Power plant night lighting scheduling method

Abstract

The invention provides a night lighting scheduling method for a power plant. According to the method, an existing security camera is combined with an improved YOLOv4-tiny algorithm to realize rapid and accurate detection of personnel, and lighting scheduling of a specific area is triggered. An infrared induction device array with a Fresnel lens is deployed in key areas such as a power distribution cabinet, signals are processed through wavelet transformation, and after cross validation is conducted on the signals and camera detection results, a full-power brightening mode is started through ZigBee transmission signals. A fuzzy Petri network control model is introduced, and an optimal illumination strategy is generated according to the position, the track and the illumination intensity of the personnel. An improved MFCC algorithm is combined with an HMM to recognize a voice instruction, when a person enters a forbidden zone, warning voice is played, the definition is guaranteed through multi-channel Wiener filtering, meanwhile, a differential privacy protection LBP algorithm is used for analyzing facial expressions, and a second-level alarm is triggered when the facial expressions are abnormal. And a self-adaptive model based on deep reinforcement learning is constructed, scheduling is optimized through a PPO algorithm and an epsilon-greedy exploration mechanism, and power consumption minimization is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and in particular to a method for scheduling nighttime lighting in power plants. Background Technology

[0002] Traditional power plant nighttime lighting systems typically operate at full power, resulting in significant electricity waste; however, reducing the lighting power would disrupt the normal work of inspection personnel.

[0003] In existing technologies, some power plants have attempted to control lighting through timed switches or manual control, but these methods cannot dynamically adjust the lighting based on actual personnel activity, making it difficult to strike a balance between energy conservation and practicality. Furthermore, for special equipment areas such as distribution cabinets, there is a lack of effective personnel safety warning mechanisms, posing a risk of electric shock.

[0004] Therefore, there is an urgent need for a method that can intelligently adjust lighting based on people's location and provide safety warnings. Summary of the Invention

[0005] The main objective of this invention is to provide a method for scheduling nighttime lighting in power plants, which addresses the problem that some power plants attempt to control lighting through timed switches or manual control, but are unable to dynamically adjust the lighting based on actual personnel activity, making it difficult to achieve a balance between energy saving and practicality.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a method for scheduling nighttime lighting in power plants, the method comprising: S1. Utilize existing security cameras and combine them with the improved YOLOv4-tiny algorithm to monitor personnel positions in real time. When personnel enter a specific patrol area, the lighting scheduling mechanism for that area is triggered. S2. Deploy an array of infrared sensors with Fresnel lenses at key equipment locations and process the signals using a wavelet transform algorithm. After cross-verification of the infrared and camera detection results, transmit the signal to the central controller via ZigBee to activate the full-power brightening mode for the area's lights. S3. Introduce a lighting control model based on fuzzy Petri nets, taking personnel position, movement trajectory, and ambient light intensity as inputs, and generate the optimal lighting control strategy through fuzzy inference rules; S4. An improved MFCC algorithm combined with HMM is used to recognize voice commands; when personnel enter the restricted area, a warning voice is automatically played, and multi-channel Wiener filtering technology is used to ensure clarity; at the same time, the differential privacy-preserving LBP algorithm is used to analyze personnel's facial expressions, and a level 2 alarm is triggered when an anomaly is detected; S5. Construct an adaptive lighting scheduling model based on deep reinforcement learning, adopt the PPO algorithm, and dynamically adjust the lighting threshold and response strategy through the ε greedy exploration mechanism to minimize long-term power consumption.

[0007] In the preferred embodiment, in step S1: Video streams are acquired from existing security cameras in the power plant, and adaptive grayscale correction is performed on each frame of the image; An improved depthwise separable convolutional structure is used to extract features from the preprocessed image. First, depthwise convolution is performed on the input feature map. Where H and W are the image height and width, and C is the number of channels, with each channel corresponding to a 3×3 convolution kernel. The calculation formula is: ; in Output is a depthwise convolution; Then perform point convolution, using a 1×1 convolution kernel. C' represents the number of output channels, calculated as follows: ; Introducing the Ghost module to generate redundant features: for feature maps output by depthwise separable convolutions. Additional features are generated through linear transformation. The calculation formula is: ; in Here, P is the weighting parameter, and P is the number of neighboring pixels. For the transformation coefficients, This is the pixel offset; An improved anchor box matching mechanism is adopted for the generated candidate boxes. x, y are the center point coordinates, w, h are the width and height, and the bounding box is the center point coordinates. The matching degree is calculated as follows: ; in , IoU is the intersection-union ratio. As the center point of the true bounding box, when When the match is successful, it is considered a successful match. Coordinates of the detected personnel images Converted to physical coordinates of the factory area through perspective transformation The transformation formula is: , where K is the camera intrinsic parameter matrix, [R|t] is the extrinsic parameter matrix, and Z is the distance from the person to the camera; Preset the physical coordinate range of a specific patrol area ,when and If personnel continuously enter the area for more than 0.5 seconds, the lighting dispatch mechanism for that area will be triggered. ,otherwise .

[0008] In the preferred solution, the parameters of the existing security cameras in the power plant are first calibrated to ensure coverage of all patrol areas that need to be monitored, and the video streams collected by the cameras are preprocessed in real time using OpenCV software. The NVIDIA Jetson Nano was selected as the edge computing module, and the improved YOLOv4-tiny algorithm was deployed using the TensorFlow framework. The algorithm's depthwise separable convolution and Ghost module were used for feature extraction and object detection. A Python program was used to calculate candidate bounding box matching and locate people. The program's formulas are based on the formula... ; A coordinate transformation program was written in C++ to convert image coordinates into physical coordinates of the factory area based on the perspective transformation formula. The physical coordinate range of the patrol area was preset. When personnel are detected entering and the duration condition is met, a trigger signal is sent to the lighting control module with Arduino Mega as the core via the MQTT protocol. This module adjusts the lighting equipment in the corresponding area through the relay module. At the same time, Docker is used to encapsulate the algorithm running environment.

[0009] In the preferred embodiment, step S2 further includes the following steps: N pyroelectric infrared sensors with Fresnel lenses are deployed in an equilateral triangle pattern at the locations of key equipment such as power distribution cabinets. The lens focal length is set to 8-12mm and the viewing angle is adjusted to 50°-60° to ensure that the detection range covers an area with a radius of 5-8 meters centered on the key equipment and that the overlap rate of the detection range of adjacent equipment is ≥30%. The infrared sensing device acquires the raw infrared radiation signal sequence s(t) at a sampling frequency of 1 kHz, and performs baseline drift removal on the signal, as shown in the formula: Where M is the size of the sliding window; The preprocessed signal was decomposed into three levels using the db4 wavelet to obtain approximation coefficients. and detail coefficient Calculate the noise reduction threshold , For noise standard deviation estimation, where N is the signal length, the effective signal is reconstructed using inverse wavelet transform. ,in The detail coefficient after noise reduction; Calculate the energy value of the reconstructed signal within a 100ms detection window. When E≥ the offline training threshold, the corresponding infrared device outputs a detection signal dᵢ=1; otherwise, dᵢ=0, where i is the device number. Cross-validation of infrared and camera detection results is performed. Let the camera detection result be C, where 1 indicates a person was detected and 0 indicates no person was detected. The overall judgment result of the infrared array is then determined. When C=1 and D≥60-70% of the infrared devices detect a person, a trigger signal T=1 is output to verify the reliability. When R ≥ 0.8, the judgment result is reliable; The signal is transmitted via a ZigBee wireless self-organizing network. The data includes the device ID, detection result, timestamp, and CRC checksum. After receiving the T=1 signal, the central controller controls the lighting driver module to start the full-power brightening mode of the lights in that area through the RS485 interface.

