Energy-saving intelligent street lamp automatic emergency response system and control method thereof

Abstract

The invention discloses an energy-saving intelligent street lamp automatic emergency response system and a control method thereof, relates to the technical field of industrial Internet of Things control, and solves the problems that an existing intelligent street lamp system is poor in dynamic scene adaptability, single in emergency response strategy and insufficient in communication stability. According to the method, a dynamic priority scheduling matrix is generated through multi-source data fusion and an adaptive weighted decision tree, and an intelligent dimming strategy is trained in combination with improved fuzzy reinforcement learning; a multi-level fuzzy control and fuzzy reasoning system is used for generating emergency parameters driven by accident levels; dynamically selecting an optimal communication link transmission instruction based on a multiple access protocol and multi-scale channel sensing; an IEEE 1588PTP protocol and Bayesian clock drift correction are adopted to guarantee time sequence consistency, and energy consumption and safety balance are optimized through multi-target reinforcement learning; the dynamic adaptive capacity, the emergency response accuracy and the communication reliability of a complex scene are remarkably improved, and collaborative optimization of energy-saving efficiency and road safety is realized.

Description

Technical Field

[0001] The present invention relates to the field of industrial Internet of Things control technology, and more particularly to an energy-saving intelligent street lamp automatic emergency response system and its control method. Background Art

[0002] Intelligent street lamps are a new generation of urban lighting facilities based on Internet of Things and artificial intelligence technologies, integrating functions such as environmental perception, dynamic dimming, and emergency response. The system collects environmental data through multi-source devices such as photosensitive sensors, radars, and cameras, and realizes remote control through 5G communication. It can automatically adjust the brightness according to pedestrian flow and weather conditions, and at the same time supports intelligent services such as traffic accident warning and equipment failure monitoring, aiming to improve energy efficiency and public safety guarantee capabilities.

[0003] Existing technical solutions achieve energy conservation and emergency response through multi-dimensional perception and adaptive control. For example, patent CN111102514A proposes an intelligent street lamp system and a street lamp lighting control method, which uses a single neuron PID control method to remotely switch or dim each street lamp and each group of street lamps, and at the same time has functions such as lamp failure alarm, parking charging, and road environment monitoring; patent CN112908039A proposes an airspace control method and an intelligent street lamp based on an intelligent street lamp, which integrates a camera device to collect images of a specified airspace and perform object detection to realize the identification, trajectory prediction, and recovery operations of aircraft; patent CN113639244A proposes a 5G communication intelligent street lamp, which provides the function that when a pedestrian passes by the street lamp, the street lamp brightens, otherwise, the street lamp dims, and provides the function of monitoring and reporting road traffic accidents; these technical solutions have significantly improved the energy-saving ability, intelligent level, and safety performance of intelligent street lamps, but there are still problems such as rigid conflict resolution logic, single emergency strategy, and poor communication stability:

[0004] First, most existing intelligent street lamp systems adopt control logic based on fixed priorities. When facing multiple control requirements (such as traffic flow detection, bicycle lane lighting, pedestrian passing detection, etc.), it is difficult to dynamically adapt to complex scenarios. For example, during the day's peak hours, the main road should be given priority for lighting, while at night with low traffic flow, the side roads require enhanced safety lighting. However, existing systems are difficult to flexibly adjust based on real-time traffic, weather, and accident situations, resulting in unreasonable resource allocation, affecting energy conservation and road safety. Secondly, the emergency response methods of existing systems are too fixed. Usually, they only respond to emergencies by increasing brightness and expanding the lighting range, lacking refined adjustment for different accident types and environmental factors. For example, for minor accidents, only moderate lighting enhancement is needed, while for serious accidents, dynamic warning lights or intelligent guidance may be required. The single strategy of existing systems easily leads to light pollution, increased energy consumption, or response lag, affecting the efficiency of emergency handling. Finally, current intelligent street lamps rely on LoRa, NB-IoT, Wi-Fi, or 5G for remote control. However, in underground roads, high-rise dense areas, or remote regions, signals are interfered with or have insufficient coverage, which may cause control delays, command loss, or response failures, affecting the stability and safety of the overall system. Summary of the Invention

[0005] In view of the deficiencies of the prior art, the present invention discloses an energy-saving intelligent street lamp automatic emergency response system and its control method, aiming to solve the problems in the background technology.

[0006] To achieve the above technical effects, the present invention adopts the following technical solutions:

[0007] An energy-saving intelligent street lamp automatic emergency response control method includes:

[0008] S1. Collect data from light sensors, radars, cameras, and environmental monitoring devices, calculate the comprehensive weights of traffic flow, pedestrian density, weather conditions, and accident levels through an adaptive weighted decision tree, and calculate the lighting demand index in combination with historical stored data, and output a dynamic priority scheduling matrix;

[0009] S2. Input the dynamic priority scheduling matrix into a deep Q network, train an intelligent dimming strategy based on an improved fuzzy reinforcement learning method, calculate the optimal brightness allocation for different road sections, and obtain time series lighting control parameters;

[0010] S3. Input the time series lighting control parameters into a fuzzy inference system, conduct an emergency level assessment in combination with accident detection data, and set a fuzzy membership function according to different accident categories to obtain emergency lighting parameters;

[0011] S4. Based on the emergency lighting parameters, adopt a multi-level fuzzy control method to calculate the lighting adjustment range, brightness change gradient, and warning light trigger conditions, and output a lighting scheduling plan;

[0012] S5. Convert the light scheduling plan into an instruction data packet, encapsulate the data using a multiple access wireless protocol, and dynamically select LoRa, NB-IoT, or 5G channels for transmission based on the channel state evaluation by a multi-scale channel state awareness method, and output a light control signal;

[0013] S6. Use the IEEE 1588PTP protocol to synchronize the timing consistency of the light control signals among the smart street lamp nodes, and use an improved Bayesian clock drift correction model for error compensation to correct the time synchronization deviation, and output a timestamp synchronization control instruction;

[0014] S7. Based on the timestamp synchronization control instruction, predict the energy consumption benefits of different dimming strategies through multi-objective reinforcement learning and sparse Bayesian optimization methods, and dynamically adjust the lighting power.

[0015] As a further technical solution of the present invention, an energy-saving smart street lamp automatic emergency response system includes: an environmental perception module for real-time monitoring and data preprocessing using a light sensor, a millimeter wave radar, an infrared sensor, and a camera;

[0016] A dynamic priority calculation module for calculating the weights of traffic, environment, and accident impact factors using an adaptive weighted decision tree, and constructing a dynamic priority scheduling matrix by combining time series trend analysis to determine the lighting requirements of each area;

[0017] A dimming strategy optimization module for training an intelligent dimming strategy using an improved fuzzy reinforcement learning combined with a deep Q network, calculating an optimal brightness allocation scheme through fuzzy value iteration, and optimizing the lighting power based on historical data;

[0018] An emergency response evaluation module for evaluating the accident level through a fuzzy inference system, and calculating the lighting response parameters corresponding to the accident category using fuzzy membership optimization and multi-layer fuzzy inference;

[0019] A light scheduling control module for calculating the light adjustment range, brightness change gradient, and warning light trigger conditions using multi-level fuzzy control, and dynamically adjusting the lighting strategy by combining state transition fuzzy optimization and fuzzy increment compensation;

[0020] A data transmission module for dynamically selecting LoRa / NB-IoT or 5G channels through a multi-scale channel state awareness and reinforcement learning channel allocation optimization method, and synchronizing the timing consistency of the smart street lamp nodes through a Bayesian clock drift correction model and a Kalman filter;

[0021] A remote management module for data storage, analysis, and online optimization and update of intelligent dimming strategies based on edge computing and cloud computing methods; the remote management module uses multi-objective reinforcement learning combined with sparse Bayesian optimization to predict the energy consumption benefits of different dimming strategies, and dynamically adjusts the lighting power through particle swarm optimization and constrained optimization of lighting control.

[0022] Based on the above technical solutions, the positive and beneficial effects of the present invention are as follows:

[0023] 1. Through the adaptive weighted decision tree in step S1 of the present invention, real-time traffic flow, pedestrian density, weather, and accident data are fused, and the weights are dynamically adjusted in combination with historical data to generate a multi-dimensional dynamic priority scheduling matrix, enabling the system to autonomously switch control logic according to time periods and scenario changes (such as giving priority to the main road during peak hours and strengthening lighting on side roads at night), so as to accurately match lighting requirements on the premise of ensuring road safety, avoid over-illumination or insufficient illumination, and significantly improve energy-saving effects and scenario adaptability.

[0024] 2. Through the fuzzy inference system in step S3 and the multi-level fuzzy control in step S4 of the present invention, different fuzzy membership functions are set according to the type of accident (such as minor accidents and serious accidents), and an emergency lighting plan including parameters such as brightness gradient adjustment and warning light trigger conditions is dynamically generated. For example, only the brightness is locally increased for minor accidents, while warning lights and dynamic guiding signals are linked for serious accidents, avoiding over-illumination of the entire road section, thereby reducing ineffective energy consumption and light pollution while improving the efficiency of emergency handling.

[0025] 3. Through the multi-scale channel state perception in step S5 to dynamically select the optimal channel (such as 5G for low-latency emergency instructions and LoRa / NB-IoT for high-coverage regular instructions), and in combination with the improved Bayesian clock drift correction model in step S6, the clock error is compensated through environmental temperature and historical deviation data to ensure the timing consistency of multi-node instruction transmission. Even in underground roads or signal-blocked areas, the integrity and real-time nature of control instructions can still be maintained through redundant channel switching and high-precision synchronization mechanisms, thereby enhancing the robustness and security of the system in extreme scenarios. Description of the Drawings

[0026] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings, where:

[0027] Figure 1It is the architecture diagram of an energy-saving intelligent street lamp automatic emergency response control method of the present invention;

[0028] Figure 2 It is the working principle framework diagram of the adaptive weighted decision tree of the present invention;

[0029] Figure 3 It is the working framework diagram of the improved fuzzy reinforcement learning method of the present invention;

[0030] Figure 4 It is the framework structure diagram of the fuzzy inference system of the present invention;

[0031] Figure 5 It is the step diagram of the multi-level fuzzy control method of the present invention;

[0032] Figure 6 It is the framework diagram of the energy-saving intelligent street lamp automatic emergency response system of the present invention. Specific implementation mode

[0033] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative work shall fall within the protection scope of the present invention.

