Inductor aging failure analysis and prediction method and system based on edge calculation

Abstract

The invention provides an inductor aging failure analysis and prediction method and system based on edge computing, and relates to the technical field of Internet of Things and edge computing, and the method comprises the steps: collecting the physical state and electrical parameter data of an inductor in real time through a multi-source sensor group; carrying out data fusion processing on the edge computing node, and extracting a dynamic feature vector related to aging failure; a dynamic failure threshold value is generated based on historical data, and real-time matching and early warning are achieved; updating a lightweight analysis model by adopting an online learning mechanism to adapt to a nonlinear aging process; the abnormal data and the feature vectors are uploaded to a cloud for failure root cause analysis and global model optimization, and optimization parameters are fed back to edge nodes; and dynamically distributing calculation tasks according to the edge node calculation load. According to the method, the problems of multi-source data fusion, dynamic threshold setting, model lightweight deployment, global optimization and the like are effectively solved, and the accuracy and the real-time performance of aging failure prediction of the inductor device are improved.

Description

Technical Field

[0001] The present invention relates to the field of Internet of Things and edge computing technologies, and in particular to an edge computing-based aging failure analysis and prediction method and system for inductor devices. Background Art

[0002] During long-term operation, inductors inevitably age due to various factors, including environmental stress, electrical load, and mechanical vibration. This can lead to performance degradation or even failure. Traditional inductor aging failure analysis relies primarily on regular offline testing and post-process repairs. This approach is not only time-consuming and labor-intensive, but also makes it difficult to capture the aging status of the device in real time, increasing the risk of unplanned downtime.

[0003] With the development of the Internet of Things and edge computing technologies, real-time online monitoring and predictive maintenance of inductor components have become possible. However, existing edge computing-based methods for analyzing and predicting inductor aging failure still have the following problems:

[0004] The fusion processing and feature extraction of multi-source heterogeneous data are currently a major challenge. Monitoring data for inductors typically includes multiple physical quantities and electrical parameters, such as temperature, vibration, current ripple rate, and equivalent series resistance. These data have varying sampling rates and dimensions, making it difficult to effectively extract features closely related to aging failures through direct analysis. Existing methods often lack effective data fusion and feature extraction mechanisms, resulting in low prediction accuracy.

[0005] Another key issue is the setting and real-time matching of dynamic failure thresholds. Because the aging process of inductors is nonlinear and uncertain, fixed failure thresholds are difficult to adapt to the needs of different aging stages. Existing methods for generating dynamic failure thresholds often rely on empirical experience or simple statistical models, which fail to accurately reflect the actual aging state of the device, resulting in delayed warnings or high false alarm rates.

[0006] The deployment and online updating of lightweight analysis models are also pressing challenges. Edge computing devices have limited resources, making it difficult to directly deploy complex, high-precision models. Furthermore, as devices age, model parameters need to be continuously adjusted to adapt to new aging characteristics. Existing methods lack model compression and online learning, making them difficult to meet real-time and accuracy requirements.

[0007] The global model optimization and task allocation mechanism is also an important factor affecting the prediction performance. In the distributed edge computing environment, how to effectively utilize the cloud resources for global model optimization, and timely feedback the optimized model parameters to the edge node is a difficulty in the current technology. At the same time, according to the computing power load state of the edge node, dynamically allocating the computing task to the adjacent edge device cluster is also the key to improve the overall performance and energy efficiency of the system. The existing method has defects in task allocation and energy efficiency management, and it is difficult to realize the optimal allocation of resources. SUMMARY

[0008] In order to solve the technical problems of multi-source heterogeneous data fusion processing and feature extraction difficulty, inaccurate dynamic failure threshold setting, insufficient lightweight analysis model deployment and online updating, and imperfect global model optimization and task allocation mechanism in the prior art, the present application provides an inductor device aging failure analysis and prediction method and system based on edge computing.

[0009] The technical scheme provided by the present application is as follows:

[0010] First aspect:

[0011] The inductor device aging failure analysis and prediction method based on edge computing provided by the present application comprises:

[0012] S1: Real-time acquisition of physical state data and electrical parameter data through a multi-source sensor group deployed on an inductor device, wherein the physical state data includes temperature, vibration acceleration frequency spectrum, and the electrical parameter data includes current ripple rate and equivalent series resistance;

[0013] S2: Fusion processing of the multi-source heterogeneous data on an edge computing node, extraction of a dynamic feature vector related to aging failure, wherein the dynamic feature vector includes at least two parameters: feature degradation rate FDR and failure confidence FCL;

[0014] S3: Generation of a dynamic failure threshold based on a historical aging data set, real-time matching of the dynamic feature vector by a lightweight analysis model, and triggering of a warning when FDR exceeds the dynamic failure threshold;

[0015] S4: Updating of the lightweight analysis model by an online learning mechanism, and adjustment of model parameters according to real-time data flow to adapt to the nonlinear aging process;

[0016] S5: Uploading of abnormal data and feature vectors of the edge node to a cloud platform, failure root cause analysis and global model optimization by the cloud, and feedback of the optimized model parameters to the edge node;

[0017] S6: Dynamic allocation of computing tasks to a neighboring edge device cluster according to the computing power load state of the edge node.

