Intelligent illumination and energy consumption optimization system for underground pipe gallery

Abstract

The invention relates to the technical field of underground pipe gallery distributed intelligent control, and discloses an underground pipe gallery intelligent illumination and energy consumption optimization system, which comprises a dynamic credibility evaluation module for analyzing multi-modal environment sensing data in real time and generating a dynamic credibility score; the edge-cloud collaborative optimization module is used for realizing distributed collaborative learning through local model fine tuning and global knowledge distillation; and the event-driven resource allocation module dynamically optimizes a lighting strategy and a communication priority based on credibility scores and emergency information, and through multi-modal data cross validation and adaptive weight adjustment, the robustness of anomaly detection is significantly improved, and misjudgment caused by sensor failure or environmental interference is effectively suppressed; the dynamic planning algorithm is used for synchronously optimizing energy consumption and equipment service life, efficient balance of safety and energy efficiency is achieved, the system is particularly suitable for complex underground environments and can automatically adapt to sudden anomalies, and the intelligent level and the energy utilization efficiency of pipe gallery illumination are remarkably improved.

Description

Technical Field

[0001] The present invention relates to an intelligent lighting and energy consumption optimization system for an underground pipe gallery, belonging to the technical field of distributed intelligent control of underground pipe galleries. Background Art

[0002] Underground utility corridors are a vital component of urban infrastructure. The intelligentization and energy management of their lighting systems directly impact operational efficiency and safety. Currently, lighting control in the complex environments of utility corridors relies primarily on fixed thresholds or single sensor data to trigger dimming. For example, illumination sensors are used to determine activity areas and adjust brightness accordingly. While these methods can achieve basic energy savings, they present the following key challenges:

[0003] 1. Insufficient dynamic perception capabilities: Existing systems struggle to integrate the spatiotemporal correlations of multimodal environmental data (such as infrared thermal imaging and vibration signals), resulting in a high rate of misjudgment of abnormal conditions. For example, traditional solutions use infrared sensors to detect human activity, but heat dissipation from equipment or condensation reflections within the tunnel can easily trigger misjudgments, requiring manual review and resulting in inefficient response. 2. Failure of edge node coordination: In distributed control, edge nodes typically only execute preset commands or upload raw data, lacking the ability to dynamically assess the credibility of local data. When some sensors fail due to aging or interference, the system still generates global policies based on untrusted data, leading to inaccurate lighting intensity and wasted communication resources. 3. Delayed emergency response: Existing technologies rely on central servers to handle emergencies (such as water seepage and equipment failure), resulting in high latency from data upload to command issuance. Although some solutions attempt to deploy lightweight models at edge nodes, static models cannot adapt to dynamic changes in the tunnel environment (such as seasonal temperature and humidity fluctuations) and still require frequent data backhaul for calibration, making it difficult to meet real-time requirements.

[0004] To address these issues, the industry has attempted to improve reliability by adding redundant sensors or shortening communication cycles. However, such improvements significantly increase hardware costs and energy consumption, and cannot fundamentally resolve the contradiction between multi-source data conflicts and rigid resource allocation. For example, while adding redundant sensors can reduce the risk of single node failure, it exacerbates communication channel congestion and, in turn, increases the transmission delay of critical data. Furthermore, edge nodes in traditional distributed architectures lack autonomous decision-making capabilities and are still limited by the computing power bottleneck of central nodes, making it difficult to achieve coordinated optimization of safety and energy efficiency in complex environments. Therefore, how to build an intelligent lighting control system that balances dynamic perception accuracy, edge collaboration efficiency, and emergency response speed has become a technical challenge that urgently needs to be overcome in the field of underground tunnel operation and maintenance. Summary of the Invention

[0005] The present invention provides an intelligent lighting and energy consumption optimization system for an underground pipeline corridor, the main purpose of which is to solve the problems of insufficient dynamic perception capability, failure of edge node coordination and delayed emergency response.

[0006] To achieve the above objectives, the present invention provides an underground pipe gallery intelligent lighting and energy consumption optimization system, the system comprising:

[0007] The dynamic credibility assessment module is configured to receive and analyze multimodal environmental perception data, including illumination data, infrared thermal imaging data, and vibration signals, from multiple edge nodes in real time. For each edge node, the dynamic credibility assessment module compares the real-time collected multimodal environmental perception data with the lightweight anomaly detection model built into the edge node to generate a data credibility score for the edge node. Furthermore, the dynamic credibility assessment module analyzes the fluctuation patterns of the historical multimodal environmental perception data of each edge node based on a sliding time window and adaptively adjusts the weight of different sensors in the data credibility score calculation to generate a dynamic credibility score.

[0008] The edge-cloud collaborative optimization module is configured to fine-tune the pre-trained illumination prediction model on the edge node side using locally collected multimodal environmental perception data, updating only the key layer parameters of the pre-trained illumination prediction model to generate a local illumination prediction model, and uploading the gradient information of the local illumination prediction model to the cloud. On the cloud side, the edge-cloud collaborative optimization module aggregates the received local illumination prediction model gradient information of each edge node to generate a lightweight global knowledge module, and periodically distributes the global knowledge module to each edge node to update the local illumination prediction model of each edge node to achieve collaborative optimization.