[0010] In the preferred embodiment, step S3 further includes the following steps: The physical coordinates (X, Y) of the factory area are mapped to a fuzzy subset {near core area, mid-distance area, edge area}. The membership formula for the near core area is: ,in The attenuation coefficient is... The coordinates of the device center are given, and the membership functions for the mid-range and edge regions correspond to the following: and ; The movement speed v and direction angle are calculated by changing the position over three consecutive frames. The mapping is a fuzzy subset of {stationary, slow-moving, high-speed moving}, and the membership formula for high-speed moving is: Low speed corresponds to 0.3-0.5 m / s, and stationary corresponds to ≤0.3 m / s; Illuminance is collected by a light sensor and mapped to a fuzzy subset {low illuminance, medium illuminance, high illuminance}. The membership formula for low illuminance is: Where L is the real-time illuminance collected by a light sensor. This is the critical value for low light intensity at night; Construct a fuzzy Petri net model: This model includes the locations, movement trajectories, and light intensity of people in the input database; intermediate inference states in the intermediate database; and lighting power levels and illuminated areas in the output database. Define 12 inference rules, each associated with a weight matrix and a trigger threshold. ; Calculate the activation degree of the input library for the transition. , Let be the weight of the transition t for the i-th input library. For the corresponding membership degree, when the activation degree When a transition is triggered, the formula for updating the membership of the output library is: ,in To assess the credibility of the changes; The centroid method is used to defuzzify the membership degree of the output library. The formula for the lighting power command value is as follows: , For preset power levels, For the corresponding membership degree; The formula for the range of illuminated areas is: ,in , v represents the movement speed.

[0011] In the preferred scheme, the fuzzy Petri net model is constructed using MATLAB's Fuzzy Logic Toolbox; Implement centroid-based deblurring using Python, based on... Generate power command, press The lighting range is determined, and control commands are sent to the lighting controller via the Modbus protocol to adjust the lighting equipment in the corresponding area. The entire algorithm environment is encapsulated using Docker.

[0012] In the preferred embodiment, step S4 further includes the following steps: Deploy 4-5 unit omnidirectional microphone arrays in key areas such as power distribution cabinets, set the sampling frequency to 16-18kHz, and use an STM32 microcontroller to collect voice signals and perform pre-emphasis processing. Improved MFCC feature extraction was achieved by using Python to call the Librosa library. The frames were divided into 25-30ms frames with a frame length and 10-12ms frame shift, and a Hanning window was added. The 39-dimensional feature vector was calculated by dynamically adjusting the bandwidth of the Mel filter. The HMM model for five types of speech commands was trained using the HTK toolkit, the state transition probabilities were updated using the Baum-Welch algorithm, and the command recognition was completed using the Viterbi algorithm. When personnel enter the restricted area, an audio power amplifier drives a speaker to play a warning message, while MATLAB's Audio Toolbox is used to implement multi-channel Wiener filtering to eliminate echoes. The algorithm crops facial regions from images captured by security cameras, implements the LBP algorithm formula for differential privacy protection using the OpenCV library, adds Laplacian noise, extracts features, and then trains an SVM classifier using the Python Scikit-learn library to analyze facial expressions. When an anomaly is detected, a secondary alarm is triggered through the ZigBee module. The entire algorithm environment is encapsulated using Docker.

[0013] In the preferred scheme, a left-to-right Hidden Markov Model is trained using five types of instructions, and the Baum-Welch algorithm is used to update the state transition probabilities. The Viterbi algorithm is used to match and recognize commands, where Let be the probability that a given state is i at time t and transitions to state j at time t+1. Let be the probability of being in state i at time t; Differential privacy-preserving LBP algorithm for facial expression analysis involves cropping a 128×128 grayscale image of the face from the image. Calculate the LBP features with Laplace noise. The expression is identified by an SVM classifier, and a level 2 alarm is triggered when an anomaly is detected. in The neighborhood radius, For noise, For sensitivity, Budget for privacy.

[0014] In the preferred embodiment, step S5 further includes the following steps: The policy network for the PPO algorithm is built in the central controller using the TensorFlow framework; The average light intensity of the factory area is collected by a light sensor, and the total power of the activated lighting area is obtained through a current transformer. Combined with the personnel location and dwell time data from the security camera in step S1, a state vector is constructed. ,in, This represents the average light intensity in the factory area. The total power of the currently activated lighting area. The distance between personnel and the nearest critical equipment. The duration of a person's stay in the current area. This represents the number of areas where lighting has been activated. Write a reward function calculation module using Python, based on... The formula generates the reward value, where, This is an electricity consumption reward item. In response to the timeliness reward program, The strategy stability reward item has the required weighting coefficient. ; During the training phase, a power plant lighting simulation environment was built using Gazebo. Every 200-400 trajectories accumulated, the network parameters were updated using the Adam optimizer. To achieve ε-greedy exploration, where This is the initial exploration rate. To minimize the exploration rate, This represents the maximum number of iterations. During deployment, the lighting threshold adjustment amount output by the model is transmitted via the OPC UA protocol. and power coefficient The signal is sent to the smart lighting controller, which controls the dimming module to adjust the power. The entire process uses Kubernetes containerized management of the algorithm components.

[0015] In the preferred scheme, a dual-network structure for the PPO algorithm is constructed, comprising a policy network and a value network: The policy network employs a 4-layer convolutional neural network. The input is a state vector S, which extracts features using a 1×1 convolutional kernel. The output layer uses a Softmax activation function to generate the action probability distribution. , For network parameters; The value network employs a 3-layer fully connected network. The input is a state vector S, with 128 and 64 neurons in the hidden layers, respectively. The output layer is the state value estimate. , For network parameters; The core optimization objective of PPO is the surrogate loss function with pruning, as shown in the formula: ; in: For the dominant function, Discount factor; This is the cutting factor; , The regularization coefficient is used. Let be the policy entropy.

[0016] This invention provides a method for power plant nighttime lighting scheduling. It utilizes existing cameras and infrared sensing devices to construct a low-cost, high-reliability intelligent lighting scheduling system. It employs a variety of less commonly used but efficient algorithms to achieve precise personnel detection, lighting control, and safety warnings. Through deep reinforcement learning, it continuously optimizes the scheduling strategy, significantly reducing power consumption while ensuring the work efficiency and safety of inspection personnel. Attached Figure Description

[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the power plant nighttime lighting scheduling method of the present invention. Detailed Implementation

[0018] Example 1 like Figure 1 As shown, a method for scheduling nighttime lighting in a power plant includes: S1. Utilize existing security cameras and combine them with the improved YOLOv4-tiny algorithm to monitor personnel positions in real time. When personnel enter a specific patrol area, the lighting scheduling mechanism for that area is triggered. S2. Deploy an array of infrared sensors with Fresnel lenses at key equipment locations and process the signals using a wavelet transform algorithm. After cross-verification of the infrared and camera detection results, transmit the signal to the central controller via ZigBee to activate the full-power brightening mode for the area's lights. S3. Introduce a lighting control model based on fuzzy Petri nets, taking personnel position, movement trajectory, and ambient light intensity as inputs, and generate the optimal lighting control strategy through fuzzy inference rules; S4. An improved MFCC algorithm combined with HMM is used to recognize voice commands; when personnel enter the restricted area, a warning voice is automatically played, and multi-channel Wiener filtering technology is used to ensure clarity; at the same time, the differential privacy-preserving LBP algorithm is used to analyze personnel's facial expressions, and a level 2 alarm is triggered when an anomaly is detected; S5. Construct an adaptive lighting scheduling model based on deep reinforcement learning, adopt the PPO algorithm, and dynamically adjust the lighting threshold and response strategy through the ε greedy exploration mechanism to minimize long-term power consumption.

