A method for realizing intelligent operation by means of park index algorithm

Abstract

The application relates to the technical field of artificial intelligence and smart park management, and specifically discloses a method for realizing intelligent operation by means of a park index algorithm, aiming to solve the problems of decision rigidity and response delay caused by the solidification of utility weights in a traditional park management system. The method comprises the following steps: collecting multi-source real-time data of logistics, security and energy and realizing space-time alignment; pre-processing and feature extraction of the data; dynamically generating utility weights of the three departments based on a lightweight long short-term memory neural network; constructing a multi-objective utility function matrix; solving a local Nash equilibrium strategy at an edge node; issuing a scheduling instruction and forming a perception-decision-execution closed loop. Through the above technical scheme, the application realizes real-time coupling of department priority and environmental variables, further reduces end-to-end control delay, significantly reduces manual intervention and resource waste, and improves the overall operation efficiency and self-adaptive ability of the park.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and smart park management technology, specifically involving a method for achieving intelligent operation by using park indicator algorithms. Background Technology

[0002] With the rapid development of smart park construction, improving operational efficiency and optimizing resource allocation through intelligent means has become an important trend. Currently, most common park management systems adopt a centralized data processing architecture and independently control and schedule subsystems such as logistics, security, and energy based on preset rules or fixed-weight algorithms.

[0003] However, such methods have significant limitations in practical applications. First, the operational goals of various departments within the park often conflict, making it difficult for static and fixed decision-making models to achieve real-time balance and coordination of multiple objectives in a dynamic operating environment. Second, the park's operating environment is complex and volatile, with frequent unforeseen events, and traditional systems lack the ability to adapt to environmental conditions, resulting in lagging and inefficient scheduling strategies. Furthermore, the reliance on centralized computing and decision-making in remote cloud environments typically involves high data communication latency, making it difficult to meet the high real-time scheduling requirements and impacting the overall system response speed and stability.

[0004] The aforementioned problems often lead to existing systems relying on frequent manual intervention when facing complex scenarios, which not only increases the operational burden but also makes it difficult to further improve resource utilization. Therefore, how to achieve a smart park operation method that can perceive the environment in real time, dynamically and adaptively adjust strategies, and quickly complete optimization decisions locally has become an urgent technical problem to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to provide a method for achieving intelligent operation by means of park indicator algorithms, which can effectively solve the problems in the background art mentioned above.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for achieving intelligent operation by leveraging park indicator algorithms includes the following specific steps:

[0008] S1. Collect multi-source sensing data of logistics, security and energy data by deploying heterogeneous sensing terminals, and send the data to the edge computing node after time synchronization;

[0009] S2. At the edge nodes, the multi-source sensing data is filtered, detected, and normalized, and then fused into a periodically aligned feature vector.

[0010] S3. Analyze the feature vectors to predict the initial weights, and use the business rule engine to correct and normalize them, outputting dynamic weight coefficients.

[0011] S4. Construct quantitative utility functions for each department based on dynamic weighting coefficients, and model cross-influences to form a utility function matrix;

[0012] S5. Based on the utility function matrix, gradient descent iterative solution is performed in the discrete policy space to obtain the local Nash equilibrium policy.

[0013] S6. Convert the Nash equilibrium strategy into a scheduling instruction and send it to the execution terminal, and update the knowledge base based on the feedback data to trigger incremental training of the model.

[0014] Furthermore, in step S1, the acquisition and time synchronization based on multi-source sensing data specifically includes:

[0015] An array of radio frequency identification (RFID) readers is deployed in the logistics operation area, with a sampling frequency set to 10 Hz, to collect the coordinates and speed data of automated guided vehicles in real time.

[0016] A cluster of high-definition cameras is set up in the security monitoring area. The video stream is compressed and transmitted using the H.265 encoding standard, and a lightweight target detection model is run on the edge side to parse the video into a quantized human flow density matrix.

[0017] Install smart sensor metering devices in the energy regulatory zone to collect power data of each power-consuming unit at a frequency of 1 Hz.

[0018] By injecting high-precision timestamps into all sensing data using a precise time protocol chip, it is ensured that the timestamp deviation when multi-source data streams arrive at the edge computing node does not exceed a set threshold.

[0019] Furthermore, in step S2, the preprocessing and feature vector fusion of multi-source sensing data specifically includes:

[0020] Edge computing nodes receive and cache heterogeneous data from logistics, security, and energy subsystems in parallel via message queue protocols, streaming media protocols, and industrial bus protocols.

[0021] A Kalman filter is applied to the logistics coordinate sequence for state estimation and smoothing. The state vector includes position and velocity.

[0022] For security video streams, real-time inference is performed through a compressed target detection network to divide each frame of the image into a grid and count the number of pedestrians, generating a pedestrian density matrix and its global density index.

[0023] Energy consumption data is normalized using a linear mapping process;

[0024] Hard synchronization of the three types of processed data is performed at fixed intervals, and a forward filling strategy is used to handle missing data.

[0025] The synchronized logistics status vector, expanded pedestrian density matrix elements, global density index, and normalized energy consumption value are concatenated in a predetermined order to form a fixed-length fused feature vector.

[0026] Furthermore, in step S3, the dynamic utility weight generation based on the long short-term memory neural network and rule engine specifically includes:

[0027] A long short-term memory neural network is constructed with input features including the rate of change of pedestrian density, energy load rate and its difference, and weather status flag bits;

[0028] The long short-term memory neural network adopts a single-layer structure with 256 hidden layer nodes. The output layer outputs the initial weights of the three departments after being normalized by the softmax function.

[0029] The long short-term memory neural network was trained using historical operational data and inferred departmental weight labels, and the model was compressed using post-training quantization techniques.

[0030] During operation, the long short-term memory neural network receives fused feature vectors at fixed intervals and performs forward inference, outputting initial predicted values ​​for the weights.

[0031] The rule engine performs gain or attenuation correction on the initial predicted value based on real-time detection of specific environmental combinations.

[0032] The corrected weights are then normalized using softmax, and the final dynamic weight coefficients are output.