[0009] The event-driven resource allocation module is configured to build a lighting strategy generator based on dynamic programming. The lighting strategy generator uses the preset minimum energy consumption as a constraint condition and generates and issues lighting strategies for the lighting intensity and emergency response speed of the underground pipeline corridor based on the data credibility score of each edge node and the received emergency information. In addition, the event-driven resource allocation module adopts a time-slotted time-division multiple access protocol to dynamically allocate communication priorities based on the data credibility score of each edge node, so that edge nodes with higher data credibility scores prioritize uploading key data, including multimodal environmental perception data and emergency information.

[0010] Preferably, the dynamic credibility assessment module further introduces a sliding time window entropy analysis in the process of calculating the dynamic credibility score to distinguish the normal operation state and the sudden abnormal state of the underground pipeline corridor, and adjusts the weight distribution logic of different sensors in the data credibility score calculation according to the distinction results, so that in the sudden abnormal state, the sensor with a higher correlation with the abnormal state is given a higher weight.

[0011] Preferably, during the process of dynamic distillation of global knowledge in the edge-cloud collaborative optimization module, the cloud side analyzes and extracts the feature importance in the local illumination prediction model of each edge node through the attention mechanism, and generates differentiated knowledge distillation strategies based on the feature importance to guide the generation and distribution of lightweight global knowledge modules.

[0012] Preferably, the event-driven resource allocation module integrates an equipment aging prediction model, which predicts the remaining life of lighting equipment and related sensors in the tunnel based on historical operating data. The lighting strategy generator further dynamically adjusts the lighting power upper limit of the lighting equipment based on the remaining life, so as to ensure safe lighting while reducing the increase in energy consumption that may be caused by equipment aging.

[0013] Preferably, the lightweight anomaly detection model built into the edge node is an improved long short-term memory neural network model. The improved long short-term memory neural network model generates node-level abnormal state prediction results by analyzing the time series characteristics of illumination data, infrared thermal imaging data and vibration signals, and serves as one of the bases for generating data credibility scores.

[0014] Preferably, the edge node is further equipped with a lightweight pulse neural network. When the lightweight pulse neural network detects that the preset illumination gradient change rate exceeds a threshold, it automatically triggers the reorganization of neuron connection weights and dynamically adjusts the complexity of the local illumination prediction model on the edge node side to improve the model inference speed while ensuring the prediction accuracy.

[0015] Preferably, the edge node is configured to use Shannon entropy to quantify the uncertainty of its own operating state. When the entropy value of the uncertainty exceeds a preset threshold, it automatically initiates a local collaboration request to the adjacent edge node to request participation in model calibration. The local collaboration request triggers the directed ripple algorithm, which only wakes up the adjacent edge nodes within the preset number of hops to participate in the model calibration.

[0016] Preferably, the edge node reuses its built-in infrared sensor data to construct a light intensity-thermal radiation coupling matrix, and extracts the water seepage feature fingerprint through singular value decomposition to distinguish between real water seepage and pipeline condensation water in the underground pipeline corridor, and uploads the distinction results as part of the emergency event information to the event-driven resource allocation module.

[0017] Preferably, the edge node is further configured with an environmental noise adaptive stripping unit, which dynamically adjusts the singular value decomposition weight parameters of the multi-spectral coupling matrix by comparing the frequency domain characteristics of the current light field data with a pre-stored typical interference spectrum feature library in real time to suppress the characteristic interference caused by non-water seepage factors.

[0018] Preferably, the environmental noise adaptive stripping unit is also integrated with a spatiotemporal correlation analysis module of the interference source. The spatiotemporal correlation analysis module dynamically corrects the priority weights of the typical interference spectrum feature library through collaborative calibration between edge nodes, specifically including: when adjacent edge nodes detect interference in the same frequency band and the interference intensity fluctuates synchronously over time, it is determined to be a systematic interference source and its spectrum feature priority is automatically increased; when the interference only exists in a single node and the intensity fluctuates randomly, it is determined to be a local instantaneous interference, its priority is automatically lowered and the light field data cache backtracking mechanism is activated to avoid residual effects.

[0019] Compared with the background technology problems, the beneficial effects of the present invention are:

[0020] 1. Through cross-validation of multimodal environmental perception data and sliding time window entropy analysis, the system can adaptively distinguish between normal operation and sudden abnormal states. The dynamic credibility scoring mechanism effectively suppresses the risk of misjudgment caused by single sensor failure or environmental interference by adjusting the sensor weight ratio in real time. Combined with the directional collaborative calibration between edge nodes based on Shannon entropy, it quickly triggers model calibration between adjacent nodes when local data confidence is insufficient, forming an anomaly detection network with dual coverage in time and space, significantly improving the robustness of data acquisition and decision reliability in complex underground environments.

[0021] 2. Edge nodes retain personalized environmental characteristics by fine-tuning parameters of key layers of local models, while the cloud uses an attention mechanism to extract common patterns across nodes to generate lightweight global knowledge modules. This architecture avoids the communication bottleneck of traditional federated learning and achieves a dynamic balance between edge-side inference speed and cloud-side knowledge generalization capabilities. When sudden anomalies cause communication interruptions, edge nodes can autonomously generate emergency strategies based on local models and historical data to ensure the system's continuous operation capabilities under extreme conditions.