[0018] Further, the calculation method of the feature degradation rate in S2 specifically comprises:

[0019] S201: wavelet packet decomposition is performed on the current ripple ratio, temperature and vibration spectrum to extract energy entropy of each frequency band;

[0020] S202: Mahalanobis distance of the multi-dimensional feature vector is calculated:

[0021]

[0022] Wherein, F t represents the feature vector at time t, μ0 represents the initial health state feature mean vector of the device, ∑ represents the health state feature covariance matrix, and Δt represents the running time difference between the current and the initial state.

[0023] Further, the calculation method of the failure confidence in S2 specifically comprises:

[0024] S211: a basic risk value is calculated based on the similarity between the real-time feature vector and the historical failure case;

[0025] S212: a probabilistic output is generated through nonlinear transformation:

[0026]

[0027] Wherein, η represents the failure sensitive factor, which is calibrated through accelerated aging test, γ represents the failure acceleration index, and FDR base represents the reference degradation rate of the same type of device.

[0028] Further, the calculation method of the dynamic failure threshold in S3 specifically comprises:

[0029]

[0030] Wherein, μ h represents the historical feature degradation rate mean, σ h represents the historical feature degradation rate standard deviation, t represents the current running time, T life represents the design life of the device, and k1 and k2 represent the Weibull distribution shape factor.

[0031] Further, the online learning mechanism in S4 adopts an exponential smoothing adaptive algorithm, and the model parameter update specifically comprises:

[0032] S401: the gradient of the loss function at the current time is calculated;

[0033] S402: the parameters are updated according to the dynamic forgetting factor:

[0034]

[0035] wherein, θ t denotes the model parameters at the current time, denotes the gradient of the loss function, and p denotes the adaptive forgetting factor.

[0036] Further, the task allocation in S6 employs an improved Hungarian algorithm, specifically including:

[0037] S601: Constructing an edge node computing capability matrix and a task cost matrix;

[0038] S602: Solving a task allocation scheme with the objective of minimizing the maximum node delay and total energy consumption:

[0039] min(max i∈N T i )+λ·∑E i

[0040] wherein, T i denotes the task processing delay of the i-th node, E i denotes the energy consumption of the i-th node, and λ denotes the energy efficiency balance coefficient, and N denotes the set of edge nodes.

[0041] Further, the lightweight analysis model also needs to be compressed, specifically including:

[0042] S021: Training a high-precision LSTM failure prediction model on a cloud platform;

[0043] S022: Removing filters with a contribution less than 5% in the convolution layer by channel pruning;

[0044] S023: Migrating the LSTM model capability to an edge-deployable lightweight CNN model through knowledge distillation.

[0045] Further, the S5 includes a failure root cause analysis step specifically including:

[0046] S501: Constructing a multi-level failure analysis model based on a fault tree;

[0047] S502: Matching the abnormal feature vector with the patterns in the failure knowledge base;

[0048] S503: Generating an analysis report including the failure location, failure type, and remaining life prediction.

[0049] Further, the mapping relationship between the dynamic feature vector and the failure mechanism in S3 specifically includes:

[0050] When the feature degradation rate is greater than 1.5 and the temperature correlation coefficient is greater than 0.7, it corresponds to insulation thermal aging;

[0051] When the high-frequency vibration energy entropy suddenly increases by more than 40%, the mechanical fatigue of the winding is caused.

[0052] The second aspect:

[0053] The inductor aging failure analysis and prediction system based on edge computing provided by the application comprises:

[0054] A processor;

[0055] A memory, wherein the memory stores computer readable instructions, and the computer readable instructions are executed by the processor to realize the inductor aging failure analysis and prediction method based on edge computing.

[0056] The technical scheme provided by the application has at least the following beneficial effects:

[0057] (1) In the application, a large number of source sensors collect data in real time and perform fusion processing at edge computing nodes to extract dynamic feature vectors closely related to aging failure, effectively solving the technical problems of multi-source heterogeneous data fusion processing and feature extraction difficulty, and improving the accuracy and real-time performance of the prediction.

[0058] (2) In the application, a dynamic failure threshold is generated based on historical aging data sets, and real-time matching and early warning are realized through a lightweight analysis model, overcoming the limitations of traditional fixed thresholds that cannot adapt to different aging stages, and significantly reducing the false positive rate and false negative rate.

[0059] (3) In the application, an online learning mechanism is used to update the lightweight analysis model, and abnormal data and feature vectors are uploaded to the cloud for global model optimization, and the computing tasks are dynamically allocated according to the edge node computing load, effectively solving the technical problems of model lightweight deployment and online update deficiency, and imperfect global optimization and task allocation mechanism. BRIEF DESCRIPTION OF DRAWINGS

[0060] In order to more clearly illustrate the technical solutions in the embodiments of the application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0061] Figure 1 The flowchart of the inductor aging failure analysis and prediction method based on edge computing provided by the application is shown in the figure.

[0062] Figure 2 The lightweight analysis model compression flowchart of the inductor aging failure analysis and prediction method based on edge computing provided by the application is shown in the figure.

[0063] Figure 3 A failure root cause analysis flowchart of the inductor device aging failure analysis and prediction method based on edge computing provided in the embodiment of the present application is shown in FIG. 1.