[0019] First, real-time visual recognition is achieved by utilizing the existing security camera network within the power plant area, combined with an improved YOLOv4-tiny algorithm. This algorithm employs a lightweight design using depthwise separable convolution and a Ghost module, significantly reducing computational complexity while maintaining high detection accuracy, enabling rapid and accurate personnel detection. When a camera detects an inspector entering a specific patrol area, it immediately triggers the lighting scheduling mechanism for that area.

[0020] At critical equipment locations such as power distribution cabinets, arrays of infrared sensors based on the pyroelectric effect are deployed, with each device equipped with a Fresnel lens to expand the detection range. A wavelet transform-based signal processing algorithm is used to perform multi-resolution analysis of the infrared sensor signals, effectively filtering out environmental interference and improving the reliability of personnel detection. When the detection results from the infrared sensors and cameras cross-validate, the signal is transmitted to the central controller via ZigBee wireless self-organizing network technology, activating the full-power brightening mode for the lights in that area.

[0021] To achieve refined lighting scheduling, this invention introduces a lighting control model based on fuzzy Petri nets. This model uses factors such as personnel location, movement trajectory, and ambient light intensity as input variables, and generates the optimal lighting control strategy through fuzzy inference rules. For example, when a person is detected to be stationary, the lighting power in surrounding non-critical areas is gradually reduced; when a person begins to move, the movement path is predicted in advance and the lights in the corresponding areas are turned on, forming dynamic lighting guidance.

[0022] For safety alerts, an improved Mel-frequency cepstral coefficient (MFCC) algorithm combined with a Hidden Markov Model (HMM) is used to accurately recognize personnel's voice commands. When the infrared sensor detects personnel entering a restricted area, the system automatically plays a preset warning message and uses echo cancellation technology based on multi-channel Wiener filtering to ensure the clarity of the warning message in complex environments. Simultaneously, a differential privacy-preserving Local Binary Pattern (LBP) algorithm is used to analyze personnel's facial expressions in real time. If abnormally tense or confused expressions are detected, a secondary alarm mechanism is triggered.

[0023] The system also constructs an adaptive lighting scheduling model based on deep reinforcement learning. It employs the Proximal Policy Optimization (PPO) algorithm to continuously learn personnel inspection habits and environmental change patterns, dynamically adjusting lighting thresholds and response strategies. By setting an ε-greedy exploration mechanism, the lighting scheduling scheme is gradually optimized while ensuring energy-saving effects, minimizing power consumption during long-term operation.

[0024] Example 2 To further illustrate with reference to Example 1, in step S1: Video streams are acquired from existing security cameras in the power plant, and adaptive grayscale correction is performed on each frame of the image; An improved depthwise separable convolutional structure is used to extract features from the preprocessed image. First, depthwise convolution is performed on the input feature map. Where H and W are the image height and width, and C is the number of channels, with each channel corresponding to a 3×3 convolution kernel. The calculation formula is: ; in Output is a depthwise convolution; Then perform point convolution, using a 1×1 convolution kernel. C' represents the number of output channels, calculated as follows: ; Introducing the Ghost module to generate redundant features: for feature maps output by depthwise separable convolutions. Additional features are generated through linear transformation. The calculation formula is: ; in Here, P is the weighting parameter, and P is the number of neighboring pixels. For the transformation coefficients, This is the pixel offset; An improved anchor box matching mechanism is adopted for the generated candidate boxes. x, y are the center point coordinates, w, h are the width and height, and the bounding box is the center point coordinates. The matching degree is calculated as follows: ; in , IoU is the intersection-union ratio. As the center point of the true bounding box, when When the match is successful, it is considered a successful match. Coordinates of the detected personnel images Converted to physical coordinates of the factory area through perspective transformation The transformation formula is: , where K is the camera intrinsic parameter matrix, [R|t] is the extrinsic parameter matrix, and Z is the distance from the person to the camera; Preset the physical coordinate range of a specific patrol area ,when and If personnel continuously enter the area for more than 0.5 seconds, the lighting dispatch mechanism for that area will be triggered. ,otherwise .

[0025] In the preferred solution, the parameters of the existing security cameras in the power plant are first calibrated to ensure coverage of all patrol areas that need to be monitored, and the video streams collected by the cameras are preprocessed in real time using OpenCV software. The NVIDIA Jetson Nano was selected as the edge computing module, and the improved YOLOv4-tiny algorithm was deployed using the TensorFlow framework. The algorithm's depthwise separable convolution and Ghost module were used for feature extraction and object detection. A Python program was used to calculate candidate bounding box matching and locate people. The program's formulas are based on the formula... ; A coordinate transformation program was written in C++ to convert image coordinates into physical coordinates of the factory area based on the perspective transformation formula. The physical coordinate range of the patrol area was preset. When personnel are detected entering and the duration condition is met, a trigger signal is sent to the lighting control module with Arduino Mega as the core via the MQTT protocol. This module adjusts the lighting equipment in the corresponding area through the relay module. At the same time, Docker is used to encapsulate the algorithm running environment.

[0026] The specific implementation method is as follows: In step S1, the improved YOLOv4-tiny algorithm requires multi-stage collaborative processing in practical applications to achieve accurate real-time monitoring of personnel and lighting triggering. The specific practical steps and mathematical formulas are as follows: First, data preprocessing is performed. After acquiring video streams from existing security cameras, adaptive grayscale correction is applied to each frame to eliminate the effects of uneven lighting at night. During the correction process, the grayscale value of each pixel in the image is first calculated. Then, the correction is performed using the following formula: ,in The mean gray level of the original image. Where is the standard deviation of the original image's grayscale, and k is the contrast adjustment coefficient. The target grayscale mean is used. Normalization enhances the grayscale difference between people and the background in the image, making it easier to distinguish people from the environment during subsequent feature extraction.

[0027] Next, the feature extraction stage is performed using an improved depthwise separable convolutional structure. For the input feature map... C represents the number of channels. First, depthwise convolution is performed, with each channel corresponding to a 3×3 convolution kernel. The calculation formula is: ,in The output is a depthwise convolution. Then, a pointwise convolution is performed using a 1×1 kernel. C' represents the number of output channels. Calculation: The purpose of these two formulas is to decompose the computational cost of standard convolution, thereby reducing computational complexity while ensuring feature extraction capabilities, enabling the algorithm to run efficiently on the edge computing units of security cameras.

[0028] Then, the Ghost module is introduced to generate redundant features for the feature maps output by depthwise separable convolutions. Additional features are generated through linear transformation. The calculation formula is: ,in Here, P is the weighting parameter, and P is the number of neighboring pixels (taken as 8). For the transformation coefficients, This represents the pixel offset. The purpose of this formula is to enrich the detailed information of the feature map and improve the ability to recognize people in different poses without adding too many parameters.

[0029] In the target detection and localization stage, an improved anchor box matching mechanism is adopted. For the generated candidate boxes... x, y are the center point coordinates, w, h are the width and height, and the bounding box is the center point coordinates. The matching degree is calculated using the following formula: ,in , IoU is the intersection-union ratio. The center point of the true bounding box.

[0030] when When the time is right, it is considered a successful match. The purpose of this formula is to comprehensively consider the overlap between the candidate box and the ground truth box and the distance between the center point, thereby improving the positioning accuracy of personnel detection at night, especially for small-sized personnel targets at a distance.

