An electrical fire hazard dynamic monitoring and early warning method based on multi-sensor fusion

Abstract

This invention discloses a dynamic monitoring and early warning method for electrical fire hazards based on multi-sensor fusion. The method deploys a multi-source heterogeneous sensor array in the target electrical circuit to simultaneously collect electrical parameters, environmental parameters, and physical state parameters. It employs an adaptive sampling frequency to synchronously collect data from each sensor. A long short-term memory network based on an attention mechanism is used to perform feature-level fusion on the preprocessed time-series data to extract fault evolution features. The fused features are input into a pre-trained Bayesian network inference model to calculate the probability of occurrence of each type of fire hazard under the current state. Finally, a dynamic threshold judgment strategy is combined to achieve a graded response from early hazard warning to emergency alarm. This invention solves the problems of strong lag and high false alarm rate in traditional single-threshold alarms, realizing the transformation of electrical fire monitoring from "steady-state monitoring" to "dynamic evolution trend prediction," significantly improving the accuracy and lead time of early warning.

Description

Technical Field

[0001] This invention relates to the field of electrical safety monitoring technology, specifically to a method for dynamic monitoring and early warning of electrical fire hazards based on multi-sensor fusion. Background Technology

[0002] With the increasing electrification of modern society, electrical equipment is widely used in industrial production, commercial buildings, and residential life. Consequently, electrical fires have become one of the major types of disasters threatening public safety. According to statistics from fire departments, electrical fires consistently account for a high proportion of all types of fire accidents, often accompanied by significant property damage and the risk of personal injury. To address this challenge, electrical fire monitoring systems are widely used both domestically and internationally. Their core monitoring methods primarily rely on residual current devices (RCDs) and temperature detectors. These systems typically install sensors in the power distribution circuit to monitor leakage current and the temperature of critical nodes in real time. Once parameters exceed preset thresholds (such as leakage current > 300mA or temperature > 85℃), an audible and visual alarm is triggered, or the power is cut off. Furthermore, some high-end systems incorporate harmonic analysis and arc fault detection technologies, attempting to capture fault characteristics at the current waveform level.

[0003] While the aforementioned traditional technologies have played a role in preventing some electrical accidents, their limitations are becoming increasingly apparent when faced with increasingly complex electrical fire hazards, mainly in the following three aspects:

[0004] The fragmentation and lag of monitoring dimensions: Traditional methods often rely on a single or a few parameters for independent judgment. For example, the residual current method is effective for grounding faults, but it is extremely insensitive to ungrounded phase-to-phase short circuits or simple overheating due to poor contact (without obvious leakage); while temperature monitoring is intuitive, thermocouples or thermistors are often installed on the surface, and the temperature only rises when heat is conducted to the sensor location. By this time, the internal insulation may have already been severely carbonized, missing the best time for intervention.

[0005] The rigidity of static thresholds and high false alarm rate: Existing systems generally use fixed alarm thresholds, which are difficult to adapt to complex and changing field environments. During the high temperatures of summer or peak load periods, the normal operating temperature of the equipment may be close to the alarm threshold, which can easily lead to false alarms; conversely, in low-temperature environments or under light load conditions, even if there is slow-developing insulation aging, parameter changes may be drowned out by noise, resulting in missed alarms.

[0006] Lack of understanding of fault evolution trends: Most current monitoring equipment remains at the "state snapshot" level, that is, judging whether the current time exceeds the standard, lacking the ability to analyze the trend of parameter changes. The occurrence of electrical fires is often a gradual physical and chemical process from electric spark, local overheating, insulation carbonization to the final ignition. Traditional technologies cannot identify this early and subtle evolutionary feature.

[0007] In summary, existing electrical fire monitoring technologies are facing a technological bottleneck in transitioning from "passive response" to "proactive prevention." The industry urgently needs a new monitoring method that can break down sensor data silos and integrate multi-dimensional information (electrical, physical, and chemical quantities). Simultaneously, to resolve the contradiction between false alarms and missed alarms, it is necessary to introduce intelligent algorithms with self-learning and adaptive capabilities, capable of dynamically adjusting judgment criteria based on environmental changes and load conditions, and using time-series data analysis techniques to predict fault development trends.

[0008] Therefore, we propose a dynamic monitoring and early warning method for electrical fire hazards based on multi-sensor fusion. Summary of the Invention

[0009] To achieve the above objectives, the present invention provides the following technical solution: a method for dynamic monitoring and early warning of electrical fire hazards based on multi-sensor fusion, comprising the following steps:

[0010] Step S1: Construct a multi-source heterogeneous sensor network, the sensor network comprising:

[0011] The electrical parameter sensing unit is configured to collect three-phase current, three-phase voltage, and voltage harmonic distortion rate.