[0064] Figure 4 A structure diagram of the inductor device aging failure analysis and prediction system based on edge computing provided in the embodiment of the present application is shown in FIG. 2. DETAILED DESCRIPTION

[0065] The technical solutions in the present application will be described below with reference to the drawings.

[0066] In the embodiments of the present application, the words such as "example", "for example" and the like are used to represent as an example, illustration or description. Any embodiment or design scheme described as "example" in the present application should not be interpreted as more preferred or more advantageous than other embodiments or design schemes. Rather, the word "example" is intended to present the concept in a specific manner. In addition, in the embodiments of the present application, the meaning expressed by "and / or" can be both, or can be one of the two.

[0067] In the embodiments of the present application, "image" and "picture" can be used interchangeably at times, and it should be pointed out that the meanings expressed are consistent when the distinction is not emphasized. "Of", "corresponding" and "corresponding" can be used interchangeably at times, and it should be pointed out that the meanings expressed are consistent when the distinction is not emphasized.

[0068] In the embodiments of the present application, sometimes the subscript such as W1 can be mistakenly used in the form of non-subscript such as W1, and the meanings expressed are consistent when the distinction is not emphasized.

[0069] In order to make the technical problems, technical solutions and advantages of the present application more clear, the following will be described in detail with reference to the drawings and specific embodiments.

[0070] Reference is made to the drawings accompanying the specification Figure 1 A flowchart of the inductor device aging failure analysis and prediction method based on edge computing provided in the embodiment of the present application is shown in FIG. 1.

[0071] The embodiment of the present application provides an inductor device aging failure analysis and prediction method based on edge computing, which can be implemented by an inductor device aging failure analysis and prediction device based on edge computing. The inductor device aging failure analysis and prediction device based on edge computing can be a terminal or a server. The processing flow of the inductor device aging failure analysis and prediction method based on edge computing can include the following steps:

[0072] S1: Real-time collection of physical state data and electrical parameter data through multi-source sensor groups deployed on the inductive device. Physical state data includes temperature, vibration acceleration frequency spectrum, and electrical parameter data includes current ripple ratio and equivalent series resistance.

[0073] Specifically, a multi-source sensor group is distributed and deployed on the surface and periphery of the inductive device, including a PT100 platinum resistance temperature sensor array (measurement point spacing ≤ 5 mm) and a three-axis MEMS (Micro-Electro-Mechanical System) vibration sensor, wherein the temperature sensor sampling frequency is set to 100 Hz, and the vibration sensor sampling frequency is set to 10 kHz. High-precision current sensors are used for electrical parameter collection, with a 1 MHz sampling rate to obtain the current ripple ratio (Current Ripple Ratio), and the equivalent series resistance (Equivalent Series Resistance, ESR) is measured by the four-wire method. The edge computing node selects an embedded industrial module, connects the data acquisition card through the PCIe interface, and executes data preprocessing in the real-time operating system (Real-Time Operating System, RTOS) environment: first, the original signal is subjected to Kalman filter noise reduction processing to eliminate electromagnetic interference in the industrial environment; second, the data stream is processed using a sliding time window mechanism (window length 500 ms, step length 100 ms); finally, the multi-source data is aligned by timestamp and stored in a ring buffer. The dynamic feature vector extraction process is completed in the inference engine of the edge node, the feature degradation rate (Feature Degradation Rate, FDR) reflects the acceleration of device performance degradation, and the failure confidence level (Failure Confidence Level, FCL) quantifies the failure probability. The dynamic failure threshold is generated by querying the threshold table issued by the cloud, which is updated every 24 hours based on the Weibull distribution model. The lightweight analysis model is deployed in a compressed format, and a hardware warning signal is triggered when the FDR exceeds the threshold for 3 consecutive periods. The online learning mechanism uses the incremental gradient descent method, and the model parameter update period is 30 minutes. Abnormal data is uploaded to the cloud platform through the MQTT protocol, and the optimized model parameters are issued to the edge node in the differential compression format. The edge device cluster is networked based on the wireless communication protocol, and the task distributor monitors the CPU utilization of each node in real time. When the load is > 80%, the feature extraction task is migrated to the adjacent node.

[0074] Further, the MQTT protocol transmission adopts a hierarchical topic structure: the first level is the area code (such as ASIA-EAST1), the second level is the device type (INDUCTOR), and the third level is the device ID (64-bit hash value). The differential compression uses the BSDiff algorithm, and the patch file generation interval is 6 hours, and the compression rate target is ≥65%. The model parameter update channel needs to reserve 10% bandwidth redundancy.

[0075] It should be noted that the edge computing module needs to meet the minimum computing power benchmark: CPU frequency ≥ 1.5 GHz (ARM Cortex-A72 architecture or similar performance), memory bandwidth ≥ 12.8 GB / s (dual-channel LPDDR4) to support real-time processing requirements. When the sensor sampling rate is full, the node memory occupancy should not exceed 70% of the total capacity. If it continues to exceed the limit, the data downsampling mechanism needs to be triggered (halve the sampling rate until the memory occupancy is ≤60%).