[0033] Furthermore, in step S305, the forced adjustment of the rule engine includes: increasing the security weight when rain is detected and the pedestrian density exceeds the first threshold; and decreasing the logistics weight when energy consumption is detected to be lower than the second threshold and the average speed of the automated guided vehicle is lower than the third threshold.

[0034] Furthermore, in step S4, constructing the departmental utility function matrix specifically includes:

[0035] Define utility function expressions for the logistics, security, and energy sectors, respectively, consisting of linear combinations of measurable operating parameters, and normalize the utility values ​​to the [0,1] interval;

[0036] The logistics utility function is defined as the weighted sum of the reciprocal of transportation time and the reciprocal of channel resource occupancy rate;

[0037] The security utility function is defined as the weighted sum of surveillance coverage and the reciprocal of risk response time;

[0038] The energy utility function is defined as the ratio of total service output to total energy consumption.

[0039] Based on the linear decay factor fitted by historical data, the cross-influence between departmental strategies is modeled, and off-diagonal elements are calculated.

[0040] The calculated main utility value and cross-effect value are filled into the matrix to form a complete 3×3 utility function matrix.

[0041] Furthermore, in step S5, solving for the local Nash equilibrium in the discrete policy space specifically includes:

[0042] The game solver receives the utility function matrix through a shared memory mapping;

[0043] Define a three-dimensional strategy space containing discrete values ​​of channel opening rate, integer values ​​of additional security personnel, and discrete values ​​of energy regulation coefficient, and initialize the strategy vector;

[0044] Using the dynamic weighted sum of the utility of each department as the optimization objective, the gradient direction is approximated in the discrete policy space using the finite difference method.

[0045] The strategy updates with an adaptive decay step size along the gradient sign direction and clamps out-of-bounds values.

[0046] The local Nash equilibrium strategy solution is output based on the convergence criterion of the maximum number of iterations or the objective function increment between consecutive iterations being less than a threshold.

[0047] Furthermore, in step S6, strategy execution and closed-loop feedback specifically include:

[0048] The local Nash equilibrium strategy vector is encapsulated into a scheduling instruction that includes channel open rate, security reinforcement personnel and energy adjustment coefficient, and then distributed through a publish-subscribe message middleware.

[0049] The automated guided vehicle controller dynamically adjusts the path planning cost weights based on the channel openness rate.

[0050] The security access control system activates corresponding backup patrol routes and adjusts access permissions based on the additional security personnel deployed.

[0051] The energy distribution unit linearly scales the output power setpoint of each substation according to the energy regulation coefficient.

[0052] The system receives the status codes returned by each execution terminal and uses them as the initial status input for the next control cycle, thus forming a closed loop.

[0053] Furthermore, step S6 also includes: writing scheduling instructions, execution status, and environmental characteristic data into a local knowledge base;

[0054] When the accumulated sample data reaches a set number, incremental training of the long short-term memory neural network is automatically triggered. During training, an elastic weight solidification strategy is used to prevent catastrophic forgetting.

[0055] Furthermore, it also includes fault tolerance steps: deploying a dual-edge node redundancy architecture, with the primary and backup nodes monitoring each other's status via periodic heartbeat packets;

[0056] When the backup node determines that the master node is faulty, it takes over the data processing and decision-making functions within a set time and reads the most recent sensing data from the local cache to ensure control continuity.

[0057] In summary, this application includes at least one of the following beneficial technical effects:

[0058] (1) Traditional park management systems use fixed weights or rules, which are difficult to cope with dynamic conflicts and sudden environmental changes among the objectives of departments such as logistics, security, and energy. This invention, through a dynamic utility function generation mechanism based on a long short-term memory neural network, can analyze multi-source data such as pedestrian density, energy consumption load, and weather conditions in real time, and adaptively adjust the decision weights of each department. At the same time, the integrated business rule engine can force intervention in extreme scenarios, so that the system decision-making can follow the intelligent rules driven by data, while ensuring operational safety and priority. This mechanism upgrades the static and rigid collaborative mode to millisecond-level dynamic adaptive adjustment, significantly improving the overall operational efficiency and robustness of the park in complex and ever-changing scenarios.

[0059] (2): This invention constructs a complete "perception-decision-execution" edge computing closed loop. By deploying core functions such as data preprocessing, dynamic weight calculation, and game equilibrium solution on local edge computing nodes within the park, and employing efficient data interaction technologies such as shared memory and zero-copy transmission, the time overhead of remote data transmission and centralized processing is significantly reduced. This solution achieves a much lower latency across the entire link from data acquisition to policy execution than traditional architectures, thereby effectively supporting rapid response to high-time-sensitivity tasks such as automated guided vehicle scheduling and real-time security deployment.

[0060] (3): This invention not only generates and executes scheduling strategies, but also forms a closed loop through execution status feedback and knowledge base accumulation. The system can use historical operating data to incrementally train the weight prediction model, achieving self-evolution and continuous optimization. More importantly, by constructing a utility function matrix that couples between departments and using game theory to solve local Nash equilibria in the discrete strategy space, the final output strategy can take into account the interests of each department and achieve a stable optimization state at the system level. This method effectively solves multi-objective conflicts, reduces the average daily number of manual interventions to an extremely low level, and significantly reduces the resource waste rate from more than 20% of traditional methods, realizing efficient and automatic optimization and allocation of park resources. Attached Figure Description

[0061] Figure 1 This is a schematic diagram of the overall architecture of the method for achieving intelligent operation by leveraging park indicator algorithms proposed in this invention; Figure 2 This is a schematic diagram illustrating the principle framework of the intelligent operation method in this invention. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0063] Currently, in the process of deepening the construction of smart parks, intelligent operation faces the problems of multi-objective conflicts and decision-making rigidity under environmental disturbances. To address these technical problems, this invention proposes a method for achieving intelligent operation using park indicator algorithms, and applies this method to a similar process.

[0064] refer to Figures 1-2 The system constructed by this invention consists of a closed-loop control link composed of perception layer devices, edge computing nodes, and execution terminals deployed in various areas of the park.

