A collaborative control method and system based on spatiotemporal holography and neural reasoning
By constructing a distributed micro-nano sensing network based on spatiotemporal holograms and neural inference, the problems of energy limitation, computational complexity, and perception ambiguity in fiber optic embedded micro-nano sensing networks were solved, achieving efficient and real-time network collaborative control and improving network lifetime and performance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- INFORMATION & COMMNUNICATION BRANCH STATE GRID JIANGXI ELECTRIC POWER CO
- Filing Date
- 2026-04-10
- Publication Date
- 2026-07-31
AI Technical Summary
Existing fiber optic embedded micro-nano sensor networks suffer from contradictions in terms of energy constraints, computational complexity, and perception ambiguity, leading to unreasonable node decisions, poor network lifespan and efficiency. Furthermore, the computational complexity is too high during large-scale heterogeneous collaboration, making it unable to adapt to the computing power of nodes. The separation of perception, communication, and decision-making results in wasted bandwidth and energy.
A collaborative control method based on spatiotemporal holography and neural reasoning is adopted to construct a distributed micro-nano sensor network. Spatiotemporal holograms are constructed through micro-nano sensor nodes. Group collaborative modeling is carried out by combining mean field game theory and neural networks. Lightweight neural networks are used for node decision-making to achieve node energy prediction and data filtering, thus forming a closed-loop collaborative control.
It achieves accurate prediction of node energy fluctuations, reduces computational complexity, improves the quality and transmission efficiency of sensing data, enhances the robustness and adaptability of the network in complex environments, and provides a unified intelligent decision-making framework.
Smart Images
Figure CN122027031B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial Internet of Things and fiber optic communication sensing technology, specifically to a collaborative control method and system based on spatiotemporal holography and neural reasoning. Background Technology
[0002] With the development of the energy internet, the demand for comprehensive, real-time, and highly reliable sensing of infrastructure such as power grids and pipelines is becoming increasingly urgent. Optical fiber embedded micro-nano sensor networks utilize optical fiber itself as a composite medium for communication, sensing, and power supply, forming existing technical architectures such as optical fiber wireless heterogeneous communication architecture, weakly coupled multi-core optical fiber transmission medium, and miniaturized low-power node technology.
[0003] However, existing solutions face three core contradictions in the construction of actual network management and control systems: the contradiction between the high uncertainty of energy under the constraints of physical laws and the static management strategy, which leads to unreasonable node decisions and poor network lifespan and efficiency; the contradiction between the "curse of dimensionality" of large-scale heterogeneous collaboration and the limited computing power of nodes, where the computational complexity of traditional algorithms increases too quickly with the number of nodes and cannot be adapted to the computing power of nodes; and the contradiction between the perceptual ambiguity in complex physical fields and the fragmented architecture of "perception-communication-decision", where nodes lack the ability to judge the value of local data, resulting in wasted bandwidth and energy. Summary of the Invention
[0004] The purpose of this invention is to provide a collaborative control method and system based on spatiotemporal holography and neural reasoning to solve the three core problems of energy limitation, computational complexity and perceptual ambiguity in the prior art.
