A cable trench intelligent humidity monitoring and remote monitoring system

Abstract

The application discloses a cable trench intelligent humidity monitoring and remote monitoring system, comprising: an intelligent acquisition terminal layer, which collects environmental temperature and humidity, cable joint temperature, partial discharge, water level and gas concentration data in real time, and performs edge preprocessing on the data; an edge computing gateway layer, which is used for receiving and converging multi-source monitoring data, running an edge intelligent analysis model to identify condensation risk and early fire characteristics, generating early warning information and uploading; a remote monitoring center layer, which is used for receiving data and early warning information uploaded by the gateway layer, running a multi-source data fusion diagnosis model to perform global situation analysis and fire probability prediction, and issuing remote control instructions. The application realizes the collaborative acquisition of multi-dimensional parameters such as temperature and humidity, condensation, water level, gas concentration, cable joint temperature and partial discharge by integrating a composite environment monitoring unit and a cable state monitoring unit in the intelligent acquisition terminal layer.

Description

Technical Field

[0001] This application belongs to the field of Internet of Things (IoT) intelligent monitoring technology, specifically relating to an intelligent humidity monitoring and remote monitoring system for cable trenches. Background Technology

[0002] As a critical infrastructure of urban power transmission networks, the internal environment of cable trenches directly affects the operational safety and service life of cables. Because cable trenches are typically located in underground enclosed or semi-enclosed spaces, they are characterized by poor ventilation, easy water accumulation, and large temperature and humidity fluctuations. Prolonged exposure to high humidity can easily lead to moisture absorption and accelerated aging of cable insulation, and in severe cases, even cause discharge, flashover, or fire accidents, posing a significant threat to the safe and stable operation of the power grid.

[0003] Currently, monitoring of cable trench environments mainly relies on manual inspections or simple point-based online monitoring devices. Manual inspections suffer from poor timeliness, high labor intensity, and difficulty in detecting hidden hazards. Existing online monitoring systems often employ single-parameter limit-exceeding alarm modes, such as monitoring humidity, temperature, or water level alone, lacking the ability to fuse and analyze multi-source data and identify correlations, making it difficult to accurately identify early risks such as condensation precursors and insulation degradation trends. Furthermore, traditional monitoring systems typically use a direct "end-to-cloud" architecture, uploading massive amounts of raw data directly to the cloud, consuming significant communication bandwidth and experiencing latency in cloud analysis, making it difficult to meet the demands of rapid on-site response. Front-end acquisition devices are mostly battery-powered, with limited battery life, making it difficult to support high-frequency, multi-parameter continuous monitoring. Simultaneously, existing systems generally lack the ability to predict humidity trends and dynamic assessment mechanisms for fire probability, offering limited early warning methods and failing to achieve a shift from "passive alarm" to "proactive prevention."

[0004] To address the aforementioned issues, there is an urgent need for a cable trench humidity monitoring and remote control system with intelligent sensing, edge computing, multi-source fusion, and predictive early warning capabilities. This system would enhance the intelligent management and control of the cable trench operating environment and effectively reduce the risk of insulation failures and fires caused by moisture intrusion. Summary of the Invention

[0005] This application provides an intelligent humidity monitoring and remote monitoring system for cable trenches, which aims to solve the problems of existing technologies lacking the ability to fuse and analyze multi-source data and to accurately identify early risks such as condensation precursors and insulation degradation trends, and failing to meet the needs of rapid on-site response.

[0006] A smart humidity monitoring and remote control system for cable trenches includes:

[0007] The intelligent data acquisition terminal layer is deployed at key monitoring points in the cable trench to collect real-time data on ambient temperature and humidity, cable joint temperature, partial discharge, water level and gas concentration, and performs edge preprocessing on the data.

[0008] The edge computing gateway layer is communicatively connected to the intelligent acquisition terminal layer. It is used to receive and aggregate multi-source monitoring data, run edge intelligent analysis models to identify condensation risks and early fire characteristics, generate early warning information and upload it.

[0009] The remote monitoring center layer is communicatively connected to the edge computing gateway layer. It is used to receive data and early warning information uploaded by the gateway layer, run a multi-source data fusion diagnostic model to perform global situation analysis and fire probability prediction, and issue remote control commands.

[0010] Optionally, the intelligent data acquisition terminal layer includes:

[0011] The composite environmental monitoring unit collects data on ambient temperature, relative humidity, water level, and gas concentration, and has a built-in damp heat aging coefficient calculation model to generate early warning information on insulation degradation.

[0012] The cable condition monitoring unit collects cable joint temperature and partial discharge signals, and dynamically adjusts the acquisition frequency according to the temperature change rate.

[0013] The low-power self-powered module includes a CT power harvesting unit and a backup battery unit, and integrates intelligent energy routing function to achieve seamless switching between main and backup power and sleep control of non-core sensors.

[0014] Optionally, the temperature, humidity, and condensation monitoring module in the composite environmental monitoring unit employs a digital temperature and humidity sensor, through... The bus connects to the MCU, which has a built-in damp heat aging index. The computational model is expressed as follows:

[0015]

[0016] in For temperature gradient, Humidity abrupt change rate, For the duration of high temperature and high humidity, , , This is a preset dynamic weighting coefficient.

[0017] Optionally, the edge computing gateway layer includes:

[0018] The multi-protocol data aggregation and protocol conversion module is used to access terminal devices with heterogeneous communication standards and complete data parsing and unified encapsulation.

[0019] The edge intelligence analysis and early warning module, with its embedded lightweight machine learning model, is used to identify condensation risks and early fire characteristics in real time.

[0020] The system self-diagnosis and health management module is used to monitor the gateway's own operating status and communication link quality, and to activate local storage and breakpoint resume mechanism when the network is interrupted.

[0021] Optionally, the edge intelligent analysis and early warning module includes:

[0022] The condensation risk warning sub-model calculates the dew point temperature based on ambient temperature, humidity and cable joint temperature using the Magnus formula to determine the condensation risk.

[0023] The fire early warning sub-model constructs an isolated forest model based on multi-dimensional feature vectors, outputs anomaly scores and performs coupling verification to generate fire risk levels and cause inferences.

[0024] Optionally, the remote monitoring center layer includes:

[0025] The data platform and GIS visualization platform are used for multi-source data aggregation, spatiotemporal correlation storage, and situational display;

[0026] The multi-source data fusion diagnostic module runs a humidity trend prediction model and a fire probability prediction model, and outputs a future humidity exceeding the limit warning and the probability of fire occurrence.

[0027] The emergency response and remote control module has a built-in rule engine that automatically generates control commands based on the risk level and sends them to the edge gateway for execution.

[0028] Optionally, the humidity trend prediction model adopts an LSTM or GRU neural network architecture, inputs historical humidity, temperature, weather forecast and time series features, outputs a humidity prediction sequence for the next 24 hours and a 95% confidence interval, and generates a humidity over-limit warning based on the prediction results.