[0065] The perception layer includes logistics RFID readers, security camera arrays, and energy sensors;

[0066] The edge computing node embeds a data preprocessing module, a dynamic utility function generator, a local game solver, and a knowledge base module.

[0067] The execution terminal includes an automated guided vehicle controller, a security access control system, and an energy distribution unit.

[0068] The entire system achieves end-to-end collaboration through a unified architecture protocol for communication with an open platform and a low-power wide-area network protocol, ensuring that the end-to-end latency from data acquisition to policy execution is controlled within 100 milliseconds.

[0069] Based on the overall architecture of this application, the present invention first constructs a real-time acquisition and high-precision synchronization system for multi-source heterogeneous data, i.e., step S1, to provide reliable data input for subsequent edge intelligent decision-making. Step S1 of this application is implemented through the following steps:

[0070] S101: Three types of heterogeneous sensing terminals are deployed in layers across the functional areas of the park: RFID reader arrays are deployed in the logistics operation area, high-definition camera clusters are installed in the security monitoring area, and intelligent sensing and metering devices are installed in the energy monitoring area. This deployment strategy forms a three-dimensional data acquisition network covering the entire park, enabling real-time and spatial sensing of transportation vehicles, personnel movement, and equipment energy consumption.

[0071] S102: The logistics subsystem deploys UHF RFID readers with positioning accuracy controlled within ±0.5 meters and a sampling frequency set to 10 Hz. It collects the two-dimensional coordinates and instantaneous speed data of all automated guided vehicles in real time. These readers communicate with the edge gateway via a low-power wide-area network protocol, with a single-hop transmission latency of less than 5 milliseconds, ensuring the real-time performance and accuracy of logistics trajectory data.

[0072] S103: The security subsystem deploys 1080P resolution cameras. The video stream is compressed using the H.265 encoding standard and then accessed to the edge node via Gigabit Ethernet. A lightweight target detection model runs on the edge side, parsing the original video into a 10×10 dimensional crowd density matrix. The heatmap generation delay is no more than 50 milliseconds, enabling real-time quantitative perception of the crowd density in the area.

[0073] S104: The energy subsystem deploys smart meters to collect active power, reactive power, and cumulative energy consumption data of each power-consuming unit at a frequency of 1 Hz. The collected raw power consumption data is mapped to the [0,1] interval through linear normalization, providing a unified dimensional benchmark for subsequent cross-subsystem data fusion.

[0074] In summary, steps S101 to S104 complete the large-scale deployment and fine-tuning of parameters of heterogeneous sensing devices, build a hardware sensing infrastructure covering the three major business domains of logistics, security and energy, and ensure the spatiotemporal consistency and dimensional comparability of multi-source data at the collection end.

[0075] S105: All sensing devices have a built-in high-precision real-time clock chip supporting the IEEE 1588 Precision Time Protocol (PTP) to achieve unified calibration of the global time base. This mechanism injects a high-precision timestamp into each data packet at the source of data generation and transmits it to the edge gateway via a low-power wide area network protocol, ensuring that heterogeneous data streams such as logistics, security, and energy are strictly aligned in the time dimension. The timestamp deviation upon arrival at the gateway does not exceed ±10 microseconds, fundamentally eliminating data fusion errors caused by time asynchrony.

[0076] S106: The edge computing node adopts an 8-core ARM architecture central processor and 32GB random access memory hardware configuration, and is physically deployed in the central computer room of the park. This deployment location maintains physical proximity to the sensing devices, ensuring that the end-to-end communication latency meets the real-time requirements, and providing localized computing power support for subsequent data preprocessing and online game solving.

[0077] Through step S1 above, this invention constructs a four-in-one data acquisition and preprocessing system at the edge of the park, encompassing "equipment deployment, parameter configuration, time synchronization, and computing power assurance." This not only enables the collection and synchronization of heterogeneous information such as logistics trajectories, pedestrian density, and energy load, but more importantly, by integrating sensing, synchronization, and computing functions into edge nodes, it reduces the latency of traditional cloud-centralized data processing from over 200 milliseconds to less than 100 milliseconds, significantly improving the system's response speed to sudden environmental changes (such as heavy rain or a surge in pedestrian traffic). This low-latency, high-synchronization-accuracy data foundation directly supports the real-time generation of dynamic utility functions and the rapid equilibrium solution of local games in subsequent steps, serving as the technical prerequisite for achieving intelligent park operation.

[0078] Following the heterogeneous data acquisition and spatiotemporal synchronization completed in step S1, step S2 aims to improve the quality and refine the features of the raw sensory data. This transforms the noisy, multidimensional source data into standardized, computable fusion feature vectors, providing clean and aligned input for the real-time generation of the dynamic utility function. The preprocessing and feature extraction process in step S2 is as follows:

[0079] S201: Edge computing nodes receive data from subsystems in parallel through three types of protocol interfaces: the logistics subsystem pushes AGV coordinate sequences via message queue protocols (such as MQTT); the security subsystem pushes encoded video streams via streaming media protocols (such as RTSP); and the energy subsystem pushes power consumption timing values ​​via industrial bus protocols (such as Modbus TCP). Internally, each node runs a corresponding protocol client service, storing the received heterogeneous data packets in a timestamp-indexed circular buffer (with a depth of approximately 5 seconds), providing a caching basis for subsequent unified parsing and 100-millisecond period alignment.

[0080] S202: For the original coordinate sequence output by the logistics subsystem, a standard Kalman filter is used for state estimation. The state vector is defined as follows: The process noise covariance matrix is ​​set as a diagonal matrix diag(0.1, 0.1, 0.05, 0.05), which is obtained based on the offline calibration of the AGV motor response delay and chassis kinematic model. The observation noise covariance is set to 0.25 square meters based on the nominal positioning error of the RFID reader (±0.5 meters) and the principle that variance = error squared. During filter initialization, the initial state is taken as the moving average of the first 5 observations, and the initial error covariance is set as an identity matrix. At the edge nodes, a prediction-update recursion is triggered every time new observation data is received, and a smooth trajectory is output for subsequent acceleration calculation.

