weighted cost functions for water supply stability, energy consumption, production-sales gap, and source-sink risk. Action decoding operations are then used to generate pump station power and valve opening adjustment commands.

[0125] The scheduling instructions are input into the physical constraint projection module. Combined with the system-set upper and lower limits of water pressure (0.25MPa~0.55MPa), flow velocity range (0.6m / s~2.4m / s), and water age constraint (<36 hours), projection calculations are performed, outputting a set of executable scheduling instructions that meet the physical constraints. The scheduling instructions are sent to the field execution controller, and dynamic simulation verification is performed synchronously in the digital twin simulation module. The system updates the simulation state every minute, calculates the deviation between the predicted state and the simulation state, and adjusts the parameters in the neural network constant differential equation model and the gating weights of the source-sink sparse estimation function accordingly.

[0126] After 7 consecutive days and 168 hours of real-time operation testing, the operational effects of this invention and traditional rule-based scheduling methods in the region were collected and compared. In terms of flow stability, the fluctuation range decreased from ±18% to ±5%; in terms of energy consumption, the energy consumption per unit of water supply decreased from 0.48 kWh / m³ to 0.38 kWh / m³; in terms of production-sales difference control, the leakage rate decreased from 9.6% to 5.3%; in the source-sink balance assessment, the average deviation of predicted source-sink distribution decreased from 12.5% ​​to 4.1%; the water pressure compliance rate increased from 89.2% to 98.3%; the number of locations exceeding the water age limit decreased by 76; the model training convergence period was controlled within 18 hours, and the inference time was within 0.8 seconds, demonstrating good real-time performance and adaptability.

[0127] Table 1: Comparison of Control Effects of Smart Water Management Systems

[0128] Indicator category Comparison project Traditional rule scheduling value Invention system scheduling value Relative improvement rate Hydraulic stability Pressure fluctuation range ±18% ±5% Promotion 72.2% Hydraulic stability Water pressure compliance rate 89.2% 98.3% Promotion 10.2% Energy consumption efficiency Unit water supply energy consumption 0.48 0.38 Reduction 20.8% Production and sales balance Leakage rate 9.6% 5.3% Reduction 44.8% Production and sales balance Source and sink distribution deviation 12.5% 4.1% Reduction 67.2% Water quality guarantee Water age over-limit point 122 46 Reduction 62.3% Real-time performance Single inference time (s) Not applicable 0.8 - Adaptive ability Model update cycle (h) Not applicable 18 -

[0129] Before implementing the intelligent water management system described in this invention, the original solution had significant shortcomings in several key indicators. In terms of water supply pressure stability, the original system's pass rate was only 88.2%, while after the implementation of this invention, this indicator improved to 97.6%, effectively reducing the impact of pressure fluctuations on end users. Regarding the production and sales difference, the original system had significant water supply losses, with a production and sales difference value as high as 12.6%, while after applying this system, this value was significantly reduced to 4.7%, indicating that the intelligent optimization in leakage control and source-sink allocation has achieved significant results.

[0130] In terms of energy consumption control, this invention optimizes pump station and valve scheduling through a multi-objective decision network, reducing the average daily energy consumption from 52.8 MWh to 38.4 MWh, achieving an energy saving effect of 27.3%. Meanwhile, regarding emergency response efficiency, due to the introduction of a continuous state prediction and digital twin simulation mechanism based on neural ordinary differential equations, the response time has been shortened from the original 42 minutes to 18 minutes, giving the system a stronger ability to handle emergencies.

[0131] Furthermore, regarding the accuracy of source and sink regulation, the sparse estimation function constructed by this system significantly improves the fineness of source and sink flow allocation, increasing the accuracy from 78.9% to 93.5%. Pump station operating efficiency has also been significantly improved, with the average number of start-ups and shutdowns decreasing from 8.2 times / day to 5.3 times / day, reflecting more stable and efficient scheduling.

[0132] Overall, by introducing technologies such as topology sensing mechanism, physical consistency modeling and dynamic feedback control, this invention not only improves scheduling accuracy and energy efficiency, but also enhances the adaptability and robustness of the entire water system to complex hydraulic conditions, demonstrating good practical application effects and promotional value.