[0029] Optionally, the fire probability prediction model employs an XGBoost or LightGBM ensemble learning algorithm, taking into input a multi-dimensional feature vector including humidity, temperature, partial discharge, load, meteorological data, static attributes, and derived features, and outputting a calibrated fire occurrence probability, which is then combined with the SHAP value for causal explanation.

[0030] Optionally, the edge computing gateway layer and the remote monitoring center layer establish a long connection through an encrypted communication protocol, supporting real-time data upload, breakpoint resume, command issuance and execution feedback.

[0031] Optionally, the system also supports closed-loop management of mobile terminal alarm push and handling. Operation and maintenance personnel can receive hierarchical alarms, view risk details, complete handling work orders, and transmit handling results back in real time through a mobile APP.

[0032] Compared with the prior art, this application has at least the following beneficial effects:

[0033] This application integrates a composite environmental monitoring unit and a cable condition monitoring unit at the intelligent data acquisition terminal layer, enabling the coordinated acquisition of multi-dimensional parameters such as temperature and humidity, condensation, water level, gas concentration, cable joint temperature, and partial discharge. The terminal incorporates a built-in damp heat aging index calculation model, which can assess insulation degradation risk in real time at the data source and dynamically adjust the partial discharge monitoring frequency based on the temperature change rate. This ensures the integrity of critical data while avoiding redundant acquisition of invalid data, thus enhancing the intelligence level of front-end sensing.

[0034] This application adopts a dual-source input scheme combining CT power extraction and lithium iron phosphate batteries, and integrates intelligent energy routing functions to achieve seamless switching between main and backup power and sleep control of non-core sensors. When the cable loses power or the load current is insufficient, the system can automatically switch to battery power and enter a deep sleep mode, retaining only the key wake-up function, which greatly extends the survival time of the terminal under conditions without external power supply and solves the technical bottlenecks of difficult power supply and short battery life of traditional monitoring equipment.

[0035] This application deploys a condensation risk early warning sub-model and a fire early warning sub-model at the edge computing gateway layer. The condensation early warning model calculates the difference between the dew point temperature and the surface temperature of the cable joint in real time based on the Magnus formula. Combined with dynamic safety thresholds and time-series confirmation mechanisms, it can issue an early warning and trigger dehumidification devices before condensation forms. The fire early warning model uses the isolated forest algorithm to detect anomalies in multi-dimensional feature vectors and identifies the "humidity-discharge" correlation features through coupling verification, effectively suppressing false alarms and improving the timeliness of on-site risk handling. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the module connection of an intelligent humidity monitoring and remote control system for cable trenches, provided as an embodiment of this application. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments.

[0038] This application provides an intelligent humidity monitoring and remote control system for cable trenches, comprising:

[0039] The system consists of an intelligent data acquisition terminal layer, an edge computing gateway layer, and a remote monitoring center layer.

[0040] The intelligent acquisition terminal layer serves as the front-end sensing and data source of the system. It adopts a modular and low-power design and is physically encapsulated in an explosion-proof shell with an IP68 protection rating. It is connected to external sensors and antennas through prefabricated aviation plugs and deployed at key monitoring points such as power joint areas, water collection wells, and intermediate joint wells in cable trenches. The intelligent acquisition terminal layer includes a composite environmental monitoring unit, a cable status monitoring unit, and a low-power self-powered module.

[0041] The composite environmental monitoring unit is uniformly scheduled by a microcontroller (MCU, model STM32L4 series) on the motherboard, and adopts a data acquisition strategy that combines time-sharing inspection and interrupt triggering.

[0042] Temperature, humidity, and condensation monitoring module: Employs a digital temperature and humidity sensor (such as SHT35 or a similar high-precision sensor), through... The module is connected to the MCU via a bus. It periodically collects ambient temperature and relative humidity data at a frequency of T = 10 seconds per cycle. The MCU embeds a damp heat aging coefficient calculation model. This model is not simply a data recording process, but rather calculates in real-time the damp heat aging index, which characterizes the aging risk of cable insulation materials, according to the following formula. :

[0043]

[0044] in, This is the temperature rise gradient relative to the reference temperature. Humidity abrupt change rate, The duration of the high temperature and high humidity condition. , , A dynamic weighting coefficient pre-set for accelerated aging tests of cable insulation. When When the preset threshold is exceeded, the system will generate and upload an insulation degradation warning message even if a fire has not yet occurred.

[0045] The water level and gas monitoring module has the following functions:

[0046] Water level monitoring employs electrode-type or pressure-type water level sensors, connected to the acquisition unit via digital or analog input interfaces. To avoid the problem of mechanical structures such as floats easily getting stuck, a non-contact photoelectric water level sensor is preferred. This sensor utilizes the principle of the difference in refractive index of light in air and liquid to accurately detect the presence or absence of water and the depth of water accumulation in the collection well.

[0047] Gas monitoring, equipped with an electrochemical gas sensor (for monitoring CO, ... Toxic gases) and catalytic combustion or infrared optical gas sensors (for monitoring) (For combustible gases). The sensor outputs a 4-20mA standard analog signal, which is converted into a digital signal by the ADC module. The MCU reads the concentration values ​​of each gas through polling and corrects the data according to the built-in temperature compensation algorithm to eliminate the influence of ambient temperature changes on the gas detection accuracy.

[0048] The cable condition monitoring unit provides the following functions:

[0049] For cable joint temperature monitoring, passive surface acoustic wave (SAW) temperature sensors or surface-mount platinum resistance thermometers (PT100 / PT1000) temperature sensors are used for cable intermediate joints or terminations. For high-voltage environments, SAW sensors are preferred, as they communicate with a nearby reader via an RF antenna, achieving electrical isolation. For low-voltage environments, PT100 sensors can be used, connected to a high-precision ADC via a three-wire or four-wire connection to eliminate measurement errors caused by conductor resistance. The monitoring frequency is once per minute during normal cable operation. When a sudden change in current load or a temperature gradient increase exceeding 5°C / min is detected, the sampling frequency is automatically increased to once per 10 seconds.

[0050] Partial discharge monitoring: High-frequency current transformers (HFCTs) or transient ground voltage (TEV) sensors are installed at cable joints or on the grounding wire to capture partial discharge signals caused by moisture and aging of the cable insulation. The acquisition unit incorporates a high-speed ADC (sampling rate ≥100MSPS) and an FPGA (Field-Programmable Gate Array) to perform real-time acquisition, filtering, and feature extraction (such as discharge amplitude, phase, and frequency) of the partial discharge signals, rather than uploading massive amounts of raw data to save bandwidth. By analyzing the correlation between the partial discharge pattern and the current ambient humidity, it is possible to accurately determine whether moisture intrusion has led to a decline in insulation performance.

