An intelligent early warning device for electrical equipment failure based on multi-modal perception

Abstract

The application relates to the technical field of online monitoring and fault early warning of electrical equipment, in particular to an intelligent electrical equipment fault early warning device based on multi-modal sensing, which comprises a sensor unit, including an infrared thermal imaging module, an ultrasonic partial discharge detection module, a vibration acceleration detection module and a high-frequency current detection module; and a working condition sensing unit, a degradation feature extraction unit, a physical degradation model unit and the like, the physical degradation equation is used to predict results, residual life prediction is provided before traditional threshold alarm, and maintenance decision is changed from passive emergency to active planning. The predicted failure time and the dominant degradation mode can be obtained by operation and maintenance personnel before the equipment is in the normal temperature range for months or even one year, cold start without fault samples is realized through particle filtering, unknown parameters in the degradation model are gradually converged through online observation, and the pain point that new substations lack historical fault data is solved.

Description

Technical Field

[0001] This invention relates to the field of online monitoring and fault early warning technology for electrical equipment, and in particular to an intelligent fault early warning device for electrical equipment based on multimodal perception. Background Technology

[0002] Existing multi-sensor online monitoring devices for electrical equipment output the following mode: multiple sensors collect physical quantities, extract statistical features, perform feature-level or decision-level fusion, and a classifier outputs discrete state labels, such as normal, attention, abnormal, and severe. This technical approach has the following shortcomings:

[0003] 1. It only diagnoses the current state and lacks the ability to predict remaining lifespan. After receiving the anomaly label, maintenance personnel cannot determine when the fault will develop to a critical state that requires a power outage, resulting in a lack of time-based basis for maintenance decisions.

[0004] Second, data-driven classifiers rely on a large number of labeled fault samples for training, while actual fault samples of power equipment are scarce, especially in the early stages of operation of newly built substations.

[0005] Third, the comprehensive score output by the multi-sensor fusion algorithm lacks physical dimensions, has insufficient interpretability, and makes it difficult for operation and maintenance personnel to trace the physical root cause of the anomaly criteria.

[0006] Fourth, traditional threshold alarm methods only trigger when the parameter exceeds the preset threshold, wasting the long degradation warning window below the threshold.

[0007] Therefore, this invention proposes an intelligent early warning device for electrical equipment faults based on multimodal perception. Summary of the Invention

[0008] The purpose of this invention is to provide an intelligent early warning device for electrical equipment faults based on multimodal perception, so as to solve the problems mentioned in the background art. The specific technical solution is as follows:

[0009] The first objective of this invention is to provide an intelligent early warning device for electrical equipment faults based on multimodal perception, comprising:

[0010] The sensor unit includes an infrared thermal imaging module, an ultrasonic partial discharge detection module, a vibration acceleration detection module, and a high-frequency current detection module;

[0011] The operating condition sensing unit is used to collect real-time load rate, ambient temperature and humidity, and cooling fan speed. It outputs the current operating condition type label through an LSTM network and uses this label to query the normal temperature baseline for each monitoring point under that operating condition. ;

[0012] The degradation feature extraction unit is used to extract the measured temperature data acquired by the infrared thermal imaging module. Subtract the normal temperature reference for the corresponding operating conditions The abnormal temperature rise ΔT is obtained; and the equivalent discharge quantity Q is extracted from the ultrasonic partial discharge detection module and the high-frequency current detection module, and the characteristic frequency amplitude A is extracted from the vibration acceleration detection module.

[0013] The physical degradation model unit deploys a joint overheating degradation model, an insulation partial discharge degradation model, and a mechanical vibration degradation model. The joint overheating degradation model uses the abnormal temperature rise ΔT and real-time load current as inputs, and calculates the joint degradation trajectory and the predicted time to reach the preset contact resistance critical value through numerical integration of the Arrhenius oxidation kinetic equation. The insulation partial discharge degradation model uses the equivalent discharge quantity Q as input, and calculates the insulation degradation trajectory and the predicted time to reach the insulation failure polymerization threshold through numerical integration of the Ekenstam polymerization degree reduction equation. The mechanical vibration degradation model uses the characteristic frequency amplitude A as input, and calculates the mechanical degradation trajectory and the predicted time to reach the mechanical damage critical value through numerical integration of the damage accumulation equation.

[0014] The multi-degradation mode competition unit is used to take the minimum value of the predicted time calculated by the joint overheating degradation model, the insulation partial discharge degradation model and the mechanical vibration degradation model as the system-level remaining lifetime, and output the degradation mode corresponding to the degradation model that makes the minimum value true as the dominant failure mode.

[0015] Preferably, the operating condition sensing unit further includes a baseline slow adaptive update mechanism: extracting the long-term trend of abnormal temperature rise ΔT at each monitoring point within a preset long-term historical time window; when the long-term trends of multiple monitoring points show synchronous linear or near-linear growth, and the growth slope is lower than a preset electrical degradation characteristic slope threshold, and the duration exceeds a preset aging judgment period, it is determined to be a baseline drift caused by physical aging of the cooling system or deterioration of heat dissipation conditions, rather than degradation of the electrical equipment itself; the operating condition sensing unit uses the long-term trend to adjust the normal temperature baseline of each monitoring point accordingly. Adaptive compensation updates are performed to ensure that the corrected ΔT reflects only the actual degradation signal of the electrical equipment itself.

[0016] The electrical degradation characteristic slope threshold is set based on the theoretical degradation rate under normal contact resistance according to the Arrhenius oxidation kinetic equation; the synchronization determination requires that no less than a preset proportion of monitoring sites within the same device show the same temperature rise trend.