[0034] In the embodiment, an energy-saving intelligent street lamp automatic emergency response control method, as Figure 1 shown, includes the following steps:

[0035] S1. Collect data from light sensors, radars, cameras, and environmental monitoring devices, calculate the comprehensive weights of traffic flow, pedestrian density, weather conditions, and accident levels through an adaptive weighted decision tree, and calculate the lighting demand index in combination with historical stored data, and output a dynamic priority scheduling matrix; among them, as Figure 2As shown in the figure, the working principle of the adaptive weighted decision tree is as follows: First, calculate the difference between the time series data of the light sensor and the historical accident distribution based on relative entropy, and screen the feature dimensions with excessive entropy offsets through the dynamic threshold within the sliding window; analyze the spatial correlation between the radar point cloud and the pedestrian density heat map through the maximum mutual information coefficient to generate a non-linear correlation coefficient matrix; then, use the C4.5 algorithm to construct a multi-branch decision tree, and divide the nodes through the double thresholds of information gain ratio and Gini index, map the traffic flow time series data, pedestrian density heat map, and weather condition quantization matrix to different decision paths respectively, and output the initial weight vector; at the same time, use the online gradient descent algorithm to use the visibility change rate gradient as the regularization term, and construct a composite objective function in combination with the illumination fluctuation variance and historical residuals to iteratively update the weight parameters; if a sudden accident level jump is detected, reconstruct the weight posterior probability distribution through the Markov chain Monte Carlo sampling method and output the lighting demand index.

[0036] S2. Input the dynamic priority scheduling matrix into the deep Q network, train the intelligent dimming strategy based on the improved fuzzy reinforcement learning method, calculate the optimal brightness allocation for different road sections, and obtain the time series lighting control parameters; as Figure 3 As shown in the figure, the working method of the improved fuzzy reinforcement learning method is as follows: First, fuzzify the continuous state of the dynamic priority scheduling matrix into discrete semantic variables through the Gaussian membership function, and construct a fuzzy rule base to generate the initial dimming strategy; then, generate a target network and an online network based on the deep Q network, calculate the Q-value expectation through the target network, and the online network updates the weights through the temporal difference error; if the state space dimension exceeds the preset threshold, add membership functions through the fuzzy rule dynamic extension mechanism to cover unknown scenarios; during the training process, the improved fuzzy reinforcement learning method uses the priority experience replay mechanism to weighted sample the historical transfer samples according to the absolute value of the TD error, and aggregates the lighting energy consumption and accident response data of multiple road sections through the federated learning framework to construct a global reward function; after the training is completed, the improved fuzzy reinforcement learning method selects the optimal brightness allocation based on the greedy strategy, outputs the time series control parameters to drive the PWM dimming module, and stores the execution results in the edge node cache pool.

[0037] S3. Input the time series lighting control parameters into the fuzzy inference system, combine the accident detection data to evaluate the emergency level, and set the fuzzy membership function according to different accident categories to obtain the emergency lighting parameters; where, as Figure 4As shown in the figure, the fuzzy inference system includes an input parsing module, an accident category classification module, a level inference module, an anomaly correction module, and an illumination parameter output module. The input parsing module is used to perform fuzzy processing on accident levels, pedestrian influence areas, traffic flow density, and environmental visibility variables by using the membership function modeling of the generalized beta function. The accident category classification module is used to classify the fuzzified accident data through fuzzy C-means clustering and output the fuzzy mapping matrix of accident categories. The level inference module is used to calculate the influence weight of accident categories on emergency lighting parameters through the fuzzy information entropy optimization method and determine the accident emergency level through the maximum membership degree decision method. The anomaly correction module is used to identify invalid or abnormal data through Mahalanobis distance weighted filtering when there are signal anomalies, data drifts, or acquisition errors in accident data, and perform time series data smoothing in combination with fuzzy Gaussian regression. The illumination parameter output module is used to generate light adjustment parameters through a rule inference engine and output the emergency lighting intensity, light diffusion range, and response duration of the accident occurrence area.

[0038] S4. Based on the emergency lighting parameters, use a multi-level fuzzy control method to calculate the light adjustment range, brightness change gradient, and warning light trigger conditions, and output the light scheduling plan. As Figure 5 shown, the working method of the multi-level fuzzy control method is as follows:

[0039] U1. Based on the accident level E and visibility V, construct a brightness membership degree distribution through a double-peak Gaussian function. The formula expression is:

[0040]

[0041] In formula (1), L represents the real-time brightness value, with the unit of lux, which is used to quantify the current illumination intensity; 120 represents the brightness center value of the first peak, corresponding to the standard deviation a1 = 15(V / 100) 0.5 , which is used to match the high-brightness warning requirements in the accident core area; 160 represents the brightness center value of the second peak, corresponding to a2 = 25(V / 100) 0.3 , which is used to cover the gradually decreasing brightness distribution in the transition area; a1 and a2 respectively represent the widths of the first peak and the second peak, which are used to control the brightness membership degree diffusion range.

[0042] U2. Based on the accident influence radius R and accident level E, extract the optimal brightness adjustment range ΔL of the current state through a non-linear gradient function. The formula expression is:

[0043]

[0044] In formula (2), represents the reference gradient term, which is used to define the maximum allowable brightness change rate, where L maxRepresents the maximum allowable brightness, L d Represents the target brightness requirement; Represents the accident coupling term, which is used to non-linearly map the accident level and the influence radius into a gradient amplification coefficient; (1 - e -t / 10 ) represents the asymptotic smoothing term, which is used to suppress the brightness mutation at the initial moment through an exponential function. t represents time, with the unit of second; 30, 200, and 10 are empirical constants, which are used to control the reference gradient, the accident coupling strength, and the smoothing speed respectively; if the detected visibility V < 50 meters, the forced gradient amplitude decays to ΔL(t)max(0.5, V / 50), and the smoothed brightness adjustment instruction is output;

[0045] U3. Construct the Pareto optimization matrix P = [S w , E w , D|] T , where the safety matrix S w = E ρ / A , ρ is the pedestrian density, and A is the accident area; the energy consumption weight t is the response time; D is the response delay penalty; the safety matrix S w , the energy consumption weight E w and the response delay penalty D are solved by multi-objective particle swarm optimization min(||P - [1, 0, 0|] T ||w), and the non-dominated solution set is output, where w = [0.6, 0.3, 0.1]E, representing the Euclidean distance weighted by the accident level, which is used to measure the deviation degree of the solution from the ideal point [1, 0, 0|] T . Then, select the actual light intensity L act = L d (1 + 0.15sigmoid(S w - E w ));

[0046] U4. Calculate the trigger condition in real time through a dynamic threshold function. The formula expression of the dynamic threshold function is:

[0047] θ(t) = 0.5 + 0.2sin(2πt / 360) + 0.1T f / C max (3)

[0048] In formula (3), θ(t) is the dynamic trigger threshold; 0.5 is the baseline threshold; 0.2sin(2πt / 360) is the periodic term, with a period of 360 seconds, which is used to simulate the natural fluctuation of day and night visibility; 0.1T f / C max is the traffic flow correction term, where T f is the real-time traffic flow, and C max is the maximum carrying capacity of the road section; if Trigger the rotating warning light and increase the brightness to 1.2L d , otherwise maintain the gradient adjustment strategy; at the same time, feedback data through the light intensity sensor; where represents the visibility change rate; if the regional brightness deviation exceeds the preset threshold, iterative correction is performed through gradient descent until the error tolerance is met.

[0049] S5. Convert the light scheduling plan into an instruction data packet, encapsulate the data using a multiple access wireless protocol, and dynamically select the LoRa, NB-IoT, or 5G channel for transmission according to the channel state evaluation using a multi-scale channel state awareness method, and output a light control signal; among them, the multiple access wireless protocol is encoded by a variable-length frame structure, and forward error correction coding is used to perform bit-level redundancy check on the instruction data packet, optimize the calculation of the data error rate using the Hamming distance, and adjust the redundancy correction ratio in combination with erasure coding; the multi-scale channel state awareness method calculates the instantaneous channel gain G of different wireless channels through a long short-term memory network combined with empirical Bayesian channel estimation c , and the calculation formula is:[[]]

[0050] G c (t) = τP r (t) + (1 + τ)G c (t - 1) τ (4)

[0051] In formula (4), G c (t) represents the channel gain at time t; P r (t) represents the received signal power at time t; τ represents the channel gain smoothing coefficient, which is used to adjust the influence weights of the current received power and the historical channel gain; G c (t - 1) τ represents the channel gain at the previous time t - 1; then, Markov channel evaluation is adopted, and the optimal channel selection probability for the future time is calculated through a weighted historical channel state fusion function, and the calculation formula is:[[]]

[0052]

[0053] In formula (5), P(C t ) represents the probability that the current channel C t is selected at time t; w i and w j represent the historical state weights of the i-th and j-th channels respectively; represents the channel gain of the i-th channel at time t - 1; represents the channel gain of the j-th channel at time t - 2; N represents the total number of available channels; if the channel state fluctuation amplitude exceeds the set threshold, the value function of the channel switching strategy is calculated by the channel allocation optimization method based on reinforcement learning, and the calculation formula is:

[0054] Q(C t , g) = (1 + g)Q(C t , g) + g(r + γmaxQ(C t+1 , g’) (6)

[0055] In formula (6), Q(C t , g) is the return value of channel C t under action g; g is the learning rate; r is the immediate reward; γ is the discount factor; maxQ(C t+1 , g’) represents the maximum Q value of the optimal channel selection decision at the next moment t + 1; based on the channel optimization strategy, the multi-scale channel state perception method dynamically allocates data packets to parallel channels through multi-path adaptive load balancing, adopts a congestion control method based on reinforcement learning, and combines Bayesian bit error rate estimation to calculate the optimal data packet retransmission strategy, and finally outputs the light control signal.