[0081] S203: For security video streams, a lightweight object detection network runs on edge nodes. Based on YOLOv5s, this model compresses the number of parameters to 40% of the original model through channel pruning and layer fusion, enabling real-time inference at 30 frames per second on an 8-core ARM architecture edge device. Each frame is divided into a 10x10 grid, and each grid cell independently counts the number of detected pedestrians, forming a 10×10 dimensional pedestrian density matrix. The global density index ρ is defined as the arithmetic mean of all elements of the matrix, i.e. .

[0082] Steps S202 and S203 above respectively complete noise suppression and structured transformation of logistics and security data. On the logistics side, a smooth motion state estimate is obtained through Kalman filtering, while on the security side, a quantified population distribution representation is obtained through compression model and mesh partitioning. Both are processed in real time under edge computing power constraints.

[0083] S204: Raw power consumption data collected by the energy subsystem (Unit: kilowatts) Normalized to the [0,1] interval using a linear mapping. The normalization formula is: ,in =0, =500 represents the maximum design load of the park. This processing eliminates the influence of dimensions on energy data, making it comparable to logistics and security characteristics on a numerical scale. If the real-time collected value exceeds the preset range, it is truncated according to the boundary value (i.e., > hour, =1.0)

[0084] S205: All three types of data undergo hard synchronization on the timeline with a basic period of 100 milliseconds. Edge nodes maintain a hash buffer with a timestamp index, using the start time of each period (e.g., 00:00:00.000, 00:00:00.100) as the key. When a data packet arrives, it is stored in the corresponding buffer slot based on its IEEE 1588 timestamp. At the end of each period, the node checks the data status of each slot: if data has arrived within the period, the sample with the timestamp closest to the midpoint of the period is taken as the representative value for that period; if no data is received within 50 milliseconds, it is considered missing, and a forward padding strategy is used to complete the data, i.e., directly copying the valid data from the same slot in the previous period. This mechanism ensures that the three features are complete and without gaps within each 100-millisecond period, providing strictly synchronized input for subsequent fusion.

[0085] S206: Three types of features output from S202 to S205 – smoothed logistics state vector (including...) The 10×10 dimensional security density matrix D, its mean ρ, and the normalized energy load value p′ are concatenated into a fixed-length feature vector in a predetermined order. Specifically, the first four floating-point numbers store the logistics status, followed by 100 floating-point numbers expanded in row-major order to store the matrix D, with ρ stored at the 105th bit and p′ at the 106th bit. This vector, along with its IEEE 1588 timestamp, is stored in the shared memory circular buffer (128 bytes per slot) of the edge node, allowing the dynamic utility function generation module to read and call it via memory mapping. The dataset is continuously updated at 100-millisecond intervals, with each cycle covering the timestamp-aligned feature vector to ensure the spatiotemporal continuity of the input data and the standardization of the calling interface.

[0086] Step S2 involves constructing a complete preprocessing chain at the edge, from "raw data access" to "noise suppression" and then to "feature fusion." This not only transforms multi-source heterogeneous data into dimensionally unified, time-aligned, and structured fused feature vectors, but more importantly, all computations are performed locally at the edge nodes, avoiding the bandwidth pressure and privacy risks associated with transmitting raw video streams or high-frequency coordinate sequences to the cloud. This preprocessing system provides clean, standardized, and low-latency input for the subsequent real-time generation of dynamic utility functions, serving as a crucial link in the overall intelligent operation methodology of the park.

[0087] Following the data cleaning and feature extraction completed in step S2, step S3 constructs a dynamic utility function generation mechanism to transform the fused features, which are unified in scale and aligned in time and space, into departmental utility weights that reflect the real-time environmental state. This mechanism achieves data-driven weight adaptation through a lightweight long short-term memory neural network and embeds business rules to ensure robustness of response under extreme scenarios. The specific implementation process is as follows:

[0088] S301: Deploy a Long Short-Term Memory (LSTM) neural network module on an edge computing node. The network's input features consist of three types of variables: the rate of change of pedestrian density Δρ / Δt, the energy load rate p' and its first-order difference Δp', and weather status flags. .in A three-class classification encoding method is used (0-sunny, 1-rainy, 2-snowy). The network input sequence length is set to 10 time steps (corresponding to the past second), and the original input dimension is 10 × 3 = 30 dimensions. To reduce the computational load on the edge side, only continuous variables are differencing, for example, the density sequence ρ(t) is converted to Δρ(t). This compresses the effective input dimension to approximately 12 dimensions (10 steps × 2 differencing variables + weather encoding). The number of LSTM hidden layer nodes is configured to 256, and after differencing compression, the total number of model parameters is controlled within 150,000, adapting to the real-time inference requirements of an 8-core ARM architecture.

[0089] S302: The network employs a single-layer LSTM structure, including an input gate, a forget gate, an output gate, and cell state units. The input layer has a dimension of 12 (corresponding to the compressed differential features of S301 + weather coding), the hidden layer has 256 nodes, and the output layer is mapped to a fully connected layer with 3 nodes. After normalization using the softmax activation function, the output outputs the utility weights (α, γ, β) for logistics, security, and energy. The input sequence length is set to 10 time steps (corresponding to the historical state within the past second), and the feature vector of each time step is input serially according to the S301 format. This structure memorizes long-term dependencies through a gating mechanism, enabling it to capture the dynamic evolution of environmental variables.

[0090] S303: During the training phase, the model uses 30 days of historical operational data to construct a sample set, and extracts 10-step feature sequences in 100-millisecond increments as input. The labels are the department weight combinations (α, γ, β) derived from the actual scheduling strategy within the next 5 seconds. The derivation rules are as follows: Calculate the resource consumption of each department within these 5 seconds—total AGV mileage (meters) for the logistics subsystem, alarm counts per minute for the security subsystem, and total energy consumption (kilowatt-hours) for the energy subsystem; calculate the proportion of each type of consumption to the total for the current period as the initial weight values, and then map them to the [0,1] interval through global min-max normalization (min / max takes the minimum / maximum value from the 30-day statistics).