[0022] 3. The event-driven resource allocation strategy deeply integrates the equipment aging prediction model with the real-time credibility score. Through the dynamic programming algorithm, it simultaneously optimizes lighting energy consumption and equipment life cycle. The time-slot communication protocol dynamically adjusts the data transmission priority based on the node credibility, so that critical data of high-trust nodes is uploaded first. While ensuring the real-time response of the system, it also reduces the redundant communication energy consumption. This mechanism effectively avoids the technical contradiction between energy efficiency optimization and security monitoring in traditional solutions.

[0023] 4. By reusing infrared sensor data to construct a light intensity-thermal radiation coupling matrix, and combining it with singular value decomposition, we extract environment-specific water seepage fingerprints. The environmental noise stripping unit uses frequency domain feature comparison and dynamic weight adjustment to precisely suppress the impact of interference factors such as dust reflection and equipment obstruction on water seepage determination. The lightweight spiking neural network built into the edge node monitors the rate of change of the illumination gradient to trigger the reorganization of neuronal connection weights, achieving adaptive adjustment of model complexity. This mechanism enables the system to maintain low-power inference during normal operation and instantly enhances the model's feature extraction capabilities when abnormal signs are detected. This creates a dynamic match between computing resources and prediction accuracy, avoiding the rigid constraints of traditional fixed-architecture models between real-time performance and energy efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a flow chart of the edge-cloud collaborative model optimization and knowledge update interaction of the present invention;

[0025] Figure 2 This is the core functional module structure and interaction logic diagram of the present invention;

[0026] Figure 3 This is a structural diagram of the collaborative model optimization between the edge nodes and the cloud based on global knowledge distillation in the present invention.

[0027] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0028] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0029] The present application provides an intelligent lighting and energy consumption optimization system for an underground pipe gallery, the system comprising:

[0030] The dynamic credibility assessment module is configured to receive and analyze multimodal environmental perception data, including illumination data, infrared thermal imaging data, and vibration signals, from multiple edge nodes in real time. For each edge node, the dynamic credibility assessment module compares the real-time collected multimodal environmental perception data with the lightweight anomaly detection model built into the edge node to generate a data credibility score for the edge node. Furthermore, the dynamic credibility assessment module analyzes the fluctuation patterns of the historical multimodal environmental perception data of each edge node based on a sliding time window and adaptively adjusts the weight of different sensors in the data credibility score calculation to generate a dynamic credibility score.

[0031] The edge-cloud collaborative optimization module is configured to fine-tune the pre-trained illumination prediction model on the edge node side using locally collected multimodal environmental perception data, updating only the key layer parameters of the pre-trained illumination prediction model to generate a local illumination prediction model, and uploading the gradient information of the local illumination prediction model to the cloud. On the cloud side, the edge-cloud collaborative optimization module aggregates the received local illumination prediction model gradient information of each edge node to generate a lightweight global knowledge module, and periodically distributes the global knowledge module to each edge node to update the local illumination prediction model of each edge node to achieve collaborative optimization.

[0032] The event-driven resource allocation module is configured to build a lighting strategy generator based on dynamic programming. The lighting strategy generator uses the preset minimum energy consumption as a constraint condition and generates and issues lighting strategies for the lighting intensity and emergency response speed of the underground pipeline corridor based on the data credibility score of each edge node and the received emergency information. In addition, the event-driven resource allocation module adopts a time-slotted time-division multiple access protocol to dynamically allocate communication priorities based on the data credibility score of each edge node, so that edge nodes with higher data credibility scores prioritize uploading key data, including multimodal environmental perception data and emergency information.

[0033] Preferably, the dynamic credibility assessment module further introduces a sliding time window entropy analysis in the process of calculating the dynamic credibility score to distinguish the normal operation state and the sudden abnormal state of the underground pipeline corridor, and adjusts the weight distribution logic of different sensors in the data credibility score calculation according to the distinction results, so that in the sudden abnormal state, the sensor with a higher correlation with the abnormal state is given a higher weight.

[0034] Preferably, during the process of dynamic distillation of global knowledge in the edge-cloud collaborative optimization module, the cloud side analyzes and extracts the feature importance in the local illumination prediction model of each edge node through the attention mechanism, and generates differentiated knowledge distillation strategies based on the feature importance to guide the generation and distribution of lightweight global knowledge modules.

[0035] Preferably, the event-driven resource allocation module integrates an equipment aging prediction model, which predicts the remaining life of lighting equipment and related sensors in the tunnel based on historical operating data. The lighting strategy generator further dynamically adjusts the lighting power upper limit of the lighting equipment based on the remaining life, so as to ensure safe lighting while reducing the increase in energy consumption that may be caused by equipment aging.

[0036] Preferably, the lightweight anomaly detection model built into the edge node is an improved long short-term memory neural network model. The improved long short-term memory neural network model generates node-level abnormal state prediction results by analyzing the time series characteristics of illumination data, infrared thermal imaging data and vibration signals, and serves as one of the bases for generating data credibility scores.

[0037] Preferably, the edge node is further equipped with a lightweight pulse neural network. When the lightweight pulse neural network detects that the preset illumination gradient change rate exceeds a threshold, it automatically triggers the reorganization of neuron connection weights and dynamically adjusts the complexity of the local illumination prediction model on the edge node side to improve the model inference speed while ensuring the prediction accuracy.