[0133] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based intelligent water management system, characterized by, The application relates to a water supply network scheduling method based on neural ordinary differential equation and source-sink sparse estimation function. The method comprises the following steps: a data acquisition and preprocessing module is used to acquire data of a water supply network, pre-process and construct a standardized water input tensor; a topological modeling module is used to construct a water supply network graph structure, extract node feature vectors and edge feature vectors, and generate a graph structure representation; a state initialization module is used to receive the standardized water input tensor and the graph structure representation, and generate a fusion state initialization vector; a state evolution calculation module is used to establish an augmented state evolution equation based on the fusion state initialization vector and perform adaptive integration; a multi-objective decision module is used to receive a predicted state sequence, construct a multi-objective cost function and generate pump station power instructions and valve adjustment instructions; a physical constraint projection module is used to project the scheduling instructions under the constraints of water pressure, flow rate and water age, and output an executable scheduling instruction set; 2. A deep learning-based intelligent water management system, characterized in that, a dynamic simulation feedback module is used to perform digital twin simulation and feed back updated neural ordinary differential equation network and source-sink sparse estimation function parameters. The modules are connected through the following methods: data of the water supply network is acquired, pre-processed and standardized to construct a water input tensor; a graph topology structure of the water supply network is constructed, node feature vectors and edge feature vectors are extracted, and a graph structure representation is generated; the standardized water input tensor and the graph structure representation are input into a state initialization network to generate a fusion state initialization vector; the fusion state initialization vector is input as an initial state of a neural ordinary differential equation, a topological perception drift function, a physical consistency gating function and a source-sink sparse estimation function are constructed, an augmented state evolution equation is established, continuous-time states are adaptively integrated, a predicted state sequence and a source-sink estimation vector are output; the predicted state sequence is input into a multi-objective decision network, a cost function including water supply stability, energy consumption, production and sales difference and source-sink risk is constructed, and a pump station power instruction set and a valve adjustment instruction set are output; the pump station power instruction set and the valve adjustment instruction set are input into a control barrier function projector, projection calculation is performed according to the upper and lower limits of water pressure, the flow rate interval and the water age constraint, and an executable scheduling instruction set is generated; 3.The smart water management system based on deep learning according to claim 2, wherein, the executable scheduling instruction set is input into a digital twin simulator for dynamic response simulation, the difference between the predicted state sequence and the simulation state sequence is calculated, and the parameters of the neural ordinary differential equation network and the gating weight of the source-sink sparse estimation function are updated according to the difference between the predicted state sequence and the simulation state sequence.

4. The intelligent water management system based on deep learning according to claim 3, characterized in that, The data of the water supply network includes flow rate, water pressure, residual chlorine, temperature and conductivity data; the preprocessing includes time alignment, abnormality elimination and normalization of the data of the water supply network. The step of inputting the standardized water input tensor and the graph structure representation into the state initialization network to generate the fusion state initialization vector comprises the following steps: the standardized water input tensor and the node feature vector are spliced in the node dimension to obtain a fusion node input feature set; the edge feature vector in the graph structure representation and the topological connection relationship are combined in the edge dimension to generate a fusion edge input feature set; the fusion node input feature set is input into a graph coding network composed of a plurality of nonlinear transformation layers and residual connections to generate an initial node representation vector set; The fusion edge input feature set is input into the edge perception coding network to perform edge feature embedding calculation to obtain an edge representation vector set; The initial node representation vector set and the edge representation vector set are jointly input into an aggregation layer constructed based on a graph attention mechanism to complete node state representation updating and output a fusion state initialization vector.

5. The intelligent water management system based on deep learning according to claim 4, characterized in that, The steps of constructing the topology-aware drift function, the physical consistency gating function, and the source-sink sparsity estimation function include: A graph Laplacian matrix of node connections is constructed according to the graph structure representation, and an edge weight matrix is generated in combination with the pipe diameter, pipe length, and roughness information contained in the edge feature vector; The fusion state initialization vector and the graph Laplacian matrix are input into a topology-aware diffusion function to perform weighted diffusion according to the connection relationship between nodes and the edge weight, and an initial state derivative vector is output; The initial state derivative vector is taken as a basic drift term in the neural ordinary differential equation; A node continuity residual vector and an energy conservation residual vector are constructed according to the flow, water pressure, and node spatial position features in the fusion state initialization vector; The continuity residual vector and the energy conservation residual vector are spliced into a residual feature set, which is input into the gating function to generate a residual gating weight; The initial state derivative vector and a theoretical state derivative calculated according to physical conservation are linearly weighted and fused to generate a physically consistent state derivative vector, with the weight being determined by the residual gating weight; The flow mutation value, water pressure fluctuation rate, and neighborhood pressure difference information of each node in the fusion state initialization vector within a time window are extracted and spliced into a source-sink feature input set, which is input into a sparse activation function to output a source-sink estimation vector; The physically consistent state derivative vector and the source-sink estimation vector are taken as joint inputs of the neural ordinary differential equation, and a continuous-time evolution process is performed using an adaptive step-size numerical integration method to output a predicted state sequence and a source-sink estimation vector.