[0091] If manually labeled conflict events (such as pedestrians and vehicles competing for road space) occur during this period, the initial value of the security weight γ is multiplied by a correction factor of 1.2, and finally, softmax normalization is applied to (α,γ,β) to ensure that the sum is 1. The loss function is MSE, the optimizer is Adam, the learning rate is 0.001, the batch size is 64, and training lasts for 50 epochs. After training, the PTQ strategy is used, and the KL divergence minimization method is used to determine the INT8 scaling factor on the calibration set (day 1 data), compressing the 32-bit floating-point weights to approximately 2 megabytes, with accuracy loss controlled within 1%.

[0092] S304: During runtime, the LSTM module receives the fused feature vector output from S206 with a 100-millisecond timer interrupt. It performs forward computation in an independent inference thread, outputting the initial predicted values ​​of the utility weight coefficients for the three departments: logistics, security, and energy. The weight results are written to a shared memory area with a mutex lock in floating-point array format (float32, 3 elements) for subsequent reading by the rule engine. If an inference takes more than 100 milliseconds, the current result is discarded and the inference state is reset to ensure normal triggering of the next cycle. This predicted value reflects the theoretically optimal weight allocation for each department under the current environmental conditions, but it has not yet incorporated business rule interventions for extreme scenarios.

[0093] S305: Embedded Business Rules Implement Forced Weight Adjustment. When a specific combination of environments is detected, the rule engine synchronously corrects the LSTM prediction results every 100 milliseconds: if the following conditions are met... If "p' < 0.3 and AGV average speed < 0.5 m / s" is satisfied, then the initial value of the security weight γ is multiplied by a gain coefficient of 1.5. If "p' < 0.3 and AGV average speed < 0.5 m / s" is satisfied, then the initial value of the logistics weight α is multiplied by a decay coefficient of 0.8. The two rules are executed sequentially (security rules take priority). After the correction is completed, softmax normalization is performed on (α, γ, β) to ensure that the weight sum is 1. The rule logic is embedded in the main loop in the form of conditional judgment code blocks and runs on the same thread as S304 to avoid asynchronous delays. This mechanism ensures that under extreme conditions such as a surge in traffic or energy shortages, the weight adjustment meets the priority requirements of park safety management and operation supply.

[0094] S306: The weight values, after being corrected according to business rules, are renormalized locally at the edge nodes using the softmax function, making α+γ+β=1 and mapping to a (0,1) probability distribution. The normalized weight coefficients are written to a pre-mapped shared memory buffer in floating-point array format, allowing the local game solver to read them directly via memory pointers, avoiding data copy overhead. After transmission, the weight data is written to the buffer along with an IEEE 1588 timestamp to ensure timing alignment with subsequent game solving. This mechanism ensures that weight adjustments both follow data-driven principles and embed key business rules, enhancing the system's robustness in extreme scenarios.

[0095] Step S3 constructs a closed-loop dynamic utility generation system at the edge, consisting of "model prediction - rule correction - normalized output". First, the edge computing power constraint problem is solved by differential compression and LSTM lightweight design. Second, offline training and online inference are used to realize the nonlinear mapping from environment state to weights. Finally, business rule embedding and normalization processing are used to ensure the controllability of response under extreme conditions.

[0096] The mechanism in step S3 upgrades the traditional static weight configuration to millisecond-level adaptive adjustment, enabling the park operation strategy to respond in real time to emergencies such as rainstorms and surges in people flow. This provides intelligent and reliable preference input for the subsequent game solver, which is the core innovation of this invention in achieving multi-objective dynamic collaboration.

[0097] Following the dynamic weight coefficients (α, γ, β) generated in step S3, the core task of step S4 is to map the abstract weights into quantifiable utility expressions for each department and construct a 3×3 utility function matrix that includes the coupling relationships between departments. This is achieved through the following steps:

[0098] S401: Based on the dynamically adjusted weighting coefficients, three types of quantifiable utility expressions are constructed. Logistics utility is expressed as a weighted sum of transportation efficiency and resource utilization; security utility as a weighted sum of monitoring coverage and response timeliness; and energy utility as a single indicator of the ratio of service output to energy consumption. Specifically, each utility function is constructed by a linear combination of measurable park operation parameters. The weighting coefficients are configured offline through departmental priorities to ensure that all three utility values ​​are normalized to the [0,1] interval after calculation, providing a unified dimensional benchmark for subsequent matrix operations.

[0099] The logistics utility function is denoted as Where L represents Logistics, defined as ,in The average transport time of all AGVs within the current 100-millisecond period is obtained by statistically averaging the path time of each AGV from the starting point to the destination. The critical channel resource occupancy rate is obtained from the percentage of time the channel was occupied within the statistical period of pressure sensors deployed on the main road. The weighting coefficients satisfy... + =1, taking it as an example , This default value is stored in the configuration file of the edge node and supports dynamic hot updates.

[0100] Security utility function, denoted as Where S represents Security, defined as ,in To ensure effective monitoring coverage, the percentage of the union of the overlapping areas of the 86 cameras' fields of view was calculated. The risk event response time is obtained by combining the access control logs and the security personnel location system to calculate the average response time over a specific period. The weighting coefficients satisfy... + =1, default value =0.5, =0.5, also supports hot updates of configuration files.

[0101] S404: The energy utility function is defined as follows: ,in The total energy consumption of the park during the period is obtained by summing the readings of 312 smart meters; To determine the total service output, a weighted calculation is performed using a preset conversion factor (e.g., 1 ton-kilometer of transport volume is converted into 10 security inspection points), and the numerator value is obtained by summing the results in real time at the edge nodes. This ratio directly reflects the service efficiency per unit of energy consumption and is a core indicator for measuring green operations.

[0102] Steps S402 to S404 completed the independent calculation of the basic utility of the three sectors. All three types of expressions are based on measurable physical parameters, and the weighting coefficients are managed through configuration files, realizing the transformation of utility calculation from qualitative to quantitative, and providing diagonal elements for matrix construction.