[0038] Preferably, the edge node is configured to use Shannon entropy to quantify the uncertainty of its own operating state. When the entropy value of the uncertainty exceeds a preset threshold, it automatically initiates a local collaboration request to the adjacent edge node to request participation in model calibration. The local collaboration request triggers the directed ripple algorithm, which only wakes up the adjacent edge nodes within the preset number of hops to participate in the model calibration.

[0039] Preferably, the edge node reuses its built-in infrared sensor data to construct a light intensity-thermal radiation coupling matrix, and extracts the water seepage feature fingerprint through singular value decomposition to distinguish between real water seepage and pipeline condensation water in the underground pipeline corridor, and uploads the distinction results as part of the emergency event information to the event-driven resource allocation module.

[0040] Preferably, the edge node is further configured with an environmental noise adaptive stripping unit, which dynamically adjusts the singular value decomposition weight parameters of the multi-spectral coupling matrix by comparing the frequency domain characteristics of the current light field data with a pre-stored typical interference spectrum feature library in real time to suppress the characteristic interference caused by non-water seepage factors.

[0041] Preferably, the environmental noise adaptive stripping unit is also integrated with a spatiotemporal correlation analysis module of the interference source. The spatiotemporal correlation analysis module dynamically corrects the priority weights of the typical interference spectrum feature library through collaborative calibration between edge nodes, specifically including: when adjacent edge nodes detect interference in the same frequency band and the interference intensity fluctuates synchronously over time, it is determined to be a systematic interference source and its spectrum feature priority is automatically increased; when the interference only exists in a single node and the intensity fluctuates randomly, it is determined to be a local instantaneous interference, its priority is automatically lowered and the light field data cache backtracking mechanism is activated to avoid residual effects.

[0042] Example 1: In an underground comprehensive pipeline corridor with a length of more than three kilometers, which is equipped with multiple branch ventilation nodes and sections with significant changes in temperature and humidity gradients, in order to realize the deployment and operation of the system described in the present invention, an edge node is first set up every hundred meters. The node integrates an illumination sensor, an infrared thermal imaging device and a micro-vibration monitor to form a multimodal perception unit. At the same time, a set of local anomaly detection models obtained by training based on an improved long short-term memory neural network is preset, and has the ability to communicate periodically with the cloud model server. The cloud deploys a global knowledge module generator and a dynamic parameter management unit to aggregate the gradient information reported by the edge nodes and uniformly control the iteration direction of the model. During typical operation, the edge nodes periodically collect local environmental perception data and judge the current state through the local model. Whether there are signs of abnormality, the dynamic credibility scoring process is started at the same time. The specific operations are as follows: first, a sliding window with a time length of five minutes is constructed to extract the fluctuation amplitude distribution characteristics of the multimodal data in the window. The sliding time window entropy values ​​of the illumination channel, thermal imaging channel and vibration channel are calculated respectively through the information entropy calculation module to evaluate the information uncertainty of each type of data channel in the current environment. Under normal conditions, the fluctuations of the illumination and thermal imaging entropy values ​​tend to be stable. If the entropy value of a certain type of channel is detected to rise rapidly or jitter greatly, it is preliminarily judged as a potential abnormality or a decrease in sensor stability, thereby triggering the credibility weight adjustment mechanism. The entropy value change rate is used as the input signal to be mapped to a weight adjustment function constructed by a set of empirical rules, and the participation coefficient of the fluctuation channel is automatically lowered to improve the stability of data fusion.

[0043] In order to improve the system's response capability to typical sudden water seepage events, a light intensity-thermal radiation coupling processing unit is installed in some nodes. The specific work flow includes: converting the heat map extracted by the infrared thermal imaging module into a quantitative thermal radiation matrix, and performing point-by-point registration with the light intensity matrix at the same position at the current moment to form a two-dimensional coupling matrix. After centralization, it enters the singular value decomposition module, extracts the principal component sequence and performs similarity comparison with the principal component template in the pre-stored water seepage sample feature database. If the similarity exceeds the set threshold, the system determines it as a water seepage event and generates an emergency event label. This label is submitted to the event-driven resource allocation module together with the original data through a high-priority channel. In the edge node, when it is detected that the illumination gradient change rate exceeds the preset threshold, which usually occurs in the scenario of personnel intrusion or sudden light outage, the weight reconstruction module of the pulse neural network will be immediately activated. Its mechanism is based on the statistical judgment of the change amplitude and historical distribution characteristics of the current input signal, and the reconstruction is carried out. The trigger calculates a redistribution strategy for the connection weights between neurons, simplifies the propagation relationship on non-critical paths, thereby shortening the model inference path while maintaining the accuracy of anomaly detection, improving the edge response speed, and avoiding the emergency hysteresis problem caused by inference delay in traditional fixed models. In addition, when the uncertainty entropy value of the operating state calculated by the edge node exceeds a set upper limit threshold, the system believes that the node's independent judgment ability in the current environment has declined, and automatically initiates a local collaborative request to the adjacent node. The collaborative request contains the current model state summary of the initiating node and the request range identifier, thereby triggering the directed ripple algorithm, which limits the request to a set of nodes with a hop count of less than three. By constructing a local propagation graph in the topology layer, only the adjacent nodes within the target hop count are activated to participate in collaborative calibration. Each awakened node returns its local model feature summary and credibility score. The edge node updates its own model structure or parameter configuration accordingly to improve the overall prediction robustness.