[0012] The environmental parameter sensing unit is configured to collect the temperature at the cable connector, the smoke concentration in the space, and the carbon monoxide gas concentration.

[0013] The physical state sensing unit is configured to collect partial discharge pulse signals on the surface of the cable insulation layer and the mechanical deformation stress of the cable.

[0014] Each sensing unit is connected to the edge computing gateway via a time-division multiplexed bus.

[0015] Step S2: Spatiotemporal reference unification and adaptive sampling, clock synchronization of the multi-source heterogeneous sensor network, and dynamic adjustment of the sampling frequency according to the fluctuation characteristics of the current signal;

[0016] Step S3: Data preprocessing and anomaly correction, performing outlier removal, normalization, and noise filtering based on unscented Kalman filtering on the acquired raw data stream;

[0017] Step S4: Deep feature-level fusion, constructing a long short-term memory network model based on the attention mechanism, inputting the preprocessed multi-dimensional time series data into the model, and extracting a deep fusion feature vector containing fault evolution features;

[0018] Step S5: Bayesian probabilistic inference. The fused feature vector output from step S4 is used as evidence input to the pre-trained Bayesian network inference engine to calculate the posterior probability distribution of the four types of hidden dangers in the current state: short circuit, overload, poor contact and insulation aging.

[0019] Step S6: Dynamic hierarchical early warning decision-making. Based on the posterior probability distribution and its slope over time, query the dynamic threshold matrix and output early warning signals of different levels or execute linkage protection actions.

[0020] Preferably, the adaptive sampling in step S2 specifically includes:

[0021] Calculate the variance σ2 of the effective current value (Irms) within the set time window in real time;

[0022] When σ2>σhigh2 (first preset threshold), the circuit determines that a transient process has occurred and increases the sampling frequency to fhigh=10kHz to capture the high-frequency harmonics and transient arc characteristics of the current.

[0023] When σ2 < σlow2 (the second preset threshold), the determination circuit is in steady-state operation, and the sampling frequency is reduced to flow = 1Hz to reduce system power consumption;

[0024] Among them, σhigh2>σlow2, and fhigh>flow.

[0025] Preferably, the unscented Kalman filtering process in step S3 specifically includes:

[0026] Establish a nonlinear state-space model that includes temperature, current, and deformation stress;

[0027] By selecting the Sigma point set through unscented transformation (UT), and propagating it through the nonlinear state equation, the predicted value is weighted and corrected using the measured value, thereby eliminating temperature measurement jitter caused by air convection disturbance and current glitches caused by electromagnetic interference.

[0028] Preferably, the Attention-LSTM model in step S4 includes an input layer, a two-layer LSTM hidden layer, a multi-head attention weight layer, and a fully connected output layer connected in sequence.

[0029] The multi-head attention weighting layer is configured to compute multiple different attention heads in parallel, and each attention head calculates the weight coefficient αt of the input data at different times through a scaled dot product attention mechanism.

[0030] By weighted summing of the outputs ht of the LSTM hidden layer at each time step, C=∑t=1Tαtht, a context vector C is generated, which is the deep fusion feature vector.

[0031] Preferably, the Bayesian network inference engine construction process in step S5 includes:

[0032] Define network nodes, including observation nodes (E={I,V,T,PD,CO}) and hypothesis nodes (H={Short,Overload,LooseContact,Aging});

[0033] The network parameters are learned using a historical failure case database, and the conditional probability table (CPT) between nodes is populated.

[0034] During the online monitoring phase, a particle filter algorithm is used to track parameter drift caused by equipment aging and update the prior probability distribution in the CPT table in real time.

[0035] Preferably, the dynamic threshold matrix in step S6 is adaptively corrected based on ambient temperature and load rate, and the specific calculation formula is as follows:

[0036] Thresholdfinal=Base_Threshold×(1+k1⋅(Tenv−Tref))×(1−k2⋅(η−ηref))

[0037] Where Base_Threshold is the base threshold, Tenv is the real-time ambient temperature, Tref is the reference temperature (25℃), η is the real-time load rate, ηref is the rated load rate, and k1 and k2 are empirical correction coefficients.

[0038] Preferably, the graded early warning signal in step S6 includes:

[0039] Level 1 Warning (Alert): When the probability of a potential hazard P(Hi) > 0.6 and the slope of P(Hi) > 0, the system logs the information and displays a yellow alert on the human-machine interface (HMI).

[0040] Level 2 warning (alarm): When P(Hi)>0.8, the system sends an SMS / APP push alarm to the user terminal through the IoT platform;

[0041] Level 3 Early Warning (Emergency): When P(Hi)>0.95 or the temperature change rate dT / dt>10℃ / min, the system immediately issues a command to the circuit breaker to perform trip protection and lock the circuit.

