A smart leakage monitoring and protection system for charging piles

Abstract

This invention relates to the field of charging pile protection technology and discloses an intelligent leakage monitoring and protection system for charging piles. The system includes a multi-source leakage sensing module, a leakage feature fusion module, a leakage risk decision-making module, and an adaptive protection execution module. The multi-source leakage sensing module simultaneously captures leakage-related signals from the charging pile's operating environment, such as gas concentration gradient distribution, abnormal frequency band energy of acoustic waves, and liquid seepage rate. The leakage feature fusion module performs cross-modal feature alignment on the multi-source signals to generate a spatiotemporally correlated fused leakage feature tensor. The leakage risk decision-making module generates a leakage type identifier, leakage source probability distribution, and protection level command based on this tensor. The adaptive protection execution module dynamically adjusts the leakage suppression strategy according to the protection level command. This system can improve the accuracy and targeted nature of leakage monitoring, ensuring the safe operation of charging piles.

Description

Technical Field

[0001] This invention relates to the field of charging pile protection technology, specifically to an intelligent leakage monitoring and protection system for charging piles. Background Technology

[0002] With the rapid development of the new energy vehicle industry, charging piles, as core energy supply facilities, are seeing a continuous increase in installation scale and usage frequency. During operation, charging piles are exposed to complex outdoor environments for extended periods, facing multiple challenges such as high temperatures, humidity, and corrosive gas corrosion. Simultaneously, issues like aging internal electrical components and damaged sealing structures gradually emerge over time, easily leading to safety hazards such as gas leaks and liquid seepage. Currently, most charging pile leak monitoring solutions on the market use single sensors for data collection. For example, relying solely on gas concentration sensors to detect combustible gas leaks or humidity sensors to determine liquid seepage. These monitoring methods have significant limitations. A single sensor cannot comprehensively capture the multi-dimensional characteristics of the leak process. When minute leaks occur or complex environmental interference occurs, false alarms or missed alarms are likely, making accurate identification and prediction of leak risks difficult.

[0003] In terms of protection implementation, existing charging piles mostly adopt single protection measures triggered by fixed thresholds, such as directly cutting off the power supply once the gas concentration exceeds the standard. This approach lacks the ability to dynamically adapt to leakage scenarios. Different types of leaks (such as electrolyte leaks, coolant leaks, and flammable gas leaks) and different leakage levels have significantly different requirements for protection strategies. Fixed protection modes either overreact in the early stages of a leak, affecting normal charging services, or fail to provide sufficient protection when the leak intensifies, leading to an expansion of safety risks. In addition, existing systems lack effective data interaction and collaborative linkage between monitoring and protection modules. Monitoring data cannot be timely and accurately converted into targeted protection commands, making the entire leak prevention and control process passive and lagging. This makes it difficult to meet the high reliability requirements for the safe operation of charging piles and poses a potential threat to equipment safety, user personal safety, and the surrounding environment. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent leakage monitoring and protection system for charging piles to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides an intelligent leakage monitoring and protection system for charging piles, the system comprising:

[0006] A multi-source leakage sensing module is used to synchronously capture leakage-related signals of the charging pile's operating environment. The leakage-related signals include gas concentration gradient distribution, acoustic abnormal frequency band energy, and liquid seepage rate.

[0007] The leakage feature fusion module is used to perform cross-modal feature alignment on the leakage correlation signals collected by the multi-source leakage sensing module to generate a spatiotemporally correlated fused leakage feature tensor.

[0008] The leakage risk decision module is used to generate leakage type identifiers, leakage source probability distributions, and protection level instructions based on the fused leakage feature tensor.

[0009] An adaptive protection execution module is used to dynamically adjust the leakage suppression strategy based on the protection level command.

[0010] Preferably, the multi-source leakage sensing module specifically includes:

[0011] A distributed gas sensing unit is deployed at the bottom of the charging pile and in the cable interface area, using a nano-gas-sensitive material array to capture the gas concentration gradient distribution.

[0012] A broadband acoustic wave acquisition unit is embedded inside the insulating shell of the charging pile, and extracts the energy of abnormal frequency bands of acoustic waves through a piezoelectric sensor group.

[0013] The microfluidic detection unit is integrated into the foundation structure of the charging pile and quantifies the liquid seepage rate based on capillary pressure sensing technology.

[0014] Preferably, the leakage feature fusion module specifically includes:

[0015] The temporal feature extraction submodule performs sliding window segmentation on the gas concentration gradient distribution to extract the concentration abrupt change slope and steady-state drift.

[0016] The frequency domain feature decomposition submodule performs wavelet packet transform on the energy of the abnormal frequency band of the acoustic wave to separate the high-frequency impact component and the low-frequency resonance component.

[0017] The flow characteristic modeling submodule constructs a seepage path topology map based on the liquid seepage rate and calculates the node velocity difference coefficient.

[0018] Preferably, the leakage feature fusion module further includes:

[0019] The cross-modal alignment submodule maps the concentration abrupt change slope, high-frequency impact component, and nodal velocity difference coefficient to a unified timestamp coordinate system;

[0020] The feature tensor generation submodule uses a multi-head attention mechanism to weight and stitch together the steady-state drift, low-frequency resonant components, and seepage path topology to output a three-dimensional fused leakage feature tensor.

[0021] Preferably, the leakage risk decision module specifically includes:

[0022] The leakage type classification unit uses a temporal convolutional network to process the fused leakage feature tensor and outputs a combustible gas leakage identifier, an electrolyte leakage identifier, or an insulating oil leakage identifier.