[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: A collaborative control method based on spatiotemporal holography and neural reasoning is applied to the scenario of fiber-optic embedded wireless heterogeneous communication (Fi-Wi), including the following steps: S1: Construct a distributed micro-nano sensor network system, which includes a central server / gateway, an optical fiber power transmission / communication backbone network, and a distributed micro-nano sensor node group. The optical fiber power transmission / communication backbone network adopts weakly coupled multi-core optical fiber (WC-MCF), and the distributed micro-nano sensor node group is embedded in or attached to the optical cable. S2: Each micro / nano sensing node constructs and updates a spatiotemporal hologram (STH). The spatiotemporal hologram is a lightweight distributed world model that encodes the joint probability distribution of key physical and information states related to network decision-making. Its core variables include the current remaining energy of the nodes. Fixed physical location of nodes along optical cables Received optical power at time t based on the optical fiber power transfer (PoF) physical model The environmental field vector is composed of local environmental physical quantities sensed by the nodes. and the estimated importance of the node to the local data to be transmitted. Among them, the received optical power The calculation formula is: ; In the formula: The optical power injected at time t; The temperature-dependent fiber attenuation coefficient, for Time Node Location Ambient temperature; For nodes Distance from the light injection end; To account for the additional loss caused by nonlinear crosstalk; each node obtains a statistical summary of the states of its neighboring nodes through local wireless communication, and combines it with its own state to encode low-dimensional latent variables using a lightweight variational autoencoder (VAE). And based on the latent variables Inferring the network mean field Low-order statistics; S3: Group Cooperative Modeling and Solution Based on Mean Field Game (MFG): S31: Define individual state Individual control of actions This includes sensing mode / frequency selection, communication decisions, and computational load allocation; S32: Define the mean field State of all nodes Based on the empirical probability distribution in the state space, each node obtains the mean field. Low-order statistics; S33: Define Nodes from Expected cost function to terminal time T: ; In the formula: R is the expected free energy; R is the control effort weight matrix; E represents the terminal cost; E represents the expected computation. S34: The coupled partial differential equations of the MFG are solved approximately using a mean-field neural network (MFNN), wherein the mean-field neural network includes a policy network. and mean field evolution network The gateway trains the MFNN offline and fine-tunes it online. Each node obtains an approximate optimal control action through the policy network based on its own state and mean field low-order statistics. S4: Each node makes individual decisions based on Neuro-Active Inference: S41: Constructing a generative model Where o is the observation and s is the hidden state. For environmental conditions, For internal state, For prior probability, Let the likelihood probability be denoted as '('). S42: State estimation is performed based on minimizing the variational free energy (VFE). The formula for calculating the variational free energy is: ; In the formula: Let KL divergence be a metric. For variational posterior probability; This is the prior probability; Let t be the observation value at time t; Indicates based on Expectation calculation; S43: Action selection is based on minimizing the expected free energy (EFE). The formula for calculating the expected free energy is: In the formula: π represents the candidate strategy; β, γ, and λ are positive weighting coefficients; The expected energy consumption for executing strategy π; The predicted distribution of environmental states after executing strategy π; The prior distribution of the desired environmental state; The expected information gain for executing strategy π; S44: Generate a set of candidate policies, calculate the expected free energy of each candidate policy through a lightweight policy evaluation neural network, and select the policy with the smallest expected free energy to be executed. S5: Repeat steps 2-4 to form a closed-loop collaborative control of "node update STH - receive mean field - execute MFG-AIF joint decision".
[0006] In a further embodiment, in step S2, the statistical summary of the neighboring node states includes the average energy of neighbors and the frequency of specific events, and the network average field... Low-order statistics include the mean and variance .
[0007] In a further embodiment, step S3, the coupled partial differential equations of the MFG include the HJB equation and the FPK equation: HJB equation: In the formula For value function, for gradient, For state The dynamic equations; FPK equation: In the formula Let be the mean field distribution function. For optimal control action, This is for divergence calculation.
[0008] In a further embodiment, in step S3, the mean field low-order statistics broadcast by the gateway include the mean. and variance Each node runs a simplified version of the policy network locally. Based on its own state and the mean ,variance The output is approximately the optimal control action.
[0009] In a further embodiment, in step S4, the neural implementation of the active neural reasoning includes an encoder neural network. Prior networks Cost prediction network, information gain prediction network, and policy evaluation network The encoder neural network is used to approximate the posterior probability of state estimation, and the prior network is based on the STH latent variables. The prior probabilities are dynamically generated using mean field information. The cost prediction network is used to predict the energy consumption of a given policy. The information gain prediction network is used to predict the information gain of a given policy. The policy evaluation network is used to output an approximation of the expected free energy.