[0042] Preferably, it also includes step S7: digital twin visualization mapping;

[0043] The fused feature vector generated in step S4 is mapped to the pre-constructed three-dimensional digital twin model of the power distribution cabinet;

[0044] In the three-dimensional model, the corresponding electrical components are rendered with different colors according to the different types of hidden dangers Hi, and the predicted temperature field distribution cloud map and insulation aging trend curve for the next 30 minutes are superimposed and displayed.

[0045] Preferably, the partial discharge sensor in the physical state sensing unit is a non-contact ultrasonic probe, installed on the outside of the cable joint shield, with its detection frequency set to 40kHz±10kHz, used to filter out power frequency electromagnetic interference and specifically for capturing partial discharge signals inside the insulation.

[0046] Preferably, the method further includes a self-learning feedback mechanism:

[0047] If the system issues an alert and the maintenance personnel do not intervene within the specified time and no actual fire occurs, the system determines it to be a false alarm and fine-tunes the weight parameters of the Attention-LSTM model and the CPT table of the Bayesian network through the backpropagation algorithm to achieve online evolution of the model.

[0048] Compared with existing technologies, this invention provides a method for dynamic monitoring and early warning of electrical fire hazards based on multi-sensor fusion, which has the following beneficial effects:

[0049] 1. This method for dynamic monitoring and early warning of electrical fire hazards based on multi-sensor fusion breaks through the limitations of traditional electrical fire monitoring, which relies solely on residual current or single-point temperature, by constructing a multi-source heterogeneous sensor array containing electrical parameters, environmental parameters, and physical state parameters. In particular, the introduction of physical state monitoring dimensions such as partial discharge and cable deformation enables the system to capture fault characteristics at the microscopic stage of insulation aging. Combined with a long short-term memory network based on an attention mechanism, this invention can not only analyze current static values ​​but also uncover the temporal evolution patterns behind the data, achieving a fundamental shift from traditional "steady-state threshold alarms" to "dynamic trend predictions." Compared to existing technologies that only trigger alarms when a fault is imminent, this invention can advance the warning time by minutes or even hours, providing users with a valuable emergency response window and truly achieving "prevention before the event."

[0050] 2. This method for dynamic monitoring and early warning of electrical fire hazards based on multi-sensor fusion addresses the high false alarm rate caused by strong electromagnetic interference and variable on-site environments in electrical environments by providing a dual optimization mechanism. On one hand, an unscented Kalman filter algorithm is used at the data input end to effectively eliminate noise fluctuations caused by air convection and electromagnetic coupling, preserving the true trends of physical quantity changes. On the other hand, a Bayesian network inference engine and a dynamic threshold matrix are introduced at the decision end, enabling the system to automatically adjust the alarm benchmark based on real-time ambient temperature and load rate. This "hardware-software combined" noise reduction and adaptive mechanism frees the system from rigid fixed thresholds, significantly reducing false alarms and missed alarms caused by seasonal temperature and humidity changes or equipment start-up and shutdown, and significantly improving the accuracy of alarms and the reliability of the system.

[0051] 3. This invention, based on multi-sensor fusion for dynamic monitoring and early warning of electrical fire hazards, not only achieves safety early warning in terms of function but also expands the operation and maintenance management mode of electrical systems at the application level. By integrating digital twin technology, the system can transform abstract electrical parameters into visualized three-dimensional model states (such as hotspot migration paths and insulation degradation curves), enabling maintenance personnel to intuitively locate fault points and understand the causes of faults. Furthermore, based on particle filtering-based online learning and model self-evolution mechanisms, the system can continuously optimize its judgment logic as equipment ages, achieving health management throughout the entire lifecycle of electrical equipment. This comprehensive solution, upgraded from simple "alarm triggering" to "diagnosis + prediction + visualization," greatly reduces the inspection burden on maintenance personnel and improves the overall intelligent management level of the power distribution system. Detailed Implementation

[0052] The technical solutions in the embodiments of the present invention will be clearly and completely described below. 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.

[0053] Example

[0054] An embodiment of a dynamic monitoring and early warning method for electrical fire hazards based on multi-sensor fusion

[0055] A method for dynamic monitoring and early warning of electrical fire hazards based on multi-sensor fusion includes the following steps:

[0056] Step S1: Construct a multi-source heterogeneous sensor network, the sensor network comprising:

[0057] The electrical parameter sensing unit is configured to collect three-phase current, three-phase voltage, and voltage harmonic distortion rate.

[0058] The environmental parameter sensing unit is configured to collect the temperature at the cable connector, the smoke concentration in the space, and the carbon monoxide gas concentration.

[0059] The physical state sensing unit is configured to collect partial discharge pulse signals on the surface of the cable insulation layer and the mechanical deformation stress of the cable.