[0103] S405: Off-diagonal element The indirect utility of the strategy in sector j to sector i is represented by a linear decay factor. Typically, this is the logistics channel openness rate. Increasing the height will reduce security coverage, therefore ,in =0.2 is an empirical coefficient representing the strength of the negative impact of logistics on security. This coefficient is obtained through offline fitting of historical conflict event data. Similarly, increased energy load limits the charging capacity of AGVs, therefore... , We set the value to 0.15. All other off-diagonal terms are calculated using the same method and updated in real time at the edge nodes each cycle.

[0104] S406: Fill the three principal utility values ​​calculated in S402-S404 into the diagonal of the matrix. The six cross-effect values ​​calculated by S405 are filled into the off-diagonal positions to form a complete utility function matrix. The matrix is ​​stored in memory as a structure array, with each element using a single-precision floating-point type, and a total size of 36 bytes. After assembly, the matrix data, along with the current period's timestamp, is encapsulated into a memory block.

[0105] S407: Matrix data is directly transferred to the local game solver via shared memory mapping. Specifically, the edge node pre-allocates a named shared memory region upon startup (e.g., created via mmap on Linux). After S406 writes the encapsulated memory block into this region, the game solver reads it directly through the pre-acquired memory pointer, without needing to execute memcpy, resulting in a transmission latency of less than 1 microsecond. This mechanism ensures that weight adjustments both follow data-driven principles and embed key business rules, enhancing the system's robustness in extreme scenarios.

[0106] Step S4 constructs a link from "utility definition → independent calculation → cross-modeling → matrix assembly → zero-copy transmission". This not only transforms departmental weights into a quantifiable benefit evaluation system, but more importantly, it realistically reflects the coupling and constraint relationships between multiple objectives within the park through cross-influence factors. The matrix-form benefit model provides a standardized input interface for the subsequent game solver, making the solution to the Nash equilibrium no longer a mechanical calculation of static weights, but a dynamic strategy evaluation based on real-time operational status. This forms the key mathematical foundation for achieving multi-departmental collaborative optimization in this invention.

[0107] In step S5, the goal is to find the local Nash equilibrium policy, which is to ensure that no sector can improve its utility by unilaterally changing its policy in the discrete policy space. The solution process is performed locally at the edge nodes using an improved fast gradient descent algorithm, ensuring real-time response. Step S5 is implemented as follows:

[0108] S501: The game solver module deployed on the same edge node receives the 3×3 utility function matrix M output by S407 via shared memory mapping. The matrix data is stored as a single-precision floating-point array (9 elements in total, row-major order). During the initialization phase, the solver converts the memory pointer into a two-dimensional array structure. It also verifies the consistency between the matrix timestamp and the current period to ensure that the calculation is based on the latest utility evaluation model.

[0109] S502: Strategy Space Defined as a three-dimensional discrete set, containing the following control variables:

[0110] Channel open rate Enumerate 21 discrete values ​​with a step size of 5%;

[0111] Increase security personnel There are a total of 21 integer values;

[0112] Energy regulation coefficient Enumerate 9 values ​​with a step size of 0.05;

[0113] Algorithm Initialization Policy Vector This initial value represents a neutral strategy with the channel half-open, moderate security, and normal energy supply, providing a balanced starting point for gradient search.

[0114] S503: Define the objective function and evaluate marginal revenue. The objective function is set as a weighted sum of the utilities of each sector: In the discrete policy space, the gradient direction is approximated using the finite difference method: for each dimension... Evaluate the change in the objective function of adjacent discrete strategies (step size ± 1). Take the direction that maximizes the increase of J as the gradient sign (if both sides decrease, the sign is 0).

[0115] S504: In each iteration k, the algorithm updates the policy along the gradient direction: The step size Adaptive decay strategy adopted: initial value The value is multiplied by a decay factor of 0.95 every 10 iterations. If the update exceeds the boundary of the discrete set, the variable is clamped to the nearest valid value (e.g., ...). (If <0%, it will be forced to 0%) to prevent invalid strategies from being generated.

[0116] S505: Convergence determination uses dual conditions:

[0117] Maximum number of iterations: Up to 200 times, to prevent infinite loops from consuming computing power.

[0118] Utility increment threshold: the increment of the objective function between 5 consecutive iterations This indicates that the strategy optimization has gotten stuck in a local plateau, and the counter is at each Reset when ≥0.001.

[0119] If any condition is met, the algorithm immediately terminates the iteration and outputs the current policy as the local Nash equilibrium solution. This solution satisfies: for any sector i, when the policies of other sectors are fixed, For all feasible This means that no one can benefit from unilateral deviation.

[0120] S506: The solver will generate the final policy vector. The parameters (channel open rate, number of security personnel, and energy coefficient) are encapsulated as a structure, and written to a message middleware (such as a publish-subscribe queue) along with their corresponding utility matrix M and timestamps. The execution terminal controller obtains scheduling instructions through subscription and drives the AGV, access control, and energy distribution units to perform actions. This solution process takes an average of 42 milliseconds on the edge node, meeting the 100-millisecond real-time requirement.

[0121] Through step S5 above, a complete local Nash equilibrium solution mechanism is implemented at the edge, encompassing "data reception → spatial definition → direction evaluation → iterative update → convergence determination → result output". This mechanism replaces continuous gradient calculation with the finite difference method in discrete space, making traditional optimization algorithms adaptable to the discrete decision-making scenarios of actual industrial parks; it compresses the solution time to 42 milliseconds through adaptive step size and dual convergence determination, far below the industry threshold of 100 milliseconds; and it decouples the solver from the execution layer through a message queue, improving system stability. This low-latency, high-convergence game-theoretic solution capability directly supports rapid strategy optimization in dynamic environments such as surges in population and energy shortages within industrial parks.

