Current sensor predictive maintenance method

Abstract

The invention relates to a predictive maintenance method for a current sensor. The method comprises the following steps: firstly, periodically collecting a current waveform, an environment temperature and equivalent stress of the current sensor; then, on one hand, performing prediction based on a multi-physical field coupling mechanism model to obtain a first state prediction value, and on the other hand, extracting multi-scale degradation characteristics and inputting the multi-scale degradation characteristics into a long-short-term memory network for prediction to obtain a second state prediction value; then, unscented Kalman filtering is adopted, weights are distributed adaptively according to prediction uncertainty of the two models, and the first state prediction value and the second state prediction value are fused into a more stable fusion state prediction value; and finally, determining the remaining service life and the average early warning amount based on the fusion health degree, and performing graded warning according to the remaining service life and the average early warning amount. According to the invention, through fusing the physical mechanism and the data driving model, accurate prediction and intelligent maintenance decision making of the state of the current sensor are realized.

Description

Technical Field

[0001] This application relates to the field of power equipment operation and maintenance technology, specifically to a predictive maintenance method for current sensors. Background Technology

[0002] With the rapid development of smart grids and new power systems, substations, as key nodes in the power grid, directly impact grid security due to the reliability of their main equipment (such as transformers). Current sensors are core components for monitoring the operating status of these devices, and their measurement accuracy and response delay directly affect the reliability of differential protection, overcurrent protection, condition assessment, and fault location.

[0003] Current methods for the operation and maintenance of substation equipment rely on periodic inspections and threshold alarms, which suffer from delayed response, high false alarm rates, and inability to predict latent faults. Existing digital twin technologies (such as multi-sensor fusion monitoring) only focus on the overall equipment and do not design specific strategies for the characteristics of current sensors (such as small current drift and electromagnetic interference sensitivity). Furthermore, purely data-driven or mechanistic models suffer from insufficient generalization and identifiability.

[0004] Therefore, the operation and maintenance methods for current sensors in substation equipment, especially for main substation equipment (such as transformers, circuit breakers, GIS, etc.), still need further innovation and improvement. Summary of the Invention

[0005] The purpose of this application is to propose a predictive maintenance method for current sensors and its system, electronic equipment and computer program products. Through the deep integration of physical mechanisms and data-driven approaches, it solves the problems of difficult early fault identification and inaccurate life prediction in traditional operation and maintenance.

[0006] To achieve the above objectives, according to the first aspect of this application, a predictive maintenance method for a current sensor is provided, comprising the following steps: Step S1: Periodically collect the current waveform of the current sensor, the ambient temperature, and the equivalent stress caused by external mechanical constraints; Step S2: Based on the current waveform, ambient temperature, equivalent stress caused by external mechanical constraints, and a preset multiphysics coupling mechanism model, the state parameters of the current sensor are predicted to obtain a first state prediction value; the first state prediction value includes the temperature of the first sensitive element of the current sensor, the equivalent stress of the first sensitive element, the first sensitivity, the first zero bias, and the first health status. Step S3: Based on the current waveform, ambient temperature, and equivalent stress caused by external mechanical constraints, extract a multi-scale degradation feature vector. Input the multi-scale degradation feature vector into a preset long short-term memory network to predict the current sensor state parameters and obtain a second state prediction value. The second state prediction value includes the second sensor temperature, the second sensor equivalent stress, the second sensitivity, the second zero bias, and the second health of the current sensor. Step S4: Using an unscented Kalman filter, weights are adaptively allocated based on the prediction uncertainty of the multiphysics coupling mechanism model and the long short-term memory network. The first state prediction value and the second state prediction value are then fused according to the adaptive weight allocation to obtain the fused state prediction value of the current sensor. The fused state prediction value includes the fused sensing element temperature, fused sensing element equivalent stress, fused sensitivity, fused zero bias, and fused health of the current sensor. Step S5: Determine the remaining service life based on the fused health status, and calculate the average early warning amount, whereby the average early warning amount is the average lead time from the first trigger threshold of the current sensor health status to actual failure. Step S6: Issue graded alarms based on the health status, remaining service life, and average early warning amount.

[0007] Preferably, in step S2, the multiphysics coupling mechanism model includes: The current-carrying heating and thermal balance model is used to calculate the Joule heating power based on the primary side current and the resistance of the sensing element, and to solve for the instantaneous temperature of the sensing element by combining the ambient temperature and heat dissipation conditions. A thermo-elastic coupled stress model is used to calculate thermal expansion stress based on temperature rise. The sensitivity and bias drift model is used to calculate the sensitivity and zero bias drift based on the instantaneous temperature and thermal expansion stress of the sensitive body. The material aging and health calculation model is used to calculate cumulative damage based on the temperature history using the Arrhenius accelerated model, and thus obtain the health status.

[0008] Preferably, the feature vector includes at least one of the following: effective current value, total harmonic distortion rate, DC bias, kurtosis, skewness, residual variance, sensor temperature, equivalent stress, sensitivity drift, zero bias drift, power factor, rate of temperature rise, relative change in resistance, equivalent service time, and noise spectrum energy.