[0060] Each sensing unit is connected to the edge computing gateway via a time-division multiplexed bus.

[0061] Step S2: Spatiotemporal reference unification and adaptive sampling, clock synchronization of the multi-source heterogeneous sensor network, and dynamic adjustment of the sampling frequency according to the fluctuation characteristics of the current signal;

[0062] Step S3: Data preprocessing and anomaly correction, performing outlier removal, normalization, and noise filtering based on unscented Kalman filtering on the acquired raw data stream;

[0063] Step S4: Deep feature-level fusion, construct a long short-term memory network model based on attention mechanism, input the preprocessed multi-dimensional time series data into the model, and extract a deep fusion feature vector containing fault evolution features;

[0064] Step S5: Bayesian probabilistic inference. The fused feature vector output from step S4 is used as evidence input to the pre-trained Bayesian network inference engine to calculate the posterior probability distribution of the four types of hidden dangers in the current state: short circuit, overload, poor contact and insulation aging.

[0065] Step S6: Dynamic hierarchical early warning decision-making. Based on the posterior probability distribution and its slope over time, query the dynamic threshold matrix and output early warning signals of different levels or execute linkage protection actions.

[0066] Specifically, the adaptive sampling in step S2 includes:

[0067] Calculate the variance σ2 of the effective current value (Irms) within the set time window in real time;

[0068] When σ2>σhigh2 (first preset threshold), the circuit determines that a transient process has occurred and increases the sampling frequency to fhigh=10kHz to capture the high-frequency harmonics and transient arc characteristics of the current.

[0069] When σ2 < σlow2 (the second preset threshold), the determination circuit is in steady-state operation, and the sampling frequency is reduced to flow = 1Hz to reduce system power consumption;

[0070] Among them, σhigh2>σlow2, and fhigh>flow.

[0071] Specifically, the unscented Kalman filtering process in step S3 is as follows:

[0072] Establish a nonlinear state-space model that includes temperature, current, and deformation stress;

[0073] By selecting the Sigma point set through unscented transformation (UT), and propagating it through the nonlinear state equation, the predicted value is weighted and corrected using the measured value, thereby eliminating temperature measurement jitter caused by air convection disturbance and current glitches caused by electromagnetic interference.

[0074] Specifically, the Attention-LSTM model in step S4 includes an input layer, a two-layer LSTM hidden layer, a multi-head attention weight layer, and a fully connected output layer, which are connected in sequence.

[0075] The multi-head attention weighting layer is configured to compute multiple different attention heads in parallel, and each attention head computes the weight coefficient αt of the input data at different times through a scaled dot product attention mechanism.

[0076] By weighted summing of the outputs ht of the LSTM hidden layer at each time step, C = ∑t=1Tαtht, a context vector C is generated, which is the deep fusion feature vector.

[0077] Specifically, the Bayesian network inference engine construction process in step S5 includes:

[0078] Define network nodes, including observation nodes (E={I,V,T,PD,CO}) and hypothesis nodes (H={Short,Overload,LooseContact,Aging});

[0079] The network parameters are learned using a historical failure case database, and the conditional probability table (CPT) between nodes is populated.

[0080] During the online monitoring phase, a particle filter algorithm is used to track parameter drift caused by equipment aging and update the prior probability distribution in the CPT table in real time.

[0081] Specifically, the dynamic threshold matrix in step S6 is adaptively corrected based on ambient temperature and load rate, and the specific calculation formula is as follows:

[0082] Thresholdfinal=Base_Threshold×(1+k1⋅(Tenv−Tref))×(1−k2⋅(η−ηref))

[0083] Where Base_Threshold is the base threshold, Tenv is the real-time ambient temperature, Tref is the reference temperature (25℃), η is the real-time load rate, ηref is the rated load rate, and k1 and k2 are empirical correction coefficients.

[0084] Specifically, the graded early warning signals in step S6 include:

[0085] Level 1 Warning (Alert): When the probability of a potential hazard P(Hi) > 0.6 and the slope of P(Hi) > 0, the system logs the information and displays a yellow alert on the human-machine interface (HMI).

[0086] Level 2 warning (alarm): When P(Hi)>0.8, the system sends an SMS / APP push alarm to the user terminal through the IoT platform;

[0087] Level 3 Early Warning (Emergency): When P(Hi)>0.95 or the temperature change rate dT / dt>10℃ / min, the system immediately issues a command to the circuit breaker to perform trip protection and lock the circuit.

[0088] Specifically, it also includes step S7: digital twin visualization mapping;

[0089] The fused feature vector generated in step S4 is mapped to the pre-constructed three-dimensional digital twin model of the power distribution cabinet;

[0090] In the 3D model, electrical components are rendered with different colors according to the different types of hazards (Hi), and the predicted temperature field distribution cloud map and insulation aging trend curve for the next 30 minutes are overlaid and displayed.