[0122] Following the local Nash equilibrium solution output in step S5, step S6 aims to transform the abstract strategy into executable instructions to drive the actions of the park's terminal devices, and to achieve a closed-loop iteration of perception-decision-execution-learning through execution feedback. Step S6 is implemented as follows:

[0123] S601: Edge nodes will use policy vectors The scheduling command is encapsulated in JSON format and contains three fields: `channel_open_rate` (channel open rate percentage), `security_staff_count` (security staff increase count), and `energy_adjust_factor` (energy adjustment factor). The command is distributed via a publish-subscribe message middleware with a communication period of 20 milliseconds. The publish topic is predefined as the campus scheduling namespace. Each terminal controller pre-subscribes to this topic, enabling one-to-many endpoint distribution of the command.

[0124] S602: The automated guided vehicle (AGV) controller parses the `channel_open_rate` field to dynamically adjust the toll cost weights in its route planning algorithm. Specifically, it multiplies the toll cost of main road nodes by a coefficient. The higher the openness rate, the lower the cost. AGVs prioritize low-cost nodes to map macro strategies to micro paths. Controller response latency is controlled within 10 milliseconds.

[0125] S603: The security access control system parses the `security_staff_count` field and activates the corresponding number of backup patrol routes. The system has a built-in pre-planned patrol route pool, with each route bound to a set of access control points. When N additional personnel are deployed, N idle routes are activated according to the round-robin load balancing principle, and the access permissions of the access control points on these routes are temporarily upgraded (e.g., changed from normally closed to normally open), realizing dynamic deployment of security forces. Access control logic adjustments are also completed within 10 milliseconds.

[0126] The energy distribution unit parses the `energy_adjust_factor` field and linearly scales the output power setting value of each substation according to the adjustment coefficient (i.e., the factor, or energy adjustment coefficient). Specifically, the original set power... Updated to The factor is the energy regulation coefficient value obtained from the analysis. The new set value is written into the holding register of the substation intelligent controller through the industrial control bus protocol to achieve second-level power redistribution with a response delay of less than 50 milliseconds.

[0127] S605: After each terminal completes its action, it returns a discrete status code to the edge node: 0 indicates successful execution, 1 indicates partial success, and 2 indicates failure. The status code is transmitted back in acknowledgment mode via a low-power wide-area network protocol to ensure reliable transmission. After receiving the status code, the edge gateway uses it as the initial status input for the next control cycle and injects it into the data acquisition process of S201, forming a closed loop.

[0128] S606: All scheduling instructions, execution status, environmental characteristics, and records of manual intervention are written into the park's operation knowledge base, which is stored locally using an embedded database (such as SQLite). Every 1000 new samples accumulated, the system automatically triggers incremental training of the LSTM dynamic utility generator: freezing the underlying network parameters, fine-tuning only the output layer weights, setting the learning rate to 0.0001, employing an elastic weight solidification strategy to prevent catastrophic forgetting, and executing training during off-peak hours at night.

[0129] S607: The system deploys a dual-edge node redundancy architecture, with the primary and backup nodes monitoring each other's status via heartbeat packets (UDP broadcast, 1-second interval). If the backup node fails to receive a heartbeat from the primary node three times consecutively, it determines that the primary node has failed, and the backup node takes over data processing and decision-making functions within 3 seconds. During the switchover, the backup node reads the sensing data from the most recent 10 seconds (approximately a 100MB circular buffer) from its local cache to ensure control continuity; the switchover process is transparent to the terminals.

[0130] Through step S6, this invention constructs a complete closed-loop control system at the edge, encompassing "command issuance → terminal execution → status feedback → knowledge accumulation → model evolution → fault tolerance assurance." This system not only transforms game-solving strategies into millisecond-level physical actions but also continuously optimizes the utility generation model through execution feedback, enabling the system to self-evolve. Simultaneously, a redundancy switching mechanism ensures uninterrupted 24 / 7 operation, forming the engineering foundation for the long-term adaptability and robustness of this invention.

[0131] To further illustrate the technical effects of the present invention, the following specific application scenario is constructed: One day, a sudden rainstorm hit the park (weather status flag). The pedestrian density in the central square area rapidly increased to 0.75. Edge nodes completed data acquisition and preprocessing within 100 milliseconds. The dynamic utility function generator detected that the condition "rainfall and ρ > 0.7" was met, increasing the security weight γ from 0.4 to 0.6, decreasing the logistics weight α from 0.3 to 0.25, and maintaining the energy weight β at 0.15. The game solver recalculated the strategy accordingly, reducing the output channel opening rate to 40% (from 70%), increasing the number of security personnel to 8 (from 3), and maintaining the energy adjustment coefficient at 1.0. After the command was issued, the automated guided vehicle slowed down and detoured through secondary channels, the security system activated emergency broadcasts and closed unnecessary entrances and exits, and the energy system prioritized power supply to lighting and drainage pumps. Fifteen minutes later, the rain stopped, pedestrian flow decreased, and the system automatically restored its normal weights. No manual intervention was triggered throughout the process, and the measured end-to-end latency was 78.3 milliseconds, far below the industry benchmark of 200 milliseconds. In a comprehensive smart park covering an area of ​​420,000 square meters, this method processes more than 200 cross-departmental collaboration requests per day, reduces the number of manual interventions to less than 2 times per day, and lowers the resource waste rate from more than 20% to less than 6%.

[0132] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

[0133] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0134] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for achieving intelligent operation using park indicator algorithms, characterized in that, Includes the following steps: S1. Collect multi-source sensing data of logistics, security and energy data by deploying heterogeneous sensing terminals, and send the data to the edge computing node after time synchronization; S2. At the edge nodes, the multi-source sensing data is filtered, detected, and normalized, and then fused into a periodically aligned feature vector. S3. Analyze the feature vectors to predict the initial weights, and use the business rule engine to correct and normalize them, outputting dynamic weight coefficients. S4. Construct quantitative utility functions for each department based on dynamic weighting coefficients, and model cross-influences to form a utility function matrix; S5. Based on the utility function matrix, gradient descent iterative solution is performed in the discrete policy space to obtain the local Nash equilibrium policy. S6. Convert the Nash equilibrium strategy into a scheduling instruction and send it to the execution terminal, and update the knowledge base based on the feedback data to trigger incremental training of the model.