[0031] Finally, region determination and trigger control are performed to determine the coordinates of the detected personnel images. Converted to physical coordinates of the factory area through perspective transformation The transformation formula is based on the camera's intrinsic parameter matrix K and extrinsic parameter matrix [R|t]: Where Z is the distance from the person to the camera. The physical coordinate range of a specific patrol area is preset. ,when and When the lighting control signal is triggered, the signal triggering condition satisfies the following equation: If personnel remain in the area for more than 0.5 seconds, otherwise... .

[0032] The system accurately maps the location of personnel in the image to the actual factory area, ensuring that the mechanism is only triggered when personnel actually enter the patrol area that requires lighting, thus avoiding false triggering.

[0033] Example 3 In further detail with reference to Embodiment 1, step S2 also includes the following steps: N pyroelectric infrared sensors with Fresnel lenses are deployed in an equilateral triangle pattern at the locations of key equipment such as power distribution cabinets. The lens focal length is set to 8-12mm and the viewing angle is adjusted to 50°-60° to ensure that the detection range covers an area with a radius of 5-8 meters centered on the key equipment and that the overlap rate of the detection range of adjacent equipment is ≥30%. The infrared sensing device acquires the raw infrared radiation signal sequence s(t) at a sampling frequency of 1 kHz, and performs baseline drift removal on the signal, as shown in the formula: Where M is the size of the sliding window; The preprocessed signal was decomposed into three levels using the db4 wavelet to obtain approximation coefficients. and detail coefficient Calculate the noise reduction threshold , For noise standard deviation estimation, where N is the signal length, the effective signal is reconstructed using inverse wavelet transform. ,in The detail coefficient after noise reduction; Calculate the energy value of the reconstructed signal within a 100ms detection window. When E≥ the offline training threshold, the corresponding infrared device outputs a detection signal dᵢ=1; otherwise, dᵢ=0, where i is the device number. Cross-validation of infrared and camera detection results is performed. Let the camera detection result be C, where 1 indicates a person was detected and 0 indicates no person was detected. The overall judgment result of the infrared array is then determined. When C=1 and D≥60-70% of the infrared devices detect a person, a trigger signal T=1 is output to verify the reliability. When R ≥ 0.8, the judgment result is reliable; The signal is transmitted via a ZigBee wireless self-organizing network. The data includes the device ID, detection result, timestamp, and CRC checksum. After receiving the T=1 signal, the central controller controls the lighting driver module to start the full-power brightening mode of the lights in that area through the RS485 interface.

[0034] In step S2, the infrared sensing and lighting control for key equipment locations requires precise scheduling through equipment deployment, signal processing, cross-validation, and signal transmission. The specific algorithm steps and mathematical formulas are as follows: First, deploy an array of infrared sensing devices. N pyroelectric infrared sensing devices with Fresnel lenses (N≥3) are deployed in an equilateral triangle around key equipment such as power distribution cabinets. The lens focal length f is set to 8mm and the viewing angle θ is adjusted to 60° to ensure that the detection range of the devices covers an area with a radius of 5 meters centered on the key equipment, and the overlap rate of the detection range of adjacent devices is ≥30% to reduce blind spots.

[0035] Next, infrared signal acquisition and preprocessing are performed. Each infrared sensor acquires ambient infrared radiation signals in real time, with a sampling frequency set to 1kHz, to obtain the original signal sequence. t is the timestamp in milliseconds. Due to interference from equipment heat dissipation and airflow at night, baseline drift removal of the signal is required first. The calculation formula is: Where M=50 is the sliding window size. The purpose of this formula is to eliminate slow baseline changes in the signal through moving averages, while preserving the rapid infrared changes caused by personnel movement.

[0036] Then, the wavelet transform algorithm is used to perform multi-resolution analysis on the preprocessed signal. The db4 wavelet is selected as the basis function, and a three-level wavelet decomposition is performed to obtain the approximation coefficients. and detail coefficient Thresholding denoising is performed on the detail coefficients. The threshold calculation formula is: ,in Here, N is the noise standard deviation estimate, and N is the signal length. The effective signal is obtained after denoising and wavelet reconstruction. The formula is: ,in This is the inverse wavelet transform. This represents the detail coefficient after denoising. It separates the effective components from noise in the signal, improving the ability to identify infrared features of people, and especially filtering out high-frequency interference generated during equipment operation.

[0037] Next, the presence of personnel is determined, and the energy value of the reconstructed signal is calculated. Where T is the detection window duration. When E≥E0, E0 is the threshold obtained through offline sample training, which is set to 0.02, it is determined that the device has detected human activity, and an output signal is generated. , where i is the device number, otherwise The purpose of this formula is to enhance the recognizability of personnel signals through energy accumulation and avoid misjudgments caused by single-pulse interference.

[0038] Then, cross-validation of the infrared and camera detection results is performed, denoted as C. This indicates that a person has been detected. This indicates that nothing was detected; the overall judgment result of the infrared array is... When satisfied and When more than 60% of the infrared devices detect a person, the cross-validation is deemed successful, and a trigger signal is output. The formula for calculating the credibility of verification is: When R ≥ 0.8, the verification result is considered reliable. This formula combines the advantages of both detection methods, reducing the risk of single-device failure or misjudgment, and improving the reliability of the trigger signal.

[0039] Finally, the signal is transmitted via a ZigBee wireless ad hoc network. The network adopts a star topology, with the central controller acting as the coordinator and the infrared sensor nodes and camera nodes acting as terminal devices. The communication frequency band is set to 2.4GHz, and the transmission rate is 250kbps. Signal transmission uses a packet-switched method with verification. Each data packet contains the device ID, detection result, timestamp, and CRC checksum. The verification formula is: , For the k-th byte of the data packet, This is an XOR operation, where L is the data packet length. The central controller receives... Upon receiving the signal, a control command is sent to the lighting driver module via the RS485 interface to activate the full-power brightness mode of the lights in that area. The purpose of this formula is to perform data integrity verification through a simple XOR operation, ensuring that the signal is not tampered with or damaged during wireless transmission.

[0040] Example 4 In further detail with reference to Embodiment 1, step S3 also includes the following steps: The physical coordinates (X, Y) of the factory area are mapped to a fuzzy subset {near core area, mid-distance area, edge area}. The membership formula for the near core area is: ,in The attenuation coefficient is... The coordinates of the device center are given, and the membership functions for the mid-range and edge regions correspond to the following: and ; The movement speed v and direction angle are calculated by changing the position over three consecutive frames. The mapping is a fuzzy subset of {stationary, slow-moving, high-speed moving}, and the membership formula for high-speed moving is: Low speed corresponds to 0.3-0.5 m / s, and stationary corresponds to ≤0.3 m / s; Illuminance is collected by a light sensor and mapped to a fuzzy subset {low illuminance, medium illuminance, high illuminance}. The membership formula for low illuminance is: Where L is the real-time illuminance collected by a light sensor. This is the critical value for low light intensity at night; Construct a fuzzy Petri net model: This model includes the locations, movement trajectories, and light intensity of people in the input database; intermediate inference states in the intermediate database; and lighting power levels and illuminated areas in the output database. Define 12 inference rules, each associated with a weight matrix and a trigger threshold. ; Calculate the activation degree of the input library for the transition. , Let be the weight of the transition t for the i-th input library. For the corresponding membership degree, when the activation degree When a transition is triggered, the formula for updating the membership of the output library is: ,in To assess the credibility of the changes; The centroid method is used to defuzzify the membership degree of the output library. The formula for the lighting power command value is as follows: , For preset power levels, For the corresponding membership degree; The formula for the range of illuminated areas is: ,in , v represents the movement speed.