[0017] Preferably, it also includes a particle filter calibration unit: in the initialization phase, multiple particles are generated, each particle carrying a set of parameter guesses for the degradation equation in the physical degradation model unit; in the prediction step, each particle independently calculates its degradation trajectory according to its own parameters; in the update step, the latest acquired value of the sensor unit is used as the observation, the likelihood between the predicted value and the observed value of each particle is calculated and the particle weight is updated; in the resampling step, high-weight particles proliferate and low-weight particles are eliminated; the moment when each particle trajectory first crosses the failure threshold constitutes the probability distribution of the remaining lifetime, and the particle filter calibration unit outputs the median and confidence interval of the remaining lifetime.

[0018] Preferably, it further includes an output unit: based on the lower limit of the confidence interval. The risk level is determined by comparing the relationship with the preset planned maintenance interval. If the risk is greater than the planned maintenance interval, it is considered low risk; if If the system's remaining lifespan is less than the planned maintenance interval and less than the system's remaining lifespan, it is classified as medium risk; if the system's remaining lifespan is less than a preset emergency threshold, it is classified as high risk.

[0019] Preferably, the infrared thermal imaging module, ultrasonic partial discharge detection module, vibration acceleration detection module, and high-frequency current detection module each have independent preset conventional alarm thresholds for their measurement parameters; when any measurement parameter exceeds the corresponding preset conventional alarm threshold, a conventional alarm is triggered independently; when any measurement parameter exceeds the preset emergency threshold, the highest level alarm is directly output without going through the physical degradation model unit and particle filter calibration unit; the risk level determination is performed in the operating state where the preset conventional alarm threshold has not been triggered.

[0020] Preferably, it also includes a sensor self-diagnostic mechanism: when the ultrasonic partial discharge detection module detects an increase in the amplitude of the partial discharge pulse signal within a preset time window, while the high-frequency current detection module does not detect the corresponding pulse current event within the same time window, and the historical correlation coefficient of the pulse amplitude of the two deviates from the preset normal range, it is determined that the ultrasonic partial discharge detection module or the high-frequency current detection module has drift or decreased sensitivity, the observation weight of the untrusted sensor in the corresponding degradation model is reduced, and a sensor abnormality alarm is output.

[0021] Preferably, the ultrasonic partial discharge detection module is further configured to transmit and receive ultrasonic pulses, and calculate the internal air temperature of the device by measuring the flight time of the ultrasonic pulses within a known geometric distance. During a preset low-load period, when the infrared thermal imaging module measures the surface temperature of the monitoring site... and When the difference is less than a preset threshold, the infrared emissivity setting value corresponding to that monitoring point is automatically corrected, so that... Approaching .

[0022] The second objective of this invention is to provide a method for predicting the remaining life of electrical equipment based on a multiphysics degradation model, comprising the following steps:

[0023] S1. Using real-time load rate, ambient temperature and humidity, and cooling fan speed as inputs, identify the current operating condition type through the LSTM network, and query the normal temperature baseline for each monitoring point under this operating condition. ;

[0024] S2. Collect the surface temperature of each monitoring point inside the equipment using an infrared thermal imaging module. Calculate abnormal temperature rise Partial discharge signals are acquired using an ultrasonic partial discharge detection module and a high-frequency current detection module, and the equivalent discharge quantity Q is extracted. Vibration signals are acquired using a vibration acceleration detection module, and the characteristic frequency amplitude A is extracted.

[0025] S3. Input the abnormal temperature rise ΔT and real-time load current into the joint overheating degradation model. The degradation equation of the joint overheating degradation model is:

[0026] ;

[0027] In the formula: This is the contact resistance, measured in Ω, with an initial value denoted as R0. The reference temperature is the normal operating temperature, in °C; ΔT is the abnormal temperature rise, in °C; Ea is the copper oxidation activation energy, in eV, with a typical range of 0.8~1.2 eV. Where kB is the Boltzmann constant, kB = 8.617 × 10⁻ 5 eV / K; β is the oxide layer growth rate constant, in Ω / s, determined by accelerated aging experiments. For real-time load current, The equivalent thermal resistance of the joint;

[0028] Solving the contact resistance using numerical integration The degradation trajectory and reaching the preset critical value Predicted time ;

[0029] S4. Input the equivalent discharge quantity Q into the insulation partial discharge degradation model. The degradation equation of the insulation partial discharge degradation model is:

[0030] ;

[0031] In the formula: The degree of polymerization of insulating paper, The basic rate constant for insulation degradation, The driving function of discharge energy on degradation rate;

[0032] The aggregation degree is solved by numerical integration. The degradation trajectory and reaching the preset failure threshold Predicted time ;

[0033] S5. Input the characteristic frequency amplitude A into the mechanical vibration degradation model, and solve the degradation trajectory of mechanical damage and the predicted time to reach the preset damage threshold by numerical integration of the damage accumulation equation. ;

[0034] S6, Take , and The minimum value is taken as the system-level remaining lifetime, and the degradation mode that makes this minimum value true is output as the dominant failure mode.

[0035] Preferably, in each degradation model numerical integration process in steps S3 to S5, an online particle filter calibration step is also included, and adaptive particle number adjustment is adopted: the initialization stage and the early degradation stage use the first particle number. When the effective sample size of the posterior distribution exceeds a preset convergence threshold, it is reduced to the number of second particles. When the ratio of degradation degree to failure threshold exceeds a preset critical ratio or the degradation rate exceeds a preset rate threshold, the number of third particles is increased. .