[0009] Preferably, in step S3, the Long Short-Term Memory network is an LSTM optimized using the Sparrow Search Algorithm (ISSA-LSTM); the ISSA is used to search for the number of layers, hidden units, learning rate, and sequence length hyperparameters of the LSTM to meet hardware constraints on latency and memory usage; the training loss function of the ISSA-LSTM includes a physical consistency residual term, which embeds the constraints of the multiphysics coupling mechanism model as soft constraints to force the prediction results of the LSTM to conform to physical laws.

[0010] Preferably, in step S4, the prediction uncertainty is the first covariance matrix of the multiphysics coupling mechanism model and the second covariance matrix of the long short-term memory network, respectively; the adaptive weight allocation is specifically as follows: the weight of the first state prediction value is proportional to the inverse of the first covariance matrix, and the weight of the second state prediction value is proportional to the inverse of the second covariance matrix.

[0011] Preferably, in step S5, determining the remaining service life includes: taking the moment when the future prediction curve of the fused health level intersects with the preset health level threshold as the failure time point, and the difference between the current moment and the failure time point is the remaining service life.

[0012] Preferably, in step S6, the graded alarm includes at least three levels: A yellow alert is triggered when the health status is in the first interval or the average early warning amount is greater than the first threshold, and a response strategy of recommended retesting or calibration is implemented. Orange alert is triggered when the health status is in the second range or the remaining service life is less than the second threshold, and a response strategy of generating a maintenance plan and increasing the sampling frequency is executed. Red Emergency: This is triggered when the health status is in the third range or the remaining lifespan is less than the third threshold, and a strong reminder and a response strategy that restricts operating conditions are implemented.

[0013] According to a second aspect of this application, a predictive maintenance system for a current sensor is provided, including a module for performing the method described in the first aspect of this application.

[0014] According to a third aspect of this application, an electronic device is provided, comprising: A communication interface used for communicating with other electronic devices; Memory is used to store computer program instructions; A processor for executing the computer program instructions to support the electronic device in implementing the method according to the first aspect of this application.

[0015] According to a fourth aspect of this application, a computer program product is provided, including computer program instructions that instruct a computer device to perform an operation corresponding to the method described in the first aspect of this application.

[0016] Compared with the prior art, this application has the following beneficial effects: (1) Improved accuracy and reliability of prediction: This application combines a multi-physics coupling mechanism model with a long short-term memory network to achieve complementary advantages of physical mechanism and data-driven approach; the mechanism model provides physical interpretation and boundary constraints, while the data-driven model can learn complex degradation patterns from massive amounts of data. The two work together to significantly improve the accuracy and reliability of state prediction. (2) Enhanced robustness of state estimation: This application adopts unscented Kalman filtering and performs adaptive weighted fusion based on the prediction uncertainty of the two models; when the prediction uncertainty of a certain model increases under a specific working condition, its weight will automatically decrease, thereby avoiding the risk of a single model failing or having excessive deviation under a specific working condition, making the fused state estimation more robust and reliable. (3) Achieved more accurate life assessment: Based on the high-precision fusion health obtained after fusion, this application can more accurately calculate the remaining useful life (RUL), providing a more reliable quantitative basis for formulating scientific maintenance plans and spare parts management; (4) It realizes refined intelligent alarm and maintenance: This application is not a simple threshold alarm, but a graded alarm based on three dimensions: health, remaining service life and average early warning amount. This enables the system to trigger different levels of maintenance strategies, from early attention and planned maintenance to emergency intervention, according to different risk levels. It realizes the transformation from passive response to proactive and refined risk management, effectively improving maintenance efficiency and equipment operation reliability. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings required in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a predictive maintenance method for a current sensor according to an embodiment of this application. Detailed Implementation

[0019] The detailed description of the accompanying drawings is intended to illustrate the present embodiments of this application and is not intended to represent only the forms in which this application can be implemented. It should be understood that the same or equivalent functions can be accomplished by different embodiments intended to be included within the spirit and scope of this application.

[0020] like Figure 1 As shown, one embodiment of this application provides a predictive maintenance method for current sensors. By integrating a physical mechanism model and a data-driven model, it enables accurate condition monitoring, lifespan prediction, and intelligent maintenance decision-making for current sensors in substation main equipment (such as transformers and circuit breakers). The method includes the following steps: Step S1: Periodically collect the current waveform of the current sensor, the ambient temperature, and the equivalent stress caused by external mechanical constraints; Specifically, current sensor arrays deployed at key locations in the substation (such as transformer bushings and circuit breaker contacts) continuously acquire the primary side current waveform at a sampling frequency of no less than 1 kHz. Simultaneously, integrated or nearby temperature and vibration / strain sensors synchronously acquire the ambient temperature at the sensor locations and the equivalent stress caused by external mechanical constraints such as equipment vibration and assembly stress. This raw data forms the basis for all subsequent analysis and predictions, collectively constituting a comprehensive view of the sensor's operating status. Data acquisition can be performed by edge computing nodes within the substation (such as STM32H7+FPGA) and high-speed, reliable communication with the sensor array is achieved via a CAN bus.