[0051] The low-power self-powered and management module is the core of ensuring the long-term stable operation of the data acquisition terminal, and it has dual-source input and intelligent energy routing functions;

[0052] The dual-source input features a CT power harvesting unit, employing a specially designed high-permeability open-type current transformer, which is sleeved on the cable body. When the cable load current exceeds 20A, the CT senses and draws power, outputting a stable 12V / 5V DC power through the back-end power management circuit (including rectification, filtering, and DC-DC regulation) to power the main board and sensors, while simultaneously charging the backup battery.

[0053] The backup battery unit uses high-energy-density, wide-temperature-range lithium iron phosphate (LFP) batteries. The battery pack (nominal voltage 12V, capacity configured according to power consumption requirements, such as 20Ah) features excellent safety and cycle life, and can seamlessly take over load power supply in the event of cable power failure or load current below the start-up threshold.

[0054] Smart energy routing specifically includes:

[0055] BMS Function Integration: The BMS chip (such as TI's BQ76940 series or similar products) monitors the voltage, current, temperature and remaining capacity (SOC, i.e. State of Charge) of the battery pack in real time, and has overcharge, over-discharge, overcurrent, short circuit and temperature protection functions.

[0056] Dynamic Power Path Management (DPPM): When the CT power harvesting unit outputs sufficient power, the energy router (integrated into the power management module) prioritizes supplying energy to the load, and excess energy is efficiently charged to the battery through the maximum power point tracking (MPPT) algorithm; when the CT power harvesting is insufficient (such as when the load current is low) or interrupted, the system switches to battery power at a speed of microseconds to ensure that the MCU and key sensors (such as partial discharge monitoring) do not crash and data is not lost.

[0057] When the communication network is interrupted and the battery level is lower than a set threshold (e.g., 20%), and there are no abnormal changes in environmental parameters, the MCU will cut off the power supply to non-core sensors (e.g., gas sensors) and enter a deep sleep mode (current consumption can be as low as tens of microamps), retaining only an ultra-low power timed wake-up function or an external interrupt (e.g., water flooding alarm signal) wake-up function to maximize the extension of the system's survival time.

[0058] The edge computing gateway layer serves as the system's regional data processing and command center. Physically, it employs an industrial-grade embedded hardware platform (preferably using an ARM Cortex-A series processor, such as the NXP i.MX8M or similar products), and is deployed inside manhole covers in cable trenches, within branch boxes, or at fire-resistant partition nodes in pipe racks. This layer boasts an IP65 protection rating and wide-temperature operating capability (-40℃). (+85℃), access the upper-level network via wired or wireless means, and establish star or mesh network connections with all intelligent acquisition terminals in the area. The edge computing gateway layer includes a multi-protocol data aggregation and protocol conversion module, an edge intelligent analysis and early warning module, and a system self-diagnosis and health management module.

[0059] The multi-protocol data aggregation and protocol conversion module is mainly responsible for solving the problems of heterogeneous communication standards (coexistence of short-range wireless and wired) and data normalization of the front-end acquisition terminals;

[0060] The gateway motherboard integrates multi-channel physical communication interfaces and onboard ZigBee (CC2530 / 2652 series) and LoRa (SX1278 / SX1262 series) RF chips, each configured with whip or suction cup antennas. As a ZigBee coordinator or LoRa gateway, the gateway can manage more than 200 terminal nodes simultaneously. Wireless communication employs an adaptive frequency hopping mechanism to avoid complex electromagnetic interference within cable trenches.

[0061] It provides two isolated RS485 bus interfaces, using the Modbus-RTU / ASCII master-slave protocol, for connecting online monitoring devices that lack wireless communication capabilities or require high stability (such as fixed partial discharge monitors). It also provides one 10 / 100M adaptive Ethernet port for local debugging or as a backup uplink.

[0062] Heterogeneous data normalization: Data frames from different interfaces and protocols first enter the protocol conversion middleware. This middleware uniformly parses private protocol messages, Modbus messages, and ZigBee / LoRa custom messages, extracts key values ​​(such as sensor ID, data type, time stamp, and numerical value), and repackages them into unified JSON or TLV format internal messages before sending them to the data processing queue.

[0063] The gateway has built-in high-capacity eMMC flash memory (≥8GB) or MicroSD card storage and establishes a circular queue database. When a communication interruption with the remote monitoring center is detected (e.g., 4G / fiber optic failure), all aggregated and preliminarily processed data (raw data and feature data) is automatically written to local storage in timestamp order. The storage strategy adopts a hierarchical mechanism: critical alarm data is permanently retained, while periodic status data is cyclically overwritten based on storage space (e.g., retained for 30 days). When the network is restored, the gateway automatically detects data gaps during the network outage and initiates a resume transmission to the central platform according to the "first-in, first-out" principle or data priority (alarm data first) to ensure data continuity and integrity. After the upload is complete, confirmed received data can be cleared from the local cache or marked as archived.

[0064] The edge intelligence analysis and early warning module is the core of the gateway layer's "intelligence." It embeds a lightweight machine learning model that has been pruned and optimized, and is responsible for online inference and immediate response to real-time data streams. Specifically, it includes:

[0065] Data preprocessing and feature engineering are performed. The aggregated data first enters a real-time sliding window (the window size can be configured according to the data type, e.g., a 5-minute window for temperature and a 1-second window for partial discharge). Within the window, the module performs data cleaning (removing outliers), normalization, and extracts time-domain and frequency-domain features. Key extracted features include: the rate of change of ambient humidity (dRH / dt), the temperature rise gradient of the cable joint (dT / dt), and the amplitude trend and phase distribution spectrum changes of the partial discharge signal.

[0066] The multi-parameter fusion early warning model, deployed within the edge intelligent analysis module, comprises a two-layer cascaded early warning model, including a condensation risk early warning sub-model based on physical mechanisms and a data-driven early fire warning sub-model. The two sub-models run in parallel, sharing the same real-time data stream. Their outputs are comprehensively analyzed by the decision fusion unit to generate a structured risk warning message, specifically including:

[0067] The core of the condensation risk warning sub-model is to use thermodynamic calculations to assess in real time the possibility of condensation on cable joints and insulation surfaces due to oversaturation of humidity, thereby issuing a warning before insulation performance deteriorates or even surface flashover occurs.

[0068] Input parameters include the real-time ambient temperature. (Unit: °C), Ambient relative humidity (Unit: %RH), Cable joint surface temperature (Unit: °C, measured by a patch or SAW temperature sensor);

[0069] The model is primarily based on the Magnus formula or its modifications, according to... RHRH calculates the dew point temperature in real time under the current environmental conditions. :

[0070]

[0071] in, The coefficients a and b are constants (for water surfaces, a = 17.27 and b = 237.7℃ are usually used; for ice surfaces, different constants are used, and the model can automatically switch according to whether TaTa is below 0℃).

[0072] Condensation risk assessment logic: Real-time calculation of the difference between the surface temperature of the cable joint and the ambient dew point temperature. ;

[0073] when Close to or below When ΔT ≤ 0, water vapor in the air will reach saturation on the cooler joint surface and condense into a liquid water film. Even if ΔT is a small positive value, condensation can still occur if there is hydrophilic contamination (such as salt deposits or ash deposits) on the cable surface.