[0036] Preferably, it also includes an emissivity self-calibration step: actively emitting ultrasonic pulses and receiving reflected echoes through the ultrasonic partial discharge detection module, and calculating the average air temperature along the sound path using the time of flight. During periods of low equipment load, when the infrared thermal imaging module measures the surface temperature of the monitoring site... and When the difference is less than a preset threshold, it is determined that the object being tested is in thermal equilibrium with the air, and the infrared emissivity setting value corresponding to the monitoring point is corrected accordingly.

[0037] It has the following beneficial effects:

[0038] 1. Providing remaining life prediction before traditional threshold alarms shifts maintenance decisions from reactive emergency response to proactive planning. Maintenance personnel can know the expected failure time and dominant degradation mode months or even a year in advance, while the equipment is still within its normal temperature range.

[0039] Second, by utilizing physical degradation equations rather than purely data-driven models, each parameter has a physical meaning, and the prediction results are interpretable and traceable. The Arrhenius equation, Ekenstam model, and Archard model are all recognized physical laws in the fields of materials science and electrical engineering.

[0040] Third, particle filtering enables cold start without fault samples. Unknown parameters in the degradation model gradually converge through online observation, solving the problem of lack of historical fault data in newly built substations.

[0041] Fourth, the operating condition awareness layer eliminates the interference of load fluctuations and environmental changes on the degradation model, ensuring that degradation predictions only respond to real equipment deterioration signals. The LSTM operating condition recognition network is trained during the equipment health phase, requiring no fault data.

[0042] Fifth, the multi-degradation mode competing architecture avoids information dilution and uninterpretability caused by sensor fusion. Each degradation mode maintains independent physical meaning, with the shortest lifespan corresponding to the most urgent failure, making the logic intuitive and traceable.

[0043] VI. Sensor self-diagnosis mechanism: Utilizes the physical redundancy between multiple sensors to automatically detect sensor drift or faults, thereby improving the reliability of the system during long-term operation. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0045] Figure 1 This is a block diagram of an intelligent early warning device for electrical equipment faults based on multimodal perception, as described in this invention.

[0046] Figure 2 This is a flowchart of the method described in Embodiment 1. Detailed Implementation

[0047] The following detailed description, in conjunction with the accompanying drawings and specific embodiments, provides a further detailed explanation of the intelligent early warning device for electrical equipment faults based on multimodal perception proposed in this invention. The advantages and features of this invention will become clearer from the following description. It should be noted that the accompanying drawings are all in a very simplified form and use non-precise proportions, and are only used to facilitate and clarify the illustration of the embodiments of this invention.

[0048] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0049] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the illustrations only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the state, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0050] Please refer to the following: An intelligent early warning device for electrical equipment faults based on multimodal perception. Figure 1 ,include:

[0051] Sensor units: infrared thermal imaging module, ultrasonic partial discharge detection module, vibration acceleration detection module, high-frequency current detection module;

[0052] Operating condition sensing unit: used to collect real-time load rate, ambient temperature and humidity, and cooling fan speed, and output the current operating condition type label through LSTM network;

[0053] Degradation feature extraction unit: Extracts abnormal temperature rise ΔT(t) and temperature rise rate, equivalent discharge quantity Q(t) and cumulative energy, vibration characteristic frequency amplitude A(t) and growth rate, and current spectrum degradation features from the time series data of each sensor;

[0054] Physical degradation model unit: Deploy corresponding physical degradation equations for different fault types, and transform degradation characteristics into degradation rate and degradation trajectory;

[0055] Multi-degradation mode competition unit: Takes the minimum value of the remaining lifetime calculated by each degradation mode as the system-level remaining lifetime, and outputs the dominant failure mode at the same time;

[0056] Particle filter calibration unit: It uses particle filtering to fuse the predicted values ​​of the physical degradation model with the real-time observations of the sensor, corrects the degradation model parameters online, and outputs the probability distribution and confidence interval of the remaining lifetime;

[0057] Output unit: Based on the comparison between the confidence lower limit of the remaining life distribution and the planned maintenance interval, output the risk level and maintenance recommendations.

[0058] The operating condition sensing unit, degradation feature extraction unit, physical degradation model unit, multi-degradation mode competition unit, and particle filter calibration unit of the device can be implemented by one or more processors executing computer programs.

[0059] Operating condition perception layer:

[0060] The LSTM network takes load rate, ambient temperature, ambient humidity, and cooling fan speed within a sliding time window as input and outputs predefined operating condition type labels. Each operating condition type is associated with a corresponding historical temperature baseline. (The average normal temperature distribution at each monitoring point when the equipment is in good condition under this operating condition).

[0061] Measured temperature collected by infrared thermal imaging module Calculate abnormal temperature rise:

[0062] ;

[0063] Temperature fluctuations caused by non-degradation factors—normal temperature rise of the joint due to increased load and ambient temperature rise—are eliminated by operating condition normalization, and only abnormal temperature rise ΔT(t) is used as a degradation signal input to the subsequent model.

[0064] Training of LSTM networks can be completed during the initial healthy operation phase of the equipment, requiring only normal operating data and no fault samples.

[0065] The physical degradation model unit includes the following three degradation equations:

[0066] Joint overheating degradation model:

[0067] This study addresses the overheating degradation of electrical contact points such as cable joints, busbar lap joints, and circuit breaker contacts due to increased contact resistance. The degradation equation is as follows:

[0068] ;

[0069] ;

[0070] in, For contact resistance, For the contact interface temperature, The activation energy for copper oxidation. Boltzmann's constant, Let be a constant that includes the oxide layer growth rate. The positive feedback mechanism described by this equation is: increased contact resistance → increased Joule heating → increased interface temperature → exponential acceleration of oxidation rate → further increase in contact resistance.