[0021] Step S2: Based on the current waveform, ambient temperature, equivalent stress caused by external mechanical constraints, and a preset multiphysics coupling mechanism model, the state parameters of the current sensor are predicted to obtain a first state prediction value; the first state prediction value includes the temperature of the first sensitive element of the current sensor, the equivalent stress of the first sensitive element, the first sensitivity, the first zero bias, and the first health status. Step S3: Based on the current waveform, ambient temperature, and equivalent stress caused by external mechanical constraints, extract a multi-scale degradation feature vector. Input the multi-scale degradation feature vector into a preset long short-term memory network to predict the current sensor state parameters and obtain a second state prediction value. The second state prediction value includes the second sensor temperature, the second sensor equivalent stress, the second sensitivity, the second zero bias, and the second health of the current sensor. Specifically, Long Short-Term Memory (LSTM) networks learn a large amount of historical data and master the nonlinear mapping relationship between these complex features and the future state of the sensor. Their output is the predicted value of each state parameter at a future time, i.e., the second state prediction value.

[0022] Step S4: Using an unscented Kalman filter, weights are adaptively allocated based on the prediction uncertainty of the multiphysics coupling mechanism model and the long short-term memory network. The first state prediction value and the second state prediction value are then fused according to the adaptive weight allocation to obtain the fused state prediction value of the current sensor. The fused state prediction value includes the fused sensing element temperature, fused sensing element equivalent stress, fused sensitivity, fused zero bias, and fused health of the current sensor. Specifically, the lower the uncertainty (the higher the certainty) of a model, the higher its weight in the fusion result. The final output fusion state prediction (including fusion-sensitive body temperature, fusion-sensitive body equivalent stress, fusion sensitivity, fusion bias, and fusion health) is more reliable than the prediction of any single model because it achieves the best balance between physical interpretability and data adaptability.

[0023] Step S5: Determine the remaining service life based on the fused health status, and calculate the average early warning amount, whereby the average early warning amount is the average lead time from the first trigger threshold of the current sensor health status to actual failure. Specifically, firstly, based on the historical data and current trends of the fused health status, the future degradation trajectory is extrapolated and predicted. Then, by finding the intersection of the predicted trajectory with the preset health status threshold (such as Hth=0.7), the expected failure time of the sensor is determined. The duration between the current moment and the failure time is the remaining service life (RUL). At the same time, the average early warning time (AWE) is statistically analyzed and calculated. This indicator is defined as the average time from the first health alarm issued by the system (i.e., the health level first falls below a certain alarm threshold) to the actual failure of the sensor. AWE is an important indicator for measuring the early warning effectiveness of the entire predictive maintenance system.

[0024] Step S6: Issue graded alarms based on the health status, remaining service life, and average early warning amount; Specifically, based on the fused health status obtained in step S4, the remaining lifetime (RUL) obtained in step S5, and the average early warning amount (AWE), an intelligent hierarchical alarm strategy is implemented. This strategy is not a single threshold alarm, but rather is divided according to risk level, for example: When the health status is in an initial decline range, or when AWE indicates that there is plenty of time for warnings, a yellow warning is triggered to remind operations and maintenance personnel to pay attention. When the health status enters the medium risk zone, or when RUL indicates that it has entered the mid-term maintenance window, an orange alert is triggered and a maintenance work order is automatically generated. When the health status drops to a dangerously low value, or when the RUL is very urgent, a red emergency alarm is triggered, which may trigger the upper-level system to restrict the operating conditions.

[0025] This embodiment combines a multiphysics coupling mechanism model with a long short-term memory network, achieving complementary advantages between physical mechanisms and data-driven approaches. The mechanism model provides physical interpretation and boundary constraints, while the data-driven model can learn complex degradation patterns from massive amounts of data. The two work together to significantly improve the accuracy and reliability of state prediction. The method in this embodiment uses unscented Kalman filtering and performs adaptive weighted fusion based on the prediction uncertainty of the two models. When the prediction uncertainty of a certain model increases under a specific working condition, its weight will automatically decrease, thereby avoiding the risk of a single model failing or having excessive deviation under a specific working condition, making the fused state estimation more robust and reliable. Based on the high-precision fusion health obtained after fusion, the method in this embodiment can more accurately calculate the remaining useful life (RUL), providing a more reliable quantitative basis for formulating scientific maintenance plans and managing spare parts; The method in this embodiment is not a simple threshold alarm, but a graded alarm that combines three dimensions: health status, remaining service life, and average early warning amount. This enables the system to trigger different levels of maintenance strategies, from early attention and planned maintenance to emergency intervention, based on different risk levels. This realizes the transformation from passive response to proactive and refined risk management, effectively improving maintenance efficiency and equipment reliability.

[0026] In some embodiments, the multiphysics coupling mechanism model in step S2 includes: The current-carrying heating and thermal equilibrium model is used to calculate the Joule heating power based on the primary current waveform and the temperature-dependent resistance of the sensing element. By combining boundary conditions such as ambient temperature and convective heat transfer coefficient, the thermal equilibrium differential equation is solved to predict the first sensing element temperature; the specific formula is shown below: Sensor Sensing Body Resistance The current-carrying heating power varies with temperature as follows: (1) Joule heating power (instantaneous), unit W. The heat power generated by current passing through the resistance of a sensitive element; : Instantaneous value of the primary (measured) current, in amperes (A); The equivalent resistance of the sensor element varies with temperature. The function, in Ω; Reference temperature The equivalent resistance is given below, in Ω. Temperature coefficient of resistance, unit ; Reflects the increase in temperature The relative change in resistance; : The current operating temperature of the sensor; Reference temperature / Calibration temperature; Equivalent heat capacity—heat dissipation model: (2) in, For equivalent heat capacity, The convective heat transfer coefficient and area, Ambient temperature, This indicates solar radiation or a heat source coupled with an adjacent heating element. This indicates the instantaneous temperature of the sensor's sensing element, that is, the actual operating temperature inside the sensor (or the sensing chip).