[0056] S6. Use the IEEE 1588 PTP protocol to synchronize the timing consistency of the light control signals between the smart street lamp nodes, and use the improved Bayesian clock drift correction model for error compensation, correct the time synchronization deviation, and output the timestamp synchronization control instruction; the specific working principle is: use the master-slave clock synchronization mechanism to calculate the local clock deviation of each smart street lamp node, and use the two-way timestamp exchange method to record the sending and receiving times of the master clock and the slave clock, calculate and separate the synchronization signal transmission delay and processing delay based on the round-trip delay to obtain the initial clock offset; then, calculate the clock drift trend based on the least squares linear regression method, and use the Kalman filter to estimate the short-term clock error to generate the error compensation parameter; if the detected synchronization error exceeds the set threshold, calculate the clock drift distribution through variational Bayesian inference, and combine the particle filter method to iteratively update the clock frequency error to adaptively adjust the clock synchronization step size by drift compensation; finally, smooth the clock adjustment amount based on the exponentially weighted moving average to optimize the global time synchronization stability, and output the timestamp synchronization control instruction.

[0057] S7. Based on the timestamp synchronization control instruction, predict the energy consumption benefits of different dimming strategies through multi-objective reinforcement learning and sparse Bayesian optimization methods, and dynamically adjust the lighting power. Among them, the working principle of the multi-objective reinforcement learning and sparse Bayesian optimization methods is as follows: construct a multi-objective joint reward function of energy consumption, brightness, and lifespan based on the deep deterministic policy gradient algorithm, and use a dynamic weight allocation mechanism to generate the Q-value matrix of different dimming strategies in real time; then, perform high-dimensional sparse sampling on the policy space through the sparse Bayesian optimization method, and model the nonlinear relationship between the power adjustment amount of each node and the energy consumption benefit based on Gaussian process regression. If the predicted benefit variance exceeds the preset threshold, use Monte Carlo sampling to perform local optimization of the policy parameters under the confidence interval constraint; at the same time, fuse the real-time energy consumption data transmitted by multiple channels in S5 and the clock synchronization deviation compensation amount in S6, dynamically adjust the lighting power duty cycle through a PID fuzzy controller, use an experience replay buffer pool to store historical state, action, and reward tuples, and update the policy network weights through temporal difference error backpropagation.

[0058] In step S1 of the above embodiment, the relative entropy feature screening mechanism is based on the statistical difference measure of Kullback-Leibler divergence, and identifies the significant deviation features (such as sudden light drop, abnormal radar reflection) between the light time series data and the historical accident distribution through a dynamic threshold (such as the 3σ criterion) within a sliding window, and screens out the abnormal periods strongly associated with environmental mutations, providing highly sensitive inputs for subsequent decision-making.

[0059] The maximum mutual information coefficient spatial correlation analysis uses a non-linear correlation measurement method to map the radar point cloud coordinates and pedestrian heat map density to a grid binning space, calculates the maximum mutual information coefficient (MIC) to quantify the traffic-pedestrian interaction risk, and identifies the spatial dependence relationship of high-risk areas such as intersections and sidewalks, injecting spatial dimension logic into the weight allocation.

[0060] The construction of the C4.5 multi-criterion decision tree adopts a double splitting criterion of information gain ratio (IGR) and Gini index, and performs optimal threshold segmentation on continuous features (traffic flow, visibility) and discrete features (weather grade) respectively. For example, when the traffic flow > 120 vehicles / minute and the visibility < 100m, trigger the "high load and low visibility" branch to generate an initial weight vector; the gradient regularization dynamic optimization design fuses a composite objective function of the real-time visibility change rate gradient and historical residuals, and iteratively updates the weight parameters through online stochastic gradient descent (SGD) to suppress the weight oscillation caused by sudden environmental disturbances (such as heavy rain), ensuring the long-term policy stability; 5. Bayesian posterior probability reconstruction: For the accident level jump scenario, use the Markov chain Monte Carlo (MCMC) sampling method to generate a candidate parameter chain, and calculate the approximation of the posterior probability distribution through the Metropolis-Hastings acceptance rate to solve the parameter mismatch problem of traditional static models under extreme events.

[0061] In the actual implementation process, the light sensors of each street lamp node continuously collect the ambient brightness changes, and combine the traffic monitoring radar and camera data to convert the pedestrian density and vehicle flow into a spatial heat map. At the same time, the meteorological station data is accessed to obtain the real-time weather condition parameters. After being preprocessed by the edge computing node, the data is input into the adaptive weighted decision tree for analysis, and the dynamic priority scheduling matrix is calculated in real time.

[0062] During implementation, during peak hours, this method can determine the road lighting demand index according to the decision path, automatically enhance the brightness of the main road, and improve driving safety; during off-peak hours, it dynamically adjusts the non-essential lighting areas in combination with the historical data trend to reduce energy consumption. When an emergency occurs, such as road closure, bad weather or traffic accident, the system will trigger the Markov chain Monte Carlo sampling mechanism to quickly correct the lighting demand index to adapt to the new road condition requirements. For example, if it is detected that the pedestrian density in the accident area has increased, the system will give priority to increasing the lighting intensity of the pedestrian passage and improve the visibility of surrounding vehicles by intelligently regulating the flashing mode of warning lights. In addition, in the low-traffic state at night, this method can smoothly adjust the lighting power through the gradient constraint mechanism to avoid visual interference caused by light fluctuations, and at the same time optimize the lighting coverage to ensure the best visibility with the lowest power consumption.

[0063] Compared with the traditional fixed-rule dimming method, this method has stronger adaptability and real-time adjustment ability, can automatically optimize the lighting strategy according to environmental changes, and avoid resource waste. In case of an emergency, the system can quickly adjust the lighting strategy through Bayesian posterior update and Monte Carlo sampling to improve road safety. In addition, combined with information entropy screening, maximum mutual information coefficient calculation and gradient optimization, this method improves the accuracy and robustness of data fusion, ensures the stability of the dimming strategy in complex environments, and makes the intelligent street lamp system more energy-efficient.

[0064] In the step S2 of the above embodiment, the Gaussian membership fuzzy processing uses the Gaussian membership function to perform fuzzy mapping on the continuous state variables (such as traffic flow, visibility) in the dynamic priority scheduling matrix, and converts the continuous numerical values into discrete semantic labels (such as "high traffic", "low visibility"). By adjusting the mean and standard deviation of the Gaussian function, it adaptively matches the data distribution characteristics of different scenarios (such as the non-linear attenuation of visibility in rainy and foggy weather), and provides a semantic input space for the fuzzy rule base.

[0065] The double-network architecture of the Deep Q-Network (DQN) adopts a separate design of a target network and an online network. The target network calculates the Q-value expectation (based on the Bellman equation), and the online network updates the network weights through the temporal difference (TD) error. The double-network mechanism alleviates the training instability caused by the fluctuation of the target value in traditional Q-learning. At the same time, a lag update strategy (synchronize parameters every N steps) is introduced to avoid the oscillation problem in the policy iteration process.

[0066] When the dimension of the state space exceeds a preset threshold (such as a new unknown accident type or extreme weather), the fuzzy rule dynamic extension mechanism adds new membership functions and rule branches through the dynamic extension of the fuzzy rule base. For example, if a drone strike (not a preset scenario) is detected, a membership function of "low-altitude object approaching" is automatically generated, and fuzzy rules (such as triggering a red light strobe warning) are derived based on historical similar event data.

[0067] The prioritized experience replay mechanism prioritizes historical transition samples (state-action-reward) based on the absolute value of the TD error. Samples with high TD errors (corresponding to policy blind spots or critical states) have an increased weighted sampling probability, accelerating the learning efficiency of the model for critical scenarios. Through importance sampling weight correction, the problems of sample priority and distribution shift are balanced.

[0068] Federated learning global reward modeling aggregates the lighting energy consumption and accident response data (such as brightness deviation, emergency delay) of multi-segment edge nodes to construct a global reward function. If the brightness distribution of a certain segment deviates from the safety threshold (such as <50 Lux or >150 Lux), a penalty coefficient is superimposed; conversely, if energy conservation is achieved within the safe range, a positive reward is superimposed. The federated framework ensures data security through gradient sparsification and differential privacy techniques.

[0069] After training is completed, the ε-greedy strategy is used to select the optimal brightness distribution (ε decays with the number of training rounds) to balance exploration and exploitation. The execution results (dimming parameters and actual energy consumption) are stored in the edge node cache pool, and an offline policy iteration thread is periodically triggered to fine-tune the network parameters using historical data, forming an online-offline hybrid optimization closed-loop.

[0070] In actual implementation, in step S2, parameters such as traffic flow (0 - 200 vehicles / minute) and visibility (0 - 1000 m) in the dynamic priority scheduling matrix are first input into the Gaussian membership function to generate semantic labels such as "low traffic - high visibility" and "high traffic - low visibility". For example, when the visibility is <50 m, the standard deviation of the Gaussian function is compressed to 0.5 times the original value to enhance the semantic discrimination of low visibility. Then, the target network (delayed update) and the online network (real - time update) are deployed. The fuzzified semantic labels are input, and the brightness allocation Q - value for each road segment is output. After every 1000 steps of training, the target network parameters are synchronized, and the weights of the online network are updated by backpropagation through the TD error (the difference between the current Q - value and the target Q - value). When an undefined scenario is detected (such as visibility <10 m due to hail weather), a membership function (mean = 10 m, standard deviation = 2 m) and the corresponding rule branch (such as "extremely low visibility → brightness increased to 180 Lux") are automatically added, and the rule weights are initialized with historical similar data (such as heavy rain scenarios). Each edge node uploads local lighting data (brightness, energy consumption) to the federal center. The central node calculates the global average brightness deviation and accident response delay, and dynamically adjusts the parameters of the reward function. For example, if the global brightness deviation rate >15%, the penalty coefficient is increased to 2 times to force the model to optimize the brightness allocation accuracy. The trained DQN model outputs time - series dimming parameters (such as the brightness of a road segment gradually changing from 80 Lux to 120 Lux), which drive the PWM dimming module to execute. The execution results (actual brightness, energy consumption) are stored in the edge node cache pool. An offline batch process is triggered every 24 hours, and historical data is used to fine - tune the model parameters to reduce the online inference delay.