[0041] In the preferred scheme, the fuzzy Petri net model is constructed using MATLAB's Fuzzy Logic Toolbox; Implement centroid-based deblurring using Python, based on... Generate power command, press The lighting range is determined, and control commands are sent to the lighting controller via the Modbus protocol to adjust the lighting equipment in the corresponding area. The entire algorithm environment is encapsulated using Docker.

[0042] S3 Steps: Detailed Explanation of Lighting Control Model Based on Fuzzy Petri Nets: A lighting control model based on fuzzy Petri nets (FPN) is constructed, taking the real-time location of personnel, movement trajectory characteristics, and ambient light intensity as input variables. Dynamic lighting control strategies (such as power adjustment and regional lighting sequence) are generated through multi-layer fuzzy inference rules, so as to achieve fine-grained scheduling of "lights follow people when they move and lights are adjusted when people are still", taking into account both energy saving and inspection efficiency.

[0043] Step 1: Fuzzification of input variables Personnel Location (P): The physical coordinates (X, Y) of the factory area are mapped to a fuzzy subset {near core area, mid-distance area, edge area}, and the impact of location on lighting requirements is quantified through a membership function. Let the coordinate range of the core area be... ( For equipment center, Membership formula: ; in The attenuation coefficient is the factor that increases with distance from the core region. The closer it is to 1, the closer it is to 0; the membership functions of the mid-range region and the edge region are similar, only the k value is adjusted, which is 0.4 and 0.2 respectively.

[0044] By transforming discrete coordinates into continuous fuzzy variables, the correlation between personnel location and lighting requirements can be quantified.

[0045] Movement trajectory characteristics: The movement speed $v$ and direction angle are calculated based on the positional changes over three consecutive frames. Construct a fuzzy subset {stationary, slow-moving, high-moving}. Velocity membership formula: ; The membership functions for low speed and stationary conditions are designed to decrease linearly, with a threshold of 0.3-0.5 m / s for low speed and ≤0.3 m / s for stationary conditions.

[0046] Transforming motion states into inferable fuzzy variables provides a basis for "predicting path lighting".

[0047] Ambient light intensity (L): Real-time illuminance (unit: lux) collected by a light sensor, fuzzy subset {low illuminance, medium illuminance, high illuminance}, membership formula: ; in , (Critical value for low illumination at night), the lower the illumination, The closer it is to 1.

[0048] Quantify the substitution effect of natural light for artificial lighting to avoid unnecessary power waste.

[0049] Step 2: Construction of the fuzzy Petri net model Places: Contains input places ( Personnel location, Movement trajectory Light intensity), intermediate storage ( : Inference intermediate state), output repository ( Lighting power rating (The area where the lights are on).

[0050] Transitions: Define 12 core reasoning rules, for example: Rule 1: If and and ,but and .

[0051] Rule 2: If and and ,but and .

[0052] Weights and Thresholds: Each transition is associated with a weight matrix W, and the weights of the input library's influence on the transition are summed to 1, triggering the threshold. A transition is triggered only when the weighted sum of the membership degrees of the input library is ≥0.6.

[0053] Step 3: Fuzzy Reasoning and Decision Output Calculate the activation degree of the input library for the transition: , Let be the weight of the transition t for the i-th input library. This corresponds to the membership degree.

[0054] when When a transition is triggered, the membership update formula for the output library is: ; in To ensure the credibility of the transition, the maximum value is taken to guarantee the priority activation of better strategies.

[0055] By fusing multiple input variables through fuzzy reasoning, the membership degree of the lighting control strategy is dynamically updated.

[0056] Step 4: Deblurring and Control Command Generation: The membership degree of the output locations is defuzzified using the centroid method to calculate the lighting power command value. : ; in Preset power levels (full power = 100%, 70%, 50%). This corresponds to the membership degree.

[0057] The illuminated area is calculated using a trajectory prediction formula: , Based on the scope, v represents the estimated time and the movement speed.

[0058] The fuzzy inference results are converted into specific lighting control parameters, which directly drive the execution module.

[0059] The fuzzy results are converted into precise control commands, enabling continuous adjustment of lighting power.

[0060] Through the above steps, the fuzzy Petri net model can generate "on-demand lighting" strategies in real time based on personnel dynamics and environmental changes. For example, when personnel are stationary, the power in non-core areas can be automatically reduced by 30%, and when they move, the path can be lit 2 meters in advance, ensuring the convenience of inspection while saving energy.

[0061] Example 5 In further detail with reference to Embodiment 1, step S4 also includes the following steps: Deploy 4-5 unit omnidirectional microphone arrays in key areas such as power distribution cabinets, set the sampling frequency to 16-18kHz, and use an STM32 microcontroller to collect voice signals and perform pre-emphasis processing. Improved MFCC feature extraction was achieved by using Python to call the Librosa library. The frames were divided into 25-30ms frames with a frame length and 10-12ms frame shift, and a Hanning window was added. The 39-dimensional feature vector was calculated by dynamically adjusting the bandwidth of the Mel filter. The HMM model for five types of speech commands was trained using the HTK toolkit, the state transition probabilities were updated using the Baum-Welch algorithm, and the command recognition was completed using the Viterbi algorithm. When personnel enter the restricted area, an audio power amplifier drives a speaker to play a warning message, while MATLAB's Audio Toolbox is used to implement multi-channel Wiener filtering to eliminate echoes. The algorithm crops facial regions from images captured by security cameras, implements the LBP algorithm formula for differential privacy protection using the OpenCV library, adds Laplacian noise, extracts features, and then trains an SVM classifier using the Python Scikit-learn library to analyze facial expressions. When an anomaly is detected, a secondary alarm is triggered through the ZigBee module. The entire algorithm environment is encapsulated using Docker.

[0062] In the preferred scheme, a left-to-right Hidden Markov Model is trained using five types of instructions, and the Baum-Welch algorithm is used to update the state transition probabilities. The Viterbi algorithm is used to match and recognize commands, where Let be the probability that a given state is i at time t and transitions to state j at time t+1. Let be the probability of being in state i at time t; Differential privacy-preserving LBP algorithm for facial expression analysis involves cropping a 128×128 grayscale image of the face from the image. Calculate the LBP features with Laplace noise. The expression is identified by an SVM classifier, and a level 2 alarm is triggered when an anomaly is detected. in The neighborhood radius, For noise, For sensitivity, Budget for privacy.

[0063] In step S4, the security alert mechanism needs to achieve multi-dimensional security protection through voice command recognition, alert voice playback, and facial expression analysis. The specific algorithm steps and mathematical formulas are as follows: First, voice commands are collected and preprocessed. Omnidirectional microphone arrays are deployed in key areas such as power distribution cabinets, with a sampling frequency of 16kHz and a quantization bit depth of 16bit to collect the voice command signals from inspection personnel. Where n is the sampling point index. To eliminate environmental noise, the signal is first pre-emphasized, using the following formula: This formula enhances the clarity of speech signals by boosting high-frequency components, laying the foundation for subsequent feature extraction.

[0064] Next, the improved MFCC algorithm is used to extract speech features. The steps are as follows: ① The pre-emphasized signal is divided into frames with a frame length of 25ms and a frame shift of 10ms. A Hanning window is added to each frame. ① m represents the number of sampling points within a frame, and N=400 represents the frame length; ② Perform a Fast Fourier Transform (FFT) on each frame signal to calculate the power spectrum. ③ Design 26 Mel filter banks, the center frequency of the i-th filter is... The filter output energy is calculated using the following formula: ,in For filter coefficients, the improvement lies in introducing dynamic bandwidth adjustment; ④ Regarding energy Taking the logarithm, we get Then perform a Discrete Cosine Transform (DCT) and take the first 13 coefficients as MFCC features. Simultaneously calculate its first-order difference. and second-order difference This results in a 39-dimensional feature vector. By specifically adjusting the filter characteristics, this improved algorithm can effectively extract speech features even in noisy environments, thus enhancing the robustness of command recognition.