[0010] In a further embodiment, in step S1, each node in the distributed micro / nano sensing node group includes an energy harvesting unit, a sensing unit, a communication unit, and a computing and storage unit. The energy harvesting unit is a photovoltaic cell used to receive light energy from a PoF dedicated optical fiber core and convert it into electrical energy. The sensing unit includes a vibration sensor, a temperature sensor, a strain sensor, and a partial discharge sensor. The communication unit supports Fi-Wi mode, receives / transmits data through optical fiber, and performs short-range wireless communication through a miniature radio frequency antenna. The computing and storage unit includes an ultra-low power microcontroller unit (MCU) and a memory used to run neural active reasoning decision-making algorithms.
[0011] In a further embodiment, in step S4, the candidate strategy set is generated by local sampling of the basic actions output by the upper-level mean field game (MFG) module. The basic actions include mid-frequency vibration monitoring, sleep mode, and low-power mode.
[0012] In a further embodiment, the system also includes a collaborative control software module, which is deployed on each node and gateway and includes a spatiotemporal hologram construction and update module, a mean field information processing module, and a neural active inference engine.
[0013] In a further proposed solution, when a sudden event is detected, the variational free energy F(t) of the affected node increases, the node automatically switches to high-frequency, multimodal sensing and reports the event alarm, the gateway updates the global mean field to the event response mode and broadcasts it, and the surrounding nodes adjust their prior and monitoring strategies according to the new mean field. After the event ends, the system returns to normal operation.
[0014] A distributed micro / nano sensor network collaborative control system based on spatiotemporal holograms and mean-field neural active reasoning includes a central server / gateway, an optical fiber power transmission / communication backbone, and a distributed micro / nano sensor node group. The central server / gateway is deployed at the beginning of the optical cable or in a data center to initialize and maintain global mean-field information, run a lightweight mean-field neural network to quickly solve the mean-field equation, and broadcast the solution to the network. It also receives summary information uploaded by nodes to update the global state portion of the spatiotemporal hologram. The optical fiber power transmission / communication backbone uses weakly coupled multi-core fiber (WC-MCF), where at least one core is used to transmit high-power energy light (PoF), and the remaining cores are used to transmit data signals. The distributed micro / nano sensor node group is embedded or attached to the optical cable at preset intervals to construct and update the spatiotemporal hologram, receive mean-field information, perform MFG-AIF joint decision-making, and execute optimal strategies.
[0015] The present invention has the following beneficial effects: This invention enables nodes to accurately predict energy fluctuations by embedding a fiber optic physical model within a spatiotemporal hologram. Combined with active reasoning to encode survival instincts, it ensures the overall survival of the network. Based on mean-field game theory and dimensionality reduction solutions from neural networks, it reduces the computational complexity of node decision-making to a level independent of network size, enabling efficient real-time collaboration in ultra-large-scale networks. Driven by the active reasoning mechanism, nodes transmit only high-value data, filter out redundant and interfering data, and improve the quality of perceived data and transmission efficiency. The spatiotemporal hologram dynamically integrates environmental information, and the equilibrium solution of the mean-field game is adjusted according to the environment, enhancing the robustness and adaptability of the network in complex environments. Active reasoning takes the minimization of free energy as its core, providing a unified and interpretable intelligent decision-making framework, which is conducive to system debugging and verification and application in security-critical scenarios. Attached Figure Description
[0016] Figure 1 This is a system block diagram of the present invention.
[0017] Figure 2 The diagram illustrates the specific steps of the method of the present invention. Detailed Implementation
[0018] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.
[0019] System overall structure construction
[0020] Construct a distributed micro / nano sensor network system as required, with the following specific configuration: The central server / gateway is deployed at the substation at the beginning of the transmission line, and is responsible for the initialization and maintenance of global mean field information, the operation of the mean field neural network, and the data reception and update functions. The fiber optic power transmission / communication backbone network uses weakly coupled seven-core fiber (WC-MCF), where the first core is designated as the energy core to transmit energy light at a wavelength of 1470nm, and the remaining six cores are signal cores to transmit communication signals in the 1550nm band. The optical cable is laid along a 500kV high-voltage transmission line. The distributed micro-nano sensing node cluster is embedded in the optical cable at 100-meter intervals. Each node includes an energy harvesting unit, a sensing unit, a communication unit, and a computing and storage unit: the energy harvesting unit is a photovoltaic cell that receives light energy from the energy core and converts it into electrical energy; the sensing unit integrates vibration, temperature, strain, and partial discharge sensors; the communication unit supports Fi-Wi mode, transmits and receives data through optical fiber, and achieves short-range wireless communication between nodes with the help of a miniature radio frequency antenna; the computing and storage unit is an ultra-low power MCU and memory used to run neural active reasoning decision-making algorithms. The collaborative control software module is deployed on each node and gateway, and includes a spatiotemporal hologram construction and update module, a mean field information processing module, and a neural active inference engine.