[0071] Compared with the prior art, this solution realizes the dynamic adaptability of the dimming strategy and the ability to handle unknown scenarios through an improved fuzzy reinforcement learning method, and solves the problem of policy rigidity of traditional fixed - rule models in complex environments; the federated learning framework and the priority experience replay mechanism improve the multi - road - segment collaborative optimization efficiency and communication reliability, and avoid the local - optimum trap caused by data deviation of a single node; the edge cache and the offline iteration mechanism further reduce the real - time computing load, and ensure the long - term operation stability and energy - saving effect of the system.

[0072] In implementation, the hardware environment for applying the improved fuzzy reinforcement learning method of this step includes a light sensor (collecting ambient light intensity), a millimeter-wave radar (monitoring vehicle flow), a thermal imaging camera (detecting pedestrian density), a meteorological monitoring station (obtaining weather parameters), an edge computing server (performing real-time data processing and dimming calculation), an intelligent street lamp control unit (executing dimming instructions), and a 5G / NB-IoT wireless communication module (transmitting control data). The hardware devices are arranged on main roads, branch road intersections, and pedestrian crossing areas. The light sensor is installed on the top of the lamp post, the millimeter-wave radar and the camera are fixed on the side of the lamp post, the meteorological monitoring station is centrally arranged at the regional center point, and the edge computing server is deployed in the urban lighting management center to achieve remote intelligent control of street lamps.

[0073] Design an experiment to compare the performance differences between the improved fuzzy reinforcement learning method (Group A) and the existing fuzzy reinforcement learning method (Group B) in the automatic emergency response control of smart street lamps. The experimental scenario is selected at the intersection of the main road and the branch road at night, and the experimental duration is 7 consecutive days, from 19:00 to 23:00 every day. Each group of methods runs five experiments, and key parameters are statistically analyzed. Group A adopts the improved fuzzy reinforcement learning method, fuzzifies the state space through the Gaussian membership function, combines the DQN double network structure to calculate the dimming strategy, and uses priority experience replay and federated learning to optimize the global reward function. Group B adopts the traditional fuzzy reinforcement learning method, uses the triangular membership function to fix the fuzzy rules, makes dimming decisions through single-network Q learning, and adopts uniform experience replay, which cannot dynamically adjust the fuzzy rules or perform cross-regional learning. The experimental data is shown in Table 1:

[0074] Table 1 Record Table of Experimental Data of the Improved Fuzzy Reinforcement Learning Method

[0075]

[0076] The experimental results show that Group A (the improved fuzzy reinforcement learning method) is significantly superior to Group B (the traditional fuzzy reinforcement learning method) in terms of energy consumption optimization, dimming strategy stability, response speed, and accident visibility improvement. Group A improves the flexibility of the dimming strategy through the dynamic fuzzy rule extension and priority experience replay mechanism, reduces the dimming fluctuation, and improves the stability of the intelligent lighting system. In addition, Group A has a fast convergence speed in the model training stage, only about 30,000 steps are required to be stable, while Group B requires more than 50,000 steps, indicating that the improved method has higher learning efficiency. Therefore, the improved fuzzy reinforcement learning method can effectively improve the intelligent level of the smart street lamp system, reduce energy consumption, and improve road safety.

[0077] In step S3 of the above embodiment, the input parsing module uses the asymmetry and flexible morphological characteristics of the Generalized Beta Distribution Function (BetaFunction) to fuzzify heterogeneous variables such as accident level, pedestrian impact area, traffic flow density, and environmental visibility. The shape parameters (α, β) of the beta function are dynamically adjusted according to the data distribution. For example, in a low visibility (<50m) scenario, the α parameter increases to enhance the sensitivity of the membership function in the low value interval, so as to more accurately map the impact of environmental mutations on the accident level.

[0078] The accident category classification module constructs a dynamic clustering space based on the fuzzy membership matrix, and divides the accident data into multiple categories (such as vehicle collisions, pedestrian falls, fires, etc.) by iteratively optimizing the objective function to minimize the within-class distance. The kernel function mapping technology is introduced to project the original data into a high-dimensional feature space, solving the problem of insufficient classification ability of traditional fuzzy C-means for non-linear data. For example, after the non-linear relationship between the smoke concentration and temperature of a fire accident is mapped by the Radial Basis Function (RBF) kernel, the clustering accuracy is increased by 25%.

[0079] The level inference module uses fuzzy information entropy to quantify the uncertainty impact of accident categories on emergency lighting parameters, and calculates the entropy weight of each accident category. Combining with the maximum membership decision rule, the accident category with the smallest entropy value (i.e., the highest certainty) is selected as the dominant factor. For example, when a fire (entropy value 0.2) and a vehicle collision (entropy value 0.5) coexist, the dimming strategy is preferentially driven by the fire accident parameters.

[0080] The anomaly correction module measures the multivariate statistical deviation degree of accident data based on the Mahalanobis Distance, and identifies signal anomalies (such as sensor drift or communication noise). The distance between the data point and the clustering center is calculated through the covariance matrix. If the distance exceeds the 95% confidence interval of the χ 2 distribution, it is determined as abnormal data and given a low weight (such as weight = 0.1) to suppress its interference with fuzzy inference. Fuzzy Gaussian regression smooths the abnormal points in the time series data, uses Gaussian Process Regression (GPR) to model the historical data distribution, and predicts the expected value under normal conditions. The time correlation of the data is described through a kernel function (such as Matern 5 / 2), and the data points that deviate from the predicted value by more than 3σ are replaced or corrected.

[0081] In specific applications, when the system detects an accident, first, the input parsing module receives traffic data from light sensors, radars, and cameras, and performs fuzzification using the membership function modeling of the generalized beta function. It fuzzifies the traffic flow density (0 - 200 vehicles per minute), sets α = 2 and β = 5, and generates membership labels of "low density", "medium density", and "high density". For example, in an environment with low visibility at night, the system is more likely to classify moderately severe accidents as high - risk events to increase the response priority of warning lights. Then, the accident category classification module classifies the accident data using the fuzzy C - means clustering algorithm, divides the accident types into minor, medium, and severe, and generates an accident category fuzzy mapping matrix. For example, a small collision involving two vehicles may be classified as a minor accident, while a multi - vehicle chain collision may be classified as a severe accident and affect a larger - scale lighting adjustment. Subsequently, the level reasoning module calculates the influence weight of the current accident category on lighting parameters based on historical accident data and real - time sensing information, and determines the final emergency level according to the maximum membership degree decision method. For example, on an urban arterial road with a high pedestrian density, even if the severity of the accident is low, the system may still increase the lighting level to reduce potential safety hazards. To ensure data accuracy, the anomaly correction module checks the data quality and corrects abnormal data. For example, if the pedestrian flow detected by the radar suddenly increases but no pedestrians are detected by the camera, the system will use Mahalanobis distance weighted filtering to check whether the data is abnormal and analyze the historical trend using fuzzy Gaussian regression to determine whether data input needs to be adjusted. Finally, the lighting parameter output module calculates the optimal lighting intensity, light diffusion range, and warning light trigger conditions for the accident area and sends the results to the intelligent street lamp control system. For example, after an accident occurs on a highway, the system may turn on high - brightness lighting within a range of 1 kilometer in front of the accident point and activate the color - changing warning lights, while on an urban road, it may adjust the lighting enhancement in the crosswalk area to ensure pedestrian safety.

[0082] Compared with the traditional accident lighting control method with fixed rules, this method has stronger adaptability and intelligence. The fuzzification process based on the generalized beta function modeling improves data adaptability, enabling the system to dynamically adjust lighting strategies; the fuzzy C - means clustering enhances the accuracy of accident classification, making the lighting response more reasonable; optimizing the weight through fuzzy information entropy improves the stability of the system in an uncertain environment; combining Mahalanobis distance weighted filtering and fuzzy Gaussian regression improves data reliability and reduces the influence of sensor noise. In addition, the lighting adjustment strategy based on the rule - based inference engine enables intelligent street lamps to perform personalized lighting control for different types of accidents, improving road safety and energy - saving effects.

[0083] In step S4 of the above embodiment, the brightness membership distribution is modeled using a bimodal Gaussian function to match the different lighting requirements of the accident area and its transition area. The traditional unimodal Gaussian function is difficult to cover both the high brightness requirement of the accident core area and the gradual brightness transition of the peripheral buffer area simultaneously. Therefore, the bimodal Gaussian model controls the light intensity of different areas through two independent peaks. The first peak is used to match the high brightness warning requirement of the accident core area, while the second peak is used to adjust the buffer brightness of the peripheral area, making the lighting distribution smoother and reducing the visual discomfort caused by sudden light changes. This method can rapidly enhance local lighting in case of emergency, while providing a reasonable brightness attenuation curve for the transition area and optimizing the overall visual experience.

[0084] Secondly, the brightness gradient optimization uses a non-linear gradient function, taking the accident influence radius and accident level as input parameters to calculate the optimal brightness adjustment range. In practical applications, the severity of an accident and its influence range often do not show a linear relationship. Therefore, through a non-linear mapping function, it can be ensured that the brightness gradient adjustment amplitude is larger in the case of a higher accident level, while the brightness change of a lower-level accident is smoother. In addition, this function introduces a progressive smoothing term to suppress sudden brightness changes in an exponential decay manner, making the light change more stable and improving the safety of driving and pedestrians. To enhance environmental adaptability, this gradient optimization mechanism also sets a low visibility threshold trigger condition. When the visibility is lower than a certain value, the system will forcefully reduce the gradient amplitude to prevent further reduction of visibility caused by excessive light adjustment in extreme weather.