[0023] The leak source localization unit analyzes the spatial correlation features in the fused leak feature tensor based on a graph neural network to generate a probabilistic thermal distribution map of the leak source.

[0024] The protection strategy decision unit matches the preset protection level instruction with the leakage type identifier and the probability thermal distribution map of the leakage source.

[0025] Preferably, the leak source location unit specifically performs the following:

[0026] A three-dimensional spatial mesh model of the charging pile is constructed, and the fused leakage feature tensor is mapped to the corresponding mesh node;

[0027] The graph convolutional layer captures the correlation of concentration gradients between adjacent nodes;

[0028] Output the set of grid nodes labeled with leakage probability values ​​to form a probability thermal distribution map of the leakage source.

[0029] Preferably, the adaptive protection execution module specifically includes:

[0030] When a gas barrier injection assembly receives a combustible gas leak indicator, it releases an inert gas envelope layer into the high-probability region of the probability thermal distribution map of the leak source.

[0031] The liquid-solid conversion component injects a polymer curing agent at key nodes in the liquid seepage path when it receives an electrolyte leakage indicator.

[0032] When the adsorption and recovery component receives an insulating oil leak indicator, it activates the negative pressure adsorption device and locates the coordinates of the leak source.

[0033] Preferably, the adaptive protection execution module further includes:

[0034] The dynamic strategy adjustment module adjusts the release concentration of the inert gas envelope, the injection flow rate of the polymer curing agent, or the power parameters of the negative pressure adsorption device in real time based on the gradient change rate of the probability thermal distribution map of the leakage source.

[0035] Preferably, the leakage risk decision module further includes:

[0036] The protection effectiveness feedback loop collects the concentration distribution of the inert gas envelope layer, the coverage area of ​​the polymer curing agent, or the recovery efficiency of the negative pressure adsorption device, and updates the gas concentration gradient distribution, acoustic abnormal frequency band energy, and liquid seepage rate in the fusion leakage characteristic tensor.

[0037] Preferably, the protection strategy decision unit is further configured to:

[0038] When the protection effectiveness feedback loop detects that the gas concentration gradient distribution has not dropped to the safety threshold, it raises the protection level instruction level and triggers the dynamic strategy adjustment module to increase the inert gas release concentration;

[0039] When a continuous increase in the liquid seepage rate is detected, an instruction to increase the injection of polymer curing agent is sent to the liquid-solid conversion component.

[0040] Compared with the prior art, the beneficial effects of the present invention are:

[0041] This intelligent leak monitoring and protection system for charging piles simultaneously captures multi-dimensional leak-related signals, such as gas concentration gradient distribution, abnormal frequency band energy of acoustic waves, and liquid seepage rate, through a multi-source leak sensing module. This breaks through the limitations of traditional single-sensor monitoring, enabling comprehensive capture of characteristic information during the leak process from different physical dimensions. It effectively avoids monitoring deviations caused by misjudgments or omissions due to single signals, making leak monitoring more comprehensive and accurate. Compared to traditional monitoring methods that can only acquire discrete data in a single dimension, the simultaneous acquisition of multi-source signals can completely present the dynamic development process of the leak, providing a richer and more reliable data source for subsequent leak characteristic analysis and risk assessment, helping the system to more accurately perceive the leak status.

[0042] The leak feature fusion module performs cross-modal feature alignment on multi-source sensing signals, generating a spatiotemporally correlated fused leak feature tensor. This enables deep integration and collaborative analysis of different types of monitoring data. In traditional systems, monitoring data are independent of each other, making it impossible to establish intrinsic correlations between data, leading to fragmented understanding of leak scenarios. Cross-modal feature alignment technology, however, can correlate and match feature data from different dimensions such as gases, sound waves, and liquids in both time and space, uncovering hidden leak patterns and characteristics. This allows the generated fused leak feature tensor to more realistically and comprehensively reflect the essential attributes of leak events, providing a more scientific basis for subsequent leak risk decisions and avoiding decision-making biases caused by isolated data.

[0043] The leakage risk decision-making module generates leakage type identifiers, leakage source probability distributions, and protection level instructions based on fused leakage feature tensors, achieving accurate classification and quantitative assessment of leakage risks. This module no longer relies on simple judgments using fixed thresholds; instead, through in-depth analysis of multi-dimensional fused features, it accurately distinguishes different types of leakage events, clarifies the possible locations and probability distributions of leakage sources, and assigns corresponding protection levels based on the severity of the leakage, making risk decisions more targeted and refined. This decision-making approach effectively avoids the blindness of protective measures caused by the inability to distinguish leakage types and locate leakage sources in traditional systems, making subsequent protective actions more targeted and improving the ability to control leakage risks.

[0044] The adaptive protection execution module dynamically adjusts leakage suppression strategies based on protection level commands, achieving precise adaptation between protection measures and leakage scenarios. Traditional fixed protection modes cannot flexibly adjust according to actual leakage conditions, while this module can automatically select and execute matching protection strategies based on protection level commands output by the risk decision module, targeting different leakage types, leakage levels, and leakage source locations. For example, it can implement local sealing and early warning for trace electrolyte leaks, and emergency power cut-off and ventilation for large-scale flammable gas leaks. This dynamically adaptable protection method minimizes the impact on the normal operation of charging piles while ensuring safety, avoiding service interruptions caused by over-protection. It also ensures timely upgrades to protection measures when leakage risks escalate, effectively curbing leak spread, reducing the likelihood of safety accidents, and comprehensively improving the safety and reliability of charging pile operation, providing stronger safety guarantees for equipment, users, and the surrounding environment. Attached Figure Description

[0045] Figure 1 This is a timing diagram of the intelligent leakage monitoring and protection system for charging piles described in this invention;

[0046] Figure 2 This is a flowchart illustrating the working principle of the multi-source leakage sensing module.