[0071] Real-time value of contact resistance By combining Joule's laws:

[0072] ;

[0073] From the measured temperature Inversion, in which Real-time load current detected by the low-frequency channel of the high-frequency current detection module. The equivalent thermal resistance of the joint is... The ambient temperature.

[0074] when Reaching the preset threshold (Typically the initial contact resistance) When the positive feedback enters the uncontrollable thermal runaway stage (when the remaining lifetime is 10 times that of the previous value), it is considered a joint failure.

[0075] ;

[0076] in The remaining life calculated using the joint overheating model. This is the preset critical value for contact resistance. This is the current assessment moment;

[0077] Insulation partial discharge degradation model:

[0078] This paper addresses the insulation degradation caused by partial discharge in transformer oil-paper insulation and switchgear epoxy resin insulation. The degradation equation is the partial discharge corrected form of the Ekenstam model:

[0079] ;

[0080] Wherein, DP represents the degree of polymerization of the insulating paper, and the new insulating paper... Recognized failure threshold Q represents the equivalent discharge quantity, retrieved in real-time from the high-frequency channels of the ultrasonic partial discharge detection module and the high-frequency current detection module. f(Q) is the driving function of discharge energy on the degradation rate, calibrated using accelerated aging experimental data. The degradation rate is calculated as DP decreases to... The moment of insulation failure is given, and the remaining lifetime is calculated by integration.

[0081] Mechanical vibration degradation model:

[0082] For mechanical faults such as motor bearing wear, loose windings, and broken rotor bars, the vibration characteristic frequency amplitude A(t) and the degree of mechanical damage D are mapped using the factory vibration-damage calibration curve.

[0083] ;

[0084] Damage evolution can be modeled using an exponential growth model. or power-law model ,in C and m are fitted online using historical data. The failure time is defined as the moment when D reaches the vibration limit or bearing fatigue limit set by the corresponding national standard.

[0085] For electromagnetic-mechanical coupling faults such as broken rotor bars and air gap eccentricity, the degradation characteristics also originate from the stator current spectrum extracted by the high-frequency current detection module through MCSA (Motor Current Characteristic Analysis). Sideband amplitude growth rate.

[0086] Multi-degradation mode competing unit:

[0087] The same electrical device may exhibit multiple degradation modes simultaneously. This device does not employ multi-sensor feature fusion; instead, it calculates the remaining lifetime for each degradation mode independently.

[0088] ;

[0089] System-level remaining lifetime Take the minimum of the three:

[0090] ;

[0091] Simultaneously, the degradation mode that makes the above minimum value true is output as the dominant failure mode. Maintenance personnel can then identify the weakest link in the equipment and its estimated failure time, rather than relying on an untraceable overall score.

[0092] Furthermore, the operating condition sensing unit also includes a baseline slow adaptive update mechanism. In actual operation, the slow deterioration process of the switchgear, such as the wear of the cooling fan bearings and the accumulation of dust on the dust filter, will cause the overall base temperature of the entire cabinet to slowly rise over a period of several months to several years under the same operating conditions (same load rate, ambient temperature, and cooling wind speed). If not corrected, this part of the temperature rise will be incorrectly included in the abnormal temperature rise ΔT. After being amplified by the exponential acceleration effect of the Arrhenius oxidation kinetic equation, it will cause a significant premature deviation in the degradation prediction. The baseline slow adaptive update mechanism identifies the degradation by extracting the long-term change trend of ΔT at each monitoring point within a long-term historical time window of no less than 30 days: cooling system aging is manifested as synchronous, linear or near-linear temperature rises at multiple monitoring points with extremely low slopes (typically 0.001-0.005°C / day); while electrical connector degradation follows the Arrhenius positive feedback law, manifested as asynchronous, exponential temperature rises with continuously increasing slopes at a single point or a few points. When a baseline drift is identified, the normal temperature baseline for each monitoring point is determined using long-term trends. Perform adaptive compensation updates.

[0093] Particle filter calibration unit:

[0094] Key parameters in the physical degradation equation ( , , , (etc.) are difficult to predict precisely. This device uses online particle filter calibration:

[0095] During the initialization phase, N particles (N=500) are generated, each carrying a set of guessed values ​​for the degradation model parameters. Random perturbations are allowed in the degradation trajectories of each particle.

[0096] ;

[0097] This represents the current state value of the aggregation degree independently derived by the i-th particle in the particle filter algorithm. Specifically, this refers to the guessed value of this parameter carried by the i-th particle in the particle filter algorithm. This refers to random perturbation noise introduced into the particle calculus trajectory. With a mean of 0 and a variance of Gaussian normal distribution (defining the characteristics of disturbance noise);

[0098] In the prediction step, each particle independently extrapolates one step forward according to its own degradation parameters; in the update step, the latest measured values ​​of the sensors (such as the junction temperature measured by infrared or the partial discharge pulse measured by ultrasound) are used as observations to calculate the likelihood between the prediction and the observation of each particle, and the particle weights are updated accordingly; in the resampling step, high-weight particles multiply and low-weight particles are eliminated.

[0099] Therefore, the degradation model parameters gradually converge to their true values ​​through online observation without requiring precise prediction. The moment when each particle trajectory first crosses the failure threshold constitutes the probability distribution of the remaining lifetime, and the device outputs the median and 95% confidence interval. .