[0085] Thirdly, the multi-objective optimization of the lighting control strategy uses a Pareto optimization matrix and solves the non-dominated solution set through Particle Swarm Optimization (PSO). This method constructs a safety matrix, an energy consumption weight matrix, and a response delay penalty matrix to minimize energy consumption while meeting safety requirements, and at the same time controls the response time to make the lighting adjustment more intelligent and real-time. Among them, the Euclidean distance weighted by the accident level is used to measure the deviation of the candidate solution from the ideal lighting control scheme, ensuring that the final solution can not only meet the lighting requirements of the accident area but also avoid energy consumption waste caused by over-illumination.

[0086] Finally, the dynamic threshold triggering mechanism is used to determine in real time whether to activate the warning lights and the light enhancement mode. This method combines factors such as traffic flow, day-night light fluctuations, and visibility change rates to calculate an adaptive trigger threshold, and dynamically adjusts the warning light activation conditions according to the ratio of the real-time traffic flow to the maximum road carrying capacity. For example, when the traffic flow is high and the visibility is lower than the safety threshold, the system will automatically trigger the warning lights and increase the brightness; when the visibility gradually recovers, the system iteratively corrects the lighting parameters based on the gradient descent algorithm to ensure that the light control is always in the optimal state. In addition, the system performs closed-loop feedback through a light intensity sensor to detect the regional brightness deviation in real time, and adopts a gradient descent optimization strategy to correct the brightness output, avoiding uneven lighting caused by over-adjustment.

[0087] In the actual implementation process, when an accident occurs, the system first receives data such as the accident level, pedestrian density, traffic flow, and weather conditions, and uses a bimodal Gaussian function model to construct a brightness membership distribution to determine the lighting levels in the core high-brightness area and the peripheral buffer area of the accident area. For example, in a major highway accident, the system will provide high-brightness warning lighting for the core accident area and gradually changing lighting in the accident extension area to enhance visibility. For a minor collision accident on an urban road, the system may only enhance the lighting in the sidewalk area to reduce the energy consumption of over-illumination. Then, the system uses a non-linear gradient optimization method to calculate the optimal brightness adjustment range, and dynamically adjusts the gradient amplification coefficient in combination with the low visibility adaptive strategy. For example, in a foggy night, the system will reduce the rate of brightness adjustment to avoid a decrease in visibility caused by overly strong light; in a sunny accident environment, the system may adopt a steeper brightness gradient to quickly enhance the lighting level in the accident area and improve safety. Subsequently, the system uses the Pareto optimization matrix to solve the non-dominated solution set based on the multi-objective particle swarm optimization (PSO) algorithm to balance safety, energy consumption, and response time. For example, in the first few minutes after an accident occurs, the system may prioritize increasing the lighting intensity to improve visibility, and after the accident is handled, the system will gradually reduce the brightness to reduce unnecessary energy consumption. Finally, the system adjusts the status of the warning lights in real time through the dynamic threshold triggering mechanism, and uses a light intensity sensor closed-loop feedback mechanism to correct the brightness. For example, when the traffic flow is high, the system may trigger the warning lights in advance at a lower visibility threshold, while when the traffic flow is low, it may trigger them appropriately later to ensure that the warning lights are only lit when necessary. In addition, the system continuously detects the regional light deviation. If the light error exceeds the preset threshold, the brightness parameters are corrected through gradient descent optimization to ensure that the overall light adjustment meets the expected requirements.

[0088] Compared with the traditional fixed - brightness emergency lighting solution, the multi - level fuzzy control method can adaptively adjust the lighting strategy according to the accident level, environmental visibility, and traffic flow, improving the intelligent level of lighting control. The bimodal Gaussian brightness distribution improves the lighting coordination between the accident core area and the buffer area, reducing the visual interference caused by sudden changes in light; the non - linear gradient optimization method ensures the smoothness of brightness adjustment, avoiding the visual adaptation difficulties caused by sudden increases in light intensity; the Pareto solution method based on multi - objective particle swarm optimization realizes energy consumption optimization while ensuring safety, improving energy utilization efficiency; the dynamic threshold trigger mechanism improves the intelligence of warning light control, enabling it to adaptively adjust according to the actual road conditions and improving the accuracy of accident response.

[0089] When implemented, the hardware components of the multi - level fuzzy control method include: millimeter - wave radar (77GHz, detection range 80m): installed on the top of the lamp post, horizontally covering the main road and sidewalk; multi - spectral camera (visible light + infrared, resolution 1920×1080): at a 20° downward angle in the middle of the lamp post, monitoring pedestrians and vehicle flow; edge computing unit (NVIDIA Jetson Orin, 64GB memory): embedded in the lamp post control box, running the multi - level fuzzy control algorithm; environmental sensor cluster (visibility, temperature, humidity, PM2.5): distributed at different heights of the lamp post, with a vertical spacing of 1.5m; adjustable - color - temperature LED module (brightness 0 - 200Lux, color temperature 3000K - 6500K): integrated in the lamp head, controlled by a PWM driver board (frequency 2kHz).

[0090] Based on the above hardware, two experimental groups are designed, including:

[0091] Group A: Apply the multi - level fuzzy control method (bimodal brightness distribution, non - linear gradient optimization, dynamic threshold trigger);

[0092] Group B: Adopt the traditional single - level fuzzy control method (fixed Gaussian membership, linear gradient adjustment, static threshold).

[0093] By simulating 4 typical scenarios (haze visibility 30m + Level 3 accident, heavy rain + Level 4 accident, low traffic flow at night + Level 2 accident, high traffic flow + sudden pedestrian aggregation). Deploy the Group A / B hardware on the same road section, with the environmental parameters (visibility, accident level, traffic flow) being strictly the same. Each group conducts 5 repeated experiments, and the data collection period is 1 hour. The experimental data records are shown in Table 2:

[0094] Table 2 Experimental record table of the application of the multi - level fuzzy control method

[0095]

[0096] As can be seen from Data Table 2, the multi-level fuzzy control method (Group A) is significantly superior to the traditional single-layer method (Group B) in terms of brightness uniformity, response speed, energy consumption efficiency, and accident coverage rate. Its bimodal brightness distribution optimizes the lighting allocation between the core area and the transition area, with the brightness uniformity increased by 35%; the non-linear gradient function reduces the brightness mutation, and the gradient smoothness is improved by 56%; the dynamic threshold trigger mechanism accurately identifies high-traffic and low-visibility scenarios, with the false trigger rate reduced by 89%. Experiments have proved that this method effectively solves the problems of uneven brightness, response lag, and energy consumption waste of traditional methods in complex environments, and realizes efficient and stable emergency lighting control.

[0097] In step S5 of the above embodiment, since the data volume of the lighting control instruction varies greatly under different circumstances, for example, the data of ordinary dimming instructions is less, while the control data in the accident response mode is more, a fixed frame length may lead to bandwidth waste or increased overhead due to packet splitting. The variable-length frame structure encoding can dynamically adjust the packet length to adapt to different application scenarios and improve the transmission efficiency. The variable-length frame structure encoding and error correction mechanism dynamically encapsulate the instruction packets using the variable-length frame structure (VLF), and allocate the frame header length according to the data priority (accident instruction > dimming instruction) to improve the transmission efficiency of high-priority instructions. The forward error correction coding (FEC) generates parity bits through bit-level redundant check (such as Reed-Solomon code), calculates the minimum code distance in combination with the Hamming distance optimization algorithm, and reduces the bit error rate (BER). When the channel noise power spectral density (PSD) exceeds the threshold, the redundancy ratio of the erasure code (such as switching from 1:2 to 1:3) is adaptively adjusted to enhance the anti-burst interference ability.

[0098] In terms of channel selection, the multi-scale channel state awareness method uses a long short-term memory network (LSTM) combined with empirical Bayesian channel estimation to calculate the instantaneous channel gain of different wireless channels. The LSTM is used to analyze historical channel data and extract long-term trends, while the empirical Bayesian channel estimation is used to dynamically calculate the probability distribution of the current channel state, thereby obtaining the predicted value of the channel gain. On this basis, the system adopts a Markov channel evaluation method to construct a discrete-time Markov chain (DTMC) model, mapping the channel state (gain, bit error rate, interference level) into a state transition probability matrix. The selection probability of each channel is calculated through a weighted historical state fusion function, and the channel with the highest probability is preferentially selected. If the channel fluctuation amplitude (such as gain standard deviation > 3dB) exceeds the limit, a reinforcement learning channel allocation strategy is triggered. When the channel state fluctuates greatly or the bandwidth resources are tight, the channel switching action space is defined, and the action value function Q(s,a) is updated through the Q-learning algorithm. The immediate reward R is calculated based on the channel throughput (bps) and the bit error rate, and the discount factor balances the short-term gain and long-term stability. The optimal policy selects the channel combination that maximizes the cumulative reward. Finally, the bit error rate of different channels is calculated through Bayesian bit error rate estimation, and the data packet retransmission strategy is optimized based on the bit error rate. For example, in the case of a high bit error rate on a certain channel, the system will reduce the data packet allocation on this channel and increase the transmission weight of other channels at the same time to reduce the overall data packet loss rate.

[0099] In specific applications, the edge computing unit divides the lighting scheduling plan into variable-length data frames according to priorities. The accident instruction uses an 8-byte frame header and Reed-Solomon (255,223) coding to generate a 32-byte check bit, and the dimming instruction uses a 4-byte frame header and RS(255,191) coding to improve the anti-noise ability. The LSTM network inputs the received power sequence of the past 10 seconds (sampling rate 100Hz), predicts the channel gain in the next 1 second and dynamically adjusts the smoothing coefficient α. In the area with frequent vehicle occlusion (Doppler frequency shift > 50Hz), α is reduced from 0.6 to 0.4 to enhance the instantaneous response. The Markov model calculates the selection probabilities of 5G, LoRa, and NB-IoT channels in real time, and preferentially activates the channels with a probability > 0.65; when the 5G channel gain drops by 30%, the Q-learning strategy switches to the LoRa spread spectrum mode (SF = 12), and the reward function calculates the immediate benefit by synthesizing the throughput (2kbps) and the bit error rate (1e-4). The multi-path load balancer distributes data packets to the primary and backup channels for parallel transmission. The Bayesian bit error rate estimation module dynamically calculates the retransmission interval according to the ACK feedback, and adopts an exponential backoff strategy (initial 1 second, maximum 16 seconds) to balance the delay and reliability. The successfully transmitted instruction drives the PWM dimming module to perform brightness adjustment, and the edge node records the transmission performance data and feeds it back to the closed loop of channel selection and retransmission strategy optimization.