[0047] Figure 3 A flowchart illustrating the working principle of the leakage feature fusion module submodule;

[0048] Figure 4 A flowchart illustrating the working principle of the leak source location unit. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] Please see Figure 1 This invention provides an intelligent leakage monitoring and protection system for charging piles, the system comprising:

[0051] The system achieves comprehensive control over leakage risks in the operating environment through a systematic data acquisition, processing, and execution mechanism. It consists of four core modules: a multi-source leakage perception module, a leakage feature fusion module, a leakage risk decision module, and an adaptive protection execution module. These modules form a closed-loop control system through hardware integration and algorithm collaboration, enabling all-round monitoring and intelligent protection of the charging pile's operating status.

[0052] The multi-source leak sensing module, serving as the system's data acquisition front-end, is responsible for synchronously capturing various leak-related signals in the charging pile's operating environment. It employs a distributed sensor network layout, covering the charging pile itself and surrounding key areas. The sensed signals include three core parameters: gas concentration gradient distribution, acoustic wave anomalous frequency band energy, and liquid seepage rate. Gas concentration gradient distribution is acquired through a nano-gas-sensitive sensor array deployed at the bottom of the charging pile and in the cable interface area. The sensors are distributed in a grid topology, with spacing optimized according to a leak diffusion model, to monitor the spatial distribution characteristics of combustible gas and volatile organic compound concentrations in real time. Acoustic wave anomalous frequency band energy acquisition is achieved through a piezoelectric sensor group embedded within the charging pile's insulating shell. These sensors feature a wideband response design, focusing on capturing acoustic signals in the 20kHz-100kHz frequency band, which contains high-frequency vibration components typical of leak events. Liquid seepage rate monitoring is accomplished through a microfluidic detection unit integrated into the charging pile's foundation structure layer. Based on capillary pressure sensing principles, it quantifies the flow characteristics of the leaking liquid in real time. All sensor data is synchronously acquired via an industrial bus, achieving millisecond-level time synchronization accuracy to ensure spatiotemporal consistency of multi-source data.

[0053] The leakage feature fusion module is responsible for processing and fusing multi-source heterogeneous data. It performs cross-modal feature alignment and fusion processing on the raw signals collected by the sensing module, employing a time-space dual alignment strategy to map gas concentration signals, acoustic signals, and liquid flow signals to a unified timestamp coordinate system. An interpolation algorithm is used to solve the synchronization problem of data with different sampling rates. Subsequently, feature extraction and selection are performed: a sliding window segmentation process is used for the gas concentration gradient distribution to extract time-domain features such as the slope of concentration abrupt changes and steady-state drift; wavelet packet transform is performed on the energy of the anomalous frequency bands of the acoustic waves to separate frequency-domain features such as high-frequency impact components and low-frequency resonant components; and flow regime modeling is performed on the liquid seepage rate data to construct a seepage path topology map and calculate the node velocity difference coefficient.

[0054] A multi-head attention mechanism is used to achieve cross-modal feature weighted fusion, mapping feature vectors with different physical meanings to a unified feature space, generating a three-dimensional fused leakage feature tensor with spatiotemporal correlation. This tensor simultaneously contains information in the time, space, and feature dimensions. The leakage risk decision module performs leakage risk assessment and decision-making based on the fused feature tensor, employing a deep learning algorithm architecture and including three functional units: leakage type identification, leakage source localization, and protection strategy generation. The leakage type classification unit uses a temporal convolutional network to process the feature tensor, extracting temporal features through multi-layer causal convolution and dilated convolution, and outputting three categories of identifiers and their probability distributions: combustible gas leakage, electrolyte leakage, and insulating oil leakage. The leakage source localization unit, based on a graph neural network architecture, constructs a three-dimensional spatial grid model of the charging pile, maps the feature tensor to grid nodes, and captures the spatial correlation between nodes through graph convolutional layers, generating a leakage source probability heat map to visualize the spatial distribution of leakage probability. The protection strategy decision unit matches preset protection level instructions according to the leakage type identifier and leakage source probability distribution. The protection level has a three-level response mechanism, corresponding to leakage scenarios with different risk levels.

[0055] The adaptive protection execution module dynamically adjusts and executes leakage suppression strategies based on the protection commands output by the decision module. It includes multiple dedicated actuators: a gas barrier injection assembly for flammable gas leaks, which releases inert gas to form an envelope layer through an array of nozzles positioned in critical areas; a liquid-solid conversion assembly for electrolyte leaks, which injects a polymer curing agent at key nodes in the seepage path using a precision injection system; and an adsorption and recovery assembly for insulating oil leaks, which uses a negative pressure adsorption device in conjunction with a mobile collection mechanism to recover and process the leaked material. All execution components are equipped with adjustable control interfaces, enabling real-time adjustments to operating parameters based on changes in the leakage situation.

[0056] The system achieves continuous optimization through a closed-loop feedback mechanism. The protection effectiveness feedback loop collects execution effect data in real time, including parameters such as changes in gas concentration distribution, solidifying agent coverage area, and adsorption recovery efficiency. This data is fed back to the feature fusion module to update feature tensors, thereby influencing subsequent decision-making processes. When the protection effect is detected to be below expectations, the system automatically upgrades the protection level and adjusts execution parameters, forming a complete control closed loop of monitoring-decision-execution-feedback. The entire system adopts an industrial-grade hardware platform and a real-time operating system to ensure reliability and real-time performance in the complex operating environment of charging piles, enabling early detection, accurate location, and effective suppression of leakage risks.