[0065] Subsequently, Hidden Markov Models (HMMs) were constructed for voice command recognition. For five types of commands, including "turn on lights," "turn off lights," and "alarm," left-to-right HMM models were trained, each containing five states and a state transition probability matrix. The observation probability follows a Gaussian mixture model.

[0066] The model is trained using the Baum-Welch algorithm, and the state transition probabilities are updated using the following formula: ,in Let be the probability that a state is in state i at time t and transitions to state j at time t+1. Let be the probability of being in state i at time t. During the recognition phase, the Viterbi algorithm is used to calculate the matching score between the feature vector sequence and each model. The instruction corresponding to the model with the highest score is the recognition result. This process uses a statistical model to model the temporal changes of speech features, achieving accurate differentiation of different instructions.

[0067] In the warning voice playback and echo cancellation process, when the infrared device detects personnel entering the restricted area, the system invokes a preset warning voice message, which is then played through a speaker driven by a power amplifier. To eliminate echo interference between the speaker playback and microphone pickup, multi-channel Wiener filtering technology is employed, and a reference signal is defined. and microphone input signal The filter coefficients $w$ are updated iteratively using the following formula: ,in For error signals, The step size factor is used to iterate to... The technology cancels out echoes by updating the filter coefficients in real time, ensuring the clarity of the warning voice in noisy environments.

[0068] Simultaneously, the differential privacy-preserving LBP algorithm is activated for facial expression analysis, detecting face regions from images captured by the camera and cropping them into 128×128 pixel grayscale images. x, y are image coordinates. Traditional LBP algorithms are prone to privacy breaches. The improvement lies in introducing Laplacian noise. The formula for calculating LBP features is: ,in The neighborhood radius, , like Otherwise 0, For noise ( For sensitivity, (For privacy budget). After feature extraction, the data is input into a Support Vector Machine (SVM) classifier to identify facial expression categories such as "normal," "nervous," and "confused." When an abnormal facial expression is detected, a secondary alarm is triggered. This formula achieves a balance between security and privacy by incorporating controllable noise into the feature calculation, thus protecting user privacy while still accurately identifying abnormal facial expressions.

[0069] Finally, a multi-mechanism collaborative response is implemented. When the HMM recognizes an "alarm" command or the LBP detects an abnormal facial expression, the system synchronously triggers a level-two alarm via the ZigBee network. Simultaneously, the volume of the warning audio is increased by 20%, and the event time, location, and related data are recorded to the local database. This algorithm, through the integration of multiple technologies, forms a closed-loop safety mechanism of "command recognition - proactive warning - anomaly response," comprehensively ensuring the operational safety of inspection personnel near restricted areas.

[0070] A four-unit omnidirectional microphone array was deployed in key areas such as the power distribution cabinet, with a sampling frequency of 16kHz. The speech signal was acquired and pre-emphasis processing was performed using an STM32 microcontroller (based on the formula). Improved MFCC feature extraction was implemented using Python and the Librosa library. Frames were divided into 25ms frames with a 10ms frame shift and a Hanning window was applied. A 39-dimensional feature vector was calculated by dynamically adjusting the Mel filter bandwidth. An HMM model (containing 5 states) for 5 types of speech commands was trained using the HTK toolkit, and the state transition probabilities were updated using the Baum-Welch algorithm (based on...). The formula uses the Viterbi algorithm to complete command recognition; when personnel enter the restricted area, an audio power amplifier (such as TPA3116) drives a speaker to play a warning voice, while MATLAB's Audio Toolbox is used to implement multi-channel Wiener filtering (according to...). Formula iteratively updates coefficients. Echo cancellation; cropping facial regions from images captured by security cameras, and implementing a differential privacy-preserving LBP algorithm using the OpenCV library (based on...). The formula (with added Laplacian noise) is used to extract features. Then, an SVM classifier is trained using Python's Scikit-learn library to analyze facial expressions. When an anomaly is detected, a secondary alarm is triggered through the ZigBee module (CC2530). The entire algorithm environment is encapsulated in Docker to ensure that all modules run in tandem.

[0071] Example 5 In further detail with reference to Embodiment 1, step S5 also includes the following steps: The policy network for the PPO algorithm is built in the central controller using the TensorFlow framework; The average light intensity of the factory area is collected by a light sensor, and the total power of the activated lighting area is obtained through a current transformer. Combined with the personnel location and dwell time data from the security camera in step S1, a state vector is constructed. ,in, This represents the average light intensity in the factory area. The total power of the currently activated lighting area. The distance between personnel and the nearest critical equipment. The duration of a person's stay in the current area. This represents the number of areas where lighting has been activated. Write a reward function calculation module using Python, based on... The formula generates the reward value, where, This is an electricity consumption reward item. In response to the timeliness reward program, The strategy stability reward item has the required weighting coefficient. ; During the training phase, a power plant lighting simulation environment was built using Gazebo. Every 200-400 trajectories accumulated, the network parameters were updated using the Adam optimizer. To achieve ε-greedy exploration, where This is the initial exploration rate. To minimize the exploration rate, This represents the maximum number of iterations. During deployment, the lighting threshold adjustment amount output by the model is transmitted via the OPC UA protocol. and power coefficient The signal is sent to the smart lighting controller, which controls the dimming module to adjust the power. The entire process uses Kubernetes containerized management of the algorithm components.

[0072] In the preferred scheme, a dual-network structure for the PPO algorithm is constructed, comprising a policy network and a value network: The policy network employs a 4-layer convolutional neural network. The input is a state vector S, which extracts features using a 1×1 convolutional kernel. The output layer uses a Softmax activation function to generate the action probability distribution. , For network parameters; The value network employs a 3-layer fully connected network. The input is a state vector S, with 128 and 64 neurons in the hidden layers, respectively. The output layer is the state value estimate. , For network parameters; The core optimization objective of PPO is the surrogate loss function with pruning, as shown in the formula: ; in: For the dominant function, Discount factor; This is the cutting factor; , The regularization coefficient is used. Let be the policy entropy.

[0073] In step S5, the adaptive lighting scheduling model based on deep reinforcement learning needs to achieve long-term power consumption optimization through state quantization, policy optimization, and dynamic exploration. The specific steps are as follows: First, quantitative modeling of the state space and action space is performed, defining the operating state of the power plant's nighttime lighting system as a multi-dimensional vector. ,in: The average light intensity in the factory area (unit: lux). The total power (in kW) of the currently activated lighting area. The distance between personnel and the nearest critical equipment (in meters). The time a person spends in the current area (in seconds). This represents the number of activated lighting areas. The action space $A$ is defined as the combination of the lighting threshold adjustment and the power regulation coefficient, i.e. ,in (Adjustment value for the light trigger threshold, in lux) (Power regulation coefficient, 1.0 is the rated power).

[0074] Next, a multi-objective reward function is designed, with the minimization of long-term power consumption as the core, while also considering the timeliness and stability of lighting response. The formula for calculating the reward value $R$ is as follows: ; Among them: Electricity consumption reward items , The decision cycle is set to 5 seconds. High power consumption is punished by negative rewards; Timeliness of response reward items , The delay time for lighting to start after personnel enter the area; , The smaller the delay, the higher the reward; Strategy stability reward items , , Punishment involves frequent and drastic adjustments to behavior; The weighting coefficients satisfy ,and (Prioritize energy conservation) , .