[0100] Compared with the prior art, this solution significantly improves the communication reliability in complex environments through dynamic coding and intelligent channel decision-making, solving the problems of high bit error rate and weak anti-interference caused by the rigidity of traditional fixed coding redundancy and single-channel dependence; the collaborative optimization of channel switching and retransmission mechanisms by reinforcement learning and Bayesian estimation reduces the transmission delay and energy consumption in high-density occlusion areas; multi-scale perception and load balancing technologies effectively cope with mobile multipath effects and burst interference, ensuring the real-time and stable transmission of emergency control signals, and providing core communication guarantee for the efficient response of the intelligent street lamp system in extreme scenarios.

[0101] To verify the effectiveness of the multi-scale channel state perception method in complex interference environments, a hardware platform was built based on the actual deployment requirements of intelligent street lamps, including a multi-mode communication module (LoRa / NB-IoT / 5G), a millimeter-wave radar (60 GHz), an edge computing unit (NVIDIA Jetson AGX Orin), a signal power sensor, and a steerable antenna array. The communication module and the antenna array were deployed on the top of the lamp post (height 4 m), and the radar and the power sensor were distributed and mounted to monitor the multipath signal strength and moving obstacles. In Experiment Set A, the multi-scale channel perception method was used to dynamically select channels and optimize load balancing; in Experiment Set B, a fixed polling channel switching (LoRa→NB-IoT→5G cycle) and a static redundancy strategy (FEC r = 0.25) were used. The hardware of the two groups was deployed on the same road section. In the simulated scenarios of high-rise building occlusion multipath fading (delay spread 10 - 50 ns), vehicle movement Doppler frequency shift (0 - 100 Hz), and heavy rain electromagnetic interference, each group conducted five independent experiments, each lasting 30 minutes and recording key communication metrics: The experimental data are shown in Table 3:

[0102] Table 3 Record Table of Comparative Experimental Data of Multi-scale Channel State Perception Method

[0103]

[0104] As can be seen from Data Table 3, in Experiment Set A under the multi-scale channel state perception method, the bit error rate is stable at the order of magnitude of 10 -5 ~10 -6 , compared with Experiment Set B (10 -3 ~10 -4The (magnitude) is reduced by 1 - 2 orders of magnitude; the transmission delay is controlled within 80 - 92 ms (more than 300 ms on average in Group B), and the real-time performance is improved by 70%; the number of channel switches is significantly reduced to 2 - 4 times per hour (more than 10 times in Group B), and the energy consumption waste caused by invalid switches is reduced; the throughput (5.6 - 6.1 Mbps) and energy consumption efficiency (17.8 - 19.2 bit / J) are increased by 2 - 3 times compared with Group B. Experiments prove that the dynamic channel prediction and intelligent switching strategy can effectively solve the problems of high bit error rate, large delay, and low energy consumption efficiency of traditional solutions in high-interference environments, providing high-reliability communication guarantee for the emergency response of smart street lights.

[0105] In step S6 of the above embodiment, the time synchronization mechanism based on the IEEE 1588 PTP protocol realizes high-precision timing alignment of multiple nodes through the master-slave clock architecture. The master and slave clocks record the exact moments of signal transmission and reception through two-way timestamp exchange (Sync, Delay_Req messages), separate the transmission delay and processing delay using the round-trip delay (T1 - T4 timestamps), calculate the initial clock offset, and eliminate the influence of network jitter on the synchronization accuracy. Based on the least squares linear regression method, the historical clock deviation data is fitted to construct a long-term trend model of the clock drift rate, quantifying the inherent deviation of the crystal oscillator frequency; the Kalman filter further estimates the short-term clock error (such as the instantaneous frequency offset caused by temperature fluctuations), and iteratively updates the error covariance matrix through the state equation (clock phase error) and the observation equation (measured offset) to generate short-term compensation parameters. When the synchronization error exceeds the set threshold (e.g., > 100 ns), the variational Bayesian inference method is introduced, modeling the clock drift as a Gaussian mixture distribution, and approximating the posterior distribution parameters by maximizing the evidence lower bound (ELBO) to solve the problem of insufficient adaptability of the traditional single-peak assumption to complex drift patterns; combining the particle filter method for Monte Carlo sampling of the drift rate, screening high-probability particles based on weights in the resampling stage, and dynamically adjusting the clock frequency compensation step size. Finally, the exponential weighted moving average (EWMA) is used to smooth the clock adjustment amount, giving higher weights to recent compensation data, suppressing the interference of high-frequency noise on the global synchronization stability, and outputting the timestamp synchronization control instruction to ensure the nanosecond-level timing consistency of the multi-node dimming signal.

[0106] In specific applications, the master clock node is deployed on the central lamp post of the road section, equipped with a high-stability atomic clock (accuracy ±10 ppb), and the slave clock nodes are synchronized with the master clock through the 5G backhaul link. The edge computing unit runs the PTP protocol stack, exchanges Sync and Delay_Req messages every 100 ms, records the T1 (master transmission), T2 (slave reception), T3 (slave transmission), and T4 (master reception) timestamps, calculates the one-way transmission delay D = (T2 - T1 + T4 - T3) / 2, and fits the clock drift rate β within 24 hours based on the least squares method (such as 1.2×10 -6)。The Kalman filter takes the real-time offset (T2 - T1 - D) as input, updates the phase error estimate through a prediction-correction loop, and outputs the compensation amount every 10 ms (such as +5 ns). When the synchronization error of a slave node exceeds 100 ns continuously for 3 times, the variational Bayesian inference module is triggered. Assuming that the clock drift follows a two-component Gaussian mixture distribution, the mean and variance parameters are iteratively optimized through the EM algorithm (μ1 = 1.5×10 -6 , σ1 = 0.3×10 -6 ; μ2 = -0.8×10 -6 , σ2 = 0.2×10 -6 ), and 1000 drift rate samples are generated using particle filtering. Particles with weights > 0.01 are selected to update the compensation step (such as adjusting from +5 ns to +8 ns). The EWMA module performs a weighted average on the last 10 compensation amounts with a decay factor α = 0.2, and outputs the final adjustment instruction after smoothing, driving the PWM controller to synchronize the phase of the dimming signal. The synchronization error data of each node is uploaded to the cloud every 30 minutes to optimize the global EWMA parameters and the variational Bayesian prior distribution, forming a closed-loop adaptive optimization link.

[0107] Compared with the prior art, this solution significantly improves the time synchronization accuracy in a complex network environment through a two-way time delay separation and dynamic error compensation mechanism; the collaborative optimization of variational Bayesian inference and particle filtering effectively solves the problem of insufficient adaptability of traditional single models to complex drift patterns, enhancing the system's resistance to temperature and electromagnetic interference; the exponentially weighted moving average suppresses high-frequency noise, ensuring the stability of long-term operation, providing a highly reliable timing reference for multi-node collaborative dimming, and avoiding the cumulative error and synchronization failure risks of traditional fixed compensation strategies.

[0108] In step S7 of the above embodiments, the multi-objective reinforcement learning and sparse Bayesian optimization method realizes the collaborative optimization of energy consumption - performance - lifespan of lighting power through dynamic policy trade-off and high-dimensional parameter search. Based on the Deep Deterministic Policy Gradient (DDPG) algorithm, a multi-objective joint reward function is constructed, mapping energy consumption (power consumption), brightness (target Lux value), and lamp lifespan (cumulative working duration) into a multi-dimensional reward vector. Through a dynamic weight allocation mechanism, the weight ratios of each objective are adjusted in real time (for example, the energy consumption weight is dynamically adjusted from 0.5 to 0.3 to respond to emergencies), generating a Q-value matrix to quantify the long-term benefits of different dimming strategies. The sparse Bayesian optimization method sparsely samples the high-dimensional policy parameter space (such as brightness gradient, dimming frequency, duty cycle), uses Gaussian Process Regression (GPR) to model the non-linear relationship between power adjustment amount and energy consumption benefits, and describes the long-range correlation between parameters through a kernel function (such as Matern 5 / 2); if the predicted benefit variance exceeds a threshold (such as σ2>0.1), Monte Carlo sampling is triggered to perform local optimization of policy parameters under confidence interval (such as 95%) constraints, preferentially exploring regions with high benefits and low variances. Integrating the real-time energy consumption data of multi-channel transmission in S5 (such as channel switching energy consumption) and the clock synchronization deviation compensation amount in S6 (such as dimming delay caused by phase error), a PID fuzzy controller input vector is constructed, and the PID parameters (proportional, integral, and differential coefficients) are dynamically adjusted through a fuzzy rule base (such as "IF synchronization deviation > 50ns THEN integral coefficient increases"), and a duty cycle control signal is output to balance power accuracy and response speed. The experience replay buffer stores historical state (environmental parameters, clock deviation), action (dimming instruction), and reward (energy consumption benefit) tuples, calculates the policy network gradient through the Temporal Difference (TD) error, and updates the network weights by backpropagation using an Adaptive Momentum Optimizer (Adam), suppressing policy oscillation and accelerating convergence.