[0076] S2: The edge computing node fuses and processes the multi-source heterogeneous data, extracts the dynamic feature vector related to aging failure, and the dynamic feature vector includes at least two parameters: feature degradation rate FDR and failure confidence FCL.

[0077] In one possible implementation, the calculation method of the feature degradation rate in S2 specifically includes:

[0078] S201: Wavelet packet decomposition is performed on the current ripple rate, temperature, and vibration spectrum, and the energy entropy of each frequency band is extracted.

[0079] S202: The Mahalanobis distance of the multi-dimensional feature vector is calculated:

[0080]

[0081] F t represents the feature vector at time t, μ0 represents the initial health state feature mean vector of the device, ∑ represents the health state feature covariance matrix, and Δt represents the running time difference between the current and the initial state.

[0082] It should be noted that the wavelet packet decomposition selects db10 wavelet basis function, and the decomposition layer is 6. The specific process includes: Hilbert transform is performed on the current ripple signal to extract the envelope line; the temperature data is converted into a rate sequence by first-order difference; the vibration frequency spectrum is converted by FFT to extract the energy of the high frequency band of 1-5 kHz. The Mahalanobis distance needs to be calculated before the establishment of the health benchmark database: 72 hours of normal operation data is collected before the new device is put into use; the abnormal values outside ±3σ are removed when calculating the initial characteristic mean vector μ0; the covariance matrix Σ is updated once a quarter by using the sliding window method. The unit of time difference Δt is kilohours (khr), and unit normalization processing is required when calculating. When the local health benchmark data is insufficient, the benchmark parameters of the same type of device in the cloud are called to complete the FDR calculation. The 8 dimensions of the feature vector F t include 8 dimensions.

[0083] Further, the 8 dimensions of the feature vector F t are as follows:

[0084] Temperature rate of change: calculated by the temperature difference between adjacent sampling points, reflecting the local heat accumulation rate of the device;

[0085] Vibration main frequency energy entropy: extracted from the wavelet packet decomposition in the 1-5 kHz frequency band, quantifying the stability of the mechanical structure;

[0086] Current ripple envelope amplitude: the peak value of the envelope line after Hilbert transform, representing the power stress level;

[0087] ESR change gradient: the first-order difference value of the equivalent series resistance, indicating the degree of conductor degradation;

[0088] Temperature-current correlation coefficient: rolling calculation of the Pearson correlation coefficient between temperature and current ripple in 50 cycles;

[0089] High-frequency vibration harmonic distortion rate: the ratio of the total energy in the frequency band above 3 kHz to the fundamental frequency energy;

[0090] Thermal accumulation integral quantity: the integral of the temperature rate of change in a 10-minute window;

[0091] Maximum temperature difference between adjacent sensors: the maximum temperature difference value in the distributed temperature array.

[0092] When establishing the health benchmark database, the initial data needs to be collected under constant temperature (25℃±1℃) and rated load conditions.

[0093] Further, the temperature rate of change: calculated by the temperature difference between adjacent sampling points (interval 10ms), unit is ℃ / s;

[0094] Vibration main frequency energy entropy: Shannon entropy calculation is performed on the coefficients after wavelet packet decomposition in the 1-5 kHz frequency band;

[0095] Current ripple envelope amplitude: take the local maximum value of the envelope line after Hilbert transform, and retain the data within the 99% confidence interval;

[0096] ESR change gradient: calculated by central difference method Δt = 100 ms;

[0097] Temperature-current correlation coefficient: rolling Pearson correlation coefficient based on 50 cycles (5-second window);

[0098] High-frequency harmonic distortion rate: calculation formula Where E f is the FFT energy spectrum;

[0099] Thermal accumulation integral component: trapezoidal method integration of temperature change rate within 600 cycles (60 seconds);

[0100] Adjacent temperature difference extreme value: calculate the maximum temperature difference value in 12 distributed temperature sensors.

[0101] When establishing the health benchmark database, the rated working condition should be simulated in a constant temperature room (25±0.5℃) using a programmable load, and 72 hours of data should be continuously collected. The covariance matrix update uses an exponential smoothing method with a decay factor of 0.9:

[0102] ∑ new = α × ∑ old + (1-α) × ∑ current

[0103] Where α is the smoothing factor (value 0.8), controlling the weight of historical data, determined by grid search (test range 0.7-0.9, step 0.05), ∑ old is the old covariance matrix, Σ current is the current covariance matrix, Σ new is the output matrix, and the update period is 168 hours (every week).

[0104] In one possible implementation, the calculation method of the failure confidence in S2 specifically includes:

[0105] S211: calculate the basic risk value based on the similarity between the real-time feature vector and the historical failure case;

[0106] S212: generate a probabilistic output through a nonlinear transformation:

[0107]

[0108] Where η represents the failure sensitivity factor, calibrated through accelerated aging test, γ represents the failure acceleration index, and FDR baseIndicates the degradation rate of the same type of device reference.