[0057] Example 1: See Figure 2The implementation of the distributed gas sensing unit involves the arrangement of a nano-gas-sensitive material array and the data acquisition method. Sensor nodes made of metal oxide semiconductor materials are uniformly distributed in a grid pattern on the surface of the charging pile's bottom support platform and inside the protective housing around the cable interface. Each sensor node contains a nano-gas-sensitive material sensing element and a temperature compensation element. The sensing element uses a tin dioxide nanowire structure, and its resistance changes with the surrounding gas concentration. These sensor nodes are connected in a network via an I2C digital bus, and the bus controller polls the data of each node at a frequency of 10Hz. The calculation of the gas concentration gradient distribution is based on the concentration difference between adjacent nodes. By measuring the concentration differences of four adjacent nodes in the east, west, south, and north directions, a vector field of concentration change is formed. To eliminate measurement errors caused by changes in ambient temperature, each node is equipped with a digital temperature sensor, and a polynomial fitting algorithm is used to compensate for the raw concentration data in real time. The sensor array is powered by an independent DC power supply module to avoid electromagnetic interference with the charging pile's main power supply. During data acquisition, baseline calibration is performed on each node. A reference benchmark is established by monitoring the resistance value in clean air, and then the actual gas concentration value is calculated differentially. The collected concentration data is transmitted to the central processing unit via the CAN bus. The transmission protocol adopts a custom real-time data frame structure, and each frame of data contains node number, concentration value, temperature value and timestamp information.

[0058] The implementation of the broadband acoustic wave acquisition unit is based on the arrangement and signal processing scheme of the piezoelectric sensor array. Four PZT-5A piezoelectric ceramic sensors are embedded in the four corners of the inner wall of the charging pile's insulating shell, forming an acoustic monitoring array. The sensors are encapsulated and protected with epoxy resin and directly coupled to the shell material to ensure acoustic wave propagation efficiency. Each sensor is connected to a high-input-impedance preamplifier with a magnification factor of 100. The signal is then digitally sampled through a 24-bit analog-to-digital converter at a sampling rate of 200kHz to meet the Nyquist sampling theorem requirement. Energy extraction in the anomalous frequency band of the acoustic wave is performed using digital signal processing. The original signal is bandpass filtered, retaining frequency components from 20kHz to 100kHz. The spectral energy distribution is then calculated using a short-time Fourier transform. Anomalous frequency band monitoring focuses on energy changes in the range of 40kHz to 60kHz, a band typically associated with turbulent noise from gas leaks and the acoustic characteristics of liquid jets. The signal processing algorithm uses a sliding window approach with a window length of 1024 sampling points and an overlap rate of 50%. Power spectral density is calculated and peak energy is detected within each window. To distinguish between normal operating noise and leakage noise, the system establishes an acoustic feature database, containing typical noise spectra generated by cooling fans, relay operation, and power electronic device switching during normal operation of the charging pile. These background noise interferences are eliminated using a spectral subtraction algorithm. The collected acoustic feature data is transmitted to the processing unit via an Ethernet interface, with UDP protocol used to ensure real-time transmission.

[0059] The microfluidic detection unit relies on a network of microfluidic chips integrated into the charging pile's foundation structure. These chips, made of PDMS material using soft lithography, contain a series of micrometer-level channels and pressure sensing chambers. The microchannel network covers the surface area of ​​the foundation, with a channel width of 200 micrometers and a depth of 100 micrometers, forming capillary-like liquid pathways. A MEMS pressure sensor is integrated at each channel intersection. The sensor operates based on the piezoresistive effect, generating pressure changes as liquid flows into the channel. Liquid seepage rate is quantified by monitoring the rate of pressure change, using a differential pressure measurement method to calculate the pressure gradient between adjacent sensing points. The system acquires pressure data from each node at 1-second intervals and converts the pressure difference into a flow velocity value using the Hagen-Poiseuille equation. To accommodate liquids of varying viscosities, the system also integrates online viscosity detection, estimating liquid viscosity by measuring the pressure decay time constant. A temperature compensation algorithm corrects for changes in liquid viscosity and sensor drift caused by temperature variations in real time. The microfluidic chip connects to a data acquisition card via flexible circuitry. The acquisition card digitizes the pressure signal at 16-bit resolution and transmits it to the central processing unit via an RS485 bus. The entire microfluidic network adopts a modular design, with each module covering an area of ​​0.25 square meters. Multiple modules are spliced ​​together to form a complete monitoring network. To prevent channel blockage, the system is also designed with a backflushing mechanism, periodically injecting clean airflow into the channels to maintain unobstructed flow.