[0075] This formula uses a multi-dimensional weighted balance between energy saving and practical needs to guide the model to learn a strategy that takes into account both long-term power saving and inspection convenience.

[0076] Subsequently, a dual-network structure for the PPO algorithm is constructed, comprising a policy network and a value network: The policy network employs a 4-layer convolutional neural network (CNN). The input is a state vector $S$, and features are extracted using 1×1 convolutional kernels (64 channels in the first layer and 32 channels in the second). The output layer uses the Softmax activation function to generate the action probability distribution. , For network parameters; The value network employs a 3-layer fully connected network. The input is a state vector $S$, with 128 and 64 neurons in the hidden layers, respectively. The output layer is the state value estimate. ( (Network parameters).

[0077] The core optimization objective of PPO is the surrogate loss function with pruning, as shown in the formula: ;in: For the dominant function, Discount factor; This is the cutting factor; , The regularization coefficient is used. Encouraging exploration through policy entropy. This formula, by limiting the policy update magnitude and introducing entropy regularization, avoids the model getting trapped in local optima, thus improving training stability.

[0078] Furthermore, by combining the dynamic ε-greedy exploration mechanism to adjust the strategy, the exploration probability is increased. The value of $t$ adapts adaptively with the number of training iterations, as shown in the formula: ; in (Initial exploration rate) (Minimum exploration rate) (Maximum number of iterations). For example, the first 1000 iterations of training. The value linearly decreases from 0.9 to 0.3, and then gradually converges to 0.05, balancing the need to explore new strategies with the need to utilize known optimal strategies.

[0079] Finally, model training and adaptive scheduling are performed. The training process is as follows: ① Collect state data S every 5 seconds, generate action $A$ based on the current policy and execute it, and record the reward $R$; ② Calculate the advantage function every 200 trajectory data points. and target value ③ Use the Adam optimizer (learning rate) Minimize the loss function ④ After training converges (loss function fluctuation less than 0.01), the model dynamically outputs the lighting threshold based on the real-time state $S$. and power coefficient For example, when a high-frequency movement of people is detected in an area, the light trigger threshold is automatically reduced from 80 lux to 50 lux, while the power coefficient is stabilized at 0.8.

[0080] Through the above algorithm, the model can continuously optimize the scheduling strategy during long-term operation. For example, it can gradually increase the lighting threshold in sparsely populated areas (reducing unnecessary lighting) and pre-light the necessary inspection paths 0.5 seconds in advance (improving response speed). Ultimately, it can reduce power consumption by more than 35% compared to the traditional mode while meeting the inspection lighting requirements.

[0081] (A policy network and value network for the PPO algorithm are constructed using the TensorFlow framework). An average light intensity in the factory area is collected using a light sensor (TSL2561), and the total power of the activated lighting area is obtained through a current transformer (SCT-013). Combined with personnel location and dwell time data from cameras, a state vector is constructed. ; Write a reward function calculation module in Python, based on... The formula generates the reward value, where , , During the training phase, a power plant lighting simulation environment was built using Gazebo. Every 200 trajectories accumulated, an Adam optimizer (learning rate) was used. Update network parameters via Implementing ε-greedy exploration ( , During deployment, the lighting threshold adjustment amount output by the model is achieved via the OPCUA protocol. and power coefficient The signal is sent to the smart lighting controller, which controls the dimming module (TRIAC dimmer) to adjust the power. The entire process uses Kubernetes containerization to manage the algorithm components, ensuring real-time communication and collaborative operation with sensors and actuators.

[0082] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A method for scheduling nighttime lighting in a power plant, characterized by: The method includes: S1. Utilize existing security cameras and combine them with the improved YOLOv4-tiny algorithm to monitor personnel positions in real time. When personnel enter a specific patrol area, the lighting scheduling mechanism for that area is triggered. S2. Deploy an array of infrared sensors with Fresnel lenses at key equipment locations and process the signals using a wavelet transform algorithm. After cross-verification of the infrared and camera detection results, transmit the signal to the central controller via ZigBee to activate the full-power brightening mode for the area's lights. S3. Introduce a lighting control model based on fuzzy Petri nets, taking personnel position, movement trajectory, and ambient light intensity as inputs, and generate the optimal lighting control strategy through fuzzy inference rules; S4. An improved MFCC algorithm combined with HMM is used to recognize voice commands; when personnel enter the restricted area, a warning voice is automatically played, and multi-channel Wiener filtering technology is used to ensure clarity; at the same time, the differential privacy-preserving LBP algorithm is used to analyze personnel's facial expressions, and a level 2 alarm is triggered when an anomaly is detected; S5. Construct an adaptive lighting scheduling model based on deep reinforcement learning, adopt the PPO algorithm, and dynamically adjust the lighting threshold and response strategy through the ε greedy exploration mechanism to minimize long-term power consumption.

2. The power plant nighttime lighting scheduling method according to claim 1, Its characteristic is: In step S1: Video streams are acquired from existing security cameras in the power plant, and adaptive grayscale correction is performed on each frame of the image; An improved depthwise separable convolutional structure is used to extract features from the preprocessed image. First, depthwise convolution is performed on the input feature map. Where H and W are the image height and width, and C is the number of channels, with each channel corresponding to a 3×3 convolution kernel. The calculation formula is: ； in Output is a depthwise convolution; Then perform point convolution, using a 1×1 convolution kernel. C' represents the number of output channels, calculated as follows: ; Introducing the Ghost module to generate redundant features: for feature maps output by depthwise separable convolutions. Additional features are generated through linear transformation. The calculation formula is: ； in Here, P is the weighting parameter, and P is the number of neighboring pixels. For the transformation coefficients, This is the pixel offset; An improved anchor box matching mechanism is adopted for the generated candidate boxes. x, y are the center point coordinates, w, h are the width and height, and the bounding box is the center point coordinates. The matching degree is calculated as follows: ； in , IoU is the intersection-union ratio. As the center point of the true bounding box, when When the match is successful, it is considered a successful match. Coordinates of the detected personnel images Converted to physical coordinates of the factory area through perspective transformation The transformation formula is: , where K is the camera intrinsic parameter matrix, [R|t] is the extrinsic parameter matrix, and Z is the distance from the person to the camera; Preset the physical coordinate range of a specific patrol area ,when and If personnel continuously enter the area for more than 0.5 seconds, the lighting dispatch mechanism for that area will be triggered. ,otherwise .

3. The power plant nighttime lighting scheduling method according to claim 2, characterized in that: First, the parameters of the existing security cameras in the power plant are calibrated to ensure coverage of all patrol areas that need to be monitored. Then, the video streams captured by the cameras are preprocessed in real time using OpenCV software. The NVIDIA Jetson Nano was selected as the edge computing module, and the improved YOLOv4-tiny algorithm was deployed using the TensorFlow framework. The algorithm's depthwise separable convolution and Ghost module were used for feature extraction and object detection. A Python program was used to calculate candidate bounding box matching and locate people. The program's formulas are based on the formula... ; A coordinate transformation program was written in C++ to convert image coordinates into physical coordinates of the factory area based on the perspective transformation formula. The physical coordinate range of the patrol area was preset. When personnel are detected entering and the duration condition is met, a trigger signal is sent to the lighting control module with Arduino Mega as the core via the MQTT protocol. This module adjusts the lighting equipment in the corresponding area through the relay module. At the same time, Docker is used to encapsulate the algorithm running environment.