[0074] The model introduces a dynamic safety threshold. This threshold is not a fixed value, but rather it is adaptively adjusted based on the surface characteristics of the cable insulation material and historical operating data. Under default operating conditions, The value is set at 2.0℃ ± 0.5℃; if the system diagnoses a high level of pollution in the area where the cable is located (indirectly determined through historical partial discharge data or manually input pollution area distribution maps), the model will automatically... The temperature range has been adjusted upwards to 3.0℃~5.0℃ to account for the additional risk of contamination catalyzing the effect of pollutants on condensation.

[0075] Timing confirmation mechanism: To avoid false alarms caused by instantaneous fluctuations in sensor readings, the model incorporates a hysteresis comparator and a timer. Only when the discrimination condition is met... The consecutive establishment time exceeds the preset confirmation time. (The typical configuration value is 300 seconds, which can be adjusted according to the cable's thermal capacity characteristics) Only after the model finally confirms the existence of condensation risk;

[0076] Output and Linkage: Once the condensation risk is confirmed, the model immediately generates an "Insulation Surface Condensation Warning" message, which includes the risk location (cable joint ID), real-time ΔT value, duration, and risk level (classified as "Attention Level"). And duration "Warning Level": ΔT≤0 or the threshold is met and timeout occurs. This message is uploaded to the remote monitoring center through the gateway, and can also directly trigger the activation of local dehumidification devices (such as semiconductor refrigeration dehumidifiers) deployed in the cable trench, suppressing condensation formation at the source.

[0077] Early Fire Warning Sub-model: This sub-model models the characteristic parameter changes caused by the thermal decomposition of cable insulation materials before a fire occurs, paying particular attention to the coupling characteristics of abnormal humidity and enhanced partial discharge activity. The specific construction includes:

[0078] Feature engineering and sample construction: The model receives sliding window data from the data preprocessing module and constructs a multidimensional feature vector for each sampling time step. The selected features include, but are not limited to:

[0079] Environmental humidity change rate (% / min);

[0080] Temperature rise gradient of cable joint (℃ / min);

[0081] : Average amplitude of partial discharge (mV);

[0082] Partial discharge pulse frequency (times / second);

[0083] : Combustible gas concentration (such as CO, (unit: ppm)

[0084] : The deviation between the current load current and the historical load during the same period.

[0085] Baseline model training and updating includes the following steps:

[0086] During the initial stage of system commissioning or after equipment maintenance, a sufficiently long "normal operation window" (e.g., 7 consecutive days, with no abnormal events confirmed manually) is selected to collect the aforementioned multi-dimensional feature data to form a training set. The Isolation Forest algorithm is used to construct the baseline model. The core idea of ​​this algorithm is to partition the data space using a random hyperplane. Outliers, due to their sparse distribution and distance from the normal population, are more easily isolated, meaning their path length from the root node to the leaf node is shorter.

[0087] The model has online incremental learning capabilities. After the system has been running for a period of time, if the self-diagnosis module confirms that the newly collected data is still under normal operating conditions, it will be incorporated into the training set, and the baseline model will be updated periodically (e.g., monthly) to enable the model to adapt to normal drift caused by seasonal changes, equipment aging, etc.

[0088] Real-time reasoning and anomaly scoring, specifically including:

[0089] Anomaly score calculation, for each multidimensional feature vector arriving in real time. This sample is then fed into a pre-trained isolated forest model. The model calculates the average path length of this sample across all isolated trees. The outlier scores were obtained by normalization. The closer the score is to 1, the more likely the sample is to be an outlier;

[0090] Dynamic threshold strategy: Instead of using a fixed outlier score threshold, the model introduces a sliding window dynamic threshold. It calculates the average of all outlier scores over the past 24 hours in real time. and standard deviation Set the alarm threshold for the current moment to Where kk is the sensitivity coefficient (typically 3). 5). When At that time, it was initially determined to be "suspected abnormality";

[0091] Strong correlation verification and causal analysis include:

[0092] Coupling Feature Verification: To prevent false alarms caused by sensor malfunctions or single parameter drift, the model initiates a coupling verification module after initially identifying an anomaly. This module specifically analyzes humidity-related features in anomaly events (…). ) and discharge / gas characteristics ( The time-series correlation of ). Calculate the Pearson correlation coefficient r or mutual information, if If the value is positive (i.e., a sudden increase in humidity is accompanied by enhanced discharge), then the anomaly is confirmed to be a "humidity-discharge coupled type" anomaly, which is very likely a precursor to insulation dampness or pyrolysis.

[0093] Risk Level Classification and Causal Inference: Based on the anomaly scores and validation results, the model classifies risks into three levels:

[0094] Attention level: However, the coupling is not significant, or the abnormal duration is <60 seconds;

[0095] Warning level: Furthermore, the "humidity-discharge" coupling is significant, or the abnormal duration is >300 seconds;

[0096] Hazard level: Significantly higher than (e.g., S>0.9), and characteristic gases (CO, ...) are detected simultaneously. The concentration rose sharply;

[0097] Meanwhile, the model infers possible causes by analyzing the contribution of each feature to the anomaly score (i.e., the ranking of feature importance). For example, if the top three features in terms of contribution are "humidity change rate", "partial discharge amplitude", and "temperature rise gradient", the cause is inferred to be "moisture intrusion leading to a decrease in insulation performance, possibly accompanied by overheating risk".

[0098] The model ultimately generates a structured fire risk warning message, which includes: risk level, anomaly score, start and end time, cause inference, and related characteristic data. This message is marked as the highest priority, jumps into the gateway communication queue for priority upload, and simultaneously triggers local audible and visual alarms. It also pre-activates related automatic devices according to the risk level (e.g., for hazardous levels, it directly activates the on-site pre-programming procedure of the automatic fire extinguishing device, which is executed after central confirmation).

[0099] Edge-side instant response: To meet real-time requirements (such as coordinated sprinkler or automatic shut-off), some emergency commands that do not require central decision-making are issued directly by the gateway. When the warning model outputs a "dangerous" level, the gateway's DO digital output interface can directly drive the on-site audible and visual alarms, or send pre-action commands to the nearby drainage / dehumidification / fire extinguishing control cabinet via RS485, controlling the end-to-end response time to the millisecond level, and then executing the final operation after confirmation by the central platform;

[0100] The remote monitoring center layer, serving as the core of the system's global data aggregation, in-depth analysis, and decision-making command, is deployed in a hybrid cloud architecture on an enterprise private cloud or public cloud server cluster. It establishes long-term connections with multiple edge computing gateway layers distributed across different geographical regions via secure and encrypted VPN or dedicated lines. This layer consists of three core modules: a data platform and GIS visualization platform, a multi-source data fusion and diagnostic module, and an emergency response and remote control module. This enables a leap from "single-point monitoring" to "regional situational awareness," from "post-event analysis" to "pre-event prediction," and from "manual handling" to "automatic linkage."