[0060] Example 2: See Figure 3The time-series feature extraction submodule employs a sliding time window analysis method when processing gas concentration gradient distribution data. The window length is set to 30 seconds, with a 15-second overlap between adjacent windows to ensure continuous monitoring and prevent the omission of any abrupt changes. Within each time window, the system calculates the gas concentration change characteristics. The slope of the concentration abrupt change is obtained by taking the maximum value of the first derivative of the concentration-time curve; this value reflects the drasticness of the concentration change. The steady-state drift is obtained by calculating the root mean square deviation between the concentration data within the window and the baseline value, used to quantify the slow trend of concentration change. These calculations rely on digital signal processing algorithms, implemented in C++ and running on an embedded processor. The processed feature data is timestamped and stored in a circular buffer for subsequent fusion processing. The frequency domain feature decomposition submodule uses wavelet packet transform to process acoustic signals. Wavelet packet transform can be understood as a more refined frequency band division method. It uses the db4 wavelet basis function to perform a four-level decomposition of the acoustic signal, dividing the signal into 16 different frequency bands. The high-frequency impact component refers to the sum of wavelet packet coefficients extracted from the frequency band above 60kHz. These high-frequency components are usually associated with sudden leakage events. The low-frequency resonance component is the sum of energy extracted from the frequency band from 20kHz to 40kHz, which is often associated with persistent leakage. The entire decomposition process is accelerated by a digital signal processor, and the overlap-preservation method is used to improve computational efficiency. The processed frequency domain features are synchronized with the time domain features. The flow characteristic modeling submodule constructs a seepage path topology map for the liquid seepage data. This map consists of nodes and edges. Nodes correspond to the pressure sensing points in the microfluidic detection unit, and edges represent the connectivity between adjacent sensing points. The node velocity difference coefficient is obtained by calculating the Euclidean distance of the velocity vectors between adjacent nodes. This coefficient reflects the non-uniformity of liquid flow, and the calculation formula is:

[0061] in: This represents the coefficient of difference in flow velocity between nodes. Indicates the first The flow rate value of each node. This represents the number of adjacent nodes. This calculation process is executed in real time by a graph computing engine, which stores topological relationships based on an adjacency list data structure and updates the flow path using a breadth-first search algorithm.

[0062] The cross-modal alignment submodule is responsible for mapping feature data from different sources to a unified time coordinate system. Since the sampling rate for gas concentration data is 10Hz, the sampling rate for acoustic feature data is 100Hz, and the sampling rate for liquid flow velocity data is 1Hz, the system uses a linear interpolation algorithm to resample all data to a unified frequency of 100Hz. The interpolation algorithm is based on a timestamp synchronization mechanism, assigning a unified time index to each data point. Three features—concentration abrupt change slope, high-frequency impact component, and node velocity difference coefficient—are selected for alignment because they all reflect the transient characteristics of leakage events. A sliding time window matching strategy is used during alignment, with a window length of 100 milliseconds to ensure that the time deviation between different modal data does not exceed 10 milliseconds. The feature tensor generation submodule uses a multi-head attention mechanism to weightedly fuse steady-state drift, low-frequency resonant components, and seepage path topology. The attention mechanism can be understood as an intelligent weighting system that automatically evaluates the importance of different features and assigns corresponding weights. The mechanism comprises four independent attention heads, each responsible for handling feature relationships across different dimensions: the first head handles feature correlations in the temporal dimension, the second handles feature distributions in the spatial dimension, the third handles correlations between different physical quantities, and the fourth handles cross-modal feature interactions. The weighted feature vectors output by each attention head are concatenated into a comprehensive feature vector, which is then processed through a 3D renormalization operation to generate a fused leakage feature tensor of size 100×50×12, where 100 represents the time step, 50 represents the number of spatial grid points, and 12 represents the number of feature dimensions. The entire fusion process is accelerated on a graphics processing unit (GPU) using a CUDA parallel computing architecture to improve processing efficiency.

[0063] When some sensor data is temporarily missing, the system uses a Kalman filter algorithm to complete the data and predicts the current missing value based on historical data trends. For dimensionality mismatch issues that occur during feature fusion, the system uses zero-padding and feature scaling methods to ensure data consistency. To protect data integrity, all feature data uses cyclic redundancy check (CRC) codes for error detection, and retransmission is immediately required once an error is detected. The entire processing flow adopts a pipeline architecture design, with each submodule running in parallel and data exchanged through a shared memory area to reduce data transmission latency. The system is also equipped with a real-time monitoring interface that can visualize the entire process of feature extraction and fusion, including changes in time series curves, spectrograms, and topology maps. All algorithm parameters can be dynamically adjusted through configuration files, such as sliding window size, wavelet basis function type, and number of attention heads, enabling the system to adapt to the monitoring needs of different types of charging piles. System operating status data is recorded in real time to log files, including indicators such as processing latency, memory usage, and computational load, providing a reference for subsequent system optimization. The implementation of this embodiment ensures that the leak monitoring system can efficiently process multi-source heterogeneous data, providing accurate feature input for subsequent risk decisions.

[0064] Example 3: See Figure 4 The leakage type classification unit employs a temporal convolutional network to process and fuse leakage feature tensors. This network structure contains four causal convolutional layers, each with 64 filters and a kernel size of 3. The dilation coefficients are set to 1, 2, 4, and 8 in ascending order of layer. Causal convolution ensures that the output depends only on the input at the current and historical moments, preventing future information leakage. Batch normalization is performed after each convolutional layer. The activation operation compresses the time dimension of the output through a global average pooling layer and maps it to a three-dimensional classification space by a fully connected layer. The function generates probability distributions for three types of leak identifiers, setting a probability threshold of 0.7. When the probability of a certain type of leak exceeds this threshold, the corresponding identifier is output. The network is trained using labeled historical leak data, with the Adam algorithm chosen as the optimizer and cross-entropy loss as the loss function. An early stopping strategy is employed during training to prevent overfitting. The leak source localization unit constructs a 3D spatial grid model of the charging pile based on a graph neural network, mapping the fused leak feature tensor to spatial grid nodes with a 5cm resolution. Each node contains a 12-dimensional feature vector. The graph neural network uses a graph attention mechanism, and its node update formula is expressed as:

[0065] in: This represents the updated feature representation of node i. express Activation function Representative node The set of neighboring nodes, Represents a node With nodes Attention weights between them This represents a trainable weight matrix. This represents the original feature vector of neighboring node j. Attention weights are obtained by calculating the feature similarity between nodes, focusing on neighboring nodes with strong correlation in concentration gradients. After three layers of graph convolution operations, the output layer... The function calculates the leakage probability value for each node, generating a heatmap of leakage source probability. This heatmap is visualized using pseudo-color rendering, with red areas representing high-probability leakage sources and blue areas representing low-probability areas.