[0109] It should be noted that the basic risk value calculation adopts a dynamic time warping (DTW) algorithm: the historical failure case library is stored in a cloud time series database, and the similarity calculation between the real-time feature vector and the historical case includes Euclidean distance weighting (temperature weight 0.4, vibration weight 0.3, and electrical parameter weight 0.3) and dynamic time warping path distance. The failure sensitivity factor η is calibrated through a 125℃ / 85%RH high temperature and high humidity accelerated aging test, and the failure acceleration index γ is set according to the device type: 2.8 for power inductors and 3.2 for filter inductors. The reference degradation rate FDR base According to the statistics of the health data of the same batch of devices, the 90th percentile value of the first 1000 hours of operation data is calculated. The probabilistic output conversion needs to be limited in the interval [0, 1], and when the calculation result is greater than 0.95, a secondary early warning is triggered. The exponential term in the formula is realized by piecewise linear approximation for hardware acceleration calculation.

[0110] It should be noted that the accelerated aging test procedure is: temperature cycle range-40℃ to 125℃, humidity 85%RH±3%, single cycle length 120 minutes, total cycle number ≥500 times. The η value calibration is realized by Weibull probability graph fitting, and the goodness of fit R 2 ≥0.95.

[0111] S3: Generate a dynamic failure threshold based on a historical aging data set, and perform real-time matching on the dynamic feature vector through a lightweight analysis model. When the FDR exceeds the dynamic failure threshold, an early warning is triggered.

[0112] In one possible implementation, the calculation method of the dynamic failure threshold in S3 specifically includes:

[0113]

[0114] Wherein, μ h represents the mean of the historical feature degradation rate, σ h represents the standard deviation of the historical feature degradation rate, t represents the current running time, T life represents the device design life, and k1 and k2 represent the Weibull distribution shape factors.

[0115] It should be noted that the construction of the historical database relied on by the dynamic failure threshold calculation needs to collect the full life cycle data of more than 200 devices of the same type, and the failure modes are divided into three categories of magnetic core saturation, winding fatigue, and insulation degradation through K-means clustering. The Weibull distribution shape factor k1 takes a value of 0.6-0.8 to represent the early failure probability, and k2 takes a value of 0.2-0.4 to represent the wear period failure acceleration. When the device running time t is greater than 0.7 times the design life Tlife When k2 value is automatically increased by 0.1. Design life T life According to the IEC 62380 standard calculation, the input parameters include environmental temperature (unit: ℃), working current (unit: A), cooling mode coefficient (1.0 for air cooling, 1.2 for liquid cooling). The double buffer mechanism is used to avoid the influence of the calculation process on the real-time matching performance when updating the threshold.

[0116] Further, the cooperative adjustment rules of k1 and k2 and the failure mode weight distribution are as follows:

[0117] Early failure period (t<0.3T life ):

[0118] k2=0.25

[0119] Where σ h / μ h is the historical data variation coefficient, reflecting the batch quality dispersion.

[0120] Random failure period (0.3T life ≤t<0.7T life ):

[0121] k1=0.68,k2=0.32

[0122] Wear period (t≥0.7T life ):

[0123]

[0124] The failure mode weight is dynamically distributed according to the fault tree analysis result: if the recent insulation degradation case proportion is >60%, μ h When calculating, give 1.5 times weight to insulation related features. The double buffer of threshold update uses ping-pong operation: the foreground buffer serves real-time matching, and the background buffer receives cloud update data. The pointer is switched every 24 hours.

[0125] In one possible implementation, the mapping relationship between the dynamic feature vector in S3 and the failure mechanism specifically includes:

[0126] When the feature degradation rate is greater than 1.5 and the temperature correlation coefficient is greater than 0.7, it corresponds to insulation thermal aging;

[0127] When the high-frequency vibration energy entropy suddenly increases by more than 40%, it corresponds to winding mechanical fatigue.

[0128] It should be noted that the insulation thermal aging judgment process first calculates the Pearson correlation coefficient between temperature and current ripple rate, and triggers the judgment when 5 consecutive samples meet FDR>1.5 and correlation coefficient>0.7. The judgment of winding mechanical fatigue requires extracting the wavelet energy entropy in the 3-5kHz frequency band of the vibration spectrum, calculating the percentage change between the current value and the baseline value, and triggering the judgment when the change is >40% and lasts for more than 300ms. The judgment result is transmitted to the PLC (Programmable Logic Controller) system via the Modbus RTU protocol. The logic for triggering the pre-maintenance program includes: reducing the load current by 20%, starting forced air cooling, and recording the failure code E307. All judgment rules are stored in the form of decision trees at the edge node, and support remote update of the rule base in the cloud.

[0129] Furthermore, the transient interference suppression strategy and misjudgment recovery process are as follows:

[0130] For transient events with a current ripple rate mutation duration of less than 100 microseconds, the system automatically identifies it as switching noise interference, and the relevant data points are not included in the temperature-current correlation coefficient calculation; when the vibration energy entropy increases by more than 40% but the temperature change rate is less than 0.1℃ / second, it is determined to be an external mechanical shock event and the abnormal record is ignored. The determination latch adopts a strict time window verification: only when at least 4 cycles in 5 consecutive sampling cycles (500 millisecond window) simultaneously meet FDR>1.5 and temperature-current ripple rate correlation coefficient ρ>0.7, the trigger flag is set (mathematical expression: if ∑I(FDR i >1.5)≥4 and∑I(ρ i >0.7)≥4, then the flag = 1, otherwise the flag = 0). The edge node implements state caching through a shift register, shifting the historical state to the left every 100 milliseconds and setting the lowest bit based on whether the current double conditions are met. The number of 1s in the lower 5 bits of the register is counted to determine whether to trigger an early warning.