[0100] Furthermore, the particle filter calibration unit employs an adaptive particle number adjustment strategy. During the initial cold start phase of device commissioning, the prior distribution range of degradation model parameters is relatively wide, requiring a large particle number N1 (e.g., 500) to fully cover the parameter space. As online observation data accumulates, the effective sample size of the particle filter posterior distribution gradually increases. When it exceeds a preset convergence threshold, it indicates that the degradation model parameters have been initially locked, and the particle number can be reduced to a smaller N2 (e.g., 50-100) to significantly reduce the real-time computational load on the embedded processor. When the current value of a degradation characteristic (ΔT, Q, or A) approaches the corresponding failure threshold—for example, when an abnormal temperature rise ΔT reaches a level that makes the contact resistance approach the thermal runaway critical value, or when the degradation rate exhibits a nonlinear acceleration—the particle number is increased again to a larger N3 (e.g., 500-1000) to accurately capture nonlinear abrupt changes near the failure critical point. This strategy enables concurrent real-time operation of multiple degradation models on an embedded ARM processor.

[0101] Degraded sensor self-diagnostic mechanism:

[0102] The data fusion processing unit is also configured with a sensor self-diagnosis function: when the first physical quantity collected by the first sensor indicates an abnormality within a preset time window, and the second physical quantity collected by the second sensor that is causally related to the first physical quantity does not indicate an abnormality, the observation weight of the first sensor in the degradation model is reduced, and a sensor abnormality alarm is output.

[0103] For example, when the ultrasonic partial discharge detection module detects an increase in the partial discharge pulse signal, while the high-frequency current detection module does not detect a corresponding pulse current event within the same time window, and the correlation coefficient of the pulse amplitudes of the two deviates from the historical normal range, it is determined that one of the sensors has drifted or its sensitivity has decreased, and the observation data of the channel with higher reliability is automatically selected to drive the degradation model.

[0104] Output and Decision Units:

[0105] The output includes: median of predicted remaining lifespan. Confidence interval The dominant degradation mode, the current degradation rate, and the estimated time to approach the alarm threshold (a preset general alarm threshold, such as a connector temperature of 90°C).

[0106] The risk level determination rules are as follows:

[0107] like The planned maintenance interval was determined to be low-risk, and the scheduled maintenance was carried out accordingly.

[0108] like Planned maintenance intervals The risk level has been determined to be medium, and it is recommended to arrange maintenance in advance.

[0109] like If a preset emergency threshold (e.g., 30 days) is set, it is considered a high-risk situation, and it is recommended to shut down the power as soon as possible.

[0110] The risk level determination is performed under operating conditions where the regular alarm threshold is not triggered. The preset emergency threshold is independent of the aforementioned regular alarm threshold and is higher than the regular alarm threshold. It serves only as a fallback safety mechanism when the degradation model fails—when the sensor's measured value directly exceeds this emergency threshold, the highest level alarm is output directly without going through the degradation prediction process.

[0111] In some embodiments, the emissivity setpoint of the infrared thermal imaging module can be self-calibrated by the ultrasonic partial discharge detection module. The ultrasonic partial discharge detection module is configured to transmit and receive ultrasonic pulses, and calculates the average air temperature inside the cabinet by measuring the flight time of the ultrasonic pulses over a known distance. The data fusion processing unit collects data from each monitoring site during periods of low equipment load (such as early morning). and ,when When the measured object and air are in thermal equilibrium, the infrared emissivity setting value of the monitoring point is automatically corrected, so that the infrared temperature measurement value is close to the sound speed temperature measurement value, realizing online emissivity correction without manual calibration.

[0112] Example 1: Deployment of this device on a 10kV switchgear

[0113] This embodiment uses a 10kV metal-clad withdrawable switchgear as the monitoring object.

[0114] Hardware deployment:

[0115] Infrared thermal imaging modules (uncooled vanadium oxide focal plane detector, resolution 640×480, operating wavelength 8-14μm) are fixedly installed in the busbar compartment and cable compartment of the switchgear, respectively, covering all busbar lap surfaces, circuit breaker contacts, and cable termination joints within the cabinet. Infrared data is transmitted to the data processing unit via gigabit Ethernet.

[0116] The ultrasonic partial discharge detection module (center frequency 40kHz, bandwidth 20-100kHz) is magnetically mounted on the rear metal housing of the switch cabinet, using contact coupling to detect solid-conductive partial discharge acoustic emission signals. A built-in preamplifier outputs an analog signal that is sampled by a 16-bit ADC before being transmitted to the data processing unit.

[0117] The vibration acceleration detection module (IEPE type, sensitivity 100mV / g, frequency response 0.5-10kHz) is installed on the switch cabinet housing near the busbar support insulator via a magnetic base.

[0118] The high-frequency current detection module (split ferrite magnetic ring, bandwidth 100kHz-50MHz) is clamped to the grounding wire of the switch cabinet to detect partial discharge pulse current and load current.

[0119] Operating condition sensing sensors: current transformers collect bus load current, temperature / humidity sensors are installed in the external environment of the cabinet, and wind speed sensors are installed at the air outlet of the forced air cooling channel.

[0120] The data processing unit is deployed in the switchgear instrument room and uses an ARM Cortex-A series processor to run an embedded Linux system, executing all algorithms for condition sensing, degradation feature extraction, physical degradation model integration, multi-degradation mode competition, particle filter calibration, and output decision.