[0066] The protection strategy decision unit generates protection level instructions based on the matching results between the leak type identifier and the leak source probability heat map. The matching algorithm calculates the product of the leak type probability value and the highest probability value in the heat map. When the product value exceeds 0.5, a Level-3 protection level instruction is triggered. Level-1 instructions correspond to low-risk scenarios, issuing only a warning signal and initiating regular monitoring mode; Level-2 instructions correspond to medium-risk scenarios, initiating pre-protection measures and increasing monitoring frequency; Level-3 instructions correspond to high-risk scenarios, immediately initiating comprehensive protection measures and triggering the emergency response mechanism. The decision logic is implemented using a rule engine, supporting online updates of matching rules to adapt to different operating environments. During system operation, changes in the leak situation are continuously monitored. When the probability product value is detected to move from the low-risk range to the high-risk range, the protection level is automatically upgraded; when the situation eases, the protection level is gradually reduced. Detailed logs are recorded for all decision-making processes, including timestamps, input features, decision results, and confidence scores, providing data support for subsequent system optimization.

[0067] When multiple leakage sources occur simultaneously, the algorithm employs a non-maximum suppression strategy to handle multiple peak regions in the heatmap, ensuring accurate identification of the main leakage source. For uncertain identification results, the system introduces a confidence assessment mechanism; when the confidence level falls below a preset threshold, a multimodal data verification process is initiated. The graph neural network training uses a semi-supervised learning approach, leveraging a small amount of labeled data and a large amount of unlabeled data to improve the model's generalization ability. The real-time inference process is accelerated through model quantization technology, converting floating-point operations to fixed-point operations to improve computational efficiency while maintaining accuracy. The system also features online learning capabilities, automatically updating model parameters based on newly occurring leakage cases to gradually improve identification accuracy. All model versions undergo rigorous testing and verification to ensure no performance degradation occurs during updates. The system's real-time monitoring interface displays leakage type identification results, leakage source location heatmaps, and the protection level decision process. Operators can view the system's operating status at any time and access a manual intervention interface.

[0068] Example 4: The specific implementation of the adaptive protection execution module involves a collaborative control and parameter adjustment mechanism for multiple actuators. The gas barrier injection assembly consists of a high-pressure nitrogen storage tank, a solenoid valve array, and a nozzle group. The storage tank operates at a pressure of 20 MPa, and the output pressure is stabilized at 0.8 MPa through a pressure reducing valve. The solenoid valve adopts a two-position three-way structure with a response time of less than 50 milliseconds. The nozzle arrangement density is 4 nozzles per square meter, covering the entire bottom area of ​​the charging pile. When a combustible gas leak indicator is received, the control system calculates the required nozzle combination to be activated based on the coordinates of areas with a probability value greater than 0.8 in the probability thermal distribution map of the leak source. The initial release concentration is 10 L / min, and the nitrogen purity is required to reach 99.99% or higher, forming an inert gas envelope layer to dilute the combustible gas concentration below the lower explosive limit. The liquid-solid conversion assembly includes a polymer curing agent storage tank, a metering pump, and an injection nozzle. The curing agent uses an acrylic rapid solidification material, and the solidification time is adjustable from 5 to 30 seconds. The injection pump is driven by a stepper motor, with a minimum flow control accuracy of 0.1 mL / s and an initial flow rate setting of 5 mL / s. The system determines critical injection locations based on nodes with a velocity difference coefficient greater than 0.3 in the seepage path topology diagram, maintaining the injection pressure within the range of 0.2-0.5 MPa. The adsorption and recovery assembly consists of a vacuum pump, an oil-water separator, and a six-degree-of-freedom robotic arm. The vacuum pump achieves an ultimate vacuum of -95 kPa and a pumping speed of 50 L / min. The robotic arm's positioning accuracy reaches ±1 mm, and the end effector is equipped with an infrared vision sensor for precise leak source location. Upon receiving an insulating oil leak indicator, the system performs initial positioning using thermal coordinates, followed by secondary precise positioning using the vision sensor, and initiates the adsorption and recovery operation, with the initial adsorption power set at -50 kPa.

[0069] The dynamic strategy adjustment module monitors the gradient change rate of the probability thermal distribution map of the leakage source in real time. This rate is obtained by calculating the difference between the probability values ​​of adjacent time frames, with a sampling interval of 1 second. Based on the magnitude of the rate change, the system automatically adjusts the operating parameters of each execution component. The adjustment logic is based on a preset parameter mapping table, which establishes the correspondence between the gradient change rate and the execution parameters. For example, when the detected gradient change rate exceeds 0.1 / s, the system proportionally increases the inert gas release concentration to 15 L / min, increases the polymer curing agent injection flow rate to 8 mL / s, and increases the power of the negative pressure adsorption device to -70 kPa. During the adjustment process, a PID control algorithm is used to achieve a smooth transition and avoid system oscillation caused by sudden parameter changes. The operating status of all execution components is fed back in real time through a sensor network, including gas concentration sensors, flow meters, and pressure sensors, forming a closed-loop control loop (see Table 1).