[0109] In a specific application, an edge computing unit (such as NVIDIA Jetson AGX Xavier) runs the DDPG algorithm, inputs the real-time channel energy consumption of S5 (such as the power consumption of the 5G module is 2.5W) and the clock synchronization deviation of S6 (such as ±20ns), and initializes the weights of the multi-objective reward function (energy consumption 0.6, brightness 0.3, lifespan 0.1). Gaussian process regression sparsely samples the dimming strategy parameters (brightness 80 - 200Lux, frequency 1 - 5Hz) (10 points per parameter dimension) to generate the initial policy return surface; when the predicted variance exceeds the limit in the heavy rain scenario, Monte Carlo sampling generates 100 groups of candidate parameters in the local space (brightness 150 - 180Lux, frequency 2 - 3Hz), and filters the optimal solution within the confidence interval (brightness 170Lux, frequency 2.5Hz). The PID fuzzy controller receives the synchronization deviation data. If the deviation > 50ns, the fuzzy rule triggers the integral coefficient to increase from 0.1 to 0.3 to accelerate the elimination of the cumulative error, and outputs a PWM duty cycle signal (such as 65%) to drive the LED module. The experience replay buffer stores a set of historical data every 5 minutes (capacity 1000 sets). After calculating the TD error, the policy network is updated through the Adam optimizer (learning rate 1e-4) to suppress policy mutations. The global policy parameters are synchronized to the cloud every 24 hours, and the prior distribution is optimized through the federated learning framework to form a cross-node collaborative tuning closed-loop.

[0110] Compared with the prior art, this solution realizes the fine balance of energy consumption, brightness, and device lifespan through multi-objective dynamic weight allocation and high-dimensional parameter search, and solves the problem of unbalanced resource allocation caused by traditional single-objective optimization; the collaborative improvement of sparse Bayesian optimization and Monte Carlo sampling improves the policy search efficiency and avoids the curse of dimensionality; the fusion of PID fuzzy control and clock deviation enhances the robustness of power adjustment under complex interference, providing the energy-saving emergency response ability with high reliability and long lifespan for intelligent street lights.

[0111] In an embodiment, an energy-saving intelligent street light automatic emergency response system includes: an environmental perception module, a dynamic priority calculation module, a dimming strategy optimization module, an emergency response evaluation module, a lighting scheduling control module, a data transmission module, and a remote management module.

[0112] As Figure 6As shown in the figure, the environmental perception module collects traffic flow, pedestrian density, accident signals, and ambient light intensity data in real time through a light sensor, millimeter-wave radar, infrared sensor, and camera. After preprocessing, the data is transmitted to the dynamic priority calculation module, which generates a dynamic priority scheduling matrix (such as accident core area weight 0.7, traffic flow weight 0.2) by fusing time series trend analysis based on the adaptive weighted decision tree algorithm, and outputs it to the dimming strategy optimization module; the dimming strategy optimization module uses improved fuzzy reinforcement learning combined with a deep Q network, takes the scheduling matrix and historical energy consumption data as inputs, trains to generate a brightness allocation strategy, and transfers the preliminary parameters to the emergency response evaluation module; the emergency response evaluation module calculates the lighting response parameters corresponding to the accident level (such as red light strobing triggered by a Level 3 accident) through a fuzzy inference system, and outputs it to the lighting scheduling control module for multi-level fuzzy control to generate specific dimming instructions (brightness gradient, warning mode); the lighting scheduling control module sends the instructions to the data transmission module, which dynamically selects the LoRa / NB-IoT or 5G channel through multi-scale channel perception, synchronizes the time series of each node in combination with the Bayesian clock drift correction model, and finally transmits the instructions to the street lamp terminal for execution; the remote management module receives the energy consumption, delay, and execution deviation data fed back by each node, online updates the policy parameters through multi-objective reinforcement learning and sparse Bayesian optimization, and injects the optimization results back into the dimming strategy optimization module and the dynamic priority calculation module to form a global closed-loop of "perception - decision - execution - optimization".

[0113] In a specific implementation, the system is simulated through the finite element simulation method: when the environmental perception module detects through the camera and millimeter-wave radar that the traffic flow density in the accident area (coordinates X: 235, Y: 189) drops suddenly by 80% and the infrared sensor identifies 5 pedestrians staying, the dynamic priority calculation module immediately assigns the highest priority (0.92) to this area, triggering the emergency response evaluation module to start. The fuzzy inference system determines it as a level L3 accident by combining the accident area (200 ㎡), visibility (50 meters), and pedestrian density (0.4 people / ㎡), and outputs the emergency lighting parameters (intensity 1200 lux, coverage radius 80 meters, duration 40 minutes). The dimming strategy optimization module synchronously calculates the brightness attenuation strategy for the surrounding area (the brightness decreases by 15% every 10 meters from the accident center), and after fusing with the emergency parameters, inputs them into the lighting scheduling control module. The latter sets the brightness to gradually increase from 800 lux to 1200 lux within 15 seconds (change rate 26.7 lux / s) through multi-level fuzzy control, activates the red rotating warning light (3 Hz stroboscopic), and broadcasts the detour instruction through the NB-IoT channel. The data transmission module switches to the LoRa + 5G dual-path redundant transmission after real-time monitoring of 5G channel congestion (packet loss rate > 12%). The Bayesian clock synchronization controls the execution time deviation of each node within ±1.5 ms, ensuring that 56 street lights within 80 meters execute dimming synchronously. The remote management module records the peak energy consumption (28% increase compared to the baseline) and traffic guidance efficiency (the accident handling is shortened by 22 minutes) in this event, adjusts the upper limit of the brightness gradient to 20 lux / s through particle swarm optimization to reduce the glare risk, and pushes the updated strategy to all network nodes. Finally, the accident area achieves a high-brightness warning without glare within 15 seconds. The surrounding traffic flow is diverted orderly through the brightness gradient guidance. The total energy consumption of the system is reduced by 34% compared to the traditional full-brightness mode, and no secondary accidents are caused.

[0114] Although the specific implementation manners of the present invention are described above, those skilled in the art should understand that these specific implementation manners are only illustrative. Without departing from the principle and essence of the present invention, those skilled in the art can make various omissions, substitutions, and changes to the details of the above methods and systems. For example, combining the above method steps, thus performing substantially the same functions in a substantially the same method to achieve substantially the same results shall fall within the scope of the present invention. Therefore, the scope of the present invention is only defined by the appended claims.

Claims

1. An energy-saving intelligent street lamp automatic emergency response control method; characterized in that: Including: S1. Collect data from light sensors, radars, cameras, and environmental monitoring devices, calculate the comprehensive weights of traffic flow, pedestrian density, weather conditions, and accident levels through an adaptive weighted decision tree, and calculate the lighting demand index in combination with historical stored data, and output a dynamic priority scheduling matrix; S2. Input the dynamic priority scheduling matrix into a deep Q network, train an intelligent dimming strategy based on an improved fuzzy reinforcement learning method, calculate the optimal brightness allocation for different road sections, and obtain time-series lighting control parameters; S3. Input the time-series lighting control parameters into a fuzzy inference system, conduct an emergency level assessment in combination with accident detection data, and set fuzzy membership functions according to different accident categories to obtain emergency lighting parameters; S4. Based on the emergency lighting parameters, use a multi-level fuzzy control method to calculate the light adjustment range, brightness change gradient, and warning light trigger conditions, and output a light scheduling plan; S5. Convert the light scheduling plan into an instruction data packet, perform data encapsulation using a multiple access wireless protocol, and dynamically select LoRa, NB-IoT, or 5G channels for transmission based on channel state evaluation through a multi-scale channel state awareness method, and output a light control signal; S6. Use the IEEE 1588PTP protocol to synchronize the timing consistency of light control signals between each smart street lamp node, and use an improved Bayesian clock drift correction model for error compensation, correct the time synchronization deviation, and output a timestamp synchronization control instruction; S7. Based on the timestamp synchronization control instruction, predict the energy consumption benefits of different dimming strategies through multi-objective reinforcement learning and sparse Bayesian optimization methods, and dynamically adjust the lighting power.

2. The automatic emergency response control method for an energy-saving intelligent street lamp according to claim 1, characterized in that: The working principle of the adaptive weighted decision tree is as follows: First, calculate the difference degree between the time-series data of the light sensor and the historical accident distribution based on relative entropy, and screen the feature dimensions with entropy offset exceeding the limit through the dynamic threshold within the sliding window; and analyze the spatial correlation between the radar point cloud and the pedestrian density heat map through the maximum mutual information coefficient to generate a non-linear correlation coefficient matrix; then, use the C4.5 algorithm to construct a multi-branch decision tree, divide the nodes through the dual thresholds of information gain ratio and Gini index, map the traffic flow time-series data, pedestrian density heat map, and weather condition quantization matrix to different decision paths respectively, and output an initial weight vector; at the same time, use the online gradient descent algorithm to use the visibility change rate gradient as a regularization term, combine the light fluctuation variance and historical residuals to construct a composite objective function, and iteratively update the weight parameters; if a sudden accident level jump is detected, reconstruct the weight posterior probability distribution through the Markov chain Monte Carlo sampling method, and output the lighting demand index.

3. An automatic emergency response control method for an energy-saving intelligent street lamp according to claim 1, characterized in that: The working method of the improved fuzzy reinforcement learning method is as follows: First, fuzzify the continuous state of the dynamic priority scheduling matrix into discrete semantic variables through a Gaussian membership function, and construct a fuzzy rule base to generate an initial dimming strategy; then, generate a target network and an online network based on the deep Q network, calculate the Q-value expectation through the target network, and the online network updates the weights through the time difference error. If the state space dimension exceeds the preset threshold, a membership function is dynamically added through a fuzzy rule expansion mechanism to cover unknown scenarios; During the training process, the improved fuzzy reinforcement learning method uses a prioritized experience replay mechanism to perform weighted sampling on historical transition samples according to the absolute value of the TD error, and aggregates multi-segment lighting energy consumption and accident response data through a federated learning framework to construct a global reward function; After training, the improved fuzzy reinforcement learning method selects the optimal brightness allocation based on a greedy strategy, outputs time series control parameters to drive the PWM dimming module, and stores the execution results in the edge node cache pool.

