Electromagnetic compatibility fault diagnosis method and system based on data driving

Abstract

The invention relates to an electromagnetic compatibility fault diagnosis method and system based on data driving, and belongs to the technical field of electromagnetic compatibility (EMC), and the method comprises the steps: injecting an interference signal into a to-be-tested device, and collecting the multi-dimensional operation parameter data of the device in an electromagnetic environment; performing feature extraction on the parameter data, including calculating statistical features in a time window and extracting periodic features through frequency domain transformation; carrying out sample equalization processing on the extracted features to balance the number of fault samples and the number of normal samples; based on the equalized feature data, a machine learning algorithm is adopted to construct a fault diagnosis model; and carrying out electromagnetic compatibility fault prediction and positioning on to-be-tested equipment by using the fault diagnosis model. According to the invention, accurate acquisition of multi-dimensional parameters, automatic extraction of fault features and intelligent prediction of the model can be realized, and the efficiency and accuracy of electromagnetic compatibility problem positioning are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of electromagnetic compatibility (EMC) technology, and in particular to a data-driven EMC sensitivity diagnostic method and system, applicable to the sensitivity analysis and optimization design of complex electronic systems and digital circuit boards. Background Technology

[0002] With the increasing integration of electronic systems and the growing complexity of electromagnetic environments, electromagnetic compatibility (EMC) issues in electronic systems are becoming increasingly prominent. Addressing the frequent occurrence of EMC susceptibility phenomena due to interference in digital circuit boards (PCBs), effective EMC fault diagnosis is crucial. This paper explores data-driven EMC data acquisition and feature engineering methods, starting with conducted susceptibility excitation testing. Intelligent algorithms are used to classify and predict EMC susceptibility phenomena, providing tools and methods for quickly and accurately locating and diagnosing EMC problems. By analyzing and characterizing electromagnetic parameters that are highly correlated with fault phenomena, targeted optimization and improvement strategies can be provided for the forward design of digital PCB EMC, solving the problem of the inefficiency and cumbersome nature of traditional experience-based, trial-and-error EMC fault rectification work.

[0003] Electromagnetic compatibility (EMC) standards provide testing methods for EMC issues in equipment and systems, serving as the primary means of resolving these problems. By conducting EMC tests, weaknesses in the system can be accurately analyzed, allowing for targeted improvements and rectifications to ensure system stability and reliability. However, due to the complexity of EMC mechanisms and the instability of phenomena, following relevant standards can only prevent most common EMC problems; there is no cure for EMC faults arising from external interference and other complex conditions. Traditional EMC troubleshooting primarily relies on adhering to established standards and empirical testing. However, in complex environments, this approach has limited efficiency and effectiveness, making it difficult to accurately diagnose faults and effectively optimize designs.

[0004] Current research on fault diagnosis primarily focuses on the system's hardware and software, offering systematic solutions across various sub-fields and application scenarios. However, research specifically targeting electromagnetic compatibility (EMC) is relatively limited, with its technological development still in the exploratory and experimental stages. Early research explored fault tree-based EMC fault diagnosis methods, constructing fault tree models by studying three key elements: electromagnetic interference sources, interference coupling paths, and sensitive equipment, thus enabling EMC fault diagnosis. Further research developed online EMC fault diagnosis platform software, significantly enhancing the practicality of this technology. However, these methods typically rely on pre-modeling and correlation analysis, making it difficult to extract analytical mechanisms from large amounts of real-world data, thus limiting their application to specific scenarios. Further research has established nonlinear relationships between fault data and fault sources, proposing SVM (Support Vector Machine)-based EMC fault diagnosis methods to train SVM models for fault prediction. Simultaneously, other studies have attempted to introduce PNN-based neural networks to train fault diagnosis models. These methods have preliminarily validated the applicability and feasibility of intelligent algorithms in solving EMC fault diagnosis problems. Summary of the Invention

[0005] This invention aims to solve the following technical problems:

[0006] 1. Design a data acquisition method for collecting electromagnetic compatibility sensitive data, and use a specific signal for conducted sensitive injection to excite the electromagnetic compatibility sensitive phenomenon of electronic circuit board, so as to collect the operating parameters of key points of circuit board and form raw data for subsequent data-driven modeling; 2. Design a data-driven modeling process for diagnosing electromagnetic compatibility problems, design effective data preprocessing, feature engineering processes and modeling methods, and build a fault classification model that can identify a variety of electromagnetic susceptibility phenomena.

[0007] To achieve the above objectives, the present invention provides a data-driven electromagnetic compatibility fault diagnosis method, which includes the following steps: Injecting interference signals into the device under test and collecting multi-dimensional operating parameter data of the device in an electromagnetic environment; Feature extraction is performed on the parameter data, including calculating statistical features within the time window and extracting periodic features through frequency domain transformation; The extracted features are subjected to sample equalization processing to balance the number of faulty samples and normal samples. Based on the equalized feature data, a fault diagnosis model is constructed using machine learning algorithms; The fault diagnosis model is used to predict and locate electromagnetic compatibility faults in the device under test.