[0070] Table 1: Parameter Adjustment Logic Comparison Table

[0071]

[0072] When multiple leak sources are detected simultaneously, the system employs a priority scheduling algorithm to determine the processing order based on the leak type, hazard level, and probability. For component failures, the system has a redundancy switching function, automatically activating backup equipment when the primary component fails. Detailed logs are recorded for all parameter adjustments, including timestamps, gradient change rates, parameter values ​​before and after adjustment, and the status of the adjustment result. This data is used for subsequent system optimization and maintenance. The user interface displays the real-time operating status of each component and the parameter adjustment process, providing a manual intervention interface for handling special situations. The system periodically self-checks the performance indicators of each component, including nozzle blockage detection, pump efficiency testing, and robotic arm precision calibration, ensuring the reliability of the protection execution. The entire implementation process ensures that the leak protection system can dynamically adjust operating parameters based on real-time monitoring data, achieving precise and effective leak suppression.

[0073] Example 5: Implementation of the Protection Performance Feedback Loop. The system collects performance data in real time through a multi-source sensor network. This data includes parameters such as the concentration distribution of the inert gas envelope, the coverage area of ​​the polymer curing agent, and the recovery efficiency of the negative pressure adsorption device. Gas concentration distribution monitoring utilizes the existing distributed gas sensing unit, continuing to monitor concentration changes at each node after protection is implemented, with the sampling frequency remaining constant at 10Hz. The curing agent coverage area is indirectly calculated through pressure changes in the microfluidic detection unit. When the curing agent is injected, the liquid flow rate changes, and the curing coverage area can be deduced from the flow rate change curve. The adsorption recovery efficiency is calculated using the flow sensor and concentration sensor built into the negative pressure adsorption device, monitoring the mass flow rate and component concentration of the adsorbed substance in real time. All these performance data are sent back to the central processing unit through the same data transmission path and correlated with the original monitoring data using timestamp alignment. The system uses a sliding window averaging algorithm to update the feature tensor. Newly acquired performance data is dynamically weighted and integrated into historical data, ensuring that the feature update reflects the latest state while maintaining data stability. The feature tensor update process is calculated using the following formula:

[0074]

[0075] in: This represents the updated feature tensor. Represents the current feature tensor. These are historical data weighting coefficients, with values ​​ranging from 0 to 1. Indicates the first Data values ​​from each performance monitoring point This represents the confidence weight of the corresponding data point. This indicates the number of monitoring points. The coefficients are dynamically adjusted based on the timeliness of the data. The weight of new data gradually increases over time, while the weight of historical data older than 5 minutes decays exponentially to ensure the system's sensitivity to the current state. The entire feedback data collection cycle is synchronized with the protection execution cycle to ensure that corresponding effectiveness evaluation data is obtained after each protection action.

[0076] The protection strategy decision unit dynamically adjusts the protection level command based on the monitoring results of the feedback loop. When the gas concentration sensor detects that the concentration value after protection has not dropped to the safety threshold, the safety threshold is set to 20% of the lower explosive limit concentration of the combustible gas, and the system automatically upgrades the protection level command. The upgrade process adopts a step-by-step upgrade strategy. When upgrading from Level-2 to Level-3, the inert gas release concentration is increased by 15%. If the concentration still does not decrease, the release concentration continues to increase. Each upgrade is calculated based on the proportion of the concentration difference. At the same time, the dynamic strategy adjustment module is triggered. This module calculates the amount of inert gas that needs to be increased based on the concentration decrease rate and uses a proportional-integral-derivative control algorithm to adjust the solenoid valve opening and injection duration. For liquid seepage scenarios, when the microfluidic detection unit detects a continuous increase in flow rate (defined as a flow rate increase of more than 10% for three consecutive sampling cycles), the system sends an increase injection command to the liquid-solid conversion component. The command generation is based on the functional relationship between the flow rate change rate and the current injection flow rate, and a feedforward control algorithm is used to pre-adjust the injection parameters to avoid response delay. The injection flow rate adjustment range is determined using a lookup table method, and a preset adjustment coefficient is matched according to the flow rate growth rate to ensure that the injection flow rate and seepage velocity maintain an appropriate ratio. All adjustment commands undergo safety verification to ensure that parameter changes are within the equipment's allowable operating range, avoiding over-adjustment that could damage the equipment or cause secondary pollution.

[0077] A robust anomaly handling mechanism is established during system operation. When feedback data deviates significantly from expected results, the system initiates a root cause analysis process, sequentially checking sensor accuracy, actuator status, and communication links. For persistent anomalies, the system automatically switches to a degraded mode, employing conservative parameter settings to ensure basic protection. Detailed operation logs are recorded for all adjustments, including initial state, adjustment commands, execution results, and final state. These logs are used for subsequent system performance analysis and optimization. The system also possesses a learning function, gradually optimizing parameter adjustment strategies and improving protection efficiency by analyzing historical adjustment records and effect data. Operators can view the feedback loop's operational status through a human-machine interface, including real-time data curves, historical adjustment records, and system performance indicators, and intervene manually when necessary. The entire feedback loop design ensures that the protection system continuously optimizes operating parameters based on actual results, achieving increasingly precise leakage protection. During regular system maintenance, the feedback loop undergoes calibration testing to ensure all sensors and actuators operate within specified accuracy ranges. Maintenance records are included in the system archive for subsequent query and analysis. The weighting coefficient κ in the formula is determined through experimental data optimization, and the confidence weight... Based on the sensor accuracy level allocation, the number of monitoring points n is dynamically adjusted according to the total number of sensors actually deployed. These parameters together ensure the accuracy and reliability of feature tensor updates.