[0027] The thermo-elastic coupled stress model is used to calculate the thermal expansion caused by temperature rise based on the predicted temperature change, the linear expansion coefficient and elastic modulus of the material, and superimposed with external mechanical constraints, thereby predicting the equivalent stress of the first sensitive element of the sensor body; the specific formula is shown below: The thermal expansion stress caused by temperature rise is: (3) in For the equivalent elastic modulus, The coefficient of linear expansion is 1 / 3. Additional stress induced by external mechanical constraints (assembly, vibration).

[0028] The sensitivity and bias drift model is used to substitute the predicted temperature and stress into a first-order linearized sensitivity drift model to calculate the resulting changes in the first sensitivity and the first zero bias; the specific formulas are as follows: Temperature and stress on sensitivity With zero bias The effect is linearized to first order: , (4) The sensor at any time Sensitivity (output / measured); Reference temperature The nominal sensitivity (calibrated value) at zero stress. Temperature coefficient (for sensitivity) describes the relative effect of temperature changes on sensitivity; Stress coefficient (for sensitivity), describing equivalent mechanical stress The relative impact on sensitivity; The sensor at any time Zero bias / offset (output with no input or zero input); Reference temperature The nominal zero deviation (calibration value) at zero stress. Temperature coefficient (with respect to zero); Stress coefficient (with respect to zero bias); : The current operating temperature of the sensor; Reference / calibration temperature; The equivalent mechanical stress of the sensitive body or sensor body; Sensor output: (5) For measuring noise (Gaussian / mixed noise).

[0029] The material aging and health calculation model is used to calculate the cumulative damage to the material based on the temperature history using the Arrhenius accelerated model, and to deduce the first health level (H = 1 - cumulative damage) characterizing the aging degree of the sensor; the specific formula is shown below: Accelerated damage using Arrhenius and Miner:

[0030] The aging rate or failure rate of a material at temperature φ indicates how quickly the material's properties degrade at that temperature. Frequency factor, which is the theoretical maximum rate of reaction or aging under ideal high temperature conditions, reflects the intrinsic reactivity of the material. : refers to the energy required to trigger the aging process; Boltzmann constant, the nature constant, is used to establish a relationship between energy and temperature; T: Absolute temperature, i.e., the temperature of the material during operation or testing; (6) Health ,when Trigger maintenance and calibration.

[0031] It is the instantaneous temperature of the material or sensor at time τ.

[0032] This embodiment employs the Arrhenius accelerated aging model to describe the thermal aging rate of the sensing material under temperature conditions, and utilizes the Miner cumulative damage model to integrate the material degradation caused by thermal-stress cycling over time, thereby obtaining the cumulative damage degree D(t) and health degree H(t) of the material. This method enables quantitative assessment of the lifespan of current sensor packaging materials, wires, and adhesive layers under complex thermal-mechanical environments. In other words, "using Arrhenius accelerated aging and Miner cumulative damage" means using these two principles to jointly predict: the aging rate of the sensor under different temperatures and operating stresses, how long it can last, and when maintenance is needed. In some embodiments, the feature vector This includes at least one of the following: RMS current, total harmonic distortion, DC bias, kurtosis, skewness, residual variance, sensor temperature, equivalent stress, sensitivity drift, zero bias drift, power factor, rate of temperature rise, relative change in resistance, equivalent service time, and noise spectrum energy.

[0033] ; : Root Mean Square (RMS) reflects the magnitude of the load current; Total Harmonic Distortion (THD) measures waveform distortion. The DC component of the current signal reflects zero drift or electromagnetic offset. Kurtosis measures the sharpness of a waveform. Skewness, a measure of waveform symmetry; Residual variance reflects the difference between model predictions and actual measurements; T: Sensitive body temperature (real-time temperature); Equivalent stress or vibration intensity; Sensitivity drift; Zero bias drift; PF: Power Factor; : Rate of temperature rise; : Relative change in resistance; Equivalent service time; Noise spectrum energy; In this embodiment, the above features are divided into several categories to explain their extraction methods and physical significance in detail; I. Time-domain and frequency-domain electrical characteristics (calculated directly from the current waveform). These characteristics are mainly obtained by performing mathematical operations on the acquired current waveform I(t).

[0034] RMS current value: Extraction method: Integrate the square of the instantaneous current value I(t) within the time window T, and then take the square root.

[0035] Physical significance: It reflects the effective work capacity of the current and is directly related to the thermal load of the sensor. A continuous increase in Irms means increased Joule heating, which will accelerate sensor aging.