[0091] Specifically, the partial discharge sensor in the physical state sensing unit uses a non-contact ultrasonic probe, which is installed on the outside of the cable joint shield. Its detection frequency is set to 40kHz±10kHz to filter out power frequency electromagnetic interference and is specifically used to capture partial discharge signals inside the insulation.

[0092] Specifically, the method also includes a self-learning feedback mechanism:

[0093] If the system issues an alert and the maintenance personnel do not intervene within the specified time and no actual fire occurs, the system determines it to be a false alarm and fine-tunes the weight parameters of the Attention-LSTM model and the CPT table of the Bayesian network through the backpropagation algorithm to achieve online evolution of the model.

[0094] Through the above technical solution, this invention breaks through the limitations of traditional electrical fire monitoring, which relies solely on residual current or single-point temperature, by constructing a multi-source heterogeneous sensor array that includes electrical parameters, environmental parameters, and physical state parameters. In particular, the introduction of physical state monitoring dimensions such as partial discharge and cable deformation enables the system to capture fault characteristics at the microscopic stage of insulation aging. Combined with a long short-term memory network based on an attention mechanism, this invention can not only analyze current static values ​​but also uncover the temporal evolution patterns behind the data, achieving a fundamental shift from traditional "steady-state threshold alarms" to "dynamic trend predictions." Compared to existing technologies that only trigger alarms when a fault is imminent, this invention can advance the warning time by minutes or even hours, providing users with a valuable emergency response window and truly achieving "prevention before the event." Addressing the high false alarm rate caused by strong electromagnetic interference and variable on-site environments in electrical environments, this invention provides a dual optimization mechanism. On the one hand, an unscented Kalman filter algorithm is used at the data input end to effectively eliminate noise jitter caused by air convection and electromagnetic coupling, while retaining the true trend of physical quantity changes. On the other hand, a Bayesian network inference engine and a dynamic threshold matrix are introduced at the decision end, enabling the system to automatically adjust the alarm benchmark based on real-time ambient temperature and load rate. This "hardware-software combined" noise reduction and adaptive mechanism frees the system from rigid fixed thresholds, significantly reducing false alarms and missed alarms caused by seasonal temperature and humidity changes or equipment start-up and shutdown, and significantly improving the accuracy of alarms and the reliability of the system. This invention not only achieves safety early warning in terms of function, but also expands the operation and maintenance management mode of electrical systems at the application level. By integrating digital twin technology, the system can transform abstract electrical parameters into visualized three-dimensional model states (such as hotspot migration paths and insulation degradation curves), enabling maintenance personnel to intuitively locate fault points and understand the causes of faults. In addition, the online learning and model self-evolution mechanism based on particle filtering enables the system to continuously optimize its judgment logic as the equipment ages, achieving health management of electrical equipment throughout its entire life cycle. This comprehensive solution, which upgrades from simple "alarm triggering" to "diagnosis + prediction + visualization," greatly reduces the inspection burden on maintenance personnel and improves the overall intelligent management level of the power distribution system.

[0095] 1. System hardware architecture deployment

[0096] The system architecture in this embodiment is divided into three layers: front-end perception layer, edge computing layer, and cloud interaction layer.

[0097] Front-end perception layer:

[0098] Electrical parameter acquisition: Three open-type Hall current sensors (model ACS758, range 0-400A) and one isolated voltage transmitter are installed at the lower busbar of the circuit breaker. Importantly, the current sensors have a wide bandwidth response (0-20kHz) to capture high-frequency harmonics caused by electric arcs.

[0099] Environmental parameter acquisition: A miniature NTC thermistor (accuracy ±0.5℃) is embedded next to each key terminal (A / B / C / N), and a photoelectric smoke sensor and an electrochemical CO sensor (range 0-1000ppm) are installed at the bottom of the cabinet.

[0100] Physical condition acquisition: A piezoelectric acoustic emission sensor (center frequency 40kHz, bandwidth 20-60kHz) is attached to the outer surface of the insulation sheath of the cable terminal head for non-contact monitoring of partial discharge; at the same time, a miniature FBG (fiber grating) strain gauge is embedded at the cable fixing clamp to monitor cable deformation caused by thermal expansion and contraction or mechanical stress.

[0101] Communication method: All sensors are connected to the RS485 interface of the edge gateway via shielded twisted-pair cables, following the Modbus RTU protocol.

[0102] Edge computing layer: An industrial gateway equipped with an ARM Cortex-A72 quad-core processor and a built-in TensorFlowLite inference engine is used to perform the sampling control and data preprocessing described in claims 2 and 3.