[0008] Preferably, the multi-dimensional operating parameters include electrical parameters, signal integrity parameters, electromagnetic interference parameters, and thermal parameters. The electrical parameters include voltage, current, and power consumption; the signal integrity parameters include clock stability and data signal waveform; the electromagnetic interference parameters include common-mode / differential-mode current and conducted noise; and the thermal parameters include the temperature of key components.

[0009] Preferably, the parameter data acquisition points include power rails, DC / DC converter areas, high-speed signal transmission lines, core components, clock crystal areas, analog-digital mixed signal areas, and / or board interface boundary locations.

[0010] Preferably, feature extraction includes: Time series data is segmented using a fixed-length time window; Calculate the mean, variance, maximum, and minimum values ​​for each time window; The main frequency components are obtained by performing a discrete Fourier transform on the time series data, and periodic features are extracted by fitting trigonometric functions. All extracted features are normalized.

[0011] Preferably, the periodic feature extraction uses Fast Fourier Transform to convert the signal from the time domain to the frequency domain, identifies the main frequency components, fits a sine function model using a nonlinear optimization method, and extracts frequency information from the fitted parameters.

[0012] Preferably, the sample equalization process includes: oversampling minority class fault samples using the SMOTE method based on K-Means clustering, and deleting samples far from the classification boundary using the Euclidean distance-based undersampling method for majority class normal samples.

[0013] Preferably, feature selection includes: firstly, using statistical methods to calculate the correlation between features and the target variable for preliminary screening; then, using tree-based feature importance assessment for secondary screening; and finally, determining the optimal feature subset through cross-validation.

[0014] Preferably, the selection of machine learning algorithms includes: using random forest or gradient boosting decision tree when the data size is less than a preset threshold, using convolutional neural network, recurrent neural network or Transformer model when the data size is greater than the preset threshold, and using CNN-LSTM combined model for data with spatiotemporal correlation characteristics.

[0015] Preferably, injecting interference signals includes the following steps: Power on the board under test and bring it to a stable working state; Collect reference electromagnetic characteristic data; Connect the interference signal generator to the power port of the board using injection clamps, and inject the interference signal at the specified intensity until a sensitive phenomenon occurs. Record electromagnetic characteristic data under abnormal conditions and compare and calibrate with reference data; Gradually reduce the interference intensity until the board returns to normal, and repeat the test until no new faults occur.

[0016] The technical solution of the present invention also provides a data-driven electromagnetic compatibility fault diagnosis system, which includes: Interference signal generation module, used to generate and amplify interference signals; The data acquisition module, including measurement sensors and instruments, is used to acquire multi-dimensional electromagnetic compatibility parameters of the device under test; The data processing module is used to perform feature extraction, sample equalization, and feature selection on the collected data; The model training and diagnosis module is used to build fault diagnosis models and output diagnosis results.

[0017] Compared with the prior art, the present invention has the following significant advantages: The improved conducted sensitivity testing platform enables comprehensive and accurate acquisition of electromagnetic compatibility (EMC) parameter data for circuit boards; it provides efficient data preprocessing and feature engineering methods to enhance the accuracy and stability of fault diagnosis models; and it constructs an intelligent fault classification model that can quickly identify various electromagnetic susceptibility phenomena, providing strong support for optimized design. Specifically, this is reflected in the following three aspects: 1) Accurate and comprehensive data collection This invention comprehensively designs a data acquisition method for electromagnetic compatibility sensitive areas from the perspectives of data acquisition platform construction, board parameter and data acquisition point selection, and data acquisition test methods. It can accurately and efficiently collect board characteristic data, thereby facilitating subsequent data-driven modeling work and effectively analyzing the electromagnetic environment effect boundary of the board.

[0018] 2) Data-driven modeling is highly operable Starting from solving the problem of fault diagnosis, a data-driven fault diagnosis modeling process was designed, and the key steps and points of the modeling were comprehensively analyzed and considered. This method can effectively achieve high-precision modeling of the logical relationship between electromagnetic parameters and fault phenomena, thereby accurately classifying and identifying various EMC fault phenomena, and is practical for engineering implementation.

[0019] 3) Possesses a certain degree of versatility and applicability. It provides a systematic solution from experimental design and data acquisition to model training, which greatly improves diagnostic efficiency and is applicable to different types of digital circuit boards, with strong versatility. Attached Figure Description

[0020] Figure 1 Deployment diagram for the data acquisition platform; Figure 2 Table of waveform types for interference experiments; Figure 3 This is a schematic diagram of unbalanced data sampling. Figure 4 A diagram illustrating the data-driven modeling process. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] This invention provides a data-driven electromagnetic compatibility (EMC) fault diagnosis method and system. Through data-driven electromagnetic susceptibility testing and analysis, it discovers the correspondence between EMC faults and module / board operating parameters. By employing intelligent technology to uncover potential correlations, it achieves EMC fault diagnosis and prediction, enabling rapid problem location and rectification, thus facilitating forward design and targeted electromagnetic protection. The technical solution of this invention is described in detail below with reference to the accompanying drawings.