[0036] Total harmonic distortion: Extraction method: Perform a Fast Fourier Transform (FFT) on the current waveform I(t) to obtain its spectrum; calculate the amplitude of the fundamental frequency (usually 50 / 60Hz) and each harmonic (2nd, 3rd...nth); THD = (square root of the sum of squares of all harmonic amplitudes) / fundamental frequency amplitude.

[0037] Physical meaning: It measures the "purity" of the current waveform. An increase in THD usually indicates the presence of a nonlinear load in the power grid, or that the sensor itself has introduced additional harmonics due to performance degradation (such as magnetic core saturation or electronic circuit distortion), which is an important sign of a hidden fault.

[0038] DC bias: Extraction method: Within the time window T, integrate the instantaneous current value I(t) and calculate the average: DC_offset = (1 / T) * ∫I(t) dt; Physical meaning: The ideal AC current waveform has an average value of 0. A significant DC bias may indicate that the sensor has zero drift, is subject to interference from a strong external magnetic field, or has a fault such as half-wave rectification, which directly affects the measurement accuracy.

[0039] Power factor: Extraction method: Voltage waveform V(t) needs to be acquired synchronously. By calculating the phase difference φ between the voltage and current waveforms, PF = cos(φ). Alternatively, it can be calculated using active power P and apparent power S: PF = P / S.

[0040] Physical meaning: It reflects the efficiency of electrical energy utilization. Although it is more of a load characteristic, changes in load characteristics will affect the current waveform, thus indirectly reflecting the changes in the operating conditions of the sensor.

[0041] II. Statistical distribution characteristics (describing waveform morphology): These characteristics describe the distribution properties of current waveforms using statistical methods.

[0042] Kuroshi: Extraction method: Calculate the fourth central moment of the current waveform within the time window, Kurtosis = E[(I(t) - μ)] 4 ] / σ 4 (where μ is the mean and σ is the standard deviation); Physical meaning: Measures the "sharpness" of waveform distribution. High kurtosis may indicate the presence of transient spikes in the waveform, which may be related to switching operations, electric arcs, or noise from internal electronic components of a sensor.

[0043] Skewness: Extraction method: Calculate the third central moment of the current waveform within the time window, Skewness = E[(I(t) - μ)] 3 ] / σ 3 ; Physical meaning: It measures the "asymmetry" of waveform distribution. A non-zero skewness indicates that the waveform is asymmetrical in the positive and negative half-cycles, which may indicate nonlinear problems such as rectification failure.

[0044] 3. Model and physics-related features (requires a reference model or baseline). Extraction of these features usually requires a reference model or initial calibration values.

[0045] Residual variance: Extraction method: Using the mechanistic model (y(t) = S(t)*I(t) + b(t)), a theoretical output value y_pred(t) is calculated based on the input current I(t) and the predicted state parameters S(t) and b(t). This theoretical value is subtracted from the actual sensor measurement value y_meas(t) to obtain the residual ε(t) = y_meas(t) - y_pred(t). The variance of this residual within the time window is then calculated.

[0046] Physical meaning: It directly measures the degree of agreement between the mechanistic model and actual behavior. An increase in residual variance indicates that the model cannot fully explain the sensor's behavior, and that unknown degradation or interference has occurred.

[0047] Sensitivity drift: Extraction method: Calculate the current sensitivity S(t) by means of an online calibration procedure (such as injecting a standard signal) or by comparison with a reference sensor, and then subtract it from the initial calibration value S0, ΔS = S(t) - S0.

[0048] Physical significance: It is a direct measure of the core performance indicators of a sensor and a core characteristic of degradation.

[0049] Zero bias drift: Extraction method: Similarly, calculate the current zero bias b(t) and subtract it from the initial calibration value b0, Δb = b(t) - b0.

[0050] Physical meaning: It is also a core performance indicator, reflecting the fixed deviation of the measurement.

[0051] Relative change in resistance: Extraction method: Based on the mechanism model R(T) = R0 * (1 + α(T - T0)), the theoretical resistance value R(T) is calculated using the currently measured temperature T, and then (R(T) - R0) / R0 is calculated.

[0052] Physical significance: It reflects the changes in the physical properties of the sensitive element caused by factors such as temperature, and is a direct manifestation of thermal aging.

[0053] IV. Direct physical sensing characteristics (from other sensors).

[0054] Sensitive body temperature: Extraction method: Directly read the values ​​from the high-precision temperature sensor integrated inside or near the sensor.

[0055] Physical significance: Temperature is the most critical external factor affecting sensor performance and lifespan; almost all chemical and physical aging processes are closely related to temperature.

[0056] Equivalent stress: Extraction method: Directly read the values ​​of the strain gauges pasted on the sensor structure, or integrate and convert the signals collected by the vibration sensor to obtain the von Mises equivalent stress.

[0057] Physical meaning: It reflects the mechanical load on the sensor. Long-term or alternating stress can lead to material fatigue and structural damage.

[0058] V. Derive time-related features (reflecting trends).

[0059] Temperature rise rate: Extraction method: Calculate the difference between the current temperature T(t) and the previous temperature t-Δt, and then divide it by the time interval Δt, dT / dt ≈ (T(t) - T(t-Δt)) / Δt.

[0060] Physical significance: Drastic temperature changes are more destructive than stable high temperatures. This characteristic can capture transient thermal shocks.