[0131] The false positive recovery mechanism includes active reset and data analysis: After the warning is triggered, if the FDR is detected to be less than 1.2 and ρ is less than 0.5 for three consecutive cycles (300 milliseconds), the warning state is automatically lifted. An analysis report containing the following data is generated for each false positive:

[0132] 1) Feature vector time series 10 minutes before the false alarm (compression rate 80%)

[0133] 2) Ambient temperature fluctuation spectrum (1Hz sampling)

[0134] 3) Power supply voltage transient event recording (fluctuation >5% of nominal voltage), with the report uploaded to the cloud after differential compression for dynamic threshold optimization.

[0135] The hierarchical response strategy of the pre-maintenance procedure is bound with the failure confidence level (FCL): when FCL∈[0.8, 0.9), execute a 10% load reduction and record L1 log; when FCL∈[0.9, 0.95), start a 20% load reduction and forced air cooling and record L2 log; when FCL≥0.95, trigger the safety shutdown sequence: linear load reduction to zero power within 30 seconds → cut off the main power supply → activate the audible and visual alarm (code E307).

[0136] S4: Update the lightweight analysis model using an online learning mechanism, and adjust the model parameters according to real-time data flow to adapt to the nonlinear aging process.

[0137] In one possible implementation, the online learning mechanism in S4 adopts an exponential smoothing adaptive algorithm, and the model parameter update specifically includes:

[0138] S401: Calculate the loss function gradient at the current time;

[0139] S402: Update the parameters according to the dynamic forgetting factor:

[0140]

[0141] where θ t represents the model parameters at the current time, represents the gradient of the loss function, and ρ represents the adaptive forgetting factor.

[0142] The exponential smoothing adaptive algorithm adopts a composite loss function: 80% Huber loss + 20% cross-entropy. Gradient calculation is achieved using automatic differentiation technology, and the differential engine of the ONNX Runtime library can be called when deployed on edge nodes. The dynamic adjustment strategy of the adaptive forgetting factor ρ is divided into three levels: when FDR<0.5, ρ=0.92; when 0.5≤FDR<1.0, ρ=0.85; when FDR≥1.0, ρ=0.75. A backup volume of historical parameters is retained during model updating, and a rollback operation is performed when the validation set accuracy decreases by >5%. Inference delay testing is required after each update to ensure that the single prediction time is ≤20ms.

[0143] In one possible implementation, as shown in Figure 2 the lightweight analysis model also needs to be compressed, specifically including:

[0144] S021: Train a high-precision LSTM failure prediction model on a cloud platform;

[0145] S022: Remove filters with a contribution of less than 5% in the convolutional layer using channel pruning;

[0146] S023: Migrate the LSTM model capabilities to a lightweight CNN model that can be deployed on the edge through knowledge distillation.

[0147] Note that the LSTM (Long Short-Term Memory) model training input data dimension is 32 time steps x 8 features, and the network structure uses a double LSTM layer (128 units per layer) plus a fully connected layer. Channel pruning uses Taylor expansion to estimate filter importance, and the pruned model is fine-tuned in three stages: the first stage is trained for 50 rounds with a learning rate of 0.001, the second stage is trained for 30 rounds with a learning rate of 0.0003, and the third stage is trained for 20 rounds with a learning rate of 0.0001. The knowledge distillation temperature parameter T is 15, the student model is a MobileNetV2 variant constructed using depth separable convolution, and the output layer MSE (Mean Square Error) loss weight of the teacher model and the student model is set to 0.7. The quantization process uses a symmetric quantization strategy, and the weight range is mapped to [-127, 127], and the offset z is fixed at 0. The compressed model needs to pass the industrial four-level electromagnetic compatibility (EMC) test.

[0148] S5: upload the abnormal data and feature vectors of the edge node to the cloud platform, perform failure root cause analysis and global model optimization in the cloud, and feed back the optimized model parameters to the edge node.

[0149] In one possible implementation, as shown in FIG. 5, S5 includes a failure root cause analysis step, which specifically includes: Figure 3

[0150] S501: construct a multi-level failure analysis model based on a fault tree;

[0151] S502: match the abnormal feature vector with the patterns in the failure knowledge base;

[0152] S503: generate an analysis report including failure location, failure type, and remaining life prediction.

[0153] Note that the top event of the fault tree (Fault Tree) analysis model is defined as "inductor device functional failure", and the intermediate events include core saturation, winding open circuit, and insulation breakdown in a three-level structure. The pattern matching uses an improved cosine similarity algorithm: first, reduce the feature vector to 3D through principal component analysis, then calculate the similarity with the failure knowledge base cases, and the matching threshold is set to 0.85. The analysis report generation uses a templating engine: the failure location is positioned based on the abnormal combination of electrical parameters (such as a third winding turn-to-turn short circuit represented by ESR sudden increase + local temperature gradient > 15℃); the remaining life prediction uses Monte Carlo simulation, and the 90% confidence interval output is taken after 1000 iterations. The report is packaged in JSON format, including timestamp, device ID, failure type code, and other metadata.