[0121] Please see Figure 2 Method and process:

[0122] S1: The operating condition sensing layer collects load rate, ambient temperature and humidity, and cooling fan speed every second, inputting them into the LSTM network in a 30-minute sliding window to identify the current operating condition type and query the normal temperature baseline for each monitoring point inside the cabinet under that operating condition. The baseline slow adaptive update mechanism runs continuously in the background. Taking a switchgear that has been in operation for four years as an example: the dust filter of this cabinet is partially blocked due to long-term lack of replacement, and the wear of the cooling fan bearings has led to a decrease in actual airflow of approximately 15%. Under the same operating conditions (high load in summer, forced air cooling), the infrared thermal imaging module detected that the temperature at nine monitoring points inside the cabinet rose synchronously by approximately 2.3°C between the third and fourth years, with an average daily temperature rise of approximately 0.006°C / day, showing a near-linear increase. The system compares this trend with the theoretical degradation rate of the joint overheating degradation model (based on the Arrhenius equation, the expected temperature rise rate starts at approximately 0.02°C / day under normal contact resistance R0, and continues to accelerate with increasing temperature): the actual temperature rise rate is far lower than the degradation characteristic slope threshold, and the nine points experience synchronous temperature rise (rather than independent temperature rise at a single point), consistent with the spatial consistency characteristics of cooling system aging. Based on this, the system determines it as baseline drift and automatically updates the temperature at each monitoring point under this operating condition. An increase of 2.3°C does not affect the remaining life prediction of the joint overheating degradation model, as the corrected ΔT remains unchanged. Conversely, if an abnormal temperature rise occurs at the A-phase busbar lap joint due to increased contact resistance during the same period, and this single point rises from ΔT=0 to ΔT=6°C within 3 months, with an average daily temperature rise of 0.067°C / day and still accelerating, the system identifies that its slope has entered the degradation characteristic range and is a single-point independent temperature rise, thus determining it as a real electrical degradation signal, and normally drives the degradation prediction model.

[0123] S2. The infrared thermal imaging module acquires one frame of the entire cabinet temperature image per second and extracts the surface temperature of each monitoring point (9 ROI areas: A / B / C three-phase busbar lap surfaces, circuit breaker upper and lower contacts, and cable terminal joints). Calculate abnormal temperature rise .

[0124] S3, the degradation feature extraction unit continuously records the time series of ΔT and calculates the temperature rise rate d(ΔT) / dt; simultaneously extracts the partial discharge pulse amplitude, repetition rate and equivalent discharge quantity Q from the ultrasonic and high-frequency current modules; and extracts the amplitude series of characteristic frequencies of 100Hz and 200Hz from the vibration module.

[0125] S4. When the cumulative change of ΔT, Q, or A exceeds the preset start threshold, the physical degradation model unit starts the numerical integration of the degradation equation to calculate the degradation trajectory and the estimated time to reach the corresponding failure threshold.

[0126] S5, the particle filter calibration unit maintains 500 particles. Whenever new observation data is obtained, it performs a prediction-update-resample cycle and outputs the probability distribution of the remaining lifetime.

[0127] S6. The multi-degradation mode competing unit compares the remaining lifetime output by the joint overheating degradation model, insulation partial discharge degradation model, and mechanical vibration degradation model, takes the minimum value as the system-level prediction, and outputs the dominant failure mode.

[0128] S7. The output unit performs risk classification based on planned maintenance intervals and remaining service life distribution. For example, if the next Class C maintenance is scheduled for 180 days later, the joint overheating degradation model will output... sky, Heaven. Because The risk level is low. It is recommended to conduct inspections as planned, but please pay attention to the contact resistance of the A-phase busbar lap joint during maintenance.

[0129] S8. When the measured temperature at any monitoring point exceeds the preset normal alarm threshold (90°C), a normal temperature alarm is triggered independently. When it exceeds the preset emergency threshold (130°C), the highest level alarm is output directly without going through the degradation prediction process. The above alarm mechanism and the remaining lifetime prediction mechanism operate independently and in parallel.

[0130] Example 2: Cold Start and Convergence Process of Particle Filtering

[0131] This embodiment demonstrates the entire lifecycle operation of a particle filter under an adaptive particle number adjustment strategy.

[0132] Days 1-60 (Cold Start-up Phase, N=500): Initial Operational Stage, Key Parameter in the Joint Overheating Degradation Model: Copper Oxidation Activation Energy eV, oxidation rate constant The system covers two orders of magnitude with a logarithmically uniform distribution and an extremely wide prior distribution. The parameter space is covered by 500 particles, each with a degradation trajectory accompanied by Gaussian noise with a standard deviation of 2%. Infrared sensors collect junction temperatures daily. Because degradation has not yet established a clear trend, the predicted values ​​for each particle are highly dispersed, with remaining lifetimes ranging from 200 to 2000 days. The system outputs low risk with a wide confidence interval. During this phase, a prediction-update-resample cycle is performed daily, and the computational load of 500 particles can be controlled within milliseconds on an ARM Cortex-A processor.

[0133] Days 61-180 (convergence tracking phase, N reduced to 80): After approximately 60 observation updates, the effective sample size of the posterior distribution exceeded the preset convergence threshold. The parameters of the degenerate model have been initially locked into a relatively narrow range. Locked to 0.92-1.05 eV, (The range is reduced by approximately one order of magnitude from the initial value). The system triggers a particle number reduction operation, decreasing N from 500 to 80, reducing the computational load by approximately 84%. During this period, ΔT slowly increases from approximately 2°C to approximately 8°C, and 80 particles are sufficient to maintain continuous tracking of the degradation trajectory. The median remaining lifetime is gradually updated from approximately 800 days to approximately 500 days.