[0044] The event-driven resource allocation module in the system further integrates a lighting equipment aging prediction model. The model uses the historical lighting duration, cumulative working cycles, average temperature and humidity, and startup frequency of each node as input variables, estimates the current remaining life of each device through a regression model, maps the life value to a lighting power upper limit adjustment factor, and controls the maximum lighting intensity that can be called by the lighting strategy generator of the node; for example, when the remaining life of a device is less than 20% of the set threshold, the system limits its maximum brightness to below 70% to reduce the heat load and delay aging. At the same time, its failure probability is incorporated into the credibility score, thereby indirectly affecting its data upload priority, forming an adaptive adjustment closed loop. During the operation of the entire system, the collaborative optimization process between the edge nodes and the cloud is based on periodic fine-tuning. Every fifteen minutes, the edge nodes report the key gradient parameters of their local models to the cloud. The cloud aggregates the parameters of the entire network to generate a set of lightweight global knowledge modules. After customized configuration based on the importance ranking of each node's features, it is pushed to each node to update its local model structure. This process ensures that the model can take into account local environmental differences and achieve collaborative optimization of energy consumption control, fault response and safety assurance in the underground pipeline corridor environment. These are all extended implementation methods known to ordinary technicians in this field.

[0045] Example 2: This example combines Figures 1 to 3 , and explain the implementation of the intelligent lighting and energy consumption optimization system for underground pipe corridors. Figure 1 As shown in the figure, first, edge node A and edge node B respectively perform local data collection operations and fine-tune the key layer parameters of the pre-trained model locally to adapt to the multimodal environmental perception data of the current underground pipeline corridor and form a personalized local model. Subsequently, the two nodes send the uploaded gradient information to the cloud server through an encrypted channel. The cloud server receives the uploaded gradient information from each edge node (encrypted channel) and performs feature importance analysis based on the uploaded content to extract the common features of each node. It further performs knowledge distillation calculations to generate a lightweight global knowledge module. This module is then periodically distributed to each edge node (period 2h) to achieve edge-cloud collaborative optimization. After receiving the distributed global module, the edge node performs model parameter fusion operations and updates the local model parameters to improve prediction accuracy and environmental adaptability. In the above-mentioned interaction process, the system also specially sets up a patented core interaction process, including heartbeat detection and confirmation survival mechanism, to ensure the connection stability between edge node A, edge node B and cloud server and the system's continuous online capability. Ultimately, global knowledge is continuously accumulated into the global knowledge base, forming a complete closed-loop optimization and dynamic update mechanism to ensure the stable, efficient and safe operation of the intelligent lighting system in the complex environment of the underground pipeline corridor.

[0046] like Figure 2As shown, the dynamic credibility assessment module takes multimodal environmental perception data as input, including illumination, infrared thermal imaging, and vibration signals. It identifies potential anomalies using a lightweight anomaly detection model and, combined with a sliding time window entropy analysis method, models the temporal fluctuation patterns of the data to generate a dynamic credibility score. The edge-cloud collaborative optimization module includes four key steps: local model fine-tuning, gradient information upload, global knowledge distillation, and model parameter update. Edge nodes perform local model fine-tuning based on the local perceived environment and upload gradient information. The cloud integrates the gradient data and completes global knowledge distillation, then distributes the updated model parameters to the edge nodes, improving the overall adaptability and generalization capabilities of the system. The event-driven resource allocation module integrates device aging prediction, inferring the remaining lifespan of lighting equipment based on its operating history and pre-regulating subsequent policy generation. The lighting policy generator integrates device status and dynamic credibility scores to formulate optimal lighting intensity and response strategies. Finally, a time-slot communication scheduling mechanism enables multi-node communication priority adjustment.

[0047] like Figure 3 As shown in the figure, in the edge node part, the local model fine-tuning module includes five links: data collection, multimodal data, key layer update, gradient information and gradient upload. The edge node first obtains the multimodal data of the current environment through data collection, and updates the key layer in combination with the local model structure to adapt to the changes in environmental characteristics; then, the edge node generates the gradient information in the update process and transmits it to the cloud through the gradient upload link; in the cloud part, the global knowledge distillation module consists of three components: knowledge aggregation, feature importance analysis and attention mechanism. After the cloud receives the gradient information from multiple edge nodes, it first integrates the information through knowledge aggregation operation, and then identifies the key features through feature importance analysis. Finally, the attention mechanism is used to construct a global knowledge module with differentiated weight distribution, and it is transmitted back to the edge node to guide the update of the key layer.