[0061] Equivalent service time: Extraction method: This is a cumulative calculation based on the Arrhenius model. It converts the operating time at different temperatures to the equivalent aging time at a certain reference temperature. Equivalent service time = Δt * exp(-Ea / (k*T(t))).

[0062] Physical significance: Quantifying a complex temperature history into a single "aging clock" reflects the true extent of material damage more accurately than a simple calendar time.

[0063] VI. Frequency Domain Noise Characteristics.

[0064] Noise spectrum energy: Extraction method: Perform FFT on the current waveform; exclude the frequency band containing the fundamental wave and its main harmonics; calculate the total energy of the remaining frequency band (usually the high-frequency band).

[0065] Physical meaning: Reflects the level of random noise in the signal. An increase in noise energy may indicate a decline in the performance of electronic components, loose connections, or electromagnetic compatibility issues.

[0066] Using the above method, the original, seemingly messy current waveform and auxiliary sensor data are transformed into a 15-dimensional, physically meaningful multi-scale degradation feature vector, providing high-quality data for the subsequent ISSA-LSTM model to make accurate lifetime predictions.

[0067] In some embodiments, in step S3, the Long Short-Term Memory network is an LSTM optimized using the Sparrow Search Algorithm (ISSA-LSTM); the ISSA is used to search for the number of layers, hidden units, learning rate, and sequence length hyperparameters of the LSTM to meet hardware constraints on latency and memory usage; the training loss function of the ISSA-LSTM includes a physical consistency residual term, which embeds the constraints of the multiphysics coupling mechanism model as soft constraints to force the prediction results of the LSTM to conform to physical laws.

[0068] Specifically, the performance (prediction accuracy) of LSTM models is highly dependent on their hyperparameters, such as the number of network layers, the number of hidden units, the learning rate, and the length of the input sequence. The combination space of these parameters is enormous, making manual tuning inefficient and difficult to find the global optimum. Deploying models on edge computing nodes (such as STM32H7 + FPGA) in substations must meet strict real-time and resource constraints; the time from data input to prediction output must be extremely short (e.g., patents require end-to-end latency ≤ 3s). An overly complex model will lead to excessively long inference times, losing its warning value. Furthermore, edge nodes have limited memory, and too many model parameters can cause loading or running failures. ISSA (Sparrow Search Algorithm), as a swarm intelligence optimization algorithm, is used as an automated hyperparameter searcher. Its workflow is as follows: Define the search space: Set the range of options for each hyperparameter (number of layers, hidden units, etc.).

[0069] Define the fitness function: ISSA evaluates the quality of a set of hyperparameters not by a single criterion of "highest accuracy," but by a multi-objective function whose objectives are: Minimize prediction error (to ensure model accuracy); Satisfy Latency ≤ Lmax (to ensure the model is fast); The condition Mem ≤ Mmax is satisfied (to ensure the model is small); Iterative search: ISSA intelligently explores and develops in the hyperparameter space to find the hyperparameter combination that satisfies all constraints and minimizes the prediction error.

[0070] Delaying the inference time of the system or model, The maximum allowable delay threshold is determined by the application's real-time requirements. This refers to the system's memory usage during runtime, including model parameters and intermediate cache. The maximum available memory capacity is limited by hardware (such as edge MCUs, FPGAs, embedded GPUs).

[0071] Through ISSA optimization, the resulting ISSA-LSTM model is no longer an academic model that performs well only on cloud servers, but an engineering model that has been tailored and optimized for specific edge hardware and can be practically deployed, solving the problem of AI models being easy to use but difficult to implement.

[0072] Specifically, the loss function of the Long Short-Term Memory network is:

[0073] in, Embed the constraints of equations (1)-(6) as soft constraints (Physics-informed); It is a loss function; In the future The true / target feature vector; Model for future moments The predicted feature vector (network output); The squared L2 norm of a vector (mean square sum) is used to measure error energy. Predicting the horizon / leading step size; The physical consistency weighting coefficient is used to balance the impact of data-driven errors and physical constraint residuals. The physical consistency residual vector is used to write the mechanistic equation in the form of "=0" and calculate its bias on the predictor.

[0074] In some embodiments, in step S4, the prediction uncertainty is the first covariance matrix of the multiphysics coupling mechanism model and the second covariance matrix of the long short-term memory network, respectively; the adaptive weight allocation specifically means that the weight of the first state prediction value is proportional to the inverse of the first covariance matrix, and the weight of the second state prediction value is proportional to the inverse of the second covariance matrix.

[0075] Specifically, the prediction mechanism model employs unscented Kalman filtering (UKF) fusion. With LSTM prediction As shown in the formula below:

[0076] and They are and The weight matrix; The weight matrix adapts to the covariance of both sides:

[0077] To obtain a more robust state estimate , These are the fusion-sensitive body temperature, fusion-sensitive body equivalent stress, fusion sensitivity, fusion bias, and fusion health. For a prediction vector containing n state parameters, such as [temperature, stress, sensitivity, bias, health], its covariance matrix is ​​an n*n square matrix. The larger the variance, the more uncertain the prediction of that parameter.