[0101] The data platform and GIS visualization platform are responsible for the aggregation, governance and spatiotemporal visualization of massive heterogeneous monitoring data, providing unified data services and interactive interfaces for upper-level diagnosis and decision-making.

[0102] The platform deploys high-performance message queues (such as Kafka or RabbitMQ) to concurrently receive real-time data streams uploaded from thousands of edge gateways, including periodic status data (sampled at the second / minute level), early warning event messages, system self-check status messages, and historical supplementary data that resumes transmission after network outages.

[0103] Data cleaning and normalization: Real-time stream processing engines (such as Flink or Spark Streaming) parse, clean (remove outliers that are obviously beyond the physical range), complete (based on time series interpolation) and normalize the format of the raw data, and convert heterogeneous data reported by sensors from different manufacturers and models into an internal standard data model.

[0104] A hybrid storage strategy is employed, using time-series databases (such as InfluxDB or TimescaleDB) to store high-frequency sampled time-series data such as temperature, humidity, and partial discharge, with different retention policies (e.g., retaining raw data for 30 days and permanently retaining aggregated minute / hourly data); using relational databases (such as PostgreSQL) to store structured data such as device ledgers, user information, alarm records, and work order records; and using distributed file systems (such as HDFS or MinIO) to store unstructured data such as waveform files and log files.

[0105] The GIS visualization platform, based on a Geographic Information System (GIS) engine (such as GeoServer or ArcGIS Server), establishes a city-level spatial database of cable trenches. Every cable trench, every manhole cover, and every set of sensors is precisely associated with geographic coordinates, and topological relationships (such as upstream and downstream relationships, substations, and power supply areas) are established.

[0106] Multi-dimensional situational awareness, including:

[0107] Macro-level overview: The main interface displays a map of the entire city or region, visually presenting the average humidity level, water accumulation status, and alarm density of each cable trench section through heat maps of different colors or aggregated icons. It supports multi-dimensional filtering and drill-down by administrative region, substation, voltage level, and other criteria.

[0108] Microscopic details: Clicking on a specific monitoring point will bring up the historical trend curve of that point (supporting multi-parameter comparison, such as the correlation between humidity and partial discharge), real-time data dashboard, equipment ledger information, and a list of recent alarm events;

[0109] Thematic map analysis: Supports dynamic generation of thematic layers such as "condensation risk distribution map", "fire hazard heat map", and "equipment online rate statistics chart" to help maintenance personnel quickly identify high-risk areas;

[0110] Data push and subscription: The platform has a built-in data service bus (DSB) that provides standardized RESTful API interfaces or MQTT subscription services for upper-layer diagnostic systems and third-party systems (such as production management systems and mobile operation and maintenance apps) to realize real-time push and sharing of monitoring data;

[0111] The multi-source data fusion diagnostic module is responsible for in-depth mining and trend prediction of the aggregated global data, and for building a cable trench health assessment and fire risk prediction model based on big data analysis, as detailed below:

[0112] With feature engineering and data cube construction, the system not only relies on the monitoring data of the cable trench itself, but also accesses external meteorological data (such as hourly forecast data from the meteorological bureau: precipitation, relative humidity, wind speed, and temperature) and power grid operation data (such as regional load curves and maintenance plans from the dispatching system) through API interfaces on a regular basis.

[0113] Time series feature extraction: For each monitoring point, statistical features including long-term (e.g., the past 30 days), medium-term (e.g., the past 7 days), and short-term (e.g., the past 24 hours) features are constructed, such as mean humidity, humidity variation coefficient, maximum temperature rise rate, and cumulative number of partial discharge pulses.

[0114] The data cube organizes the aforementioned multi-source data into a multi-dimensional data cube according to the "spatial dimension (route / site) - time dimension (moment / period) - feature dimension (monitoring / meteorology / operation)" to provide upper-level algorithm models with flexible slicing and drill-down analysis;

[0115] The humidity trend prediction model, as one of the core sub-modules of the multi-source data fusion diagnostic system, is responsible for quantitatively predicting the future trend of humidity changes in the cable trench environment, providing forward-looking data support for condensation risk early warning and fire probability assessment. This module adopts a deep learning-based time-series prediction architecture, realizing autonomous learning of humidity evolution patterns and multi-step advanced prediction.

[0116] Long Short-Term Memory Neural Network (LSTM) and Gated Recurrent Unit (GRU) are selected as the core prediction engine. Both are variants of Recurrent Neural Network (RNN) and can effectively solve the gradient vanishing and gradient explosion problems that exist in traditional RNN in long sequence modeling.

[0117] LSTM Unit Design: LSTM achieves selective memorization and forgetting of information in the cell state by introducing three gating structures: a forget gate, an input gate, and an output gate. The computation process of the LSTM unit is as follows:

[0118] Forgotten Gate: This determines which information to discard from the cell's state at the previous moment;

[0119] Input Gate: , Decide which new information to store in the cell state;

[0120] Cell status update: ;

[0121] Output gate: , ;

[0122] The GRU unit, a simplified variant of LSTM, merges the forget gate and input gate into an update gate and introduces a reset gate, resulting in fewer parameters and higher training efficiency. The computation process of the GRU unit is as follows:

[0123] Update Gate: It controls the degree to which the previous state is carried over into the current state;

[0124] Reset Door: This controls the degree to which the state of the previous moment is ignored;

[0125] Current candidate status: ;

[0126] Final output: ;

[0127] Model Selection Strategy: This module supports adaptive switching between LSTM and GRU models based on different application scenarios. For scenarios with high prediction accuracy requirements and relatively abundant computing resources (such as main cable trenches), a multi-layer stacked LSTM architecture is preferred to fully explore the long-term dependencies of humidity time series. For scenarios requiring rapid response and edge deployment (such as branch cable trenches), the GRU architecture is preferred to improve computational efficiency while ensuring prediction accuracy. Existing research shows that LSTM can achieve over 98% accuracy in temperature and humidity prediction, while GRU maintains similar prediction performance with approximately 30% fewer parameters.

[0128] Data preprocessing and feature engineering specifically include:

[0129] Input feature construction: The model input is a multi-dimensional time series feature vector. ,include:

[0130] Historical humidity sequence: Ambient humidity values ​​sampled every 10 minutes over the past 48 hours (288 time steps in total);

[0131] Historical temperature series: ambient temperature and cable joint surface temperature sampled synchronously with humidity;

[0132] External meteorological characteristics: Hourly forecast data for the next 24 hours obtained through the meteorological API interface, including relative humidity forecast, probability of precipitation, wind speed, and air pressure;

[0133] Temporal context features: periodic features such as timestamp encoding (hour, day of the week, season) and holiday identifiers;

[0134] Data normalization and sliding window: Z-score normalization is performed on all continuous features to eliminate the influence of units. A sliding window method is used to construct training samples with a window length of 48 hours, a prediction step of the next 24 hours (i.e., multi-step prediction), and a sliding step of 1 hour to achieve sample augmentation;

[0135] Model training and optimization specifically include:

[0136] The training dataset was constructed using over three years of historical cable trench humidity monitoring data and corresponding historical meteorological data for model training. The training, validation, and test sets were divided chronologically in a 7:1:2 ratio to prevent future data leakage.