6. The intelligent water management system based on deep learning according to claim 5, characterized in that, The construction steps of the augmented state evolution equation include: The state derivative vector generated by the topology-aware drift function and the physically consistent state derivative vector generated by the physical consistency gating function are spliced to generate an augmented state derivative vector; The augmented state derivative vector is input into the neural ordinary differential equation in combination with the source-sink estimation vector generated by the source-sink sparsity estimation function to augment the state derivative; The augmented state derivative is integrated by an adaptive step-size numerical integrator to obtain an augmented state vector; At each time step, a gradient descent method is used to minimize the error between the augmented state derivative and the hydraulic constraints and energy conservation constraints to ensure the physical consistency of the augmented state vector; The augmented state vector and the predicted state sequence are input into a control decision process to generate a control instruction set.

7. The deep learning-based intelligent water management system according to claim 6, wherein, The construction process of the multi-objective decision network includes the following steps: The continuous water pressure sequence, flow sequence, and water age sequence of each water supply node in the predicted state sequence are extracted to construct a state time sequence tensor; The state time sequence tensor is normalized and input into a feature cross-coding layer to extract a cross-coupling feature vector set; The cross-coupling feature vector set is input into a four-branch target modeling structure to construct a water supply stability estimation branch, an energy consumption estimation branch, a production and sales difference estimation branch, and a source-sink risk score branch, respectively. In each branch, a shared weight feedforward subnetwork is used to calculate the single-objective cost output under the current state, and all single-objective cost outputs are aggregated into a multi-objective joint cost vector at the output layer; The multi-objective joint cost vector is input into a policy generation unit to perform an action decoding operation in combination with the current state time series tensor, and a pump station power instruction set and a valve adjustment instruction set are output. 8.The intelligent water management system based on deep learning according to claim 7, wherein, The construction step of the control barrier function projector includes: An initial control instruction is extracted from the pump station power instruction set and the valve adjustment instruction set, and a constraint condition set of the control instruction is constructed in combination with the water pressure, flow rate and water age information of each node in the water supply network; According to the set upper and lower limits of water pressure, the flow rate interval of the pipe section and the water age constraint, the physical constraint boundary of each control instruction is constructed to form a constraint set containing multi-dimensional restriction conditions; The initial control instruction set and the constraint set are input into the control barrier function projector, and the control barrier function projector adjusts the initial control instruction by the gradient descent method; The projection value of the control instruction after adjustment is calculated by the control barrier function projector to generate an executable scheduling instruction set that satisfies the physical constraint condition. 9.The intelligent water management system based on deep learning according to claim 8, wherein, The step of dynamic response simulation of the digital twin simulator includes: The executable scheduling instruction set is input into the physical structure simulation diagram, the simulation starting time and time step are set, the continuous time hydraulic state evolution calculation is performed based on the current state input to generate a simulation state sequence; The node pressure, pipe section flow rate and water age at each time step in the simulation state sequence are extracted to construct a state vector set consistent with the predicted state sequence; The predicted state sequence and the simulation state sequence are aligned by time step, the state difference value between each corresponding time step is calculated to generate a state difference value sequence; According to the state difference value sequence, the neural ordinary differential equation network parameters in the augmented state evolution equation are optimized and updated by error back propagation; The difference value guided adjustment update of the gating weight of the source-sink sparse estimation function is performed synchronously to generate a new round of parameter weight set.

10. The intelligent water management system based on deep learning according to claim 9, characterized in that, The step of performing difference value guided adjustment update of the gating weight of the source-sink sparse estimation function includes: The state difference value sequence is normalized to generate an amplitude weight coefficient set, and the amplitude weight of each node is mapped and matched with the corresponding gating weight position in the source-sink sparse estimation function to construct an error positioning index set; The error positioning index set is traversed, the current weight value of the gating weight at each corresponding position is extracted, and the direction consistency is judged in combination with the amplitude weight coefficient, if the error direction is consistent with the gating output direction, the current weight is maintained, otherwise the sign is flipped; The gating weight value at each position is scaled and adjusted according to the amplitude weight coefficient to obtain an updated gating weight matrix; The updated gating weight matrix is written into the corresponding position of the source-sink sparse estimation function to replace the original weight structure, and the gating weight adjustment operation of the source-sink sparse estimation function is completed.

Images