[0048] Example 3: In a newly built subway underground integrated pipeline corridor project in a certain city, in order to achieve the deployment and optimization of a highly robust intelligent lighting control system in a complex environment, a main pipeline corridor with a length of approximately 2,500 meters was selected as a demonstration application area. This area contains multiple intersections, two ventilation nodes with significant temperature and humidity fluctuations, and a section with a perennial water seepage risk. It has typical scenario characteristics and is suitable for verifying the system's performance in dynamic perception, edge collaboration, and energy consumption optimization. First, in the implementation of the dynamic credibility scoring mechanism, each edge node continuously collects local illumination, infrared thermal imaging and micro-vibration data, and establishes a sliding time window containing thirty groups of continuous data with a sampling period of three seconds. The information entropy of each type of perception data is calculated by the entropy analysis module. The entropy calculation adopts the partition probability modeling method of the signal intensity distribution to generate the data uncertainty index of the current channel; in order to improve the sensitivity of weight adjustment, the absolute value of the first-order difference of the entropy value is introduced as the dynamic signal disturbance index, and the stable period and the disturbance period are distinguished by the empirically set disturbance sensitivity threshold. When the disturbance index of a certain channel exceeds the set threshold in three consecutive rounds of calculation, the system determines that the channel data is in an abnormally unstable state, and immediately reduces its credibility weight to 30% of the original value. At the same time, the fusion weight of the remaining channels is increased, and the re-evaluation cycle begins within two minutes. This mechanism ensures the stability of the data fusion process and avoids local abnormal interference from affecting the overall perception results.

[0049] Secondly, in the lighting strategy generator, in order to balance energy consumption and response time, the optimal lighting scheduling is achieved by constructing a dynamic programming model based on state-action mapping. Specifically, the system defines the current state of the underground pipeline corridor as a combination of the credibility score vector summarized by the edge nodes and the emergency type. For example, when there is no emergency and the average score of all nodes is greater than 80%, it is defined as steady-state operation; in this state, the lighting strategy gives priority to the lowest power output and sets the power upper limit with reference to the remaining life factor of the lighting equipment. On the contrary, when there are nodes with a credibility score lower than 40% or there are event labels such as water seepage and temperature rise, the state is converted to abnormal response, and the strategy switches to high-priority brightness output. At the same time, the corresponding communication priority level of the abnormal node is temporarily increased to ensure timely data reporting; in the state transfer process, a rolling window prediction mechanism is adopted to determine the strategy queue for the next five minutes based on the current and historical states, to ensure that the lighting power switching process is forward-looking and stable.

[0050] For the anomaly detection model and pulse neural network structure within the edge node, this embodiment adopts a hierarchical parameter adjustment method to improve the adaptability of the model. The first layer of the model is a time series feature extraction unit, which processes thermal imaging and vibration data respectively through a dual-channel parallel structure; the second layer is a state judgment layer, which outputs the current state confidence score. When it is detected that the illumination change rate exceeds the set gradient threshold by 50%, the system activates the weight reconstruction module of the pulse neural network and redistributes the connection weights. The specific method is to amplify the path weight corresponding to the critical channel signal by two times and compress the non-critical path to 50% of the original value, thereby shortening the reasoning path and enhancing the abnormal response speed; the setting logic is based on the probability modeling results of illumination mutation caused by sudden entry of people or sudden extinction of light sources in actual scenarios, optimizing the model reasoning efficiency and prediction Dynamic trade-off between accuracy. In addition, in the collaborative calibration mechanism between edge nodes, the system adopts an uncertainty measurement method based on Shannon entropy. When the entropy value of a node exceeds the set threshold for three consecutive cycles and the local score fluctuates violently, it is considered that the judgment ability is insufficient and a directed collaborative request is actively initiated. The collaborative request contains the current node model state summary and the number of neighbor node hops. The system defaults to a maximum hop number of two to avoid information generalization caused by crossing too far. After the target node responds, the edge node compares its model summary with the local state vector, and obtains a temporary enhanced model by weighted averaging. It is only used for the current window judgment task to avoid permanent error migration. This collaborative strategy is mainly aimed at areas with concentrated local anomalies, such as vibration disturbances during construction or intermittent equipment failures to improve the overall perception robustness.

[0051] In the global knowledge distillation module in the cloud, this embodiment further refines the feature importance extraction method. After the cloud server receives the local model key layer gradient information uploaded by all edge nodes, it first performs principal component analysis on each group of gradients, extracts the first two principal components as feature representatives, and then calculates the mean and variance ratio of each type of feature in different nodes by comparing the gradients between nodes. This ratio is used to characterize the degree of commonality and difference of the feature. Features with high differences are given higher attention weights to construct differentiated global knowledge modules; in the downlink module, the edge node receives customized modules based on its own historical scoring performance and communication frequency. The system also retains module version control in the cloud to ensure that it can quickly roll back to the previous stable version in case of abnormal response. In order to reflect the adaptability of the system in real water seepage scenarios, this embodiment sets up an experimental exciter in the water seepage risk section to simulate the process of introducing a low-frequency continuous heat source. The edge node converts the infrared heat map into a two-dimensional thermal radiation matrix, and generates a coupling matrix synchronously with the current illumination distribution. After centralization processing, it is input into the singular value decomposition module, and the top three principal components are selected for similarity comparison with the local preset water seepage feature template. The similarity score is composed of five weighting factors, and the weights are set to 40% matching of the thermal radiation principal component, 30% matching of the illumination feature, 20% consistency of the gradient direction, and 10% similarity of the time series. If the set threshold is exceeded, it is determined as a water seepage event and marked for upload. The scoring structure is set by the expert group based on historical case experience, and has engineering practicality and reproducibility. They are all extended implementation methods that can be known to ordinary technicians in this field.