[0023] 1. Data Acquisition Method Design To construct an EMC fault diagnosis system, a dedicated data acquisition platform needs to be built to monitor the operation of equipment in complex electromagnetic environments and acquire electromagnetic compatibility characteristic data for subsequent diagnostic model design and verification. By injecting interference signals into the device under test, sensitive phenomena on the circuit board are stimulated, ultimately yielding the required experimental data.

[0024] 1.1 Basic Components of the Test Platform To fully explore the electromagnetic compatibility (EMC) characteristics of digital circuit boards, a dedicated testing platform needs to be built to monitor the device's operation in complex electromagnetic environments and acquire EMC characteristic data for subsequent diagnostic model design and verification. By injecting interference signals into the device under test, the sensitive phenomena of the circuit board are stimulated, ultimately yielding the required experimental data.

[0025] This platform is used for electromagnetic sensitivity testing of digital circuit boards under test. The testing platform mainly consists of three parts: the circuit board under test, test sensor fixtures, and post-processing. The sensor measures frequency and time domain information generated by the circuit board under test, which can be used for electromagnetic compatibility analysis. It also connects the sensor to the receiving device and provides filtering and shielding functions. The receiving device receives various electromagnetic signals collected by the sensor and transmits them according to a specified format. The post-processing unit uses a large amount of experimental data for data preprocessing and trains a fault diagnosis model. The platform mainly consists of the following parts: Interference signal generator: Generates various interference signals, including sine signals, pulse amplitude modulation signals, sine frequency modulation signals, etc., and injects them into the power port or control port of the board under test through a power amplifier.

[0026] Power amplifier: Used to increase the strength of interference signals and ensure that the interference signals can cover the electromagnetic susceptibility range of the device under test.

[0027] Measuring instruments: Used to capture electromagnetic signals from the board's response, including parameters such as voltage, current, and frequency. By transmitting multiple key locations of the board under test to devices such as oscilloscopes, the electromagnetic signals of the device under various interference conditions are comprehensively acquired.

[0028] Data transmission: Responsible for receiving signal data collected by sensors and transmitting the data to the backend data processing module.

[0029] 1.2 Design of Board Parameter Selection Method This invention aims to collect key parameters of the circuit board under test (PCB) for subsequent data-driven modeling and real-time monitoring data acquisition. Specific parameters should be researched and selected based on the characteristics of the PCB. When selecting the PCB's electromagnetic compatibility (EMC) parameters, the design should consider key dimensions such as electrical parameters, signal integrity, and electromagnetic interference, with a focus on sensitive areas and critical components. Simultaneously, high precision, high real-time performance, and scalability should be considered. Furthermore, the collected data should undergo thorough post-processing and analysis to ensure accurate assessment of the PCB's EMC characteristics and related issues.

[0030] In electromagnetic compatibility (EMC) analysis, electrical parameters are crucial foundational data used to assess whether a circuit board operates within its design specifications. Acquiring operating voltage and current data helps identify overvoltage, undervoltage, or excessive ripple on power rails, as well as whether load fluctuations cause overcurrent or undercurrent. Monitoring power consumption can determine the presence of short circuits, abnormal power consumption, or other hardware faults, which may further lead to electromagnetic interference (EMI) or electromagnetic susceptibility (EMS) issues.

[0031] Signal integrity is one of the main sources of electromagnetic interference (EMI) and a key factor in assessing electromagnetic susceptibility. Clock signal stability (including frequency skew and jitter) directly affects the generation of high-frequency EMI, while data signal waveforms (such as overshoot, undershoot, and noise levels) affect the reliability of high-speed data transmission. Monitoring crosstalk signals helps assess interference between high-speed signal lines, preventing communication errors and additional EMI problems.

[0032] Electromagnetic interference (EMI) parameters are key data that directly reflect the interference characteristics of a circuit board. By collecting common-mode current, differential-mode current, and conducted noise, the interference behavior of the circuit board on power lines and signal lines can be evaluated. Noise spectrum analysis helps to locate high-radiation frequency bands (such as clock harmonics) and identify frequency points that need optimization. Combining these data provides comprehensive support for the electromagnetic radiation analysis of the circuit board.

[0033] Thermal parameters are also an important dimension of EMC performance, especially in high-power and high-density designs. Excessive temperature can lead to component performance degradation, such as crystal drift or power module instability, indirectly affecting electromagnetic compatibility. By monitoring the temperature of high-heat areas and sensitive components, potential heat-related problems can be identified, ensuring that the entire circuit board maintains normal EMC characteristics under various operating conditions.

[0034] The above parameters directly affect the stability and electromagnetic characteristics of the circuit board. Collecting these parameters helps determine whether the system is working properly.

[0035] 1.3 Design of Collection Point Selection Principles The selection of sampling points should cover the entire circuit board while highlighting key and sensitive areas, especially core components such as power supplies, clock signals, high-speed signals, and interfaces. Simultaneously, the reasonable distribution and actual characteristics of the sampling points should be considered, and test points should be reserved in conjunction with the circuit design layout to comprehensively evaluate the EMC performance of the circuit board and promptly identify potential problems.