[0078] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0079] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart leakage monitoring and protection system for charging piles, characterized in that, include: A multi-source leakage sensing module is used to synchronously capture leakage-related signals of the charging pile's operating environment. The leakage-related signals include gas concentration gradient distribution, acoustic abnormal frequency band energy, and liquid seepage rate. The leakage feature fusion module is used to perform cross-modal feature alignment on the leakage correlation signals collected by the multi-source leakage sensing module to generate a spatiotemporally correlated fused leakage feature tensor. The leakage risk decision module is used to generate leakage type identifiers, leakage source probability distributions, and protection level instructions based on the fused leakage feature tensor. An adaptive protection execution module is used to dynamically adjust the leakage suppression strategy based on the protection level command; The leakage feature fusion module specifically includes: The temporal feature extraction submodule performs sliding window segmentation on the gas concentration gradient distribution to extract the concentration abrupt change slope and steady-state drift. The frequency domain feature decomposition submodule performs wavelet packet transform on the energy of the abnormal frequency band of the acoustic wave to separate the high-frequency impact component and the low-frequency resonance component. The flow characteristic modeling submodule constructs a seepage path topology map based on the liquid seepage rate and calculates the node velocity difference coefficient. The leakage feature fusion module also includes: The cross-modal alignment submodule maps the concentration abrupt change slope, high-frequency impact component, and nodal velocity difference coefficient to a unified timestamp coordinate system; The feature tensor generation submodule uses a multi-head attention mechanism to weight and stitch together the steady-state drift, low-frequency resonant components, and seepage path topology to output a three-dimensional fused leakage feature tensor.

2. The intelligent leakage monitoring and protection system for charging piles according to claim 1, characterized in that, The multi-source leakage detection module specifically includes: A distributed gas sensing unit is deployed at the bottom of the charging pile and in the cable interface area, using a nano-gas-sensitive material array to capture the gas concentration gradient distribution. A broadband acoustic wave acquisition unit is embedded inside the insulating shell of the charging pile, and extracts the energy of abnormal frequency bands of acoustic waves through a piezoelectric sensor group. The microfluidic detection unit is integrated into the foundation structure of the charging pile and quantifies the liquid seepage rate based on capillary pressure sensing technology.

3. The intelligent leakage monitoring and protection system for charging piles according to claim 1, characterized in that, The leakage risk decision-making module specifically includes: The leakage type classification unit uses a temporal convolutional network to process the fused leakage feature tensor and outputs a combustible gas leakage identifier, an electrolyte leakage identifier, or an insulating oil leakage identifier. The leak source localization unit analyzes the spatial correlation features in the fused leak feature tensor based on a graph neural network to generate a probabilistic thermal distribution map of the leak source. The protection strategy decision unit matches the preset protection level instruction with the leakage type identifier and the probability thermal distribution map of the leakage source.

4. The intelligent leakage monitoring and protection system for charging piles according to claim 3, characterized in that, The leak source location unit specifically performs the following: A three-dimensional spatial mesh model of the charging pile is constructed, and the fused leakage feature tensor is mapped to the corresponding mesh node; The graph convolutional layer captures the correlation of concentration gradients between adjacent nodes; Output the set of grid nodes labeled with leakage probability values ​​to form a probability thermal distribution map of the leakage source.

5. The intelligent leakage monitoring and protection system for charging piles according to claim 4, characterized in that, The adaptive protection execution module specifically includes: When a gas barrier injection assembly receives a combustible gas leak indicator, it releases an inert gas envelope layer into the high-probability region of the probability thermal distribution map of the leak source. The liquid-solid conversion component injects a polymer curing agent at key nodes in the liquid seepage path when it receives an electrolyte leakage indicator. When the adsorption and recovery component receives an insulating oil leak indicator, it activates the negative pressure adsorption device and locates the coordinates of the leak source.

6. The intelligent leakage monitoring and protection system for charging piles according to claim 5, characterized in that, The adaptive protection execution module also includes: The dynamic strategy adjustment module adjusts the release concentration of the inert gas envelope, the injection flow rate of the polymer curing agent, or the power parameters of the negative pressure adsorption device in real time based on the gradient change rate of the probability thermal distribution map of the leakage source.

7. The intelligent leakage monitoring and protection system for charging piles according to claim 6, characterized in that, The leakage risk decision-making module also includes: The protection effectiveness feedback loop collects the concentration distribution of the inert gas envelope layer, the coverage area of ​​the polymer curing agent, or the recovery efficiency of the negative pressure adsorption device, and updates the gas concentration gradient distribution, acoustic abnormal frequency band energy, and liquid seepage rate in the fusion leakage characteristic tensor.

8. The intelligent leakage monitoring and protection system for charging piles according to claim 7, characterized in that, The protection strategy decision unit is also used for: When the protection effectiveness feedback loop detects that the gas concentration gradient distribution has not dropped to the safety threshold, it raises the protection level instruction level and triggers the dynamic strategy adjustment module to increase the inert gas release concentration; When a continuous increase in the liquid seepage rate is detected, an instruction to increase the injection of polymer curing agent is sent to the liquid-solid conversion component.

Images