[0134] Days 181-350 (critical approach phase, N rises to 800): ΔT reaches approximately 8°C, corresponding to the junction contact resistance... It has grown to approximately four times its initial value, and is now close to the thermal runaway threshold. ( The ratio of the degradation rate to the pre-set critical ratio exceeds the threshold. The degradation rate enters the Arrhenius exponential acceleration range. The system triggers a particle number increase operation, raising N from 80 to 800 to capture nonlinear abrupt changes near the failure threshold at a higher resolution. By day 350, ΔT reaches approximately 18°C, with the measured temperature approaching the 90°C alarm threshold. The median remaining lifetime is predicted to be 35 days, with a lower confidence limit of 28 days. The system outputs a high-risk warning.

[0135] This example demonstrates the complete cycle of adaptive particle number adjustment: Compared to using a fixed 500 particles throughout the process, it saves approximately 84% of computational resources during the convergence tracking phase from day 61 to day 180; compared to using 50 particles throughout the process, it avoids convergence failure caused by insufficient parameter space coverage during the cold start phase, and avoids the decrease in prediction accuracy caused by particle sparsity during the critical approximation phase.

[0136] Example 3: Typical Scenario of Competition Among Multiple Degradation Modes

[0137] Taking a 110kV oil-immersed transformer that has been in operation for 8 years as an example:

[0138] The ultrasonic partial discharge detection module detected intermittent partial discharge inside the oil-paper insulation. The equivalent discharge quantity Q showed a slow upward trend (an average annual increase of about 15%). After integration by the Ekenstam degradation model, the remaining insulation lifetime was output as about 850 days.

[0139] The vibration acceleration detection module detected that the amplitude of the 100Hz characteristic frequency increased year by year, and it was estimated that the winding compression state loosened year by year. The mechanical degradation model output that the remaining mechanical life is about 1200 days.

[0140] The infrared thermal imaging module detected an abnormal temperature rise ΔT in the high-voltage bushing terminal, with an average annual increase of approximately 8°C. The overheating degradation model of the connector outputs a remaining contact degradation life of approximately 420 days.

[0141] The minimum value for the multi-degradation mode competing unit is 420 days, resulting in a system-level remaining lifetime of 420 days. The dominant failure mode is contact degradation of the high-voltage bushing terminals. Based on this, maintenance personnel should prioritize contact resistance testing and tightening of the high-voltage bushing terminals, rather than being caught in a trade-off between three potential problems.

[0142] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An intelligent early warning device for electrical equipment faults based on multimodal perception, characterized in that, include: The sensor unit includes an infrared thermal imaging module, an ultrasonic partial discharge detection module, a vibration acceleration detection module, and a high-frequency current detection module; The operating condition sensing unit is used to collect real-time load rate, ambient temperature and humidity, and cooling fan speed. It outputs the current operating condition type label through an LSTM network and uses this label to query the normal temperature baseline for each monitoring point under that operating condition. ; The degradation feature extraction unit is used to extract the measured temperature data acquired by the infrared thermal imaging module. Subtract the normal temperature reference for the corresponding operating conditions The abnormal temperature rise ΔT is obtained; and the equivalent discharge quantity Q is extracted from the ultrasonic partial discharge detection module and the high-frequency current detection module, and the characteristic frequency amplitude A is extracted from the vibration acceleration detection module. The physical degradation model unit deploys a joint overheating degradation model, an insulation partial discharge degradation model, and a mechanical vibration degradation model. The joint overheating degradation model uses the abnormal temperature rise ΔT and real-time load current as inputs, and calculates the joint degradation trajectory and the predicted time to reach the preset contact resistance critical value through numerical integration of the Arrhenius oxidation kinetic equation. The insulation partial discharge degradation model uses the equivalent discharge quantity Q as input, and calculates the insulation degradation trajectory and the predicted time to reach the insulation failure polymerization threshold through numerical integration of the Ekenstam polymerization degree reduction equation. The mechanical vibration degradation model uses the characteristic frequency amplitude A as input, and calculates the mechanical degradation trajectory and the predicted time to reach the mechanical damage critical value through numerical integration of the damage accumulation equation. The multi-degradation mode competition unit is used to take the minimum value of the predicted time calculated by the joint overheating degradation model, the insulation partial discharge degradation model and the mechanical vibration degradation model as the system-level remaining lifetime, and output the degradation mode corresponding to the degradation model that makes the minimum value true as the dominant failure mode.

2. The apparatus according to claim 1, characterized in that, The operating condition sensing unit also includes a baseline slow adaptive update mechanism: extracting the long-term trend of abnormal temperature rise ΔT at each monitoring point within a preset long-period historical time window. When the long-term change trends of multiple monitoring points show synchronous linear or near-linear growth, and the growth slope is lower than the preset electrical degradation characteristic slope threshold, and the duration exceeds the preset aging judgment period, it is determined that the reference drift is caused by physical aging of the cooling system or deterioration of heat dissipation conditions, rather than degradation of the electrical equipment itself; the operating condition sensing unit uses the long-term change trends to determine the normal temperature reference of each monitoring point accordingly. Adaptive compensation updates are performed to ensure that the corrected ΔT reflects only the actual degradation signal of the electrical equipment itself. The electrical degradation characteristic slope threshold is set based on the theoretical degradation rate under normal contact resistance according to the Arrhenius oxidation kinetic equation; the synchronization determination requires that no less than a preset proportion of monitoring sites within the same device show the same temperature rise trend.

3. The apparatus according to claim 1, characterized in that, It also includes a particle filter calibration unit: in the initialization phase, multiple particles are generated, each carrying a set of parameter guesses for the degradation equation in the physical degradation model unit; in the prediction step, each particle independently calculates its degradation trajectory according to its own parameters; in the update step, the latest acquisition value of the sensor unit is used as the observation, the likelihood between the predicted value and the observed value of each particle is calculated and the particle weight is updated; in the resampling step, high-weight particles proliferate and low-weight particles are eliminated; the moment when each particle trajectory first crosses the failure threshold constitutes the probability distribution of the remaining lifetime, and the particle filter calibration unit outputs the median and confidence interval of the remaining lifetime.