[0036] The selection of sampling points should first cover critical circuit areas of the circuit board, including power rails, DC / DC converters, and high-speed signal transmission lines (such as PCIe, DDR, etc.). These areas are the core of the circuit board's operation and are also high-risk areas for electromagnetic interference (EMI) and electromagnetic susceptibility (EMS) problems. In addition, areas such as antenna interfaces and RF paths in RF circuits, as well as the vicinity of core components (such as CPU, FPGA, memory, etc.), should also be key sampling targets, as they are crucial to the overall electromagnetic performance.

[0037] Data collection should prioritize sensitive locations, as these are often areas where EMI and EMS issues are concentrated. For example, antenna interfaces, power and signal line interfaces, and clock crystal oscillator areas are typically major sources of electromagnetic interference. Simultaneously, coupling areas between power and ground planes, as well as noise-sensitive areas such as analog circuits and high-speed signal paths, should also be key areas for data collection. Data from these sensitive areas can provide important clues for assessing the circuit board's immunity to interference.

[0038] To ensure comprehensive and balanced data acquisition, the distribution of acquisition points needs to consider both global and local characteristics. For large-area circuit boards, acquisition points should be evenly distributed across major areas, covering different layers such as power, ground, and signal layers. Simultaneously, acquisition should be tailored to different circuit characteristics: for digital circuits, the focus should be on high-speed signal integrity; for analog circuits, the focus should be on low-frequency noise; and for mixed-signal areas, the interference between digital and analog signals needs to be analyzed.

[0039] Finally, the selection of sampling points must be considered in conjunction with the actual circuit layout design. Priority should be given to areas prone to radiation or coupling, such as critical nodes of the switching power supply and both ends of filters, as well as the vicinity of high-speed clock driver chips. Simultaneously, the interfaces and boundary locations of the circuit board are the main pathways for electromagnetic interference to enter and exit the board; therefore, the signal and noise characteristics of these areas should be the focus of sampling. Furthermore, reasonably reserving test points, especially along critical paths and in sensitive areas, can improve the targeting and operability of the sampling.

[0040] 1.4 Other Key Considerations for Data Acquisition To obtain effective and reliable data, the following important principles must be considered: For example, the data acquisition frequency and resolution must be rationally selected based on the characteristics of the target signal to ensure accurate capture of high-speed signals and transient noise, while also considering the system's real-time performance and synchronization. The acquisition of external environmental parameters focuses on the electromagnetic, mechanical, temperature, humidity, and power supply factors of the circuit board, helping to analyze whether the board's electromagnetic compatibility (EMC) performance in actual application scenarios meets design requirements. Combining these two aspects allows for a comprehensive evaluation of the board's EMC performance and potential problems.

[0041] 1) Sampling frequency High-speed signal requirements: For high-speed signals (such as PCIe, DDR, USB, etc.), the sampling frequency needs to be at least 10 times the signal bandwidth to accurately capture signal characteristics and noise. For GHz-level high-speed signals, the sampling frequency should reach tens of GHz.

[0042] Low-frequency signals: For low-frequency signals such as power supply ripple and voltage fluctuations, the sampling frequency can be appropriately reduced, but it is necessary to ensure that the complete signal cycle is captured. It is generally recommended to use a frequency within 5-10 times the bandwidth of the target signal.

[0043] Transient signals: Some sudden noises (such as ESD discharge or power surge) occur for a very short time and require high-speed ADCs or oscilloscopes for ultra-high frequency sampling to capture transient waveforms.

[0044] 2) Sampling resolution Voltage and current signals: To accurately detect minute fluctuations, the ADC resolution should meet the signal's dynamic range requirements. A 12-bit or 16-bit resolution ADC is generally recommended; higher resolution devices (such as 24-bit) can be used if necessary.

[0045] Noise and weak signals: When acquiring low-amplitude noise signals, high-resolution ADCs can provide more accurate results and avoid interference from quantization errors.

[0046] Data transmission and storage: High resolution and high frequency sampling generate a large amount of data, so a trade-off must be made between accuracy and storage requirements, and an appropriate resolution and compression strategy must be selected.

[0047] 3) Real-time performance and synchronization of the data acquisition system Real-time acquisition: The analysis of dynamic EMC problems (such as electromagnetic susceptibility transient effects) requires the acquisition system to have real-time data transmission and processing capabilities.

[0048] Multi-channel synchronous sampling: For multi-point acquisition (such as simultaneous monitoring of power rails, signal lines, and interfaces), it is necessary to ensure time synchronization of each acquisition point to avoid signal distortion due to sampling delay.

[0049] 4) External electromagnetic environment Background electromagnetic noise: This involves collecting the electromagnetic noise level of the external environment, especially industrial noise and radio interference near the test area. It can be measured using a spectrum analyzer or field strength meter.

[0050] Ambient field strength: Measure the electromagnetic field strength around the circuit board, especially in a 3m or 10m radio anechoic chamber, to analyze the board's performance in radiation or immunity tests.

[0051] Specific frequency interference: Assess whether there are interference sources in the environment that are the same as or close to the operating frequency of the board (such as wireless devices such as Wi-Fi, 5G, and Bluetooth).