[0021] Implementation of the construction and updating of spacetime holograms (STH)
[0022] Each node constructs and updates the spacetime hologram during operation in the following manner: Core variable acquisition: Current remaining energy of the node Real-time data acquisition via energy harvesting unit; physical location. Fixed coordinates are preset during node deployment; environmental field vector. Data is obtained by collecting local environmental physical quantities such as temperature, electromagnetic intensity, and vibration spectrum characteristics from sensing units; data value estimation. Initial values are generated and dynamically updated by a neural active inference engine; Received power prediction Calculation: According to the formula calculate; in Initial optical power injected into the gateway, The initial value is preset as the baseline value of the fiber attenuation coefficient, and is subsequently corrected through updated data broadcast by the gateway. The distance between the node and the light injection end (calculated based on the deployment spacing and initial position). Calculated based on a pre-defined nonlinear crosstalk loss model; Mean field statistic inference: Nodes obtain statistical summaries such as the average energy of neighboring nodes and the frequency of specific events through short-range wireless communication. The neighbor summary input is used to encode low-dimensional latent variables using a pre-defined lightweight variational autoencoder (VAE). And based on this latent variable, the network mean field can be directly inferred. mean and variance .
[0023] Group collaboration based on mean-field game (MFG)
[0024] Modeling parameter definition: Individual state ,in The remaining energy of the node. Low-dimensional latent variables of spatiotemporal holograms; individual control actions In the middle, the sensing mode / frequency selection is set to discrete options such as mid-frequency vibration monitoring and low-frequency monitoring; the communication decision includes sleep, listening, transmission and corresponding transmission power levels; and the computational load allocation is clearly defined as the specific computation cycle of the inference algorithm; mean field Low-order statistics (mean) and variance The gateway calculates the data and then broadcasts it to each node. Expected cost function application: Nodes according to formula
[0025] ; The evaluation cost is given by R, which is a preset control effort weight matrix. The terminal cost is set to encourage nodes to eventually maintain energy above a safe threshold. Mean Field Neural Network (MFNN) Training and Solving: The gateway trains the policy network based on simulation scenario data and historical operation data. and mean field evolution network After offline training and system deployment, the gateway performs online fine-tuning of the MFNN at preset intervals based on real-time network data to minimize HJB residuals and FPK evolution errors; the average value received by each node from the gateway broadcast is... and variance Based on one's own situation Through a simplified policy network running locally The output is approximately the optimal control action.
[0026] Individual decision-making based on neuro-active inference
[0027] Generative model building: Generative models In the diagram, 'o' represents the sensor reading and the received message. This represents the actual vibration modes and other environmental conditions. This refers to the node's internal state, such as its health and energy sufficiency. Provided by spatiotemporal holograms and prior networks, A preset sensor noise model is used; State estimation: Nodes are based on current observations Through encoder neural network Minimize the variational free energy to achieve hidden state inference; Action selection: Based on the basic actions output by the mean-field game module, a candidate strategy set is generated through local sampling. Candidate strategies include increasing / decreasing monitoring frequency and switching sensing modes. The EnergyCost(π) of each candidate strategy is predicted using a cost prediction network, and the InfoGain(π) is predicted using an information gain prediction network. Then, the results are calculated according to the formula: Calculate the expected free energy, where β, γ, and λ are preset positive weighting coefficients; policy evaluation network. Based on the above results, an approximate value of the expected free energy is output, and the node selects the strategy with the smallest value to execute; Neural Implementation: Encoder Neural Network Prior networks The cost prediction network, information gain prediction network, and policy evaluation network are all pre-trained lightweight neural networks, adapted to the operating requirements of ultra-low power MCUs.