[0078] For example, when two models are predicting the temperature of a sensor, the mechanistic model (based on physical equations) is very good at handling this condition, giving a prediction of "85°C ± 0.5°C" with low uncertainty and small covariance. The LSTM model also gives a prediction, but its uncertainty is slightly larger, such as "85.2°C ± 1°C". The fusion model will calculate that the mechanistic model has a higher weight, so the final fused temperature will be very close to 85°C.

[0079] For example, a sudden current surge creates complex dynamics that exceed the descriptive power of a simple mechanistic model. Its prediction might be "92°C ± 5°C," with a dramatic increase in uncertainty and covariance. An LSTM model, having learned similar patterns from historical data, might predict "93.5°C ± 1.5°C," with relatively lower uncertainty. The fusion process automatically reduces the weight of the mechanistic model and increases the weight of the LSTM model, resulting in a fused temperature closer to 93.5°C, thus avoiding the significant errors introduced by the mechanistic model in its unsuitable range.

[0080] In some embodiments, in step S5, determining the remaining useful life includes: taking the moment when the future prediction curve of the fused health level intersects with a preset health level threshold as the failure time point, and the difference between the current moment and the failure time point is the remaining useful life.

[0081] Specifically, firstly, the future degradation trajectory curve is extrapolated and predicted using the obtained fused health sequence; then, the predicted curve is compared with a preset health threshold representing sensor failure or maintenance requirements. When the predicted curve first intersects the threshold line, the corresponding moment is defined as the predicted failure time point; finally, the time difference between the current moment and this failure time point is the remaining service life (RUL) of the current sensor.

[0082] In some embodiments, in step S6, the graded alarm includes at least three levels: A yellow alert is triggered when the health status is in the first range or the average early warning amount exceeds the first threshold, and a response strategy of recommended retesting or calibration is implemented. Specifically, a yellow alert (early attention) is a low-risk, long-term alarm that focuses on the slackness of the warning. A yellow alert is triggered when the health status has just begun to decline, or when the system determines through the average early warning amount (AWE) that we have ample time to respond. At this time, the system does not require immediate action, but recommends that operations and maintenance personnel perform low-cost, preventative operations such as retesting or calibration to confirm the problem and delay degradation. An orange alert is triggered when the health status is in the second range or the remaining useful life (RUL) is less than the second threshold, and response strategies such as generating maintenance plans and increasing sampling frequency are implemented. Specifically, the orange alert (planned maintenance) is a medium-risk, medium-cycle alarm; it begins to focus on the urgency of failure. When the health status enters the moderate degradation range, or the remaining useful life (RUL) shortens to the window period requiring planning, the orange alert is triggered. At this time, the system will automatically implement more proactive risk management measures such as generating maintenance work orders and increasing data sampling frequency to prepare for the upcoming maintenance. A red emergency is triggered when the health status is in the third range or the remaining useful life is less than the third threshold, and a strong alert and operating condition restriction response strategy are implemented. Specifically, a red emergency (immediate intervention) is a high-risk, short-cycle alarm; it is entirely focused on the imminent risk of failure. A red emergency is triggered when the health status drops to a dangerous level or the remaining useful life (RUL) is very short. At this time, the system will take the most drastic response, such as issuing a strong alert to notify all relevant personnel, and may even link with the upper-level control system to restrict equipment operating conditions to avoid catastrophic failure. Through this hierarchical approach, the alarm system is no longer a simple binary judgment of normal / abnormal, but has become an intelligent decision support system that can communicate with operations and maintenance personnel. It clearly conveys how serious the problem is and what we should do, achieving a leap from passive response to proactive and refined risk management.

[0083] Another embodiment of this application provides a predictive maintenance system for a current sensor, including a module for performing the predictive maintenance method for a current sensor described in the above embodiments.

[0084] It should be noted that the system provided in this embodiment can be used to execute the methods described in the above embodiments. Therefore, the contents not described in detail in this embodiment can be obtained by referring to the contents of the methods in the above embodiments, and will not be repeated here.

[0085] Another embodiment of this application provides an electronic device, including: A communication interface used for communicating with other electronic devices; Memory is used to store computer program instructions; A processor is configured to execute the computer program instructions to support the electronic device in implementing the methods described in the embodiments above.

[0086] In this embodiment, the memory mainly includes a program storage area and a data storage area. The program storage area can store the operating device, applications required for at least one function, etc., and the data storage area can store related data, etc. Furthermore, the memory can be a high-speed random access memory, or a non-volatile memory, such as a plug-in hard disk, a smart media card (SMC), a secure digital card (SD), and a flash card, or other volatile solid-state storage devices.

[0087] The processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor, or the processor can be any conventional processor. The processor is the control center of the electronic device and uses various interfaces and lines to connect the various parts of the electronic device.

[0088] Another embodiment of this application provides a computer program product including computer program instructions that instruct a computer device to perform operations corresponding to the methods described in the above embodiments.

[0089] Specifically, the computer program product includes a series of computer program instructions, which are codes written in a computer program. These instructions define how to perform specific operations. The computer program instructions are designed to be loaded onto a computer device and instruct the device to perform specific operations, which refer to the various steps in the predictive maintenance method for current sensors described in the above embodiments. In this way, the computer program product of this embodiment provides a complete software solution that can run on various computer devices to implement the predictive maintenance method for current sensors described in the above embodiments.