[0052] 5) Grounding and shielding measures Ground potential fluctuations: Monitor whether the ground potential of the circuit board is stable and whether there are abnormal fluctuations in the ground loop current, which may lead to coupling noise or ground loop interference.

[0053] Shielding effectiveness: Analyze the shielding effect of the shielding layer against external electromagnetic interference, including the integrity of the conductive shielding material and the quality of the shielding grounding connection.

[0054] 6) Other environmental conditions Mechanical environment: Considering the effects of vibration and shock, mechanical vibration or shock may cause some contact points to loosen, thereby introducing electromagnetic noise or sensitivity issues.

[0055] Temperature and humidity conditions: Changes in external temperature and humidity may alter the electromagnetic properties of the circuit board materials (such as capacitance and impedance). It is necessary to monitor the ambient temperature and humidity parameters in real time and analyze their impact on EMC performance.

[0056] External power supply quality: Monitor the quality of the power supply input to the circuit board (such as ripple and spike noise) to determine whether the stability of the power supply affects the EMC performance of the board.

[0057] 1.5 Data Acquisition Experimental Methods and Procedures To analyze the electromagnetic compatibility characteristics of circuit boards, it is necessary to induce various operating states of the boards through interference injection and other methods to collect a large amount of operational parameter data, thereby facilitating research and analysis. This invention considers using conducted susceptibility testing for interference injection, aiming to enable the digital circuit board to stably reproduce a certain state and collect the operating state parameters at that time, supporting subsequent data-driven modeling work.

[0058] To effectively stimulate the sensitivity of the circuit board, the testing of its sensitivity primarily follows the 5+X test standard. In the 5+X test method, two of the five typical signals are theoretically sinusoidal frequency modulation (10% frequency deviation) and sinusoidal frequency modulation (90% frequency deviation). Two fixed frequency deviation values ​​are used in the test. The port cables for the interference signals include various external cables such as complete power cables and Ethernet cables. The waveforms to be used are as follows: Figure 2 As shown.

[0059] When conducting sensitive phenomenon tests, the methods for determining abnormalities of the board under test include, but are not limited to: displaying water ripples, black screen, network interruption, peripheral I / O port abnormality, system crash, or other easily observable abnormal phenomena. The method for determining the abnormal state should be determined based on the specific functional performance of the board.

[0060] The interference injection experiment and data acquisition were completed by performing the following steps.

[0061] Step 1: Power on the board to bring it to a stable working state.

[0062] Step 2: Monitor the status of the circuit board to ensure that the data acquisition of measuring instruments such as oscilloscopes is working properly.

[0063] Step 3: Obtain the electromagnetic characteristic data at this time as a benchmark while observing the board's working performance in real time.

[0064] Step 4: Connect the signal generator and power amplifier together to the power port on the board under test using a current injection clamp.

[0065] Step 5: Inject various electromagnetic interference signals specified in the test into the power port in a certain amount until the board shows an abnormal condition in the sensitivity criteria.

[0066] Step 6: Compare and calibrate the electromagnetic characteristic data obtained at this time with the electromagnetic characteristic data during normal operation, and use an oscilloscope to store the voltage data of each test point.

[0067] Step 7: Reduce the interference field strength until the board returns to normal operation.

[0068] Step 8: Repeat steps 4-6 until the board no longer experiences new faults.

[0069] 2. Design of Data-Driven Modeling Methods 2.1 Design of Data Preprocessing Methods When obtaining electromagnetic compatibility (EMC) related parameter data from the board under test, the data obtained directly from the monitoring equipment usually has a number of problems, such as being messy, missing, having too many dimensions, class imbalance, and insufficient feature information. Therefore, it is necessary to carry out comprehensive and detailed steps such as feature extraction, feature selection, visualization analysis, and data balancing to construct a high-quality EMC dataset, and then train an effective fault diagnosis model.

[0070] Electromagnetic compatibility-related time series data obtained from the board are used to extract features from the time series data of each monitoring point, including the mean, variance, maximum and minimum values, and to obtain periodic features by applying trigonometric function fitting.

[0071] 1) Time window design A fixed time window length is chosen to divide the data into segments within the time series data. The selection of the time window length requires comprehensive consideration of the data characteristics and the analysis objectives. Shorter time windows can capture data changes more precisely, but may be more sensitive to noise. Longer time windows can smooth the data, but may lead to the loss of some details. Overlapping time windows can provide more information overlap, but increase computational cost. Largely spaced time windows may result in information loss. By appropriately selecting the time window length and setting overlap and intervals, it is possible to ensure that the extracted features accurately reflect the changing characteristics of the data.

[0072] 2) Extraction of mean, variance, maximum and minimum values For the data at each monitoring point, perform the following steps: calculate the mean, variance, maximum, and minimum values ​​within a given time window. These statistics provide basic information about the data distribution, reflecting the trend and magnitude of fluctuations in the data within the time window.

[0073] The mean represents the central tendency of the data distribution, indicating the overall level of the data within a time window. Variance measures the dispersion of the data, reflecting its volatility. By identifying the maximum and minimum values ​​of the data within a time window, we can understand the extreme cases within that window. These statistics provide a basic understanding of the overall performance of the monitored data within a time window, depicting its trends, volatility, and extreme cases.