[0154] ​S6: dynamically allocate computing tasks to the adjacent edge device cluster according to the edge node computing power load state.

[0155] In a possible implementation, the task allocation in S6 adopts an improved Hungarian algorithm, specifically including:

[0156] S601: construct an edge node computing capability matrix and a task cost matrix;

[0157] S602: solve a task allocation scheme with the objective of minimizing the maximum node delay and the total energy consumption:

[0158] min(max i∈N T i )+λ·∑E i

[0159] wherein T i represents the task processing delay of the i-th node, E i represents the energy consumption of the i-th node, λ represents an energy efficiency balance coefficient, and N represents the edge node set.

[0160] It should be noted that the hardware acceleration implementation of the improved Hungarian algorithm includes a computing capability matrix test: the matrix multiplication throughput test adopts a 1024x1024 floating point matrix, and the memory bandwidth test is obtained through DDR4 read-write stress test. The construction rule of the task cost matrix is that the feature extraction task consumes 1.2 GB of memory and takes 15 ms of calculation time, and the model inference task consumes 800 MB of memory and takes 8 ms of calculation time. The objective function solving adopts a branch and bound method, and the constraint conditions include that the maximum delay of a single node is ≤50 ms and the total energy consumption of the cluster is ≤20 W. The energy efficiency balance coefficient λ is dynamically adjusted through a fuzzy PID controller, the controller input is the average temperature of the cluster (unit: ℃) and the power supply voltage fluctuation rate, and the response time of the output λ value is 500 ms.

[0161] Further, the input-output mapping relationship and constraint processing mechanism of the fuzzy PID controller are as follows:

[0162] The input quantization factors include a temperature deviation e T = (T avg - 40) / 10, domain [-2, 2], and a voltage fluctuation rate e V = ΔV / V nom x 100, domain [-15, 15]. The fuzzy rule table is as follows:

[0163]

[0164]

[0165] The constraint violation processing includes that when max T i> 50ms, λ is forced to be below 0.2;

[0166] When the total energy consumption exceeds the limit, suspend non-critical tasks (such as historical data backup).

[0167] The fuzzy output of the energy efficiency balance coefficient λ needs to be de-fuzzied: the barycenter method is used to calculate the exact value, the output domain is [0.1, 0.8], and the resolution is 0.05. When the power supply voltage fluctuation rate is > 8%, λ is forced to be locked at 0.3 and the local degradation mode is started (only the feature extraction basic task is executed).

[0168] When solving by branch and bound method, α-β pruning optimization is used: when the objective function value of the current branch is greater than 1.2 times the optimal solution, the branch search is terminated.

[0169] Referring to the description attached Figure 4 , the structure schematic diagram of the inductor aging failure analysis and prediction system based on edge computing provided by the embodiment of the present application is shown.

[0170] The present application also provides an inductor aging failure analysis and prediction system based on edge computing, which is applied to the inductor aging failure analysis and prediction method based on edge computing and comprises:

[0171] The processor 201.

[0172] The memory 202, the memory 202 stores computer readable instructions, and the computer readable instructions are executed by the processor 201, realizing the inductor aging failure analysis and prediction method based on edge computing as in the method embodiment.

[0173] The inductor aging failure analysis and prediction system based on edge computing 20 provided by the present application can execute the inductor aging failure analysis and prediction method based on edge computing described above, and realize the same or similar technical effects. To avoid repetition, the present application will not be described again.

[0174] The technical scheme provided by the embodiment of the present application has at least the following beneficial effects:

[0175] (1) In the present application, too many source sensor groups collect data in real time and perform fusion processing at the edge computing node, extract dynamic feature vectors closely related to aging failure, effectively solve the technical problems of multi-source heterogeneous data fusion processing and feature extraction difficulty, and improve the accuracy and real-time performance of the prediction.

[0176] (2) In the present application, the dynamic failure threshold is generated based on the historical aging data set, and real-time matching and early warning are realized through a lightweight analysis model, overcoming the limitations of traditional fixed threshold that cannot adapt to different aging stages, significantly reducing the false positive rate and false negative rate.

[0177] (3) In the present application, an online learning mechanism is adopted to update the lightweight analysis model, and abnormal data and feature vectors are uploaded to the cloud for global model optimization, and the computing tasks are dynamically distributed according to the edge node computing power load, effectively solving the technical problems of insufficient model lightweight deployment and online update, and imperfect global optimization and task allocation mechanism.

[0178] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0179] The following points need to be explained:

[0180] (1) The drawings of the embodiments of the present application only involve the structures involved in the embodiments of the present application, and other structures can refer to the usual design.

[0181] (2) For the sake of clarity, the thickness of the layers or regions is exaggerated or reduced in the drawings used to describe the embodiments of the present application, that is, these drawings are not drawn according to the actual proportion. It can be understood that when elements such as layers, films, regions or substrates are referred to as being located "on" or "under" another element, the element can be "directly" located "on" or "under" another element or there can be an intermediate element.

[0182] (3) In the case of no conflict, the embodiments of the present application and the features in the embodiments can be combined with each other to obtain new embodiments.