[0103] Cloud-based interaction layer: Alibaba Cloud IoT platform, responsible for storing historical data, training large models, and providing Web / App access interfaces.

[0104] 2. Adaptive sampling and spatiotemporal benchmark unification

[0105] To resolve the contradiction between the data flood caused by high-frequency sampling and the loss of details in low-frequency sampling, this embodiment implements the following adaptive sampling strategy:

[0106] Reference synchronization: The gateway broadcasts a time synchronization pulse to all sensors every 100ms to ensure that the sampling time error of all sensors is less than 1ms.

[0107] Steady-state monitoring (low frequency): In the default mode, the system collects all parameters at a frequency of flow=1Hz. At this time, the system power consumption is low and it is only used for routine data recording.

[0108] Transient capture (high frequency): The gateway monitors the current waveform in real time. Let I(t) be the instantaneous current, and calculate the current variance σ2 within the sliding window (window width 1s).

[0109] If σ2 > (0.1⋅Irated)2 (the first threshold, corresponding to severe load fluctuations or arcing), the gateway immediately pulls the GPIO pin high, triggering all analog sensors to enter "high-speed mode", and the sampling frequency instantly switches to fhigh = 12.8kHz.

[0110] This high-speed mode lasts for 5 seconds, or until σ2 falls below (0.02⋅Irated)2.

[0111] Technical benefits: When capturing intermittent arcs caused by poor contact, the 12.8kHz sampling rate is sufficient to cover the Nyquist frequency range of 0-6.4kHz, ensuring that high-frequency harmonic components are not lost.

[0112] 3. Data Preprocessing and Unscented Kalman Filtering

[0113] The site environment contains a high-power frequency converter, resulting in severe electromagnetic interference. To prevent misdiagnosis, the following steps should be performed:

[0114] Outlier removal: Set physical limits, such as if the temperature reading jumps by more than 50°C instantaneously, it is judged as an interference signal and the data from the previous cycle is used to replace it.

[0115] UKF filtering:

[0116] Establish the state vector Xk=[TA,TB,TC,Irms,Strain]T (representing the temperature, effective current value, and strain of phases A / B / C, respectively).

[0117] Because temperature conduction exhibits nonlinear hysteresis, linear Kalman filtering is ineffective. This embodiment employs unscented Kalman filtering (UKF).

[0118] Sigma point generation: At each sampling time, 2n+1 Sigma points are generated based on the current state estimate Xk−1 (n is the state dimension).

[0119] State prediction: These points are propagated through nonlinear state transition equations (including a simplified model of the heat conduction differential equation).

[0120] Measurement update: The sensor's measured values ​​are compared with the predicted values, the residuals are calculated, and the state estimate is corrected using the Kalman gain calculated by the unscented transform.

[0121] Result: After UKF processing, the temperature curve fluctuation (±3℃) caused by the electromagnetic interference of the frequency converter was smoothed into a stable gradual curve (±0.2℃), which truly reflects the heat accumulation process.

[0122] 4. Attention-LSTM Deep Feature Fusion

[0123] This step is run on the AI ​​acceleration unit of the edge gateway.

[0124] Data Construction: Data from the past 60 seconds is used to construct a sequence matrix with 60 time steps. Each time step has a feature dimension of 10 (3 channels of temperature + 3 channels of current + 3 channels of voltage harmonics + 1 channel of strain).

[0125] Model structure:

[0126] Input layer: Receives the matrix described above.

[0127] LSTM layer: A double-layer stacked LSTM is used, with 128 neurons in each layer. The first LSTM layer outputs the hidden state ht for all time steps, which is used to capture long short-term memory.

[0128] Multi-head attention layer:

[0129] Set 4 attention points.

[0130] Each head independently computes the Query, Key, and Value matrix. For example, one head focuses on "current mutation" while another head focuses on "temperature-strain correlation".

[0131] The weights are calculated using the scaled dot product attention formula: αt,i=softmax(dkQKT).

[0132] The LSTM output ht is weighted and summed to obtain four context vectors C1, C2, C3, and C4.

[0133] Concatenation and Fully Connected Layer: The four context vectors are concatenated into a 512-dimensional vector, which is then passed through a Dropout layer to prevent overfitting. Finally, it is compressed into a 128-dimensional deep fusion feature vector Vf through a fully connected layer.

[0134] 5. Bayesian Network Inference and Dynamic Threshold Determination

[0135] Bayesian network construction:

[0136] Node definition: Observation node E={Iharm,Tterm,PD,Strain} (taken from key activation values ​​in Vf); assume node H={Short,Loose,Aging}.

[0137] Network topology: Iharm points to Short, Tterm and Strain point to Loose, and PD points to Aging.