[0074] 3) Periodic Feature Extraction The periodic feature extraction stage aims to analyze the periodic variation characteristics of the parameters to better understand their frequency components. This process involves Discrete Fourier Transform (FFT), frequency component selection, and trigonometric function fitting. The time-series data of the parameters is subjected to Discrete Fourier Transform (FFT) to transform the signal from the time domain to the frequency domain. The Discrete Fourier Transform is performed according to the following steps: a) Sampling and constructing time-domain signals: Treat the time-series data of the parameters as a discrete time-domain signal and determine the sampling rate, i.e., the number of data points collected per unit time.

[0075] b) Apply Fast Fourier Transform: Use the FFT algorithm to perform a Discrete Fourier Transform on the constructed time-domain signal to obtain the frequency domain representation of the signal, which includes the amplitude and phase information of various frequency components.

[0076] c) Selecting frequency components: Analyze the results of the Fourier transform to identify the main frequency components, which represent the periodic characteristics of the parameters.

[0077] Frequency information is obtained by fitting selected frequency components, thereby calculating periodic characteristics. A sine function is used as the fitting model. This model has three parameters: amplitude, frequency, and phase. The selected principal frequency components are fitted using a nonlinear optimization method (such as least squares). The optimization goal is to make the model fit the data, finding the optimal amplitude, frequency, and phase. Frequency information is extracted from the fitted parameters. For a sine function, the frequency is the number of periods within each period, i.e., the reciprocal of the period.

[0078] By following these steps, we can obtain the periodic characteristics of the parameters, which helps us to understand the periodic changes of the signal more deeply. These characteristics can provide valuable information for further analysis and applications.

[0079] 4) Feature normalization After feature extraction, the feature values ​​obtained from each monitoring point are normalized to ensure their consistency in scale. This avoids problems that may be caused by scale differences between different features and provides a more reliable data foundation for subsequent analysis and modeling. This paper uses min-max normalization for data scaling, mapping the feature values ​​to the range [0,1].

[0080] 2.2 Sample Balancing and Feature Selection Processing Design Fault samples exhibiting electromagnetic susceptibility are typically far fewer than normal samples. Accurate identification of fault samples belonging to the minority class is more valuable than identifying normal samples belonging to the majority class. Furthermore, the time from a critical board malfunctioning to causing damage is highly unpredictable, sometimes as short as a few minutes. Traditional machine learning algorithms rely on the fundamental assumption that the number of training samples in each class is equal, and therefore the cost of misclassification is also equal. However, the imbalanced distribution of sample data severely impacts the accuracy of minority class classification. This is especially true in fault detection and prediction, where minority samples are the primary focus.

[0081] The SMOTE

[15] algorithm can be used for data oversampling. The undersampling method does not randomly discard the majority class, but removes samples that do not contribute much to the classification, thus avoiding the loss of important information. For samples with a high proportion of imbalance, the classification accuracy of minority class samples is not high. Further equalization processing of sample data is carried out by combining the undersampling method based on Euclidean distance and the oversampling method based on K-Means

[17] SMOTE based on cluster distribution to reconstruct sample data. Imbalanced data resampling is as follows: Figure 3 As shown, the steps are as follows: 1) Classify the monitoring status data sample set D of each board according to the sample attributes. The minority class sample set is defined as S, and the majority class sample set is defined as M.

[0082] 2) Based on the empirical formula for resampling in binary classification of imbalanced datasets, as shown in formulas (1) and (2), set the resampling ratio. Wherein, α is the oversampling size of the minus class, α≥1, β is the undersampling size of the max class, 0≤β≤10, and R is the ratio of the number of minority class samples to the number of majority class samples.

[0083] (1) (2) 3) Use the K-Means SMOTE method, which generates samples based on cluster distribution, to oversample minority class samples.

[0084] 4) An undersampling method based on Euclidean distance is used to undersample majority class samples in non-boundary regions. This method identifies boundary samples and deletes majority class samples that are far from the classification boundary and have little impact on classification. The training set preserves safe samples that have a large classification influence and are close to the classification boundary.

[0085] 5) Combine the updated minority class samples and majority class samples into a new balanced sample set. The new balanced sample set achieves both class balance and maintains the data distribution characteristics of the original dataset.

[0086] Feature selection refers to filtering out the features that have the greatest impact on model classification from all the features in the original data, thereby compressing the feature space dimension and improving the model's learning efficiency. In other words, by selecting features, we can obtain "few but excellent" features with a low probability of classification error in the dataset, which can reduce the complexity of the subsequent model.

[0087] 1) Relevant feature selection based on feature selection function Feature selection functions are a statistical feature selection technique. In feature selection functions, the correlation or other statistical indicators between each feature and the target variable are first calculated. This is achieved through methods such as calculating correlation coefficients, mutual information, and chi-square tests. After obtaining the scores for each feature, they are sorted from highest to lowest score, and then the top k features with the highest scores are selected as key features. To ensure that the selected number of features achieves optimal accuracy in the model, cross-validation is used. Cross-validation with different subsets of data helps find a number of features that perform well on both the training and validation sets, resulting in higher accuracy.