[0052] Example 4: The intelligent lighting and energy consumption optimization system described in the present invention is deployed in a certain underground integrated pipeline corridor intersection hub area. To improve the dynamic response capability and abnormal robustness of this area under frequent environmental disturbance conditions, the system, based on the original edge node configuration, focuses on introducing a dynamic credibility scoring optimization mechanism, a multimodal fusion weight adaptive control process, and a refined implementation path of the light intensity-thermal radiation coupling matrix identification module. During the system initialization phase, all edge nodes load a local anomaly detection model. This model is composed of an improved long-short-term memory neural network. Its structure includes four layers of components: input gate, forget gate, output gate, and state unit. It is used to process the change pattern in time series data and is particularly suitable for analyzing the complex characteristic behavior under the interaction of illumination fluctuations, thermal radiation changes, and vibration signals. The node collects illumination, thermal imaging, and vibration data every three seconds, and every thirty sets of continuous data form a sliding time window. The data of each channel in the window is first fitted with an interval probability distribution. The signal amplitude is divided into eight segments using equal-width binning. The signal entropy value of the channel is calculated by counting the occurrence frequency of each segment. The entropy value reflects the degree of uncertainty of the channel data in the current time period. At the same time, in order to achieve adaptive linkage between credibility scoring and fusion weight adjustment, in this embodiment, the system introduces an entropy disturbance index as a basis for sensitivity judgment, specifically the average of the absolute values ​​of the first-order differences of the entropy values ​​in three consecutive sliding windows. If the average value exceeds the preset disturbance threshold by 15%, the current fluctuation of the channel is judged to be abnormal, and the weight adjustment module is triggered. The logic of weight adjustment is as follows: the basic weight is set to the average value. When a channel is judged to be abnormally disturbed, its weight is adjusted to 30% of the original value. At the same time, the remaining channels are increased by the same proportion. The system maintains this weight structure within two minutes and then enters a re-evaluation cycle. If the channel returns to stability, the weight will gradually increase, otherwise it will remain in a reduced state. This mechanism effectively improves the stability and sensitivity of multimodal data fusion.

[0053] In the scenario of responding to sudden water seepage incidents, the node's built-in light intensity-thermal radiation coupling processing module converts the infrared thermal map into a two-dimensional thermal radiation matrix after data collection, and uses the pixel coordinates as the index to pair it point by point with the illumination matrix at the same time to form a light intensity-thermal radiation coupling matrix. The matrix is ​​first centralized, that is, the average value of each column of data is subtracted to remove the overall brightness offset. It is then input into the singular value decomposition module to extract the first three principal component vectors to represent the main energy distribution characteristics at that moment; the system compares the principal component combination with the standard fingerprint in the locally preset water seepage feature template library, and determines the degree of matching through cosine similarity calculation. When the similarity is higher than 80%, the system determines it as a possible water seepage incident and automatically generates a label. At the same time, the communication priority of the node is increased to give priority to uploading key data including the original matrix, principal component vectors and judgment results.

[0054] In terms of adaptive updating of local models at edge nodes, the system sets up a pulse neural network structure to respond to sudden states, where the monitoring item is the rate of change of illumination per unit time. If the value exceeds 40%, it is considered a significant mutation, usually caused by personnel entering or light source failure. The system immediately starts the weight reconstruction mechanism, which redistributes the internal connection weights of the model according to the strategy of amplifying key channels and compressing redundant paths, while the weights of non-critical paths connected to vibration and thermal imaging channels are compressed to 50%, thereby focusing computing resources on the input dimensions with the highest recognition value in the current cycle, shortening the reasoning path and improving the response speed. Taking into account that some nodes may have independent To judge unstable situations, this embodiment introduces an uncertainty propagation mechanism based on Shannon entropy. When the overall information entropy of the edge node is higher than the set threshold and the credibility score drops significantly (for example, below 40%) within three consecutive sampling periods, the system automatically determines that its discrimination ability has declined and triggers a directed collaborative request. The collaborative request contains a summary of the current node status (including the summary of the principal components of its current model, the statistical characteristics of the sliding window and the score record), and sets the number of collaborative propagation hops to two to prevent information generalization. After receiving the request, the adjacent node returns its local model summary. After weighted averaging, the initiating node generates a set of temporary enhanced models. This model is only used for current window judgment to avoid long-term deviations due to misfusion.

[0055] On the cloud side, in the global knowledge module generation process, principal component analysis is first performed on the key layer gradients uploaded from multiple edge nodes. After extracting representative feature components, the mean-variance ratio index of the feature between different nodes is calculated. This index measures the commonality and difference of the feature in the entire network; for features with larger variance, the system assigns a higher attention coefficient to guide the distillation module to build a more differentiated knowledge representation. The global knowledge module finally generated on the cloud is sent back to each edge node in a customized manner. The nodes receive module versions of different complexities according to the stability of historical scores. Low-score volatility nodes receive complete modules, and high-volatility nodes only receive basic general modules to avoid misleading updates. These are all extended implementation methods known to ordinary technicians in this field.