[0138] Parameter learning: The network is trained using a historical fault database (containing 1000 real fault cases) to determine the conditional probability table (CPT). For example, when PD=High, P(Aging|PD=High)=0.9.

[0139] Dynamic threshold calculation:

[0140] Assume the current ambient temperature Tenv = 35℃ (summer afternoon) and the load rate η = 80%.

[0141] The base threshold is 0.7.

[0142] Substitute into the formula:

[0143] Thresholdfinal=0.7×(1+0.01×(35−25))×(1−0.1×(0.8−0.5))

[0144] Thresholdfinal=0.7×1.1×0.97≈0.747

[0145] Interpretation: Due to the increase in ambient temperature, the system has slightly relaxed the alarm threshold (from 0.7 to 0.747) to prevent false alarms caused by high temperatures; at the same time, due to the high load rate, it has been tightened.

[0146] 6. Tiered early warning and digital twin mapping

[0147] Reasoning result: Assume that the system calculates P(Loose|Vf)=0.88 and the temperature slope dT / dt=2℃ / min.

[0148] Decision execution:

[0149] Since 0.88 > Thresholdfinal(0.747) and the slope is greater than 0, the system determines it as a "Level 2 warning (alarm)".

[0150] Action 1 (Local): The red LED on the gateway control cabinet flashes and the buzzer emits intermittent beeps.

[0151] Action 2 (Remote): Send a JSON data packet to the cloud platform via the 4G module: {"alert_level": 2, "type": "LooseContact", "location": "Phase_A_Terminal"}.

[0152] Action 3 (Interlocking): If the probability exceeds 0.95, the gateway directly outputs a dry contact signal to cut off the circuit breaker coil power supply.

[0153] Digital twin visualization:

[0154] After receiving the data, the cloud platform drives the Unity 3D engine to build a digital twin model.

[0155] On the screen, the material of the A-phase terminal in the 3D model of the distribution cabinet changed from metallic to dark red, and a bubble popped up displaying: "Prediction: The temperature will reach 90℃ in 30 minutes."

[0156] Meanwhile, the right panel displays the insulation aging index curve of the circuit over the past 24 hours, visually demonstrating the performance degradation process.

[0157] 7. Self-learning feedback mechanism

[0158] Closed-loop verification: The system records this alarm.

[0159] Manual feedback: 15 minutes later, on-site maintenance personnel arrived and found that the A-phase terminal was indeed loose, and then tightened it. The system did not trip.

[0160] Model fine-tuning: The operations and maintenance personnel click "Confirm Potential Hazard" on the App. The system adds this sample (feature vector Vf + label "Loose") to the training set.

[0161] Online evolution: During the off-peak hours that evening, the edge gateway used an incremental learning algorithm to fine-tune the last fully connected layer of the Attention-LSTM model, so that when encountering similar features again, the model can converge to the correct probability output more quickly.

[0162] 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 multi-sensor fusion-based electrical fire hazard dynamic monitoring and early warning method, characterized in that: Includes the following steps: Step S1: Construct a multi-source heterogeneous sensor network, the sensor network comprising: The electrical parameter sensing unit is configured to collect three-phase current, three-phase voltage, and voltage harmonic distortion rate. The environmental parameter sensing unit is configured to collect the temperature at the cable connector, the smoke concentration in the space, and the carbon monoxide gas concentration. The physical state sensing unit is configured to collect partial discharge pulse signals on the surface of the cable insulation layer and the mechanical deformation stress of the cable. Each sensing unit is connected to the edge computing gateway via a time-division multiplexed bus. Step S2: Spatiotemporal reference unification and adaptive sampling, clock synchronization of the multi-source heterogeneous sensor network, and dynamic adjustment of the sampling frequency according to the fluctuation characteristics of the current signal; Step S3: Data preprocessing and anomaly correction, performing outlier removal, normalization, and noise filtering based on unscented Kalman filtering on the acquired raw data stream; Step S4: Deep feature-level fusion, constructing a long short-term memory network model based on the attention mechanism, inputting the preprocessed multi-dimensional time series data into the model, and extracting a deep fusion feature vector containing fault evolution features; Step S5: Bayesian probabilistic inference. The fused feature vector output from step S4 is used as evidence input to the pre-trained Bayesian network inference engine to calculate the posterior probability distribution of the four types of hidden dangers in the current state: short circuit, overload, poor contact and insulation aging. Step S6: Dynamic hierarchical early warning decision-making. Based on the posterior probability distribution and its slope over time, query the dynamic threshold matrix and output early warning signals of different levels or execute linkage protection actions.