4. The apparatus according to claim 3, characterized in that, It also includes an output unit: based on the lower limit of the confidence interval. The risk level is determined by comparing the relationship with the preset planned maintenance interval. If the risk is greater than the planned maintenance interval, it is considered low risk; if If the interval is less than the planned maintenance interval and less than the remaining lifespan of the system, it is classified as medium risk. If the remaining lifespan of the system is less than a preset emergency threshold, it is determined to be high risk.

5. The apparatus according to claim 4, characterized in that, The infrared thermal imaging module, ultrasonic partial discharge detection module, vibration acceleration detection module, and high-frequency current detection module each have independent preset conventional alarm thresholds for their measurement parameters. When any measurement parameter exceeds the corresponding preset conventional alarm threshold, a conventional alarm is triggered independently. When any measurement parameter exceeds the preset emergency threshold, the highest level alarm is directly output without going through the physical degradation model unit and particle filter calibration unit. The risk level determination is performed under the operating state where the preset conventional alarm threshold has not been triggered.

6. The apparatus according to claim 1, characterized in that, It also includes a sensor self-diagnostic mechanism: when the ultrasonic partial discharge detection module detects an increase in the amplitude of the partial discharge pulse signal within a preset time window, while the high-frequency current detection module does not detect the corresponding pulse current event within the same time window, and the historical correlation coefficient of the pulse amplitude of the two deviates from the preset normal range, it is determined that the ultrasonic partial discharge detection module or the high-frequency current detection module has drift or decreased sensitivity, the observation weight of the untrusted sensor in the corresponding degradation model is reduced, and a sensor abnormality alarm is output.

7. The apparatus according to claim 1, characterized in that, The ultrasonic partial discharge detection module is also configured to transmit and receive ultrasonic pulses, and calculate the internal air temperature of the device by measuring the flight time of the ultrasonic pulses over a known geometric distance. ; During a preset low-load period, when the infrared thermal imaging module measures the surface temperature of the monitoring site... and When the difference is less than a preset threshold, the infrared emissivity setting value corresponding to that monitoring point is automatically corrected, so that... Approaching .

8. A method for intelligent early warning of electrical equipment faults based on multimodal perception, characterized in that, Includes the following steps: S1. Using real-time load rate, ambient temperature and humidity, and cooling fan speed as inputs, identify the current operating condition type through the LSTM network, and query the normal temperature baseline for each monitoring point under this operating condition. ; S2. Collect the surface temperature of each monitoring point inside the equipment using an infrared thermal imaging module. Calculate abnormal temperature rise Partial discharge signals are acquired using an ultrasonic partial discharge detection module and a high-frequency current detection module, and the equivalent discharge quantity Q is extracted. Vibration signals are acquired using a vibration acceleration detection module, and the characteristic frequency amplitude A is extracted. S3. Input the abnormal temperature rise ΔT and real-time load current into the joint overheating degradation model. The degradation equation of the joint overheating degradation model is: ； In the formula: This is the contact resistance, measured in Ω, with an initial value denoted as R0. The reference temperature is the normal operating temperature, in °C; ΔT is the abnormal temperature rise, in °C; Ea is the copper oxidation activation energy, in eV, with a typical range of 0.8~1.2 eV. Where kB is the Boltzmann constant, kB = 8.617 × 10⁻ 5 eV / K; β is the oxide growth rate constant, in Ω / s, determined by accelerated aging experiments. For real-time load current, The equivalent thermal resistance of the joint; Solving the contact resistance using numerical integration The degradation trajectory and reaching the preset critical value Predicted time ; S4. Input the equivalent discharge quantity Q into the insulation partial discharge degradation model. The degradation equation of the insulation partial discharge degradation model is: ； In the formula: The degree of polymerization of insulating paper, The basic rate constant for insulation degradation, The driving function of discharge energy on degradation rate; The aggregation degree is solved by numerical integration. The degradation trajectory and reaching the preset failure threshold Predicted time ; S5. Input the characteristic frequency amplitude A into the mechanical vibration degradation model, and solve for the degradation trajectory of mechanical damage and the predicted time to reach the preset damage threshold by numerical integration of the damage accumulation equation. ; S6, Take , and The minimum value is taken as the system-level remaining lifetime, and the degradation mode that makes this minimum value true is output as the dominant failure mode.

9. The method according to claim 8, characterized in that, In the numerical integration process of each degradation model in steps S3 to S5, an online particle filter calibration step is also included, and adaptive particle number adjustment is adopted: the initialization stage and the early stage of degradation use the first particle number. When the effective sample size of the posterior distribution exceeds a preset convergence threshold, it is reduced to the number of second particles. When the ratio of degradation degree to failure threshold exceeds a preset critical ratio or the degradation rate exceeds a preset rate threshold, the number of third particles is increased. .

10. The method according to claim 8, characterized in that, It also includes an emissivity self-calibration step: the ultrasonic partial discharge detection module actively emits ultrasonic pulses and receives reflected echoes, and uses the time of flight to calculate the average air temperature over the sound path. ; During periods of low equipment load, when the infrared thermal imaging module measures the surface temperature of the monitoring site... and When the difference is less than a preset threshold, it is determined that the object being tested is in thermal equilibrium with the air, and the infrared emissivity setting value corresponding to the monitoring point is corrected accordingly.

Images