[0137] The loss function and optimizer are as follows: mean squared error (MSE) is used as the loss function, Adam optimizer is used for parameter updates, the initial learning rate is set to 0.001, and an early stopping mechanism is introduced to prevent overfitting. Training is terminated when the validation set loss does not decrease for 10 consecutive epochs.

[0138] Multi-step prediction strategy: A sequence-to-sequence (Seq2Seq) prediction architecture is adopted. The encoder encodes the input sequence into a context vector, and the decoder generates the predicted values ​​for the next 24 hours step by step. To further improve prediction accuracy, an attention mechanism can be introduced into the encoder-decoder structure, allowing the model to dynamically focus on key time steps in the input sequence during the decoding process.

[0139] Rolling forecasts and uncertainty quantification specifically include:

[0140] Rolling prediction mechanism, the model uses the current time as the basis for prediction. Starting from this point, input the measured data from the past 48 hours. and weather forecast data for the next 24 hours A single inference outputs a humidity prediction sequence for the next 24 hours. Every hour, the model re-executes predictions based on the latest measured data, achieving a rolling update.

[0141] Confidence interval estimation employs Monte Carlo Dropout, an approximate Bayesian inference method. During the prediction phase, the Dropout layer remains active, and N forward propagations (N=100) are performed to obtain N sets of prediction results. The mean and standard deviation at each time step are calculated to generate predicted values ​​and 95% confidence intervals. The width of the confidence interval reflects the uncertainty of the prediction, providing a risk assessment basis for subsequent early warning decisions.

[0142] The core criterion for humidity exceeding the limit early warning logic is: the predicted sequence. Perform time-step scanning and count those that continuously exceed the humidity threshold. (Default value 85%RH, can be adaptively adjusted according to cable type, insulation class, and historical operating data) Maximum duration ;

[0143] Warning trigger conditions: If there are consecutive exceedances The time period, and that time period occurs within the next 6 hours (i.e. ), and duration Exceeding the preset duration threshold (The default value is 120 minutes, which can be dynamically adjusted according to the season and regional pollution level), then it is judged as "high risk of humidity exceeding the limit";

[0144] The module generates structured early warning messages, including: risk level (divided into "Attention Level": exceeding limits within the next 24 hours but not within the next 6 hours; "Warning Level": exceeding limits within the next 6 hours and the duration of the exceedance is <). "High-risk level": Exceeding limits within the next 6 hours and the duration of exceeding limits is ≥ The system includes the predicted period of exceeding limits, the predicted peak humidity value, the confidence interval width, and related characteristics (such as the predicted temperature for the same period and the condensation risk index). This warning message serves as input to the emergency response and remote control module, triggering corresponding preventative control measures (such as starting the dehumidifier fan in advance and increasing the frequency of partial discharge monitoring).

[0145] The cable trench fire probability prediction model, as another core sub-model of the multi-source data fusion diagnostic system, is responsible for real-time assessment and dynamic prediction of the probability of fires and insulation faults in cable trenches. This model adopts an ensemble learning architecture based on gradient boosting trees, integrates multi-dimensional spatiotemporal feature data, and constructs a classification model with strong generalization ability and high interpretability. It achieves a leap from "single-parameter limit alarm" to "multi-factor fusion risk assessment," specifically including the following processes:

[0146] (I) Model Selection and Input Feature Construction: This module selects Extreme Gradient Boosting (XGBoost) and Lightweight Gradient Boosting Machine (LightGBM) as the core algorithm engines. Both are ensemble learning algorithms based on decision trees, which iteratively train weak classifiers and combine them with weights through a boosting strategy, and have the following advantages:

[0147] Handling high-dimensional sparse features: It can automatically handle missing values ​​without additional imputation;

[0148] Strong nonlinear fitting capability: It captures complex interaction relationships between features through the combination of tree models;

[0149] Good interpretability: It can output the ranking of feature importance, which facilitates risk tracing;

[0150] High training efficiency: LightGBM introduces one-sided gradient sampling (GOSS) and mutually exclusive feature binding (EFB) techniques, which significantly outperforms the traditional Boosting algorithm in training speed when dealing with massive amounts of data;

[0151] Preferably, depending on the data scale and scenario requirements, the module supports adaptive switching between XGBoost and LightGBM: for scenarios with high feature dimensions and large sample size (such as centralized deployment at the provincial level), LightGBM is preferred to improve training efficiency; for core areas requiring higher accuracy, XGBoost is preferred and supplemented with hyperparameter grid search optimization.

[0152] Input feature construction: The model uses multidimensional feature vectors extracted from the data cube. As input, the feature dimensions comprise 8 major categories and over 60 dimensions, specifically including:

[0153] Humidity characteristics: current humidity value, average humidity over the past 1 / 6 / 12 / 24 hours, maximum humidity, humidity coefficient of variation, peak humidity change rate, and cumulative duration of humidity exceeding 85%RH;

[0154] Temperature characteristics: current cable joint temperature, temperature rise rate (5-minute / 30-minute gradient), temperature difference between joint and ambient, highest temperature in the past 24 hours;

[0155] Partial discharge characteristics: current partial discharge amplitude, pulse frequency, discharge phase distribution characteristics (statistical parameters of PRPS spectrum), and partial discharge activity index (cumulative value of discharge energy) over the past 24 hours.

[0156] Load characteristics: current load current, load change rate, average load over the past 24 hours, peak load;

[0157] Meteorological characteristics: average ambient humidity (external) over the past 3 hours, cumulative rainfall over the past 24 hours, and rainfall probability forecast for the next 6 hours;

[0158] Static attribute characteristics: cable service life, insulation material type (XLPE / oil paper, etc.), cable cross-sectional specifications, number of historical defects (such as external force damage, intermediate joint failure records), and number of days since maintenance records were made.

[0159] Temporal context features: season (spring / summer / autumn / winter), time period (peak / valley / flat), whether it is a holiday;

[0160] Derivative features: Humidity prediction features (predicted peak humidity and probability of exceeding limits in the next 6 hours), condensation risk index (a sequence of ΔT values ​​calculated and reported by the edge gateway);

[0161] (II) Sample Construction and Imbalance Handling Strategies

[0162] Sample label definition: "Whether a fire / insulation failure occurred" is used as the binary label. Positive samples (y=1) are defined as events such as cable trench fires, cable joint breakdowns, insulation erosion, and surface flashovers that have been manually confirmed; negative samples (y=0) are defined as time slices randomly selected from normal operating periods that have no fault records within 3 months before and after the event.