2. The method according to claim 1, characterized in that: The adaptive sampling in step S2 specifically includes: Calculate the variance σ2 of the effective current value (Irms) within the set time window in real time; When σ2>σhigh2 (first preset threshold), the circuit determines that a transient process has occurred and increases the sampling frequency to fhigh=10kHz to capture the high-frequency harmonics and transient arc characteristics of the current. When σ2 < σlow2 (the second preset threshold), the determination circuit is in steady-state operation, and the sampling frequency is reduced to flow = 1Hz to reduce system power consumption; Among them, σhigh2>σlow2, and fhigh>flow.

3. The method for dynamic monitoring and early warning of electrical fire hazards based on multi-sensor fusion according to claim 1, characterized in that: The unscented Kalman filtering process in step S3 specifically involves: Establish a nonlinear state-space model that includes temperature, current, and deformation stress; By selecting the Sigma point set through unscented transformation (UT), and propagating it through the nonlinear state equation, the predicted value is weighted and corrected using the measured value, thereby eliminating temperature measurement jitter caused by air convection disturbance and current glitches caused by electromagnetic interference.

4. The method for dynamic monitoring and early warning of electrical fire hazards based on multi-sensor fusion according to claim 1, characterized in that: The Attention-LSTM model in step S4 includes an input layer, a two-layer LSTM hidden layer, a multi-head attention weight layer, and a fully connected output layer, which are connected in sequence. The multi-head attention weighting layer is configured to compute multiple different attention heads in parallel, and each attention head calculates the weight coefficient αt of the input data at different times through a scaled dot product attention mechanism. By weighted summing of the outputs ht of the LSTM hidden layer at each time step, C=∑t=1Tαtht, a context vector C is generated, which is the deep fusion feature vector.

5. The method for dynamic monitoring and early warning of electrical fire hazards based on multi-sensor fusion according to claim 1, characterized in that: The Bayesian network inference engine construction process in step S5 includes: Define network nodes, including observation nodes (E={I,V,T,PD,CO}) and hypothesis nodes (H={Short,Overload,LooseContact,Aging}); The network parameters are learned using a historical failure case database, and the conditional probability table (CPT) between nodes is populated. During the online monitoring phase, a particle filter algorithm is used to track parameter drift caused by equipment aging and update the prior probability distribution in the CPT table in real time.

6. The method for dynamic monitoring and early warning of electrical fire hazards based on multi-sensor fusion according to claim 1, characterized in that: The dynamic threshold matrix in step S6 is adaptively corrected based on ambient temperature and load rate. The specific calculation formula is as follows: Thresholdfinal=Base_Threshold×(1+k1⋅(Tenv−Tref))×(1−k2⋅(η−ηref)) Where Base_Threshold is the base threshold, Tenv is the real-time ambient temperature, Tref is the reference temperature (25℃), η is the real-time load rate, ηref is the rated load rate, and k1 and k2 are empirical correction coefficients.

7. The method for dynamic monitoring and early warning of electrical fire hazards based on multi-sensor fusion according to claim 1, characterized in that: The graded early warning signals in step S6 include: Level 1 Warning (Alert): When the probability of a potential hazard P(Hi) > 0.6 and the slope of P(Hi) > 0, the system logs the information and displays a yellow alert on the human-machine interface (HMI). Level 2 warning (alarm): When P(Hi)>0.8, the system sends an SMS / APP push alarm to the user terminal through the IoT platform; Level 3 Early Warning (Emergency): When P(Hi)>0.95 or the temperature change rate dT / dt>10℃ / min, the system immediately issues a command to the circuit breaker to perform trip protection and lock the circuit.

8. The method for dynamic monitoring and early warning of electrical fire hazards based on multi-sensor fusion according to claim 1, characterized in that: It also includes step S7: digital twin visualization mapping; The fused feature vector generated in step S4 is mapped to the pre-constructed three-dimensional digital twin model of the power distribution cabinet; In the three-dimensional model, the corresponding electrical components are rendered with different colors according to the different types of hidden dangers Hi, and the predicted temperature field distribution cloud map and insulation aging trend curve for the next 30 minutes are superimposed and displayed.

9. The method for dynamic monitoring and early warning of electrical fire hazards based on multi-sensor fusion according to claim 1, characterized in that: The partial discharge sensor in the physical state sensing unit is a non-contact ultrasonic probe, installed on the outside of the cable joint shield. Its detection frequency is set to 40kHz±10kHz to filter out power frequency electromagnetic interference and is specifically used to capture partial discharge signals inside the insulation.

10. The method for dynamic monitoring and early warning of electrical fire hazards based on multi-sensor fusion according to claim 1, characterized in that: The method also includes a self-learning feedback mechanism: If the system issues an alert and the maintenance personnel do not intervene within the specified time and no actual fire occurs, the system determines it to be a false alarm and fine-tunes the weight parameters of the Attention-LSTM model and the CPT table of the Bayesian network through the backpropagation algorithm to achieve online evolution of the model.