[0028] Closed-loop coordinated control operation
[0029] The system implements closed-loop collaborative control according to the following process: Every second, the node acquires local observation data and updates the spatiotemporal hologram; The latest average field statistics received by the gateway broadcast; Basic action suggestions are obtained based on the mean-field game strategy network; Run neural active reasoning, evaluate the expected free energy of candidate strategies, and select the optimal strategy to perform sensing, communication or hibernation operations; The gateway continuously collects summary information uploaded by nodes, updates the mean field statistics, and periodically fine-tunes the MFNN and fiber optic physical model parameters. When a sudden event is detected, the variational free energy F(t) of the affected node increases, automatically switches to high-frequency, multimodal sensing and reports an event alarm. The gateway updates the global mean field to the event response mode and broadcasts it. Surrounding nodes adjust their prior and monitoring strategies according to the new mean field. After the event ends, the system returns to normal operation.
[0030] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A spatiotemporal holographic and neural inference based collaborative control method, characterized in that, Includes the following steps: S1: Construct a distributed micro-nano sensor network system. The distributed micro-nano sensor network system includes a central server / gateway, an optical fiber power transmission / communication backbone network, and a distributed micro-nano sensor node group. The optical fiber power transmission / communication backbone network adopts weakly coupled multi-core optical fiber, and the distributed micro-nano sensor node group is embedded in or attached to the optical cable. S2: Each micro-nano sensing node constructs and updates a spatiotemporal hologram, which is a lightweight distributed world model that encodes the joint probability distribution of key physical and information states related to network decision-making. S3: Group collaboration modeling and solution based on mean-field game theory; S4: Each micro / nano sensing node makes individual decisions based on neural active reasoning; S5: Repeat steps S2-S4 to form a closed-loop collaborative control of "node update STH-receive mean field-execute MFG-AIF joint decision". Step S3 is as follows: S31: Define individual state Individual control of actions This includes sensing mode / frequency selection, communication decisions, and computational load allocation. The remaining energy of the node. For the low-dimensional latent variables of the spatiotemporal hologram, T is the matrix / vector transpose symbol; S32: Define the mean field State of all nodes Based on the empirical probability distribution in the state space, each node obtains the mean field. Low-order statistics; S33: Define Nodes from Expected cost function to terminal time T: ； In the formula: For the first Each node from the current time The expected cost function to the terminal time T; For mathematical expectation calculation, Let T be the current time and T be the end time. for Time of the first The instantaneous expected free energy of each node, Let R be the control action vector of the node, and let R be the control effort weight matrix. For the quadratic cost term of the control action, For the first The terminal cost function of each node at terminal time T; S34: The coupled partial differential equations of the MFG are solved approximately using a mean-field neural network, which includes a policy network. and mean field evolution network The gateway trains and fine-tunes the mean-field neural network offline and online. Each node obtains an approximately optimal control action through the policy network based on its own state and low-order mean-field statistics. It is the first The control action vector of each node It is the state vector of the node.
2. The collaborative control method based on spatiotemporal holography and neural reasoning according to claim 1, characterized in that, Step S4 is as follows: S41: Constructing a generative model In the formula Let s be the observable and s be the hidden state. Let be the prior probability distribution of the environmental state. Let be the prior probability distribution of the internal states. Observation given hidden state s The conditional probability distribution; S42: State estimation is performed based on minimizing the variational free energy (VFE). The formula for calculating the variational free energy is: In the formula for The variational free energy at time, Let KL be the relative divergence. for Hide state at all times The variational posterior probability distribution, for Hide state at all times The prior probability distribution, for The hidden state at all times Based on variational posterior distribution Mathematical expectation operation, for Observations at a given time; S43: Action selection is based on minimizing the expected free energy. The formula for calculating the expected free energy is: ； In the formula: β, γ, λ are weighting coefficients. The energy cost required to execute strategy π To measure the divergence between two probability distributions, For the environment state under policy π The variational posterior prediction distribution, The prior preference distribution for environmental states. The information gain resulting from the execution of strategy π; S44: Generate a set of candidate policies, calculate the expected free energy of each candidate policy through a lightweight policy evaluation neural network, and select the policy with the smallest expected free energy to be executed.