[0163] Sample construction strategies include:

[0164] Positive sample construction: Extracting event occurrence times from historical accident logs ,by Using the reference point, cut forward. Feature data within a time period are used as positive samples. The observation window is set to 7 days by default, but can be adjusted according to the fault type. Each fault event can generate multiple positive samples (e.g., one sample per day) to increase the number of positive samples.

[0165] Negative sample construction: Randomly select time points from the normal operation period of the entire network, remove the 30-day buffer period before and after the failure event, and construct an initial set of negative samples of the same order of magnitude;

[0166] Handling Imbalanced Positive and Negative Samples: To address the extreme imbalance issue of cable trench fires being a low-probability event with far fewer positive samples than negative samples, this module introduces a dual processing mechanism:

[0167] SMOTE oversampling: Employing the Synthetic Minority Oversampling Technique (SMOTE), this technique selects the nearest neighbors of positive samples in the feature space and synthesizes new positive samples through linear interpolation, rather than simply copying them, thus avoiding overfitting. In its implementation, for each positive sample... Randomly select one of its k nearest neighbors of the same type. Synthesize new samples ,in It is a random number;

[0168] Cost-sensitive learning: During model training, class weights are introduced into the objective function, assigning a higher penalty coefficient to misclassification of positive samples. Both XGBoost and LightGBM support setting the weight ratio of positive and negative samples through the scale_pos_weight parameter, which defaults to the ratio of the number of negative samples to the number of positive samples, making the model pay more attention to the correct classification of minority class samples;

[0169] (III) Model Training and Probability Output

[0170] Training-Validation-Test Split: The dataset is divided into a training set (first 60% of the time period), a validation set (middle 20% of the time period), and a test set (last 20% of the time period) in chronological order to ensure that future data is not introduced during model evaluation;

[0171] Hyperparameter optimization: Key hyperparameters of the model are tuned using Bayesian optimization or random search, including: maximum tree depth (max_depth), learning rate (learning_rate), subsample ratio (subsample), feature sampling ratio (colsample_bytree), L1 / L2 regularization coefficients, etc. The AUC (Area Under Curve) on the validation set is used as the optimization objective.

[0172] Probabilistic Calibration: The scores output by the original XGBoost / LightGBM are not strictly probability values ​​and exhibit distribution shifts. This module introduces **Platt scaling or isotonic regression** to perform probabilistic calibration on the output scores, ensuring the calibrated output... It is closer to the true probability distribution;

[0173] Real-time inference: After the model is deployed, an inference calculation is triggered every hour for each monitoring point, using the feature vector at the current moment. Input: calibrated fire probability ;

[0174] (iv) Risk level classification and interpretability output

[0175] Risk level classification: based on The size of the fire hazard is used to classify fire risk into four levels, corresponding to different levels of attention and response strategies:

[0176] Level I (Low Risk): The system only records logs and does not generate proactive alarms;

[0177] Level II (Attention Level): The system highlights the area in yellow on the monitoring interface to alert maintenance personnel to pay attention to it.

[0178] Level III (Warning Level): The system generates an early warning message and pushes it to the team leader, suggesting that monitoring be strengthened or on-site inspections be arranged;

[0179] Level IV (High Risk): The system immediately triggers the emergency response procedure, notifies relevant personnel, and prepares to activate the fire extinguishing equipment;

[0180] Feature importance ranking and causal analysis, specifically including:

[0181] Global Explanation: After the model is trained, it outputs a feature importance ranking map based on information gain or number of splits, which helps operations and maintenance personnel understand which factors contribute the most to fire risk (such as humidity features, which often rank highly).

[0182] Local Explanation: For each specific early warning event, the SHAP (SHapley Additive Explanations) value is used to explain the current sample. The SHAP value is based on the Shapley value in game theory, quantifying the marginal contribution of each feature to the final prediction result. For samples with high current Pfire, the model automatically calculates the SHAP value of each feature, sorts them by absolute value, and generates a readable natural language description;

[0183] This explanatory information is pushed along with the early warning message to provide decision-making support for operations and maintenance personnel.

[0184] (v) Continuous Model Optimization and Version Management

[0185] Incremental learning mechanism: The system has online incremental learning capabilities. Newly generated fault cases are automatically or semi-automatically added to the training set after manual confirmation. A warm-start approach is used, incrementally training is performed on the existing model, updating only a portion of the tree structure or adding weak learners, avoiding the computational overhead of training from scratch. LightGBM natively supports incremental learning (using the `refit` or `continue_train` parameter), and XGBoost can achieve similar results by loading historical models and continuing training.

[0186] Periodic retraining: In addition to incremental learning, the system performs a full retraining every quarter, using the latest accumulated data to build a brand new model in order to capture long-term drift in data distribution.

[0187] A / B testing and canary release of versions:

[0188] Parallel Deployment: After the new version model is trained, it is not directly replaced by the old version. Instead, it is deployed in parallel using an A / B testing architecture. The old and new models perform inference calculations on the same real-time data stream simultaneously, but the output of the old model is still used for actual early warning decisions, while the output of the new model is only used for performance comparison.

[0189] Performance evaluation: Compare the performance metrics of the new and old models over an observation period (e.g., 30 days), including: accuracy, recall, F1 score, AUC, and false positive rate (highly monitored). If the new model significantly reduces the false positive rate (e.g., a relative decrease of more than 10%) without reducing the recall, then the new model is considered superior to the old model.

[0190] Gray-scale flow switching: A gradual flow switching strategy is adopted, initially allowing the new model to handle 5% of the decision traffic. After observing no anomalies for a week, the flow is gradually increased to 30%, then 70%, and finally completely replacing the old model. Throughout the process, the old model serves as a backup and can be quickly rolled back at any time.

[0191] Feedback loop: All on-site handling results (such as "confirmed false alarm", "confirmed real alarm", "handling completed") are recorded and fed back to the sample database as the basis for label correction in subsequent model optimization, forming a complete closed loop of "monitoring-early warning-handling-feedback-optimization";

[0192] The emergency response and remote control module is responsible for translating diagnostic results into specific control commands and handling procedures, achieving a closed loop from "intelligent sensing" to "precise control." Specifically, it includes:

[0193] Emergency Response Plan Rule Engine: The system has a built-in visual rule engine that allows operations and maintenance personnel to configure differentiated emergency response plans based on risk level, event type, and location. Example rule: IF (Fire Probability Level == Level IV AND Area == "Core Urban Area") THEN (Execute Plan A: Issue instructions to activate automatic fire suppression systems + Notify area manager + Notify fire control room);

[0194] It comes pre-set with a variety of typical emergency response plan templates, such as "Condensation Handling Plan" (start dehumidifying fan, enhance monitoring), "Water Accumulation Handling Plan" (start drainage pump, notify emergency repair), and "High-Risk Fire Plan" (start fire extinguishing equipment, cut off non-fire power supply, start evacuation instructions, notify 119), etc., and supports user customization and modification;

[0195] Remote command issuance and closed-loop control: Command encapsulation and routing: When the system is diagnosed as being in a high-risk state and triggers the emergency plan, the command generation service encapsulates standardized control commands (JSON format or Modbus TCP packets) according to the plan content. These commands include the target device ID (e.g., the No. 3 drainage pump in a cable trench), action type (start / stop / query), and action parameters (runtime, speed). The commands are then sent to the corresponding edge computing gateway via a message queue.