[0088] 2) Tree-based feature selection Tree-based feature selection methods utilize algorithms such as decision trees or random forests to evaluate the importance of features. The core idea of ​​this approach is to construct decision trees or random forests to measure the contribution of each feature to the model. The greater the information gain, the more important the feature plays in the prediction.

[0089] Feature selection functions and tree-based relevance feature selection methods provide effective tools for selecting key features during the modeling process. These methods reduce dimensionality, noise, and optimize the number of features, thereby improving model performance and interpretability.

[0090] 2.3 Design of Modeling Algorithm Traditional fault diagnosis methods often rely on expert experience and are unable to effectively handle large and complex datasets. Therefore, leveraging machine learning algorithms is considered to improve diagnostic efficiency and accuracy. Rapid and accurate fault diagnosis is crucial for ensuring system reliability. Data-driven machine learning and deep learning algorithms are introduced to construct intelligent diagnostic models for different fault scenarios. Algorithm selection must consider data characteristics, problem complexity, and application scenarios, comprehensively taking into account the suitability of traditional machine learning and deep learning methods.

[0091] 1) Algorithm type considerations The choice of data modeling algorithm depends on the specific circumstances. It mainly includes the following two categories: a) Traditional machine learning algorithms: Random Forest is suitable for non-linear data and has good feature interpretation capabilities. Gradient Boosting Decision Tree (GBDT) is suitable for small to medium-sized data and performs well in handling non-linear relationships. Support Vector Machine (SVM) is suitable for both linear and non-linear problems, especially small-scale high-dimensional data.

[0092] b) Deep Learning Algorithms: Convolutional Neural Networks (CNNs) excel at extracting spatial features from multidimensional signals and can be used for frequency domain analysis or image-based data. Recurrent Neural Networks (RNNs) and their variants (LSTM, GRU) are suitable for time-series data and can capture long-term dependencies. Transformers perform exceptionally well on sequence data and are suitable for high-dimensional, complex, and dynamic data. Combination models (such as CNN-LSTM) combine the spatial feature extraction of CNNs with the temporal feature capture of LSTMs, making them suitable for multimodal and complex fault scenarios.

[0093] 2) Algorithm selection principles The choice of which specific algorithm to use should be based on factors such as the amount of data, data characteristics, and task objectives. a) Small-scale data: Traditional machine learning algorithms (such as random forest, support vector machine, etc.) are preferred because they have short training time and low data requirements.

[0094] b) Large-scale data: Choosing deep learning models (such as CNN, RNN, Transformer, etc.) can fully uncover the deep patterns in complex data.

[0095] c) Linear characteristic relationship: suitable for linear regression or logistic regression models.

[0096] d) Nonlinear relationships or high-dimensional data: Use tree models (such as gradient boosting decision trees, XGBoost) or deep learning models.

[0097] e) Time series data: Choose a recurrent neural network (such as LSTM, GRU) or a time series Transformer.

[0098] f) Spatial-temporal joint characteristics: Using combined models such as CNN-LSTM and Bi-LSTM, the complex interaction between space and time is captured.

[0099] In summary, for small-scale datasets, random forests or GBDT combined with feature engineering can be used to quickly build high-performance classification models. For large-scale time series data, LSTM or Transformers can be used, combined with data augmentation methods (such as SMOTE) to improve the model's sensitivity to the minority class. Multimodal data fusion combined with a CNN-LSTM architecture can extract frequency and time domain features, improving the fault classification capability in complex scenarios.

[0100] 3) Optimization and Enhancement Model Enhance model performance and optimize model hyperparameters through grid search, random search, or Bayesian optimization. Combine multiple algorithms (such as Bagging or Boosting) to implement ensemble models and improve classification robustness.

[0101] 2.4 Data Modeling Process The data-driven modeling of this invention includes the following main steps: Step 1: Conducted Sensitivity Testing and Data Acquisition. Using a conducted sensitivity testing platform, different types of interference signals are injected into the digital circuit board under test to elicit sensitive phenomena in the device. During this process, electromagnetic characteristic data of the board under different interference conditions are collected, including multiple parameters such as voltage, frequency, and radiation intensity.

[0102] Step 2: Data cleaning and preprocessing. The collected electromagnetic compatibility data is cleaned, denoised, and normalized to ensure data scale consistency. Simultaneously, key features are extracted through time-domain and frequency-domain analysis, ready for input into the classification model.

[0103] Step 3: Feature selection and sample balancing. Feature selection methods (such as information gain-based or tree-based feature importance ranking) are used to select the most relevant features, reducing data dimensionality and redundant information. Furthermore, SMOTE and undersampling methods are used to balance the samples, improving the classification model's ability to identify minority class fault samples.

[0104] Step 4: Build and train the fault classification model.