3. The collaborative control method based on spatiotemporal holography and neural reasoning according to claim 1, characterized in that, In step S2, the statistical summary of neighboring node states includes the average energy of neighbors and the frequency of specific events, as well as the network average field. Low-order statistics include the mean and variance .
4. The collaborative control method based on spatiotemporal holography and neural reasoning according to claim 1, characterized in that, In step S3, the coupled partial differential equations of the mean-field game include the HJB equation and the FPK equation: HJB equation: ; In the formula This is the first-order partial derivative of the value function with respect to time t. For at any time ,state The value function under, for gradient, For state The dynamic equation, for The mean field distribution function of the network at time t, This is the dot product of the gradient vector and the dynamic equation; FPK equation: In the formula The first partial derivative of the mean field distribution function with respect to time t is... The mean field distribution function is... For optimal control action, For divergence calculation, This is the product of the mean field distribution and the state evolution.
5. The collaborative control method based on spatiotemporal holography and neural reasoning according to claim 2, characterized in that, In step S4, the neural implementation of the active neural reasoning includes an encoder neural network. Prior networks Cost prediction network, information gain prediction network, and policy evaluation network The encoder neural network is used to approximate the posterior probability of state estimation, and the prior network is based on the low-dimensional latent variables of the spatiotemporal hologram. The prior probabilities are dynamically generated using mean field information. The cost prediction network is used to predict the energy consumption of a given policy. The information gain prediction network is used to predict the information gain of a given policy. The policy evaluation network is used to output an approximation of the expected free energy.
6. The collaborative control method based on spatiotemporal holography and neural reasoning according to claim 1, characterized in that, In step S1, each node in the distributed micro-nano sensing node group includes an energy harvesting unit, a sensing unit, a communication unit, and a computing and storage unit. The energy harvesting unit is a photovoltaic cell used to receive light energy from the PoF dedicated optical fiber core and convert it into electrical energy. The sensing unit includes a vibration sensor, a temperature sensor, a strain sensor, and a partial discharge sensor. The communication unit supports Fi-Wi mode, receives / transmits data through optical fiber, and performs short-range wireless communication through a miniature radio frequency antenna. The computing and storage unit includes an ultra-low power microcontroller unit and a memory used to run neural active reasoning decision-making algorithms.
7. The collaborative control method based on spatiotemporal holography and neural reasoning according to claim 2, characterized in that, In step S44, the candidate strategy set is generated by local sampling of the basic actions output by the upper-level mean field game module. The basic actions include mid-frequency vibration monitoring, sleep mode, and low-power mode.
8. The collaborative control method based on spatiotemporal holography and neural reasoning according to claim 1, characterized in that, The micro-nano sensor network system also includes a collaborative control software module, which is deployed on each node and gateway and includes a spatiotemporal hologram construction and update module, a mean field information processing module, and a neural active inference engine.
9. A distributed micro / nano sensor network collaborative control system based on spatiotemporal holograms and mean-field neural active reasoning, characterized in that, The method for implementing any one of claims 1-8 includes a central server / gateway, an optical fiber power transmission / communication backbone network, and a distributed micro-nano sensing node group; the central server / gateway is deployed at the beginning of the optical cable or in a data center, and is used to initialize and maintain global mean field information, run a lightweight mean field neural network to quickly solve the mean field equation, broadcast the solution results to the network, and receive summary information uploaded by nodes to update the global state part in the spatiotemporal hologram; The optical fiber power transmission / communication backbone network adopts weakly coupled multi-core optical fiber, of which at least one core is used to transmit high-power energy light, and the remaining cores are used to transmit data signals. The distributed micro-nano sensing node group is embedded or attached to the optical cable at a preset interval, and is used to construct and update spatiotemporal holograms, receive mean field information, perform MFG-AIF joint decision-making, and execute the optimal strategy.