[0196] The system employs a four-step handshake protocol: "command-confirmation-execution-feedback". Upon receiving a command, the gateway first replies "received"; after executing the device action, it reports the "execution result"; if no confirmation is received within a timeout or execution fails, the central platform will retry or escalate the alarm.

[0197] For high-risk actions (such as activating fire extinguishing devices), the system can link with the video surveillance cameras in the area, automatically preset positions and record on-site video, which can be manually confirmed by remote duty personnel before the final action is executed, or the system can automatically record video after the action is executed for future reference, realizing three-dimensional verification of "remote signaling-remote measurement-remote viewing";

[0198] The closed loop for mobile terminal alarms and handling includes:

[0199] Multi-channel alarm push: The platform uses mobile push services (such as JPush or APNs) to push alarm information to the mobile APP or SMS platform of maintenance personnel in a tiered manner. Alarm tiering strategy: Level I-II risks are only pushed to the system logs and web interface; Level III risks are pushed to the team leader; Level IV risks are pushed to the team leader, department leader, and supervisor.

[0200] Mobile-based work orders: Alarm information is presented as structured cards on mobile devices, including: risk level, occurrence time, specific location (GIS map link), risk description, cause inference, and suggested remedial measures. Maintenance personnel can perform operations such as "confirm receipt," "on-site check-in," "remediation completed," and "request review" on their mobile devices. Operation records are transmitted back to the central platform in real time, forming a closed-loop management system from "alarm triggering" to "remediation archiving."

[0201] Voice broadcast and AI outbound calling: For high-risk alarms, the system can automatically trigger AI voice outbound calling to notify on-duty personnel through telephone voice broadcast, avoiding missed reports due to silent phones or unread messages.

[0202] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A smart humidity monitoring and remote control system for cable trenches, characterized in that, include: The intelligent data acquisition terminal layer is deployed at key monitoring points in the cable trench to collect real-time data on ambient temperature and humidity, cable joint temperature, partial discharge, water level, and gas concentration, and performs edge preprocessing on the data. The edge computing gateway layer is communicatively connected to the intelligent acquisition terminal layer. It is used to receive and aggregate multi-source monitoring data, run edge intelligent analysis models to identify condensation risks and early fire characteristics, generate early warning information and upload it. The remote monitoring center layer is communicatively connected to the edge computing gateway layer. It is used to receive data and early warning information uploaded by the gateway layer, run a multi-source data fusion diagnostic model to perform global situation analysis and fire probability prediction, and issue remote control commands.

2. The intelligent humidity monitoring and remote control system for cable trenches according to claim 1, characterized in that, The intelligent data acquisition terminal layer includes: The composite environmental monitoring unit collects data on ambient temperature, relative humidity, water level, and gas concentration, and has a built-in damp heat aging coefficient calculation model to generate early warning information on insulation degradation. The cable condition monitoring unit collects cable joint temperature and partial discharge signals, and dynamically adjusts the acquisition frequency according to the temperature change rate. The low-power self-powered module includes a CT power harvesting unit and a backup battery unit, and integrates intelligent energy routing function to achieve seamless switching between main and backup power and sleep control of non-core sensors.

3. The intelligent humidity monitoring and remote control system for cable trenches according to claim 2, characterized in that, The temperature, humidity, and condensation monitoring module in the composite environmental monitoring unit uses a digital temperature and humidity sensor. The bus connects to the MCU, which has a built-in damp heat aging index. The computational model is expressed as follows: in For temperature gradient, Humidity abrupt change rate, For the duration of high temperature and high humidity, , , This is a preset dynamic weighting coefficient.

4. The intelligent humidity monitoring and remote control system for cable trenches according to claim 1, characterized in that, The edge computing gateway layer includes: The multi-protocol data aggregation and protocol conversion module is used to access terminal devices with heterogeneous communication standards and complete data parsing and unified encapsulation. The edge intelligence analysis and early warning module, with its embedded lightweight machine learning model, is used to identify condensation risks and early fire characteristics in real time. The system self-diagnosis and health management module is used to monitor the gateway's own operating status and communication link quality, and to activate local storage and breakpoint resume mechanism when the network is interrupted.

5. The intelligent humidity monitoring and remote control system for cable trenches according to claim 4, characterized in that, The edge intelligent analysis and early warning module includes: The condensation risk warning sub-model calculates the dew point temperature based on ambient temperature, humidity and cable joint temperature using the Magnus formula to determine the condensation risk. The fire early warning sub-model constructs an isolated forest model based on multi-dimensional feature vectors, outputs anomaly scores and performs coupling verification to generate fire risk levels and cause inferences.

6. The intelligent humidity monitoring and remote control system for cable trenches according to claim 1, characterized in that, The remote monitoring center layer includes: The data platform and GIS visualization platform are used for multi-source data aggregation, spatiotemporal correlation storage, and situational display; The multi-source data fusion diagnostic module runs a humidity trend prediction model and a fire probability prediction model, and outputs a future humidity exceeding the limit warning and the probability of fire occurrence. The emergency response and remote control module has a built-in rule engine that automatically generates control commands based on the risk level and sends them to the edge gateway for execution.

7. The intelligent humidity monitoring and remote control system for cable trenches according to claim 6, characterized in that, The humidity trend prediction model adopts an LSTM or GRU neural network architecture, inputs historical humidity, temperature, weather forecast and time series features, outputs the humidity prediction sequence for the next 24 hours and a 95% confidence interval, and generates a humidity over-limit warning based on the prediction results.

8. The intelligent humidity monitoring and remote control system for cable trenches according to claim 6, characterized in that, The fire probability prediction model uses XGBoost or LightGBM ensemble learning algorithms. The input includes a multi-dimensional feature vector containing humidity, temperature, partial discharge, load, meteorological data, static attributes, and derived features. The output is a calibrated fire occurrence probability, which is combined with SHAP values ​​for causal explanation.

9. The intelligent humidity monitoring and remote control system for cable trenches according to claim 1, characterized in that, The edge computing gateway layer and the remote monitoring center layer establish a long connection through an encrypted communication protocol, supporting real-time data upload, breakpoint resume, command issuance and execution feedback.

10. The intelligent humidity monitoring and remote control system for cable trenches according to claim 1, characterized in that, The system also supports closed-loop management of mobile terminal alarm push and handling. Operation and maintenance personnel can receive hierarchical alarms, view risk details, complete handling work orders, and transmit handling results back in real time through a mobile APP.

Images