4. The power plant nighttime lighting scheduling method according to claim 1, characterized in that: Step S2 also includes the following steps: N pyroelectric infrared sensors with Fresnel lenses are deployed in an equilateral triangle pattern at the locations of key equipment such as power distribution cabinets. The lens focal length is set to 8-12mm and the viewing angle is adjusted to 50°-60° to ensure that the detection range covers an area with a radius of 5-8 meters centered on the key equipment and that the overlap rate of the detection range of adjacent equipment is ≥30%. The infrared sensing device acquires the raw infrared radiation signal sequence s(t) at a sampling frequency of 1 kHz, and performs baseline drift removal on the signal, as shown in the formula: Where M is the size of the sliding window; The preprocessed signal was decomposed into three levels using the db4 wavelet to obtain approximation coefficients. and detail coefficient Calculate the noise reduction threshold , For noise standard deviation estimation, where N is the signal length, the effective signal is reconstructed using inverse wavelet transform. ,in The detail coefficient after noise reduction; Calculate the energy value of the reconstructed signal within a 100ms detection window. When E≥ the offline training threshold, the corresponding infrared device outputs a detection signal dᵢ=1; otherwise, dᵢ=0, where i is the device number. Cross-validation of infrared and camera detection results is performed. Let the camera detection result be C, where 1 indicates a person was detected and 0 indicates no person was detected. The overall judgment result of the infrared array is then determined. When C=1 and D≥60-70% of the infrared devices detect a person, a trigger signal T=1 is output to verify the reliability. When R ≥ 0.8, the judgment result is reliable; The signal is transmitted via a ZigBee wireless self-organizing network. The data includes the device ID, detection result, timestamp, and CRC checksum. After receiving the T=1 signal, the central controller controls the lighting driver module to start the full-power brightening mode of the lights in that area through the RS485 interface.

5. The power plant nighttime lighting scheduling method according to claim 1, characterized in that: Step S3 also includes the following steps: The physical coordinates (X, Y) of the factory area are mapped to a fuzzy subset {near core area, mid-distance area, edge area}. The membership formula for the near core area is: ,in The attenuation coefficient is... The coordinates of the device center are given, and the membership functions for the mid-range and edge regions correspond to the following: and ; The movement speed v and direction angle are calculated by changing the position over three consecutive frames. The mapping is a fuzzy subset of {stationary, slow-moving, high-speed moving}, and the membership formula for high-speed moving is: Low speed corresponds to 0.3-0.5 m / s, and stationary corresponds to ≤0.3 m / s; Illuminance is collected by a light sensor and mapped to a fuzzy subset {low illuminance, medium illuminance, high illuminance}. The membership formula for low illuminance is: L is the real-time illuminance collected by a light sensor. This is the critical value for low light intensity at night; Construct a fuzzy Petri net model: This model includes the locations, movement trajectories, and light intensity of people in the input database; intermediate inference states in the intermediate database; and lighting power levels and illuminated areas in the output database. Define 12 inference rules, each associated with a weight matrix and a trigger threshold. ; Calculate the activation degree of the input library for the transition. , Let be the weight of the transition t for the i-th input library. For the corresponding membership degree, when the activation degree When a transition is triggered, the formula for updating the membership of the output library is: ,in To assess the credibility of the changes; The centroid method is used to defuzzify the membership degree of the output library. The formula for the lighting power command value is as follows: , For preset power levels, For the corresponding membership degree; The formula for the range of illuminated areas is: ,in , v represents the movement speed.

6. A method for scheduling nighttime lighting in a power plant according to claim 5, characterized in that: A fuzzy Petri net model was constructed using MATLAB's Fuzzy Logic Toolbox; Implement centroid-based deblurring using Python, based on... Generate power command, press The lighting range is determined, and control commands are sent to the lighting controller via the Modbus protocol to adjust the lighting equipment in the corresponding area. The entire algorithm environment is encapsulated using Docker.

7. The power plant nighttime lighting scheduling method according to claim 1, characterized in that: Step S4 also includes the following steps: Deploy 4-5 unit omnidirectional microphone arrays in key areas such as power distribution cabinets, set the sampling frequency to 16-18kHz, and use an STM32 microcontroller to collect voice signals and perform pre-emphasis processing. Improved MFCC feature extraction was achieved by using Python to call the Librosa library. The frames were divided into 25-30ms frames with a frame length and 10-12ms frame shift, and a Hanning window was added. The 39-dimensional feature vector was calculated by dynamically adjusting the bandwidth of the Mel filter. The HMM model for five types of speech commands was trained using the HTK toolkit, the state transition probabilities were updated using the Baum-Welch algorithm, and the command recognition was completed using the Viterbi algorithm. When personnel enter the restricted area, an audio power amplifier drives a speaker to play a warning message, while MATLAB's Audio Toolbox is used to implement multi-channel Wiener filtering to eliminate echoes. The algorithm crops facial regions from images captured by security cameras, implements the LBP algorithm formula for differential privacy protection using the OpenCV library, adds Laplacian noise, extracts features, and then trains an SVM classifier using the Python Scikit-learn library to analyze facial expressions. When an anomaly is detected, a secondary alarm is triggered through the ZigBee module. The entire algorithm environment is encapsulated using Docker.

8. A method for scheduling nighttime lighting in a power plant according to claim 7, characterized in that: A left-to-right Hidden Markov Model (HMM) is trained using five types of instructions, and the Baum-Welch algorithm is used to update the state transition probabilities. The Viterbi algorithm is used to match and recognize commands, where Let be the probability that a given state is i at time t and transitions to state j at time t+1. Let be the probability of being in state i at time t; Differential privacy-preserving LBP algorithm for facial expression analysis involves cropping a 128×128 grayscale image of the face from the image. Calculate the LBP features with Laplace noise. The expression is identified by an SVM classifier, and a level 2 alarm is triggered when an anomaly is detected. in The neighborhood radius, For noise, For sensitivity, Budget for privacy.

9. The power plant nighttime lighting scheduling method according to claim 1, characterized in that: Step S5 also includes the following steps: The policy network for the PPO algorithm is built in the central controller using the TensorFlow framework; The average light intensity of the factory area is collected by a light sensor, and the total power of the activated lighting area is obtained through a current transformer. Combined with the personnel location and dwell time data from the security camera in step S1, a state vector is constructed. ,in, This represents the average light intensity in the factory area. The total power of the currently activated lighting area. The distance between personnel and the nearest critical equipment. The duration of a person's stay in the current area. This represents the number of areas where lighting has been activated. Write a reward function calculation module using Python, based on... The formula generates the reward value, where, This is an electricity consumption reward item. In response to the timeliness reward program, The strategy stability reward item has the required weighting coefficient. ; During the training phase, a power plant lighting simulation environment was built using Gazebo. Every 200-400 trajectories accumulated, the network parameters were updated using the Adam optimizer. To achieve ε-greedy exploration, where The initial exploration rate, To minimize the exploration rate, This represents the maximum number of iterations. During deployment, the lighting threshold adjustment amount output by the model is transmitted via the OPC UA protocol. and power coefficient The signal is sent to the smart lighting controller, which controls the dimming module to adjust the power. The entire process uses Kubernetes containerized management of algorithm components.

10. A method for scheduling nighttime lighting in a power plant according to claim 9, characterized in that: Construct a dual-network structure for the PPO algorithm, including a policy network and a value network: The policy network employs a 4-layer convolutional neural network. The input is a state vector S, which extracts features using a 1×1 convolutional kernel. The output layer uses a softmax activation function to generate the action probability distribution. , For network parameters; The value network employs a 3-layer fully connected network. The input is a state vector S, with 128 and 64 neurons in the hidden layers, respectively. The output layer is the state value estimate. , For network parameters; The core optimization objective of PPO is the surrogate loss function with pruning, as shown in the formula: ; in: For the dominant function, Discount factor; This is the cutting factor; , The regularization coefficient is used. Let be the policy entropy.

Images