2. The method for achieving intelligent operation using park indicator algorithms according to claim 1, characterized in that, In step S1, the acquisition and time synchronization based on multi-source sensing data specifically includes: An array of radio frequency identification (RFID) readers is deployed in the logistics operation area, with a sampling frequency set to 10 Hz, to collect the coordinates and speed data of automated guided vehicles in real time. A cluster of high-definition cameras is set up in the security monitoring area. The video stream is compressed and transmitted using the H.265 encoding standard, and a lightweight target detection model is run on the edge side to parse the video into a quantized human flow density matrix. Install smart sensor metering devices in the energy regulatory zone to collect power data from each electricity-consuming unit; By injecting high-precision timestamps into all sensing data using a precise time protocol chip, it is ensured that the timestamp deviation when multi-source data streams arrive at the edge computing node does not exceed a set threshold.

3. The method for achieving intelligent operation using park indicator algorithms according to claim 1, characterized in that, In step S2, the preprocessing and feature vector fusion of multi-source sensing data specifically include: Edge computing nodes receive and cache heterogeneous data from logistics, security, and energy subsystems in parallel via message queue protocols, streaming media protocols, and industrial bus protocols. A Kalman filter is applied to the logistics coordinate sequence for state estimation and smoothing. The state vector includes position and velocity. For security video streams, real-time inference is performed through a compressed target detection network to divide each frame of the image into a grid and count the number of pedestrians, generating a pedestrian density matrix and its global density index. Energy consumption data is normalized using a linear mapping process; Hard synchronization of the three types of processed data is performed at fixed intervals, and a forward filling strategy is used to handle missing data. The synchronized logistics status vector, expanded pedestrian density matrix elements, global density index, and normalized energy consumption value are concatenated in a predetermined order to form a fixed-length fused feature vector.

4. The method for achieving intelligent operation using park indicator algorithms according to claim 1, characterized in that, In step S3, the dynamic utility weight generation based on the long short-term memory neural network and rule engine specifically includes: A long short-term memory neural network is constructed with input features including the rate of change of pedestrian density, energy load rate and its difference, and weather status flag bits; The long short-term memory neural network adopts a single-layer structure with 256 hidden layer nodes. The output layer outputs the initial weights of the three departments after being normalized by the softmax function. The long short-term memory neural network was trained using historical operational data and inferred departmental weight labels, and the model was compressed using post-training quantization techniques. During operation, the long short-term memory neural network receives fused feature vectors at fixed intervals and performs forward inference, outputting initial predicted values ​​for the weights. The rule engine performs gain or attenuation correction on the initial predicted value based on real-time detection of specific environmental combinations. The corrected weights are then normalized using softmax, and the final dynamic weight coefficients are output.

5. The method for achieving intelligent operation using park indicator algorithms according to claim 4, characterized in that, In step S305, the forced adjustment of the rule engine includes: increasing the security weight when rain is detected and the pedestrian density exceeds the first threshold; and decreasing the logistics weight when energy consumption is detected to be lower than the second threshold and the average speed of the automated guided vehicle is lower than the third threshold.

6. The method for achieving intelligent operation using park indicator algorithms according to claim 1, characterized in that, In step S4, constructing the departmental utility function matrix specifically includes: Define utility function expressions for the logistics, security, and energy sectors, respectively, consisting of linear combinations of measurable operating parameters, and normalize the utility values ​​to the [0,1] interval; The logistics utility function is defined as the weighted sum of the reciprocal of transportation time and the reciprocal of channel resource occupancy rate; The security utility function is defined as the weighted sum of surveillance coverage and the reciprocal of risk response time; The energy utility function is defined as the ratio of total service output to total energy consumption. Based on the linear decay factor fitted by historical data, the cross-influence between departmental strategies is modeled, and off-diagonal elements are calculated. The calculated main utility value and cross-effect value are filled into the matrix to form a complete 3×3 utility function matrix.

7. The method for achieving intelligent operation using park indicator algorithms according to claim 1, characterized in that, Step S5, solving for the local Nash equilibrium in the discrete policy space, specifically includes: The game solver receives the utility function matrix through a shared memory mapping; Define a three-dimensional strategy space containing discrete values ​​of channel opening rate, integer values ​​of additional security personnel, and discrete values ​​of energy regulation coefficient, and initialize the strategy vector; Using the dynamic weighted sum of the utility of each department as the optimization objective, the gradient direction is approximated in the discrete policy space using the finite difference method. The strategy updates with an adaptive decay step size along the gradient sign direction and clamps out-of-bounds values. The local Nash equilibrium strategy solution is output based on the convergence criterion of the maximum number of iterations or the objective function increment between consecutive iterations being less than a threshold.

8. The method for achieving intelligent operation using park indicator algorithms according to claim 4, characterized in that, In step S6, strategy execution and closed-loop feedback specifically include: The local Nash equilibrium strategy vector is encapsulated into a scheduling instruction that includes channel open rate, security reinforcement personnel and energy adjustment coefficient, and then distributed through a publish-subscribe message middleware. The automated guided vehicle controller dynamically adjusts the path planning cost weights based on the channel openness rate. The security access control system activates corresponding backup patrol routes and adjusts access permissions based on the additional security personnel deployed. The energy distribution unit linearly scales the output power setpoint of each substation according to the energy regulation coefficient. The system receives the status codes returned by each execution terminal and uses them as the initial status input for the next control cycle, thus forming a closed loop.

9. The method for achieving intelligent operation using park indicator algorithms according to claim 8, characterized in that, Step S6 also includes: writing scheduling instructions, execution status, and environmental characteristic data into a local knowledge base; When the accumulated sample data reaches a set number, incremental training of the long short-term memory neural network is automatically triggered. During training, an elastic weight solidification strategy is used to prevent catastrophic forgetting.

10. The method for achieving intelligent operation using park indicator algorithms according to claim 1, characterized in that, It also includes fault tolerance steps: deploying a dual-edge node redundancy architecture, with the primary and backup nodes monitoring each other's status via periodic heartbeat packets; When the backup node determines that the master node is faulty, it takes over the data processing and decision-making functions within a set time and reads the most recent sensing data from the local cache to ensure control continuity.

Images