[0056] It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Finally, it should be noted that the above embodiments are only intended to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An intelligent lighting and energy consumption optimization system for underground pipe corridors, characterized in that: The system comprises: The dynamic credibility assessment module is configured to receive and analyze multimodal environmental perception data, including illumination data, infrared thermal imaging data, and vibration signals, from multiple edge nodes in real time. For each edge node, the dynamic credibility assessment module compares the real-time multimodal environmental perception data with the lightweight anomaly detection model built into the edge node to generate a data credibility score for the edge node. Furthermore, the dynamic credibility assessment module analyzes the fluctuation patterns of the historical multimodal environmental perception data of each edge node based on a sliding time window and adaptively adjusts the weight of different sensors in the calculation of the data credibility score. The edge-cloud collaborative optimization module is configured to, on the edge node side, fine-tune the pre-trained illumination prediction model using locally collected multimodal environmental perception data, update only the key layer parameters of the pre-trained illumination prediction model, generate a local illumination prediction model, and upload the gradient information of the local illumination prediction model to the cloud. On the cloud side, the edge-cloud collaborative optimization module aggregates the received local illumination prediction model gradient information of each edge node to generate a lightweight global knowledge module, and periodically distributes the global knowledge module to each edge node to update the local illumination prediction model of each edge node. The event-driven resource allocation module is configured to build a lighting strategy generator based on dynamic programming. The lighting strategy generator uses the preset minimum energy consumption as a constraint condition and generates and issues lighting strategies for the lighting intensity and emergency response speed of the underground pipeline corridor based on the data credibility score of each edge node and the received emergency information. In addition, the event-driven resource allocation module adopts a time-slotted time-division multiple access protocol to dynamically allocate communication priorities based on the data credibility score of each edge node, so that edge nodes with higher data credibility scores prioritize uploading key data, including multimodal environmental perception data and emergency information.

2. The underground pipe gallery intelligent lighting and energy consumption optimization system according to claim 1 is characterized in that: In the process of calculating the dynamic credibility score, the dynamic credibility assessment module further introduces sliding time window entropy analysis to distinguish the normal operation state and sudden abnormal state of the underground pipeline corridor. Based on the distinction results, the weight distribution logic of different sensors in the data credibility score calculation is adjusted. In the case of a sudden abnormal state, sensors with a higher correlation with the abnormal state are assigned a higher weight.

3. The underground pipe gallery intelligent lighting and energy consumption optimization system according to claim 2 is characterized in that: During the dynamic distillation of global knowledge in the edge-cloud collaborative optimization module, the cloud side analyzes and extracts the feature importance in the local illumination prediction model of each edge node through the attention mechanism, and generates differentiated knowledge distillation strategies based on the feature importance to guide the generation and distribution of lightweight global knowledge modules.

4. The underground pipe gallery intelligent lighting and energy consumption optimization system according to claim 3 is characterized in that: The event-driven resource allocation module integrates an equipment aging prediction model, which predicts the remaining life of lighting equipment and related sensors in the corridor based on historical operating data. The lighting strategy generator further dynamically adjusts the lighting power upper limit of the lighting equipment based on the remaining life.

5. The underground pipe gallery intelligent lighting and energy consumption optimization system according to claim 1 is characterized in that: The lightweight anomaly detection model built into the edge node is an improved long-short-term memory neural network model. The improved long-short-term memory neural network model generates node-level abnormal state prediction results by analyzing the time series characteristics of illumination data, infrared thermal imaging data, and vibration signals, and serves as one of the bases for generating data credibility scores.

6. The underground pipe gallery intelligent lighting and energy consumption optimization system according to claim 5 is characterized in that: The edge node is further equipped with a lightweight spiking neural network. When the lightweight spiking neural network detects that the preset illumination gradient change rate exceeds a threshold, it automatically triggers the reorganization of neuron connection weights and dynamically adjusts the complexity of the local illumination prediction model on the edge node side.

7. The underground pipe gallery intelligent lighting and energy consumption optimization system according to claim 1 is characterized in that: The edge node is configured to use Shannon entropy to quantify the uncertainty of its own operating status. When the entropy value of the uncertainty exceeds the preset threshold, it automatically initiates a local collaboration request to the adjacent edge node. The local collaboration request triggers the directed ripple algorithm, which only wakes up the adjacent edge nodes within the preset number of hops to participate in model calibration.

8. The underground pipe gallery intelligent lighting and energy consumption optimization system according to claim 1 is characterized in that: The edge node reuses its built-in infrared sensor data to construct a light intensity-thermal radiation coupling matrix, and extracts the water seepage feature fingerprint through singular value decomposition to distinguish between real water seepage and pipeline condensation in the underground pipeline corridor. The distinction results are uploaded to the event-driven resource allocation module as part of the emergency event information.

9. The underground pipe gallery intelligent lighting and energy consumption optimization system according to claim 8 is characterized in that: The edge node is further configured with an environmental noise adaptive stripping unit, which dynamically adjusts the singular value decomposition weight parameters of the multi-spectral coupling matrix by comparing the frequency domain characteristics of the current light field data with a pre-stored typical interference spectrum feature library in real time.

10. The underground pipe gallery intelligent lighting and energy consumption optimization system according to claim 9 is characterized in that: The environmental noise adaptive stripping unit also integrates a spatiotemporal correlation analysis module for interference sources. This module dynamically corrects the priority weights of the typical interference spectrum feature library through collaborative calibration between edge nodes. Specifically, when adjacent edge nodes detect interference in the same frequency band and the interference intensity fluctuates synchronously over time, it is determined to be a systematic interference source and its spectrum feature priority is automatically increased; when the interference only exists in a single node and the intensity fluctuates randomly, it is determined to be local instantaneous interference, its priority is automatically lowered, and the light field data cache backtracking mechanism is activated to avoid residual effects.

Images