[0183] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. The edge computing-based inductor device aging failure analysis and prediction method is characterized by: include: S1: A multi-source sensor group deployed on the inductor collects physical state data and electrical parameter data in real time. The physical state data includes temperature and vibration acceleration spectrum, and the electrical parameter data includes current ripple rate and equivalent series resistance. S2: fusing the multi-source heterogeneous data at the edge computing node to extract a dynamic feature vector related to aging failure, where the dynamic feature vector includes at least two parameters: a feature degradation rate (FDR) and a failure confidence level (FCL); S3: Generate a dynamic failure threshold based on the historical aging data set, match the dynamic feature vector in real time through a lightweight analysis model, and trigger an alarm when the FDR exceeds the dynamic failure threshold; S4: updating the lightweight analysis model using an online learning mechanism, adjusting model parameters according to real-time data streams to adapt to the nonlinear aging process; S5: Upload the abnormal data and feature vectors of the edge node to the cloud platform, which performs failure root cause analysis and global model optimization, and feeds the optimized model parameters back to the edge node; S6: Dynamically allocate computing tasks to adjacent edge device clusters based on the computing load status of edge nodes.

2. The method for analyzing and predicting aging failure of inductor devices based on edge computing according to claim 1, characterized in that: The calculation method of the characteristic degradation rate in S2 specifically includes: S201: performing wavelet packet decomposition on the current ripple rate, temperature and vibration spectrum to extract the energy entropy of each frequency band; S202: Calculate the Mahalanobis distance of the multidimensional feature vector: Among them, F t represents the eigenvector at time t, μ0 represents the mean vector of the initial health state characteristics of the device, ∑ represents the health state characteristic covariance matrix, and Δt represents the running time difference between the current and initial states.

3. The method for analyzing and predicting aging failure of inductor devices based on edge computing according to claim 1, characterized in that: The calculation method of the failure confidence in S2 specifically includes: S211: Calculate the basic risk value based on the similarity between the real-time feature vector and historical failure cases; S212: Generate probabilistic output through nonlinear transformation: Among them, η represents the failure sensitivity factor, which is calibrated by accelerated aging test, γ represents the failure acceleration index, and FDR base Indicates the baseline degradation rate of similar devices.

4. The method for analyzing and predicting aging failure of inductor devices based on edge computing according to claim 1, characterized in that: The calculation method of the dynamic failure threshold in S3 specifically includes: Among them, μ h represents the mean historical feature degradation rate, σ h represents the standard deviation of the historical characteristic degradation rate, t represents the current running time, T life represents the design life of the device, k1 and k2 represent the Weibull distribution shape factors.

5. The method for analyzing and predicting aging failure of inductor devices based on edge computing according to claim 1, characterized in that: The online learning mechanism in S4 adopts an exponential smoothing adaptive algorithm, and its model parameter update specifically includes: S401: Calculate the gradient of the loss function at the current moment; S402: Update parameters according to the dynamic forgetting factor: Among them, θ t represents the model parameters at the current moment, represents the gradient of the loss function, and ρ represents the adaptive forgetting factor.

6. The method for analyzing and predicting aging failure of inductor devices based on edge computing according to claim 1, characterized in that: The task allocation in S6 adopts the improved Hungarian algorithm, which specifically includes: S601: Constructing edge node computing capability matrix and task cost matrix; S602: Solve the task allocation solution with the goal of minimizing the maximum node delay and total energy consumption: min(max i∈N T i )+λ·∑E i Among them, T i represents the task processing delay of the i-th node, E i represents the energy consumption of the i-th node, λ represents the energy efficiency balance coefficient, and N represents the set of edge nodes.

7. The method for analyzing and predicting aging failure of inductor devices based on edge computing according to claim 1, characterized in that: The lightweight analysis model also needs to be compressed, specifically including: S021: Train a high-precision LSTM failure prediction model on the cloud platform; S022: Use channel pruning to remove filters in the convolutional layer whose contribution is less than 5%; S023: Migrate LSTM model capabilities to edge-deployable lightweight CNN models through knowledge distillation.

8. The method for analyzing and predicting aging failure of inductor devices based on edge computing according to claim 1, characterized in that: The step S5 includes failure root cause analysis, specifically including: S501: Construct a multi-level failure analysis model based on a fault tree; S502: Matching the abnormal feature vector with the pattern in the failure knowledge base; S503: Generate an analysis report including failure location, failure type and remaining life prediction.

9. The method for analyzing and predicting aging failure of inductor devices based on edge computing according to claim 1, characterized in that: The mapping relationship between the dynamic feature vector in S3 and the failure mechanism specifically includes: When the characteristic degradation rate is greater than 1.5 and the temperature correlation coefficient is greater than 0.7, it corresponds to insulation thermal aging; When the high-frequency vibration energy entropy suddenly increases by more than 40%, it corresponds to winding mechanical fatigue.

10. The inductor device aging failure analysis and prediction system based on edge computing is characterized by: include: processor; A memory having computer-readable instructions stored thereon, wherein when the computer-readable instructions are executed by the processor, the edge computing-based inductor device aging failure analysis and prediction method according to any one of claims 1 to 9 is implemented.

Images