[0105] Step 4.1: Model Selection. Choose a suitable model based on the characteristics of the data. Possible algorithms include: Traditional machine learning models: Random Forest, Gradient Boosting Decision Tree (GBDT), etc. Deep learning models: One-dimensional Convolutional Neural Network (1D-CNN), Long Short-Term Memory Network (LSTM), Transformer, etc. Combined models: such as CNN-LSTM, Bi-LSTM, etc., used to capture spatiotemporal features.

[0106] Step 4.2: Model Training. Input the processed data into the selected model and train the model using the training set. Model performance can be evaluated using cross-validation techniques.

[0107] Step 4.3: Hyperparameter tuning. Use techniques such as grid search (GridSearchCV), random search, or Bayesian optimization to find the optimal combination of hyperparameters for the model.

[0108] Step 5: Testing and Optimization Feedback. Using the trained classification model, the system classifies the diagnostic data samples to determine the type of electromagnetic susceptibility. Based on the classification results, the system provides corresponding optimization suggestions to help designers make targeted improvements to electromagnetic compatibility.

[0109] 3. Alternative Solutions Alternative to conducted susceptibility testing: In addition to conducted susceptibility testing, the system of this invention can also be combined with radiation susceptibility testing methods to comprehensively evaluate the electromagnetic compatibility of equipment.

[0110] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A data-driven electromagnetic compatibility fault diagnosis method, characterized in that, Includes the following steps: Injecting interference signals into the device under test and collecting multi-dimensional operating parameter data of the device in an electromagnetic environment; Feature extraction is performed on the parameter data, including calculating statistical features within the time window and extracting periodic features through frequency domain transformation; The extracted features are subjected to sample equalization processing to balance the number of faulty samples and normal samples. Based on the equalized feature data, a fault diagnosis model is constructed using machine learning algorithms; The fault diagnosis model is used to predict and locate electromagnetic compatibility faults in the device under test.

2. The data-driven electromagnetic compatibility fault diagnosis method according to claim 1, characterized in that, Multi-dimensional operating parameters include electrical parameters, signal integrity parameters, electromagnetic interference parameters, and thermal parameters. Electrical parameters include voltage, current, and power consumption; signal integrity parameters include clock stability and data signal waveforms; electromagnetic interference parameters include common-mode / differential-mode current and conducted noise; and thermal parameters include the temperature of key components.

3. The data-driven electromagnetic compatibility fault diagnosis method according to claim 1, characterized in that, The parameter data acquisition points include the power rail, DC / DC converter area, high-speed signal transmission line, core components, clock crystal area, analog-digital mixed signal area and / or board interface boundary location.

4. The data-driven electromagnetic compatibility fault diagnosis method according to claim 1, characterized in that, Feature extraction includes: Time series data is segmented using a fixed-length time window; Calculate the mean, variance, maximum, and minimum values ​​for each time window; The main frequency components are obtained by performing a discrete Fourier transform on the time series data, and periodic features are extracted by fitting trigonometric functions. All extracted features are normalized.

5. The data-driven electromagnetic compatibility fault diagnosis method according to claim 4, characterized in that, Periodic feature extraction includes: using Fast Fourier Transform to convert the signal from the time domain to the frequency domain, identifying the main frequency components, fitting a sine function model using a nonlinear optimization method, and extracting frequency information from the fitted parameters.

6. The data-driven electromagnetic compatibility fault diagnosis method according to claim 1, characterized in that, The sample equalization process includes: oversampling minority class fault samples using the SMOTE method based on K-Means clustering, and undersampling majority class normal samples using the Euclidean distance-based method to remove samples far from the classification boundary.

7. A data-driven electromagnetic compatibility fault diagnosis method according to claim 1 or 6, characterized in that, Feature selection includes: first, preliminary screening by calculating the correlation between features and the target variable using statistical methods; then, secondary screening by using feature importance assessment based on a tree model; and finally, determining the optimal feature subset through cross-validation.

8. The data-driven electromagnetic compatibility fault diagnosis method according to claim 1, characterized in that, The selection of machine learning algorithms includes: using random forest or gradient boosting decision tree when the data size is less than a preset threshold, using convolutional neural network, recurrent neural network or Transformer model when the data size is greater than a preset threshold, and using CNN-LSTM combined model for data with spatiotemporal correlation characteristics.

9. The data-driven electromagnetic compatibility fault diagnosis method according to claim 1, characterized in that, Injecting interference signals includes the following steps: Power on the board under test and bring it to a stable working state; Collect reference electromagnetic characteristic data; Connect the interference signal generator to the power port of the board using injection clamps, and inject the interference signal at the specified intensity until a sensitive phenomenon occurs. Record electromagnetic characteristic data under abnormal conditions and compare and calibrate with reference data; Gradually reduce the interference intensity until the board returns to normal, and repeat the test until no new faults occur.

10. A system applying the data-driven electromagnetic compatibility fault diagnosis method according to any one of claims 1-9, characterized in that, include: Interference signal generation module, used to generate and amplify interference signals; The data acquisition module, including measurement sensors and instruments, is used to acquire multi-dimensional electromagnetic compatibility parameters of the device under test; The data processing module is used to perform feature extraction, sample equalization, and feature selection on the collected data; The model training and diagnosis module is used to build fault diagnosis models and output diagnosis results.

Images