4. An automatic emergency response control method for an energy-saving intelligent street lamp according to claim 1, characterized in that: The fuzzy inference system includes an input parsing module, an accident category classification module, a level inference module, an anomaly correction module, and a lighting parameter output module; the input parsing module is used to perform fuzzy processing on variables such as accident level, pedestrian influence area, traffic flow density, and environmental visibility using generalized beta function membership modeling; the accident category classification module is used to classify the fuzzified accident data through fuzzy C-means clustering and output an accident category fuzzy mapping matrix; the level inference module is used to calculate the influence weight of the accident category on the emergency lighting parameters through a fuzzy information entropy optimization method and determine the accident emergency level through the maximum membership degree decision method; the anomaly correction module is used to identify invalid or abnormal data through Mahalanobis distance weighted filtering when there are signal anomalies, data drift, or acquisition errors in the accident data, and perform time series data smoothing in combination with fuzzy Gaussian regression; the lighting parameter output module is used to generate lighting adjustment parameters through a rule inference engine and output the emergency lighting intensity, lighting diffusion range, and response duration of the accident occurrence area.

5. The automatic emergency response control method for an energy-saving intelligent street lamp according to claim 1, characterized in that: The working method of the multi-level fuzzy control method is as follows: U1. Based on the accident level E and visibility V, construct a brightness membership distribution through a double-peak Gaussian function, and the formula expression is: In formula (1), L represents the real-time brightness value in lux, which is used to quantify the current light intensity; 120 represents the brightness center value of the first peak, corresponding to the Gaussian function standard deviation a1 = 15 (V / 100) 0.5 , which is used to match the high-brightness warning requirements in the accident core area; 160 represents the brightness center value of the second peak, corresponding to a2 = 25 (V / 100) 0.3 , which is used to cover the gradually decreasing brightness distribution in the transition area; a1 and a2 respectively represent the widths of the first peak and the second peak, which are used to control the brightness membership diffusion range; U2. Based on the accident impact radius R and accident level E, extract the optimal brightness adjustment range ΔL of the current state through a non-linear gradient function, and the formula expression is: In formula (2), represents the reference gradient term, which is used to define the maximum allowable brightness change rate, where L max represents the maximum allowable brightness, and L d represents the target brightness requirement; represents the accident coupling term, which is used to nonlinearly map the accident level and the influence radius into a gradient amplification coefficient; (1 - e -t / 10 ) represents the progressive smoothing term, which is used to suppress the brightness mutation at the initial moment through an exponential function. t represents time, with the unit of seconds; 30, 200, and 10 are empirical constants, which are used to control the reference gradient, the accident coupling strength, and the smoothing speed respectively; if the detected visibility V < 50 meters, the forced gradient amplitude decays to ΔL(t)max(0.5, V / 50), and a smoothed brightness adjustment instruction is output; U3. Construct the Pareto optimization matrix P = [S w , E w , D|] T , where the safety matrix S w = E ρ / A, ρ is the pedestrian density, A is the accident area; the energy consumption weight t is the response time; D is the response delay penalty; the security matrix S w , the energy consumption weight E w and the response delay penalty D are solved by multi-objective particle swarm optimization min(||P - [1, 0, 0] T ||w), and the non-dominated solution set is output, where w = [0.6, 0.3, 0.1]E, representing the Euclidean distance weighted by the accident level, used to measure the deviation of the solution from the ideal point [1, 0, 0|] T . Then, select the actual light intensity L act = L d (1 + 0.15sigmoid(S w - E w )); U4. Calculate the trigger condition in real time through a dynamic threshold function, and the formula expression of the dynamic threshold function is: θ(t) = 0.5 + 0.2sin(2πt / 360) + 0.1T f / C max (3) In formula (3), θ(t) is the dynamic trigger threshold; 0.5 is the baseline threshold; 0.2sin(2πt / 360) is the periodic term, and the period is 360 seconds, which is used to simulate the natural fluctuation of day and night visibility; 0.1T f / C max is the traffic flow correction term, where T f is the real-time traffic flow, and C max is the maximum carrying capacity of the road section; if triggers the rotating warning light and increases the brightness to 1.2L d , otherwise maintain the gradient adjustment strategy; at the same time, feedback data through the light intensity sensor; where represents the visibility change rate; If the regional brightness deviation exceeds the preset threshold, iterative correction is performed through gradient descent until the error tolerance is met.

6. A method for automatically controlling the emergency response of an energy-saving intelligent street lamp according to claim 1, characterized in that: The multiple access wireless protocol is encoded through a variable-length frame structure, and forward error correction coding is used to perform bit-level redundant verification on the instruction data packet, optimize the calculation of the data bit error rate using the Hamming distance, and adjust the redundancy correction ratio in combination with erasure coding.

7. An automatic emergency response control method for an energy-saving intelligent street lamp according to claim 1, characterized in that: The multi-scale channel state awareness method calculates the instantaneous channel gain G of different wireless channels through a long short-term memory network combined with empirical Bayesian channel estimation c , and the calculation formula is as follows: G c G(t) = τP r G(t)+(1 + τ)G(t - 1) c (t - 1) τ (4) In formula (4), G c (t) represents the channel gain at time t; P r (t) represents the received signal power at time t; τ represents the channel gain smoothing coefficient, which is used to adjust the influence weights of the current received power and the historical channel gain; G c (t - 1) τ represents the channel gain at the previous time t - 1; then, Markov channel evaluation is adopted, and the optimal channel selection probability at the future time is calculated through the weighted historical channel state fusion function. The calculation formula is as follows: In formula (5), P(C t ) represents the probability that the current channel C t is selected at time t; w i and w j represent the historical state weights of the i-th and j-th channels respectively; represents the channel gain of the i-th channel at time t - 1; represents the channel gain of the j-th channel at time t - 2; N represents the total number of available channels; if the channel state fluctuation amplitude exceeds the set threshold, the value function of the channel switching strategy is calculated by the channel allocation optimization method based on reinforcement learning, and the calculation formula is: Q(C t , g) = (1 + g)Q(C t , g) + g(r + γmaxQ(C t+1 , g’) (6) In formula (6), Q(C t , g) is the return value of channel C t under action g; g is the learning rate; r is the immediate reward; γ is the discount factor; maxQ(C t+1 , g’) represents the maximum Q value of the optimal channel selection decision at the next moment t + 1; based on the channel optimization strategy, the multi-scale channel state awareness method dynamically allocates data packets to parallel channels through multi-path adaptive load balancing, adopts a congestion control method based on reinforcement learning, and combines Bayesian bit error rate estimation to calculate the optimal data packet retransmission strategy, and finally outputs a light control signal.

8. A method for automatically controlling emergency response of an energy-saving intelligent street lamp according to claim 1, characterized in that: The working principle of S6 is as follows: The master-slave clock synchronization mechanism is adopted to calculate the local clock deviation of each intelligent street lamp node, and the bidirectional timestamp exchange method is used to record the sending and receiving times of the master clock and the slave clock. The round-trip delay is used to calculate and separate the synchronization signal transmission delay and the processing delay to obtain the initial clock offset. Then, based on the least squares linear regression method, the clock drift trend is calculated, and the Kalman filter is used to estimate the short-term clock error to generate the error compensation parameter. If the detected synchronization error exceeds the set threshold, the clock drift distribution is calculated through variational Bayesian inference, and the clock frequency error is iteratively updated by combining the particle filter method to adaptively adjust the clock synchronization step size with drift compensation. Finally, based on the exponentially weighted moving average, the clock adjustment amount is smoothed to optimize the global time synchronization stability, and the timestamp synchronization control instruction is output.

9. A method for automatically controlling the emergency response of an energy-saving intelligent street lamp according to claim 1, characterized in that: In S7, the working principles of the multi-objective reinforcement learning and the sparse Bayesian optimization method are as follows: Based on the deep deterministic policy gradient algorithm, a multi-objective joint reward function for energy consumption, brightness, and lifespan is constructed, and the Q-value matrix of different dimming strategies is generated in real time by using the dynamic weight allocation mechanism. Then, the sparse Bayesian optimization method is used to perform high-dimensional sparse sampling on the policy space, and the nonlinear relationship between the power adjustment amount of each node and the energy consumption benefit is modeled based on Gaussian process regression. If the predicted benefit variance exceeds the preset threshold, Monte Carlo sampling is used to perform local optimization on the policy parameters under the confidence interval constraint. At the same time, the real-time energy consumption data of multi-channel transmission in S5 and the clock synchronization deviation compensation amount in S6 are fused, and the lighting power duty cycle is dynamically adjusted through a PID fuzzy controller. The historical state, action, and reward tuples are stored in the experience replay buffer pool, and the weights of the policy network are updated through the backpropagation of the temporal difference error.

10. An energy-saving intelligent street lamp automatic emergency response system, characterized in that: Applied to an energy-saving intelligent street lamp automatic emergency response control method described in any one of claims 1-9, it includes: An environmental perception module, which is used to perform real-time monitoring and data preprocessing by using a light sensor, a millimeter-wave radar, an infrared sensor, and a camera. A dynamic priority calculation module, which is used to calculate the weights of traffic, environment, and accident impact factors by using an adaptive weighted decision tree, and construct a dynamic priority scheduling matrix by combining time series trend analysis to determine the lighting requirements of each area. A dimming strategy optimization module, which is used to train an intelligent dimming strategy by using improved fuzzy reinforcement learning combined with a deep Q network, calculate the optimal brightness allocation scheme through fuzzy value iteration, and optimize the lighting power based on historical data. An emergency response evaluation module, which is used to evaluate the accident level through a fuzzy inference system, and calculate the lighting response parameters corresponding to the accident category by using fuzzy membership optimization and multi-layer fuzzy inference. A lighting scheduling control module, which is used to calculate the lighting adjustment range, brightness change gradient, and warning light trigger conditions by using multi-level fuzzy control, and dynamically adjust the lighting strategy by combining state transition fuzzy optimization and fuzzy incremental compensation. A data transmission module, which is used to dynamically select LoRa / NB-IoT or 5G channels through a multi-scale channel state awareness and reinforcement learning channel allocation optimization method, and synchronize the timing consistency of smart street lamp nodes through a Bayesian clock drift correction model and Kalman filtering; A remote management module, which is used to perform data storage, analysis, and online optimization and update of intelligent dimming strategies based on edge computing and cloud computing methods; the remote management module uses multi-objective reinforcement learning combined with sparse Bayesian optimization to predict the energy consumption benefits of different dimming strategies, and dynamically adjusts the lighting power through particle swarm optimization and constrained optimization of lighting control.

Images