[0090] The various embodiments of this application have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used in these embodiments is chosen to best explain the principles, practical applications, or technological improvements to the embodiments in the market, or to enable other those skilled in the art to understand the embodiments disclosed herein.

Claims

1. A predictive maintenance method for a current sensor, characterized in that, include: Step S1: Periodically collect the current waveform of the current sensor, the ambient temperature, and the equivalent stress caused by external mechanical constraints; Step S2: Based on the current waveform, ambient temperature, equivalent stress caused by external mechanical constraints, and a preset multiphysics coupling mechanism model, the state parameters of the current sensor are predicted to obtain a first state prediction value; the first state prediction value includes the temperature of the first sensitive element of the current sensor, the equivalent stress of the first sensitive element, the first sensitivity, the first zero bias, and the first health status. Step S3: Based on the current waveform, ambient temperature, and equivalent stress caused by external mechanical constraints, extract a multi-scale degradation feature vector. Input the multi-scale degradation feature vector into a preset long short-term memory network to predict the current sensor state parameters and obtain a second state prediction value. The second state prediction value includes the second sensor temperature, the second sensor equivalent stress, the second sensitivity, the second zero bias, and the second health of the current sensor. Step S4: Using unscented Kalman filtering, weights are adaptively allocated based on the prediction uncertainty of the multiphysics coupling mechanism model and the long short-term memory network, and the first state prediction value and the second state prediction value are fused based on the adaptive weight allocation to obtain the fused state prediction value of the current sensor. The predicted fusion status values ​​include the fusion sensor temperature, fusion sensor equivalent stress, fusion sensitivity, fusion bias, and fusion health of the current sensor. Step S5: Determine the remaining service life based on the fused health status, and calculate the average early warning amount, whereby the average early warning amount is the average lead time from the first trigger threshold of the current sensor health status to actual failure. Step S6: Issue graded alarms based on the health status, remaining service life, and average early warning amount.

2. The method according to claim 1, characterized in that, In step S2, the multiphysics coupling mechanism model includes: The current-carrying heating and thermal balance model is used to calculate the Joule heating power based on the primary side current and the resistance of the sensing element, and to solve for the instantaneous temperature of the sensing element by combining the ambient temperature and heat dissipation conditions. A thermo-elastic coupled stress model is used to calculate thermal expansion stress based on temperature rise. The sensitivity and bias drift model is used to calculate the sensitivity and zero bias drift based on the instantaneous temperature and thermal expansion stress of the sensitive body. The material aging and health calculation model is used to calculate cumulative damage based on the temperature history using the Arrhenius accelerated model, and thus obtain the health status.

3. The method according to claim 1, characterized in that, The eigenvector includes at least one of the following: RMS current, total harmonic distortion, DC bias, kurtosis, skewness, residual variance, sensor temperature, equivalent stress, sensitivity drift, zero bias drift, power factor, rate of temperature rise, relative change in resistance, equivalent service time, and noise spectrum energy.

4. The method according to claim 1, characterized in that, In step S3, the Long Short-Term Memory network is optimized using the Sparrow Search Algorithm (ISSA). The ISSA is used to search for the number of layers, hidden units, learning rate, and sequence length hyperparameters of the LSTM to meet hardware constraints on latency and memory usage. The training loss function of the Long Short-Term Memory network includes a physical consistency residual term, which embeds the constraints of the multiphysics coupling mechanism model as soft constraints to force the prediction results of the LSTM to conform to physical laws.

5. The method according to claim 1, characterized in that, In step S4, the prediction uncertainty refers to the first covariance matrix of the multiphysics coupling mechanism model and the second covariance matrix of the long short-term memory network, respectively; the adaptive weight allocation specifically means that the weight of the first state prediction value is proportional to the inverse of the first covariance matrix, and the weight of the second state prediction value is proportional to the inverse of the second covariance matrix.

6. The method according to claim 1, characterized in that, In step S5, determining the remaining service life includes: taking the moment when the future prediction curve of the fused health level intersects with the preset health level threshold as the failure time point, and the difference between the current moment and the failure time point is the remaining service life.

7. The method according to claim 1, characterized in that, In step S6, the graded alarm includes at least three levels: A yellow alert is triggered when the health status is in the first interval or the average early warning amount is greater than the first threshold, and a response strategy of recommended retesting or calibration is implemented. Orange alert is triggered when the health status is in the second range or the remaining service life is less than the second threshold, and a response strategy of generating a maintenance plan and increasing the sampling frequency is executed. Red Emergency: This is triggered when the health status is in the third range or the remaining lifespan is less than the third threshold, and a strong reminder and a response strategy that restricts operating conditions are implemented.

8. A predictive maintenance system for a current sensor, characterized in that, Includes a module for performing the method according to any one of claims 1 to 7.

9. An electronic device, characterized in that, include: A communication interface used for communicating with other electronic devices; Memory is used to store computer program instructions; A processor for executing the computer program instructions to support the electronic device in implementing the method according to any one of claims 1 to 7.

10. A computer program product, characterized in that, It includes computer program instructions that instruct a computer device to perform an operation corresponding to the method described in any one of claims 1 to 7.

Images